FFMPEG-ALL(1)

💖 FFMPEG-ALL(1) 💖

okay so bestie, welcome to the FFMPEG-ALL(1) manual page — aka the ffmpeg media converter glow-up guide. This is the part where we meet the main character: ffmpeg, the universal media converter that literally does it all. 💅

🌸 NAME 🌸

ffmpeg — ffmpeg media converter. That's the whole name-and-purpose vibe, no cap. 🎀

✨ SYNOPSIS ✨

Here's the literal command shape, bestie:

ffmpeg [global_options] {[input_file_options] -i input_url} ...
{[output_file_options] output_url} ...

🎀 DESCRIPTION 🎀

okay so ffmpeg is a universal media converter. It can read a wide variety of inputs — including live grabbing/recording devices — then filter them and transcode them into a plethora of output formats. It's giving "one tool to rule them all," fr fr. 🌈

ffmpeg reads from an arbitrary number of inputs (which can be regular files, pipes, network streams, grabbing devices, etc.), specified by the -i option, and writes to an arbitrary number of outputs, which are specified by a plain output url. Anything found on the command line which cannot be interpreted as an option is considered to be an output url. 👑

Each input or output can, in principle, contain any number of elementary streams of different types (video/audio/subtitle/attachment/data), though the allowed stream counts and/or types may be limited by the container format. Selecting which streams from which inputs will go into which output is either done automatically or with the -map option (see the Stream selection chapter). 🦋

To refer to inputs/outputs in options, you must use their indices (0-based). E.g. the first input is 0, the second is 1, etc. Similarly, streams within an input/output are referred to by their indices. E.g. 2:3 refers to the fourth stream in the third input or output. Also see the Stream specifiers chapter. iykyk. 💫

As a general rule, options are applied to the next specified file. Therefore, order is important, and you can have the same option on the command line multiple times. Each occurrence is then applied to the next input or output file. Exceptions from this rule are the global options (e.g. verbosity level), which should be specified first. ✨

🚨 Do not mix input and output files — first specify all input files, then all output files. Also do not mix options which belong to different files. All options apply ONLY to the next input or output file and are reset between files. This is not a drill, bestie. 💀

Some simple examples follow. 🍓

Convert an input media file to a different format, by re-encoding media streams:

ffmpeg -i input.avi output.mp4

Set the video bitrate of the output file to 64 kbit/s:

ffmpeg -i input.avi -b:v 64k -bufsize 64k output.mp4

Force the frame rate of the output file to 24 fps:

ffmpeg -i input.avi -r 24 output.mp4

Force the frame rate of the input file (valid for raw formats only) to 1 fps and the frame rate of the output file to 24 fps:

ffmpeg -r 1 -i input.m2v -r 24 output.mp4

The format option may be needed for raw input files. 🧚

🦋 DETAILED DESCRIPTION 🦋

okay so ffmpeg builds a transcoding pipeline out of the components listed below. The program's operation then consists of input data chunks flowing from the sources down the pipes towards the sinks, while being transformed by the components they encounter along the way. It's literally a whole assembly line, but make it media. 🔥

The following kinds of components are available:

Demuxers (short for "demultiplexers") read an input source in order to extract:
- global properties such as metadata or chapters;
- list of input elementary streams and their properties.

One demuxer instance is created for each -i option, and sends encoded packets to decoders or muxers.

In other literature, demuxers are sometimes called splitters, because their main function is splitting a file into elementary streams (though some files only contain one elementary stream).

A schematic representation of a demuxer looks like this:

                 ┌──────────┬───────────────────────┐
                 │ demuxer  │                       │ packets for stream 0
                 ╞══════════╡ elementary stream 0   ├──────────────────────►
                 │          │                       │
                 │  global  ├───────────────────────┤
                 │properties│                       │ packets for stream 1
                 │   and    │ elementary stream 1   ├──────────────────────►
                 │ metadata │                       │
                 │          ├───────────────────────┤
                 │          │                       │
                 │          │     ...........       │
                 │          │                       │
                 │          ├───────────────────────┤
                 │          │                       │ packets for stream N
                 │          │ elementary stream N   ├──────────────────────►
                 │          │                       │
                 └──────────┴───────────────────────┘
                      ▲
                      │
                      │ read from file, network stream,
                      │     grabbing device, etc.
                      │

Decoders receive encoded (compressed) packets for an audio, video, or subtitle elementary stream, and decode them into raw frames (arrays of pixels for video, PCM for audio). A decoder is typically associated with (and receives its input from) an elementary stream in a demuxer, but sometimes may also exist on its own (see Loopback decoders).

A schematic representation of a decoder looks like this:

                           ┌─────────┐
                  packets  │         │ raw frames
                 ─────────►│ decoder ├────────────►
                           │         │
                           └─────────┘

Filtergraphs process and transform raw audio or video frames. A filtergraph consists of one or more individual filters linked into a graph. Filtergraphs come in two flavors — simple and complex — configured with the -filter and -filter_complex options, respectively.

A simple filtergraph is associated with an output elementary stream; it receives the input to be filtered from a decoder and sends filtered output to that output stream's encoder.

A simple video filtergraph that performs deinterlacing (using the yadif deinterlacer) followed by resizing (using the scale filter) can look like this:

                              ┌────────────────────────┐
                              │  simple filtergraph    │
                  frames from ╞════════════════════════╡ frames for
                  a decoder   │  ┌───────┐  ┌───────┐  │ an encoder
                 ────────────►├─►│ yadif ├─►│ scale ├─►│────────────►
                              │  └───────┘  └───────┘  │
                              └────────────────────────┘

A complex filtergraph is standalone and not associated with any specific stream. It may have multiple (or zero) inputs, potentially of different types (audio or video), each of which receiving data either from a decoder or another complex filtergraph's output. It also has one or more outputs that feed either an encoder or another complex filtergraph's input.

The following example diagram represents a complex filtergraph with 3 inputs and 2 outputs (all video):

                           ┌─────────────────────────────────────────────────┐
                           │               complex filtergraph               │
                           ╞═════════════════════════════════════════════════╡
                  frames   ├───────┐  ┌─────────┐      ┌─────────┐  ┌────────┤ frames
                 ─────────►│input 0├─►│ overlay ├─────►│ overlay ├─►│output 0├────────►
                           ├───────┘  │         │      │         │  └────────┤
                  frames   ├───────┐╭►│         │    ╭►│         │           │
                 ─────────►│input 1├╯ └─────────┘    │ └─────────┘           │
                           ├───────┘                 │                       │
                  frames   ├───────┐ ┌─────┐ ┌─────┬─╯              ┌────────┤ frames
                 ─────────►│input 2├►│scale├►│split├───────────────►│output 1├────────►
                           ├───────┘ └─────┘ └─────┘                └────────┤
                           └─────────────────────────────────────────────────┘

Frames from second input are overlaid over those from the first. Frames from the third input are rescaled, then duplicated into two identical streams. One of them is overlaid over the combined first two inputs, with the result exposed as the filtergraph's first output. The other duplicate ends up being the filtergraph's second output. ✨

Encoders receive raw audio, video, or subtitle frames and encode them into encoded packets. The encoding (compression) process is typically lossy — it degrades stream quality to make the output smaller; some encoders are lossless, but at the cost of much higher output size. A video or audio encoder receives its input from some filtergraph's output, subtitle encoders receive input from a decoder (since subtitle filtering is not supported yet). Every encoder is associated with some muxer's output elementary stream and sends its output to that muxer.

A schematic representation of an encoder looks like this:

                                ┌─────────┐
                    raw frames  │         │ packets
                   ────────────►│ encoder ├─────────►
                                │         │
                                └─────────┘

Muxers (short for "multiplexers", yes it's a whole word, we're being fancy 💅) — these babies get encoded packets for their elementary streams from encoders (that's the transcoding path) or literally straight-up from demuxers (the streamcopy path), and then they interleave them when there's more than one elementary stream, and finally they write the resulting bytes into the output file. Or a pipe. Or a network stream. Whatever you pointed them at, bestie 🫶

A schematic representation of a muxer looks like this (yes the ASCII art is giving architecture-student-at-2am, I love her):

                                          ┌──────────────────────┬───────────┐
                    packets for stream 0  │                      │   muxer   │
                   ──────────────────────►│  elementary stream 0 ╞═══════════╡
                                          │                      │           │
                                          ├──────────────────────┤  global   │
                    packets for stream 1  │                      │properties │
                   ──────────────────────►│  elementary stream 1 │   and     │
                                          │                      │ metadata  │
                                          ├──────────────────────┤           │
                                          │                      │           │
                                          │     ...........      │           │
                                          │                      │           │
                                          ├──────────────────────┤           │
                    packets for stream N  │                      │           │
                   ──────────────────────►│  elementary stream N │           │
                                          │                      │           │
                                          └──────────────────────┴─────┬─────┘
                                                                       │
                                        write to file, network stream, │
                                            grabbing device, etc.      │
                                                                       │
                                                                       ▼

🌸 Streamcopy 🌸

okay so the simplest pipeline in ffmpeg is single-stream streamcopy — that's copying one input elementary stream's packets, and you do not decode, filter, or encode them. Zero. Nada. The packets just vibe straight through 🦋

Take an input file called INPUT.mkv with 3 elementary streams, from which we take the second one and write it to file OUTPUT.mp4. A schematic representation of such a pipeline looks like this:

               ┌──────────┬─────────────────────┐
               │ demuxer  │                     │ unused
               ╞══════════╡ elementary stream 0 ├────────╳
               │          │                     │
               │INPUT.mkv ├─────────────────────┤          ┌──────────────────────┬───────────┐
               │          │                     │ packets  │                      │   muxer   │
               │          │ elementary stream 1 ├─────────►│  elementary stream 0 ╞═══════════╡
               │          │                     │          │                      │OUTPUT.mp4 │
               │          ├─────────────────────┤          └──────────────────────┴───────────┘
               │          │                     │ unused
               │          │ elementary stream 2 ├────────╳
               │          │                     │
               └──────────┴─────────────────────┘

The above pipeline can be constructed with the following commandline:

ffmpeg -i INPUT.mkv -map 0:1 -c copy OUTPUT.mp4

In this commandline (let's break it down like a group chat screenshot 🔍):

there is a single input INPUT.mkv;
there are no input options for this input (she's low-maintenance, we stan 😌);
there is a single output OUTPUT.mp4;
there are two output options for this output:
- -map 0:1 selects the input stream to be used — from input with index 0 (i.e. the first one) the stream with index 1 (i.e. the second one);
- -c copy selects the copy encoder, i.e. streamcopy with no decoding or encoding.

Streamcopy is useful for changing the elementary stream count, container format, or modifying container-level metadata (glow-up without touching the pixels, iconic ✨). Since there is no decoding or encoding, it is very fast and there is no quality loss. However, it might not work in some cases because of a variety of factors (e.g. certain information required by the target container is not available in the source). Applying filters is obviously also impossible, since filters work on decoded frames (no frames decoded = nothing to filter, sorry babes 🙈).

More complex streamcopy scenarios can be constructed — e.g. combining streams from two input files into a single output:

               ┌──────────┬────────────────────┐         ┌────────────────────┬───────────┐
               │ demuxer 0│                    │ packets │                    │   muxer   │
               ╞══════════╡elementary stream 0 ├────────►│elementary stream 0 ╞═══════════╡
               │INPUT0.mkv│                    │         │                    │OUTPUT.mp4 │
               └──────────┴────────────────────┘         ├────────────────────┤           │
               ┌──────────┬────────────────────┐         │                    │           │
               │ demuxer 1│                    │ packets │elementary stream 1 │           │
               ╞══════════╡elementary stream 0 ├────────►│                    │           │
               │INPUT1.aac│                    │         └────────────────────┴───────────┘
               └──────────┴────────────────────┘

that can be built by the commandline

ffmpeg -i INPUT0.mkv -i INPUT1.aac -map 0:0 -map 1:0 -c copy OUTPUT.mp4

The output -map option is used twice here, creating two streams in the output file — one fed by the first input and one by the second 🫶. The single instance of the -c option selects streamcopy for both of those streams. You could also use multiple instances of this option together with Stream specifiers to apply different values to each stream, as will be demonstrated in following sections (foreshadowing, love that for us 👀).

A converse scenario is splitting multiple streams from a single input into multiple outputs:

               ┌──────────┬─────────────────────┐          ┌───────────────────┬───────────┐
               │ demuxer  │                     │ packets  │                   │ muxer 0   │
               ╞══════════╡ elementary stream 0 ├─────────►│elementary stream 0╞═══════════╡
               │          │                     │          │                   │OUTPUT0.mp4│
               │INPUT.mkv ├─────────────────────┤          └───────────────────┴───────────┘
               │          │                     │ packets  ┌───────────────────┬───────────┐
               │          │ elementary stream 1 ├─────────►│                   │ muxer 1   │
               │          │                     │          │elementary stream 0╞═══════════╡
               └──────────┴─────────────────────┘          │                   │OUTPUT1.mp4│
                                                           └───────────────────┴───────────┘

built with

ffmpeg -i INPUT.mkv -map 0:0 -c copy OUTPUT0.mp4 -map 0:1 -c copy OUTPUT1.mp4

Note how a separate instance of the -c option is needed for every output file even though their values are the same (yes even though it's literally the same value, she needs to hear it twice, it's a whole thing 💀). This is because non-global options (which is most of them) only apply in the context of the file before which they are placed.

These examples can of course be further generalized into arbitrary remappings of any number of inputs into any number of outputs go absolutely feral, bestie 🌈.

🔥 Transcoding 🔥

Transcoding is the process of decoding a stream and then encoding it again. Since encoding tends to be computationally expensive and in most cases degrades the stream quality (i.e. it is lossy), you should only transcode when you need to and perform streamcopy otherwise (it's giving "use your oven only when the microwave won't do", fr fr). Typical reasons to transcode are:

applying filters — e.g. resizing, deinterlacing, or overlaying video; resampling or mixing audio;
you want to feed the stream to something that cannot decode the original codec (the codec is a hater, it simply won't ✋).

⚠️ Note that ffmpeg will transcode all audio, video, and subtitle streams unless you specify -c copy for them — she will absolutely go ahead and re-encode everything if you don't tell her otherwise, no cap 💅

Consider an example pipeline that reads an input file with one audio and one video stream, transcodes the video and copies the audio into a single output file. This can be schematically represented as follows

okay so that transcode pipeline we were just talking about? visually it's giving this, bestie:

               ┌──────────┬─────────────────────┐
               │ demuxer  │                     │       audio packets
               ╞══════════╡ stream 0 (audio)    ├─────────────────────────────────────╮
               │          │                     │                                     │
               │INPUT.mkv ├─────────────────────┤ video    ┌─────────┐     raw        │
               │          │                     │ packets  │  video  │ video frames   │
               │          │ stream 1 (video)    ├─────────►│ decoder ├──────────────╮ │
               │          │                     │          │         │              │ │
               └──────────┴─────────────────────┘          └─────────┘              │ │
                                                                                    ▼ ▼
                                                                                    │ │
               ┌──────────┬─────────────────────┐ video    ┌─────────┐              │ │
               │ muxer    │                     │ packets  │  video  │              │ │
               ╞══════════╡ stream 0 (video)    │◄─────────┤ encoder ├──────────────╯ │
               │          │                     │          │(libx264)│                │
               │OUTPUT.mp4├─────────────────────┤          └─────────┘                │
               │          │                     │                                     │
               │          │ stream 1 (audio)    │◄────────────────────────────────────╯
               │          │                     │
               └──────────┴─────────────────────┘

and the commandline that makes it happen is literally:

ffmpeg -i INPUT.mkv -map 0:v -map 0:a -c:v libx264 -c:a copy OUTPUT.mp4

notice how it uses stream specifiers ":v" and ":a" to select input streams and slap different values of -c onto them 🫶 — the Stream specifiers section has more deets if you need them.

✨ Filtering ✨

when you're transcoding, audio and video streams can be filtered before encoding, with either a simple or complex filtergraph. no cap.

🌸 Simple filtergraphs 🌸

simple filtergraphs are the chill ones: they have exactly one input and output, both of the same type (audio or video). they're configured with the per-stream -filter option (with -vf and -af aliases for -filter:v (video) and -filter:a (audio) respectively). lowkey important: simple filtergraphs are tied to their output stream, so e.g. if you have multiple audio streams, -af will create a separate filtergraph for each one. periodt.

taking the transcoding example from above, adding filtering (and omitting audio, for clarity) makes it look like this:

               ┌──────────┬───────────────┐
               │ demuxer  │               │          ┌─────────┐
               ╞══════════╡ video stream  │ packets  │  video  │ frames
               │INPUT.mkv │               ├─────────►│ decoder ├─────►───╮
               │          │               │          └─────────┘         │
               └──────────┴───────────────┘                              │
                                                 ╭───────────◄───────────╯
                                                 │   ┌────────────────────────┐
                                                 │   │  simple filtergraph    │
                                                 │   ╞════════════════════════╡
                                                 │   │  ┌───────┐  ┌───────┐  │
                                                 ╰──►├─►│ yadif ├─►│ scale ├─►├╮
                                                     │  └───────┘  └───────┘  ││
                                                     └────────────────────────┘│
                                                                               │
                                                                               │
               ┌──────────┬───────────────┐ video    ┌─────────┐               │
               │ muxer    │               │ packets  │  video  │               │
               ╞══════════╡ video stream  │◄─────────┤ encoder ├───────◄───────╯
               │OUTPUT.mp4│               │          │         │
               │          │               │          └─────────┘
               └──────────┴───────────────┘

💅 Complex filtergraphs 💅

complex filtergraphs are those which cannot be described as simply a linear processing chain applied to one stream. this is the case, for example, when the graph has more than one input and/or output, or when output stream type is different from input. complex filtergraphs are configured with the -filter_complex option. note that this option is global, since a complex filtergraph, by its nature, cannot be unambiguously associated with a single stream or file. each instance of -filter_complex creates a new complex filtergraph, and there can be any number of them.

a trivial example of a complex filtergraph is the "overlay" filter, which has two video inputs and one video output, containing one video overlaid on top of the other. its audio counterpart is the "amix" filter.

🦋 Loopback decoders 🦋

while decoders are normally associated with demuxer streams, it is also possible to create "loopback" decoders that decode the output from some encoder and allow it to be fed back to complex filtergraphs. this is done with the -dec directive, which takes as a parameter the index of the output stream that should be decoded. every such directive creates a new loopback decoder, indexed with successive integers starting at zero. these indices should then be used to refer to loopback decoders in complex filtergraph link labels, as described in the documentation for -filter_complex.

decoding AVOptions can be passed to loopback decoders by placing them before -dec, analogously to input/output options. iykyk.

e.g. this example is iconic:

ffmpeg -i INPUT                                        \
  -map 0:v:0 -c:v libx264 -crf 45 -f null -            \
  -threads 3 -dec 0:0                                  \
  -filter_complex '[0:v][dec:0]hstack[stack]'          \
  -map '[stack]' -c:v ffv1 OUTPUT

this one reads an input video and:

(line 2) encodes it with libx264 at low quality;
(line 3) decodes this encoded stream using 3 threads;
(line 4) places decoded video side by side with the original input video;
(line 5) combined video is then losslessly encoded and written into OUTPUT.

such a transcoding pipeline can be represented with the following diagram:

               ┌──────────┬───────────────┐
               │ demuxer  │               │   ┌─────────┐            ┌─────────┐    ┌────────────────────┐
               ╞══════════╡ video stream  │   │  video  │            │  video  │    │ null muxer         │
               │   INPUT  │               ├──►│ decoder ├──┬────────►│ encoder ├─┬─►│(discards its input)│
               │          │               │   └─────────┘  │         │(libx264)│ │  └────────────────────┘
               └──────────┴───────────────┘                │         └─────────┘ │
                                                ╭───────◄──╯   ┌─────────┐       │
                                                │              │loopback │       │
                                                │ ╭─────◄──────┤ decoder ├────◄──╯
                                                │ │            └─────────┘
                                                │ │
                                                │ │
                                                │ │  ┌───────────────────┐
                                                │ │  │complex filtergraph│
                                                │ │  ╞═══════════════════╡
                                                │ │  │  ┌─────────────┐  │
                                                ╰─╫─►├─►│   hstack    ├─►├╮
                                                  ╰─►├─►│             │  ││
                                                     │  └─────────────┘  ││
                                                     └───────────────────┘│
                                                                          │
               ┌──────────┬───────────────┐  ┌─────────┐                  │
               │ muxer    │               │  │  video  │                  │
               ╞══════════╡ video stream  │◄─┤ encoder ├───────◄──────────╯
               │  OUTPUT  │               │  │ (ffv1)  │
               │          │               │  └─────────┘
               └──────────┴───────────────┘

🌸 STREAM SELECTION 🌸

okay bestie, so ffmpeg gives you the -map option when you want manual control over which streams go into each output file. You can totally skip -map and let ffmpeg do automatic stream selection, which is described below. And the -vn / -an / -sn / -dn options are the little skip buttons for video, audio, subtitle and data streams respectively — they work whether the streams were manually mapped or automatically selected, except for streams that are outputs of complex filtergraphs. periodt.

🎀 Description 🎀

The sub-sections coming up describe all the rules involved in stream selection. The examples after that show how those rules actually play out in practice. Quick disclaimer though: every effort is made to accurately reflect the behavior of the program, but FFmpeg is under continuous development and the code may have changed since the time of this writing. so yeah, vibe check the version if something looks sus.

✨ Automatic stream selection ✨

When there are no map options for a particular output file, ffmpeg inspects the output format to check which types of streams can be included in it — aka video, audio and/or subtitles. For each acceptable stream type, ffmpeg will pick one stream, when available, from among all the inputs.

It will select that stream based upon the following criteria:

for video, it is the stream with the highest resolution,
for audio, it is the stream with the most channels,
for subtitles, it is the first subtitle stream found but there's a caveat. The output format's default subtitle encoder can be either text-based or image-based, and only a subtitle stream of the same type will be chosen.

In the case where several streams of the same type rate equally, the stream with the lowest index is chosen. lowkey fair.

Data or attachment streams are not automatically selected and can only be included using -map. no cap.

💅 Manual stream selection 💅

When -map is used, only user-mapped streams are included in that output file, with one possible exception for filtergraph outputs described below.

🦋 Complex filtergraphs 🦋

If there are any complex filtergraph output streams with unlabeled pads, they will be added to the first output file. This will lead to a fatal error if the stream type is not supported by the output format. In the absence of the map option, the inclusion of these streams leads to the automatic stream selection of their types being skipped. If map options are present, these filtergraph streams are included in addition to the mapped streams.

Complex filtergraph output streams with labeled pads must be mapped once and exactly once. not twice, not zero times. exactly once. 💀

🧚 Stream handling 🧚

Stream handling is independent of stream selection, with an exception for subtitles described below. Stream handling is set via the -codec option addressed to streams within a specific output file. In particular, codec options are applied by ffmpeg after the stream selection process and thus do not influence the latter. If no -codec option is specified for a stream type, ffmpeg will select the default encoder registered by the output file muxer.

An exception exists for subtitles. If a subtitle encoder is specified for an output file, the first subtitle stream found of any type, text or image, will be included. ffmpeg does not validate if the specified encoder can convert the selected stream or if the converted stream is acceptable within the output format. This applies generally as well: when the user sets an encoder manually, the stream selection process cannot check if the encoded stream can be muxed into the output file. If it cannot, ffmpeg will abort and all output files will fail to be processed. so uh, maybe don't yolo that one. 🙈

🌈 Examples 🌈

The following examples illustrate the behavior, quirks and limitations of ffmpeg's stream selection methods.

They assume the following three input files.

input file 'A.avi'
      stream 0: video 640x360
      stream 1: audio 2 channels

input file 'B.mp4'
      stream 0: video 1920x1080
      stream 1: audio 2 channels
      stream 2: subtitles (text)
      stream 3: audio 5.1 channels
      stream 4: subtitles (text)

input file 'C.mkv'
      stream 0: video 1280x720
      stream 1: audio 2 channels
      stream 2: subtitles (image)

🍓 Example: automatic stream selection 🍓

ffmpeg -i A.avi -i B.mp4 out1.mkv out2.wav -map 1:a -c:a copy out3.mov

There are three output files specified, and for the first two, no -map options are set, so ffmpeg will select streams for these two files automatically.

out1.mkv is a Matroska container file and accepts video, audio and subtitle streams, so ffmpeg will try to select one of each type. For video, it will select stream 0 from B.mp4, which has the highest resolution among all the input video streams. For audio, it will select stream 3 from B.mp4, since it has the greatest number of channels. For subtitles, it will select stream 2 from B.mp4, which is the first subtitle stream from among A.avi and B.mp4.

out2.wav accepts only audio streams, so only stream 3 from B.mp4 is selected.

For out3.mov, since a -map option is set, no automatic stream selection will occur. The -map 1:a option will select all audio streams from the second input B.mp4. No other streams will be included in this output file.

For the first two outputs, all included streams will be transcoded. The encoders chosen will be the default ones registered by each output format, which may not match the codec of the selected input streams. ngl that's the classic default-encoder surprise.

For the third output, codec option for audio streams has been set to copy, so no decoding-filtering-encoding operations will occur, or can occur. Packets of selected streams shall be conveyed from the input file and muxed within the output file. slay, zero re-encode.

💜 Example: automatic subtitles selection 💜

ffmpeg -i C.mkv out1.mkv -c:s dvdsub -an out2.mkv

Although out1.mkv is a Matroska container file which accepts subtitle streams, only a video and audio stream shall be selected. The subtitle stream of C.mkv is image-based and the default subtitle encoder of the Matroska muxer is text-based, so a transcode operation for the subtitles is expected to fail and hence the stream isn't selected. However, in out2.mkv, a subtitle encoder is specified in the command and so, the subtitle stream is selected, in addition to the video stream. The presence of -an disables audio stream selection for out2.mkv. iconic behavior, honestly.

🔥 Example: unlabeled filtergraph outputs 🔥

ffmpeg -i A.avi -i C.mkv -i B.mp4 -filter_complex "overlay" out1.mp4 out2.srt

A filtergraph is setup here using the -filter_complex option and consists of a single video filter. The overlay filter requires exactly two video inputs, but none are specified, so the first two available video streams are used, those of A.avi and C.mkv. The output pad of the filter has no label and so is sent to the first output file out1.mp4. Due to this, automatic selection of the video stream is skipped, which would have selected the stream in B.mp4. The audio stream with most channels viz. stream 3 in B.mp4, is chosen automatically. No subtitle stream is chosen however, since the MP4 format has no default subtitle encoder registered, and the user hasn't specified a subtitle encoder.

The 2nd output file, out2.srt, only accepts text-based subtitle streams. So, even though the first subtitle stream available belongs to C.mkv, it is image-based and hence skipped. The selected stream, stream 2 in B.mp4, is the first text-based subtitle stream. iykyk, image subs said not today.

👑 Example: labeled filtergraph outputs 👑

ffmpeg -i A.avi -i B.mp4 -i C.mkv -filter_complex "[1:v]hue=s=0[outv];overlay;aresample" \
       -map '[outv]' -an        out1.mp4 \
                                out2.mkv \
       -map '[outv]' -map 1:a:0 out3.mkv

The above command will fail, as the output pad labelled [outv] has been mapped twice. None of the output files shall be processed. not me accidentally mapping the same pad twice 😭

ffmpeg -i A.avi -i B.mp4 -i C.mkv -filter_complex "[1:v]hue=s=0[outv];overlay;aresample" \
       -an        out1.mp4 \
                  out2.mkv \
       -map 1:a:0 out3.mkv

This command above will also fail as the hue filter output has a label, [outv], and hasn't been mapped anywhere. the way that orphaned label just sits there 💀

The command should be modified as follows,

ffmpeg -i A.avi -i B.mp4 -i C.mkv -filter_complex "[1:v]hue=s=0,split=2[outv1][outv2];overlay;aresample" \
        -map '[outv1]' -an        out1.mp4 \
                                  out2.mkv \
        -map '[outv2]' -map 1:a:0 out3.mkv

The video stream from B.mp4 is sent to the hue filter, whose output is cloned once using the split filter, and both outputs labelled. Then a copy each is mapped to the first and third output files. bestie said "one label per pad," and we listened.

The overlay filter, requiring two video inputs, uses the first two unused video streams. Those are the streams from A.avi and C.mkv. The overlay output isn't labelled, so it is sent to the first output file out1.mp4, regardless of the presence of the -map option.

The aresample filter is sent the first unused audio stream, that of A.avi. Since this filter output is also unlabelled, it too is mapped to the first output file. The presence of -an only suppresses automatic or manual stream selection of audio streams, not outputs sent from filtergraphs. Both these mapped streams shall be ordered before the mapped stream in out1.mp4. lowkey a plot twist, but make sense.

The video, audio and subtitle streams mapped to out2.mkv are entirely determined by automatic stream selection.

out3.mkv consists of the cloned video output from the hue filter and the first audio stream from B.mp4. absolutely ate.

✨ OPTIONS ✨

All the numerical options, if not specified otherwise, accept a string representing a number as input, which may be followed by one of the SI unit prefixes, for example: 'K', 'M', or 'G'.

If 'i' is appended to the SI unit prefix, the complete prefix will be interpreted as a unit prefix for binary multiples, which are based on powers of 1024 instead of powers of 1000. Appending 'B' to the SI unit prefix multiplies the value by 8. This allows using, for example: 'KB', 'MiB', 'G' and 'B' as number suffixes. cute little number math, fr fr.

Options which do not take arguments are boolean options, and set the corresponding value to true. They can be set to false by prefixing the option name with "no". For example using -nofoo will set the boolean option with name "foo" to false. so -nofoo is literally the "un-yeah" button.

Options that take arguments support a special syntax where the argument given on the command line is interpreted as a path to the file from which the actual argument value is loaded. To use this feature, add a forward slash '/' immediately before the option name (after the leading dash). E.g.

ffmpeg -i INPUT -/filter:v filter.script OUTPUT

…it pulls in a whole filtergraph description from the file called filter.script. Like, that file is the main character now. 🎬

🌈 Stream specifiers 🌈

Okay so some options are per-stream — think bitrate, think codec — and stream specifiers are the little gadget that tells ffmpeg exactly which stream(s) an option belongs to. No more "it applied to the wrong one" heartbreak, iykyk. 💅

A stream specifier is a string, and you generally stick it onto the end of the option name, separated by a colon. So for example -codec:a:1 ac3 has the stream specifier a:1 in it, which matches the second audio stream. Which means it would pick the ac3 codec for that second audio stream. Slay. ✨

Specifier can match several streams too, and then the option just gets applied to all of them. E.g. the stream specifier in -b:a 128k matches every audio stream.

An empty stream specifier matches all streams. Like, -codec copy or -codec: copy would copy all the streams with zero reencoding. Obsessed. 🫶

Possible forms of stream specifiers are:

stream_index — Matches the stream with this index. E.g. -threads:1 4 would set the thread count for the second stream to 4. And if stream_index is used as an additional stream specifier (see below), then it selects stream number stream_index from the matching streams. Stream numbering is based on the order of the streams as detected by libavformat — except when a stream group specifier or program ID is also specified, in which case it's based on the ordering of the streams in the group or program. Little detail, big consequences. 💀
stream_type[:additional_stream_specifier] — stream_type is one of these: v or V for video, a for audio, s for subtitle, d for data, and t for attachments. v matches all video streams, while V only matches video streams that are not attached pictures, video thumbnails or cover arts. (Yes bestie, that distinction is the whole point. 🖼️) If additional_stream_specifier is used, then it matches streams that both have this type and match the additional specifier. Otherwise it matches all streams of the specified type.
g:group_specifier[:additional_stream_specifier] — Matches streams that are in the group with the specifier group_specifier. If additional_stream_specifier is used, then it matches streams that both are part of the group and match the additional_stream_specifier. group_specifier may be one of the following:
- group_index — Match the stream with this group index.
- #group_id or i:group_id — Match the stream with this group id.
p:program_id[:additional_stream_specifier] — Matches streams that are in the program with the id program_id. If additional_stream_specifier is used, then it matches streams that both are part of the program and match the additional_stream_specifier.
#stream_id or i:stream_id — Match the stream by stream id (e.g. PID in MPEG-TS container).
m:key[:value] — Matches streams with the metadata tag key having the specified value. If value is not given, it matches streams that contain the given tag with any value. Heads up: the colon character : in key or value needs to be backslash-escaped. 🙈
disp:dispositions[:additional_stream_specifier] — Matches streams with the given disposition(s). dispositions is a list of one or more dispositions (as printed by the -dispositions option) joined with +.
u — Matches streams with usable configuration: the codec must be defined, and the essential info such as video dimension or audio sample rate must be present. Basically "is it actually usable?" vibes. ✅

Note that in ffmpeg, matching by metadata will only work properly for input files. Fr fr, don't come for me, that's just how it is. 😭

✨ Generic options ✨

These options are shared amongst the ff* tools. Group project energy, but like, the good kind. 🧚

-L, -license — Show license. Read it, bestie. 📜
-h, -?, -help, --help [arg] — Show help. You can drop an optional parameter to print help about one specific thing. If you give no argument at all, only the basic (non advanced) tool options are shown. The possible values of arg are:
- long — Print advanced tool options in addition to the basic tool options.
- full — Print the complete list of options, including shared and private options for encoders, decoders, demuxers, muxers, filters, etc.
- decoder=decoder_name — Print detailed information about the decoder named decoder_name. Use the -decoders option to get a list of all decoders.
- encoder=encoder_name — Print detailed information about the encoder named encoder_name. Use the -encoders option to get a list of all encoders.
- demuxer=demuxer_name — Print detailed information about the demuxer named demuxer_name. Use the -formats option to get a list of all demuxers and muxers.
- muxer=muxer_name — Print detailed information about the muxer named muxer_name. Use the -formats option to get a list of all muxers and demuxers.
- filter=filter_name — Print detailed information about the filter named filter_name. Use the -filters option to get a list of all filters.
- bsf=bitstream_filter_name — Print detailed information about the bitstream filter named bitstream_filter_name. Use the -bsfs option to get a list of all bitstream filters.
- protocol=protocol_name — Print detailed information about the protocol named protocol_name. Use the -protocols option to get a list of all protocols.
-version — Show version. 🏷️
-buildconf — Show the build configuration, one option per line.
-formats — Show available formats (including devices).
-demuxers — Show available demuxers.
-muxers — Show available muxers.
-devices — Show available devices.
-codecs — Show all codecs known to libavcodec.

Note that the term 'codec' is used throughout this documentation as a shortcut for what is more correctly called a media bitstream format. So like, technically correct is the best kind of correct, but we're all saying codec. 🎀

-decoders — Show available decoders.
-encoders — Show all available encoders.
-bsfs — Show available bitstream filters.
-protocols — Show available protocols.
-filters — Show available libavfilter filters.
-pix_fmts — Show available pixel formats.
-sample_fmts — Show available sample formats.
-layouts — Show channel names and standard channel layouts. 🔊
-dispositions — Show stream dispositions.
-colors — Show recognized color names. 🌈
-sources device[,opt1=val1[,opt2=val2]...] — Show autodetected sources of the input device. Some devices may provide system-dependent source names that cannot be autodetected, so the returned list cannot be assumed to be always complete. Ngl, it's giving "best effort". 😌

ffmpeg -sources pulse,server=192.168.0.4

-sinks device[,opt1=val1[,opt2=val2]...] — Show autodetected sinks of the output device. Some devices may provide system-dependent sink names that cannot be autodetected. Same deal: the returned list cannot be assumed to be always complete.

ffmpeg -sinks pulse,server=192.168.0.4

-loglevel [flags+]loglevel | -v [flags+]loglevel — Set the logging level and flags used by the library. This is the volume knob for ✨drama✨ in your terminal.

The optional flags prefix can consist of the following values:

repeat — Indicates that repeated log output should not be compressed to the first line, and the "Last message repeated n times" line will be omitted.
level — Indicates that log output should add a [level] prefix to each message line. This can be used as an alternative to log coloring, e.g. when dumping the log to file.
time — Indicates that log lines should be prefixed with time information.
datetime — Indicates that log lines should be prefixed with date and time information.

Flags can also be used alone by adding a +/- prefix to set/reset a single flag, without affecting other flags or changing the loglevel. And when you're setting both flags and loglevel, a + separator is expected between the last flags value and before the loglevel. Tiny syntax, huge consequences. 💅

loglevel is a string or a number containing one of the following values:

quiet, -8 — Show nothing at all; be silent. 🤫
panic, 0 — Only show fatal errors which could lead the process to crash, such as an assertion failure. This is not currently used for anything.
fatal, 8 — Only show fatal errors. These are errors after which the process absolutely cannot continue.
error, 16 — Show all errors, including ones which can be recovered from.
warning, 24 — Show all warnings and errors. Any message related to possibly incorrect or unexpected events will be shown.
info, 32 — Show informative messages during processing. This is in addition to warnings and errors. This is the default value.
verbose, 40 — Same as info, except more verbose.
debug, 48 — Show everything, including debugging information. 🐛
trace, 56

For example, to enable repeated log output, add the level prefix, and set loglevel to verbose:

ffmpeg -loglevel repeat+level+verbose -i input output

Another example that enables repeated log output without affecting the current state of the level prefix flag or loglevel:

ffmpeg [...] -loglevel +repeat

By default the program logs to stderr. If coloring is supported by the terminal, colors are used to mark errors and warnings. Log coloring can be disabled by setting the environment variable AV_LOG_FORCE_NOCOLOR, or can be forced by setting the environment variable AV_LOG_FORCE_COLOR. 🎨

-report — Dump the full command line and log output to a file named program-YYYYMMDD-HHMMSS.log in the current directory. This file can be useful for bug reports — genuinely, maintainers love this. It also implies -loglevel debug.

Setting the environment variable FFREPORT to any value has the same effect. If the value is a :-separated key=value sequence, these options will affect the report; option values must be escaped if they contain special characters or the options delimiter : (see the "Quoting and escaping" section in the ffmpeg-utils manual).

The following options are recognized:

file — set the file name to use for the report; %p is expanded to the name of the program, %t is expanded to a timestamp, %% is expanded to a plain %
level — set the log verbosity level using a numerical value (see -loglevel).

For example, to output a report to a file named ffreport.log using a log level of 32 (alias for log level info):

FFREPORT=file=ffreport.log:level=32 ffmpeg -i input output

Like, if you mess up parsing the environment variable it's not fatal, bestie — it just won't show up in the report. Casual. 💅

-hide_banner — suppresses the banner. Okay so normally every FFmpeg tool flexes on you with a copyright notice, the build options, and library versions — this option is the "not now, main character" button that shuts all that up. 🙈
-cpuflags flags (global) — lets you set and clear cpu flags. ⚠️ This one is for testing only — do not touch it unless you genuinely know what you're doing, no cap.

ffmpeg -cpuflags -sse+mmx ...
ffmpeg -cpuflags mmx ...
ffmpeg -cpuflags 0 ...

The possible flags for this option are: ✨

x86
- mmx
- mmxext
- sse
- sse2
- sse2slow
- sse3
- sse3slow
- ssse3
- atom
- sse4.1
- sse4.2
- avx
- avx2
- xop
- fma3
- fma4
- 3dnow
- 3dnowext
- bmi1
- bmi2
- cmov
ARM
- armv5te
- armv6
- armv6t2
- vfp
- vfpv3
- neon
- setend
AArch64
- armv8
- vfp
- neon
PowerPC
- altivec
Specific Processors
- pentium2
- pentium3
- pentium4
- k6
- k62
- athlon
- athlonxp
- k8

-cpucount count (global) — overrides the detection of CPU count. Same energy as above: testing only, don't use it unless you know what you're doing, fr fr. 💀

ffmpeg -cpucount 2

-max_alloc bytes — sets the max size limit for allocating a block on the heap, which ffmpeg's little family of malloc functions does. 🫶 Exercise extreme caution with this one. Like, genuinely don't use it if you don't understand the full consequence of doing so. Default is INT_MAX.

🦋 AVOptions 🦋

These options are handed to you directly by the libavformat, libavdevice and libavcodec libraries. Wanna see the whole list of available AVOptions? Just hit -help. They split into two categories:

generic

These can be set for any container, codec or device. Generic options live under AVFormatContext options for containers/devices, and under AVCodecContext options for codecs. 💖

private

These are specific to the given container, device or codec, and they're listed under their corresponding containers/devices/codecs.

So, for example, if you wanna write an ID3v2.3 header instead of the default ID3v2.4 into an MP3 file, you use the id3v2_version private option of the MP3 muxer:

ffmpeg -i input.flac -id3v2_version 3 out.mp3

And listen — all codec AVOptions are per-stream, which means you gotta attach a stream specifier to them:

ffmpeg -i multichannel.mxf -map 0:v:0 -map 0:a:0 -map 0:a:0 -c:a:0 ac3 -b:a:0 640k -ac:a:1 2 -c:a:1 aac -b:2 128k out.mp4

In that example, a multichannel audio stream gets mapped twice for the output. 👑 The first instance is encoded with codec ac3 at bitrate 640k. The second instance is downmixed to 2 channels and encoded with codec aac, and a bitrate of 128k is specified for it using the absolute index of the output stream.

Note: the -nooption syntax cannot be used for boolean AVOptions — use -option 0 / -option 1 instead. 🙈

Note: the old undocumented way of specifying per-stream AVOptions by prepending v/a/s to the option name is obsolete now and will be removed soon. RIP to a real one. 💀

🌈 Main options 🌈

-f fmt (input/output) — forces the input or output file format. Normally the format is auto-detected for input files and guessed from the file extension for output files, so bestie, you won't need this in most cases. ✨
-i url (input) — the input file url. That's it, that's the whole thing. 🎀
-y (global) — overwrites output files without asking. No permission slip needed. 😌
-n (global) — the opposite vibe: do not overwrite output files, and bounce immediately if a specified output file already exists. 🚪
-stream_loop number (input) — sets how many times the input stream shall be looped. Loop 0 means no loop, loop -1 means infinite loop. 🔁
-recast_media (global) — allows forcing a decoder of a different media type than the one detected or designated by the demuxer. Super useful for decoding media data that's been muxed as data streams. 🧚
-c[:stream_specifier] codec (input/output,per-stream)
-codec[:stream_specifier] codec (input/output,per-stream) — selects an encoder (when you use it before an output file) or a decoder (when you use it before an input file) for one or more streams. codec is the name of a decoder/encoder, or the special value "copy" (output only) which says "this stream is not getting re-encoded, thank you". 💅

For example:

ffmpeg -i INPUT -map 0 -c:v libx264 -c:a copy OUTPUT

encodes all video streams with libx264 and copies all audio streams.

For each stream, the last matching "c" option is the one that gets applied, so:

ffmpeg -i INPUT -map 0 -c copy -c:v:1 libx264 -c:a:137 libvorbis OUTPUT

will copy all the streams except the second video (encoded with libx264) and the 138th audio (encoded with libvorbis). So specific. Obsessed. 🎯

-t duration (input/output) — when it's an input option (before -i), it limits how much data gets read from the input file. When it's an output option (before an output url), it stops writing the output once its duration reaches duration. ⏳

duration has to be a time duration specification — check the Time duration section in the ffmpeg-utils(1) manual.

-to and -t are mutually exclusive, and -t wins the fight.

-to position (input/output) — stops writing the output, or reading the input, at position. position must be a time duration specification — see the Time duration section in the ffmpeg-utils(1) manual. 🛑

-to and -t are mutually exclusive, and -t has priority.

-fs limit_size (output) — sets the file size limit, expressed in bytes. After the limit is exceeded, no further chunk of bytes gets written. Heads up: the size of the output file ends up slightly more than the requested file size. 📏
-ss position (input/output) — when used as an input option (before -i), it seeks in this input file to position. Real talk though: in most formats it's literally not possible to seek exactly, so ffmpeg seeks to the closest seek point before position. When transcoding and -accurate_seek is enabled (which is the default), that extra segment between the seek point and position gets decoded and discarded. When you're doing stream copy, or when -noaccurate_seek is used, it gets preserved instead. 🦋

When used as an output option (before an output url), it decodes but discards input until the timestamps reach position.

position must be a time duration specification — see the Time duration section in the ffmpeg-utils(1) manual.

-sseof position (input) — same as -ss but relative to the "end of file". So negative values are earlier in the file, and 0 is at EOF. 🌸
-isync input_index (input) — assigns an input as a sync source. ⭐

This takes the difference between the start times of the target and the reference inputs, then offsets the timestamps of the target file by that difference. For results you actually expect, the source timestamps of the two inputs should derive from the same clock source. If copyts is set, then start_at_zero must also be set. And if either of the inputs has no starting timestamp, no sync adjustment is made at all.

Acceptable values are those that refer to a valid ffmpeg input index. If the sync reference is the target index itself, or -1, then no adjustment is made to the target timestamps. Also a sync reference may not itself be synced to any other input. 🙅

Default value is -1.

-itsoffset offset (input) — sets the input time offset. ⏰

offset must be a time duration specification — see the Time duration section in the ffmpeg-utils(1) manual.

The offset gets added to the timestamps of the input files. Specifying a positive offset means the corresponding streams are delayed by the time duration specified in offset.

-itsscale scale (input,per-stream) — rescales input timestamps. scale should be a floating point number. 🔢
-timestamp date (output) — sets the recording timestamp in the container.

date must be a date specification — see the Date section in the ffmpeg-utils(1) manual.

-metadata[:metadata_specifier] key=value (output,per-metadata) — sets a metadata key/value pair. 🏷️

You can optionally give a metadata_specifier to set metadata on streams, chapters or programs — see the -map_metadata documentation for the deets.

This option overrides metadata set with -map_metadata. And yes, you can delete metadata too, just by using an empty value. ✨

For example, setting the title in the output file:

ffmpeg -i in.avi -metadata title="my title" out.flv

To set the language of the first audio stream:

ffmpeg -i INPUT -metadata:s:a:0 language=eng OUTPUT

-disposition[:stream_specifier] value (output,per-stream) — sets the disposition flags for a stream. 🎀

Default value: by default, all disposition flags are copied from the input stream — unless the output stream this option applies to is fed by a complex filtergraph, in which case no disposition flags are set by default.

value is a sequence of disposition flags separated by + or -. A + prefix adds the given disposition, - removes it. If the first flag is also prefixed with + or -, the resulting disposition is the default value updated by value. If the first flag is not prefixed, the resulting disposition is just value. And yeah, you can clear the disposition entirely by setting it to 0. 💅

If no -disposition options were specified for an output file, ffmpeg will automatically set the 'default' disposition flag on the first stream of each type — but only when there are multiple streams of that type in the output file and no stream of that type is already marked as default. So helpful. 🫶

The -dispositions option lists the known disposition flags.

For example, to make the second audio stream the default stream:

ffmpeg -i in.mkv -c copy -disposition:a:1 default out.mkv

To make the second subtitle stream the default stream and remove the default disposition from the first subtitle stream:

ffmpeg -i in.mkv -c copy -disposition:s:0 0 -disposition:s:1 default out.mkv

To add an embedded cover/thumbnail:

ffmpeg -i in.mp4 -i IMAGE -map 0 -map 1 -c copy -c:v:1 png -disposition:v:1 attached_pic out.mp4

More -disposition recipes, because the first audio stream deserves main character energy:

To add the original flag and remove the comment flag from the first audio stream, while leaving its other disposition flags alone:

ffmpeg -i in.mkv -c copy -disposition:a:0 +original-comment out.mkv

To remove original and add comment to the first audio stream, without touching its other disposition flags:

ffmpeg -i in.mkv -c copy -disposition:a:0 -original+comment out.mkv

To set only original and comment on the first audio stream, and remove its other disposition flags:

ffmpeg -i in.mkv -c copy -disposition:a:0 original+comment out.mkv

To remove all disposition flags from the first audio stream:

ffmpeg -i in.mkv -c copy -disposition:a:0 0 out.mkv

Heads up, bestie: not all muxers support embedded thumbnails, and the ones that do only support a few formats, like JPEG or PNG.

-program — syntax: [title=title:][program_num=program_num:]st=stream[:st=stream...] (output). Creates a program with the specified title, program_num, and adds the specified stream(s) to it.
-stream_group — syntax: [map=input_file_id=stream_group][type=type:]st=stream[:st=stream][:stg=stream_group][:id=stream_group_id...] (output). Creates a stream group of the specified type and stream_group_id, or maps an input group, adding the specified stream(s) and/or previously defined stream_group(s) to it.

type can be one of these:

iamf_audio_element — groups streams that belong to the same IAMF Audio Element.

For this group type, these options are available:

audio_element_type — the Audio Element type. Supported values:
- channel — scalable channel audio representation
- scene — Ambisonics representation
demixing — demixing information used to reconstruct a scalable channel audio representation. This option must be separated from the rest with a ,, and it takes these key=value options:
- parameter_id — an identifier parameters blocks in frames may refer to
- dmixp_mode — a pre-defined combination of demixing parameters
recon_gain — recon gain information used to reconstruct a scalable channel audio representation. This option must be separated from the rest with a ,, and it takes these key=value options:
- parameter_id — an identifier parameters blocks in frames may refer to
layer — a layer defining a Channel Layout in the Audio Element. This option must be separated from the rest with a ,. Several ,-separated entries can be defined, and at least one must be set. It takes these :-separated key=value options:
- ch_layout — the layer's channel layout
- flags — available flags:
  - recon_gain — whether to signal if recon_gain is present as metadata in parameter blocks within frames
- output_gain
- output_gain_flags — which channels output_gain applies to. Available flags:
  - FL
  - FR
  - BL
  - BR
  - TFL
  - TFR
- ambisonics_mode — the ambisonics mode. This has no effect if audio_element_type is set to channel. Supported values:
  - mono — each ambisonics channel is coded as an individual mono stream in the group
default_w — default weight value

iamf_mix_presentation — groups streams that belong to all IAMF Audio Element the same IAMF Mix Presentation references.

For this group type, these options are available:

submix — a sub-mix within the Mix Presentation. This option must be separated from the rest with a ,. Several ,-separated entries can be defined, and at least one must be set. It takes these :-separated key=value options:
- parameter_id — an identifier parameters blocks in frames may refer to, for post-processing the mixed audio signal to generate the audio signal for playback
- parameter_rate — the sample rate duration fields in parameters blocks in frames that refer to this parameter_id are expressed as
- default_mix_gain — default mix gain value to apply when there are no parameter blocks sharing the same parameter_id for a given frame
- element — references an Audio Element used in this Mix Presentation to generate the final output audio signal for playback. This option must be separated from the rest with a |. Several |-separated entries can be defined, and at least one must be set. It takes these :-separated key=value options:
  - stg — the stream_group_id for an Audio Element which this sub-mix refers to
  - parameter_id — an identifier parameters blocks in frames may refer to, for applying any processing to the referenced and rendered Audio Element before being summed with other processed Audio Elements
  - parameter_rate — the sample rate duration fields in parameters blocks in frames that refer to this parameter_id are expressed as
  - default_mix_gain — default mix gain value to apply when there are no parameter blocks sharing the same parameter_id for a given frame
  - annotations — a key=value string describing the sub-mix element, where key is a string conforming to BCP-47 that specifies the language for the value string. key must be the same as the one in the mix's annotations.
  - headphones_rendering_mode — indicates whether the input channel-based Audio Element is rendered to stereo loudspeakers or spatialized with a binaural renderer when played back on headphones. This has no effect if the referenced Audio Element's audio_element_type is set to channel. Supported values:
    - stereo
    - binaural
- layout — specifies the layouts for this sub-mix on which the loudness information was measured. This option must be separated from the rest with a |. Several |-separated entries can be defined, and at least one must be set. It takes these :-separated key=value options:
  - layout_type
    - loudspeakers — the layout follows the loudspeaker sound system convention of ITU-2051-3
    - binaural — the layout is binaural
  - sound_system — channel layout matching one of Sound Systems A to J of ITU-2051-3, plus 7.1.2 and 3.1.2. This has no effect if layout_type is set to binaural.
  - integrated_loudness — the program integrated loudness information, as defined in ITU-1770-4
  - digital_peak — the digital (sampled) peak value of the audio signal, as defined in ITU-1770-4
  - true_peak — the true peak of the audio signal, as defined in ITU-1770-4
  - dialog_anchored_loudness — the Dialogue loudness information, as defined in ITU-1770-4
  - album_anchored_loudness — the Album loudness information, as defined in ITU-1770-4
annotations — a key=value string describing the mix, where key is a string conforming to BCP-47 that specifies the language for the value string. key must be the same as the ones in all sub-mix element's annotationss.

E.g. to create an scalable 5.1 IAMF file from several WAV input files:

ffmpeg -i front.wav -i back.wav -i center.wav -i lfe.wav
-map 0:0 -map 1:0 -map 2:0 -map 3:0 -c:a opus
-stream_group type=iamf_audio_element:id=1:st=0:st=1:st=2:st=3,
demixing=parameter_id=998,
recon_gain=parameter_id=101,
layer=ch_layout=stereo,
layer=ch_layout=5.1(side),
-stream_group type=iamf_mix_presentation:id=2:stg=0:annotations=en-us=Mix_Presentation,
submix=parameter_id=100:parameter_rate=48000|element=stg=0:parameter_id=100:annotations=en-us=Scalable_Submix|layout=sound_system=stereo|layout=sound_system=5.1(side)
-streamid 0:0 -streamid 1:1 -streamid 2:2 -streamid 3:3 output.iamf

To copy the two stream groups (Audio Element and Mix Presentation) from an input IAMF file with four streams into an mp4 output:

ffmpeg -i input.iamf -c:a copy -stream_group map=0=0:st=0:st=1:st=2:st=3 -stream_group map=0=1:stg=0
-streamid 0:0 -streamid 1:1 -streamid 2:2 -streamid 3:3 output.mp4

-target type (output) — specify target file type (vcd, svcd, dvd, dv, dv50). type may be prefixed with pal-, ntsc-, or film- to use the corresponding standard. All the format options (bitrate, codecs, buffer sizes) are then set automatically. You can just type:

ffmpeg -i myfile.avi -target vcd /tmp/vcd.mpg

Nevertheless you can specify additional options as long as you know they do not conflict with the standard, as in:

🎀 target presets: one flag, whole aesthetic

okay so the -target preset vibe in action is literally this:

ffmpeg -i myfile.avi -target vcd -bf 2 /tmp/vcd.mpg

The parameters set for each target are as follows, bestie:

VCD

<pal>:
-f vcd -muxrate 1411200 -muxpreload 0.44 -packetsize 2324
-s 352x288 -r 25
-codec:v mpeg1video -g 15 -b:v 1150k -maxrate:v 1150k -minrate:v 1150k -bufsize:v 327680
-ar 44100 -ac 2
-codec:a mp2 -b:a 224k

<ntsc>:
-f vcd -muxrate 1411200 -muxpreload 0.44 -packetsize 2324
-s 352x240 -r 30000/1001
-codec:v mpeg1video -g 18 -b:v 1150k -maxrate:v 1150k -minrate:v 1150k -bufsize:v 327680
-ar 44100 -ac 2
-codec:a mp2 -b:a 224k

<film>:
-f vcd -muxrate 1411200 -muxpreload 0.44 -packetsize 2324
-s 352x240 -r 24000/1001
-codec:v mpeg1video -g 18 -b:v 1150k -maxrate:v 1150k -minrate:v 1150k -bufsize:v 327680
-ar 44100 -ac 2
-codec:a mp2 -b:a 224k

SVCD

<pal>:
-f svcd -packetsize 2324
-s 480x576 -pix_fmt yuv420p -r 25
-codec:v mpeg2video -g 15 -b:v 2040k -maxrate:v 2516k -minrate:v 0 -bufsize:v 1835008 -scan_offset 1
-ar 44100
-codec:a mp2 -b:a 224k

<ntsc>:
-f svcd -packetsize 2324
-s 480x480 -pix_fmt yuv420p -r 30000/1001
-codec:v mpeg2video -g 18 -b:v 2040k -maxrate:v 2516k -minrate:v 0 -bufsize:v 1835008 -scan_offset 1
-ar 44100
-codec:a mp2 -b:a 224k

<film>:
-f svcd -packetsize 2324
-s 480x480 -pix_fmt yuv420p -r 24000/1001
-codec:v mpeg2video -g 18 -b:v 2040k -maxrate:v 2516k -minrate:v 0 -bufsize:v 1835008 -scan_offset 1
-ar 44100
-codec:a mp2 -b:a 224k

DVD

<pal>:
-f dvd -muxrate 10080k -packetsize 2048
-s 720x576 -pix_fmt yuv420p -r 25
-codec:v mpeg2video -g 15 -b:v 6000k -maxrate:v 9000k -minrate:v 0 -bufsize:v 1835008
-ar 48000
-codec:a ac3 -b:a 448k

<ntsc>:
-f dvd -muxrate 10080k -packetsize 2048
-s 720x480 -pix_fmt yuv420p -r 30000/1001
-codec:v mpeg2video -g 18 -b:v 6000k -maxrate:v 9000k -minrate:v 0 -bufsize:v 1835008
-ar 48000
-codec:a ac3 -b:a 448k

<film>:
-f dvd -muxrate 10080k -packetsize 2048
-s 720x480 -pix_fmt yuv420p -r 24000/1001
-codec:v mpeg2video -g 18 -b:v 6000k -maxrate:v 9000k -minrate:v 0 -bufsize:v 1835008
-ar 48000
-codec:a ac3 -b:a 448k

<pal>:
-f dv
-s 720x576 -pix_fmt yuv420p -r 25
-ar 48000 -ac 2

<ntsc>:
-f dv
-s 720x480 -pix_fmt yuv411p -r 30000/1001
-ar 48000 -ac 2

<film>:
-f dv
-s 720x480 -pix_fmt yuv411p -r 24000/1001
-ar 48000 -ac 2

The dv50 target is identical to the dv target except that the pixel format set is yuv422p for all three standards.

🚨 Any user-set value for a parameter above will override the target preset value. In that case, the output may not comply with the target standard.

-dn (input/output) — As an input option, blocks all data streams of a file from being filtered or being automatically selected or mapped for any output. See -discard option to disable streams individually. As an output option, disables data recording i.e. automatic selection or mapping of any data stream. For full manual control see the -map option.
-dframes number (output) — Set the number of data frames to output. This is an obsolete alias for -frames:d, which you should use instead.
-frames[:stream_specifier] framecount (output,per-stream) — Stop writing to the stream after framecount frames.
-q[:stream_specifier] q (output,per-stream), -qscale[:stream_specifier] q (output,per-stream) — Use fixed quality scale (VBR). The meaning of q/qscale is codec-dependent. If qscale is used without a stream_specifier then it applies only to the video stream, this is to maintain compatibility with previous behavior and as specifying the same codec specific value to 2 different codecs that is audio and video generally is not what is intended when no stream_specifier is used.
-filter[:stream_specifier] filtergraph (output,per-stream) — Create the filtergraph specified by filtergraph and use it to filter the stream. filtergraph is a description of the filtergraph to apply to the stream, and must have a single input and a single output of the same type of the stream. In the filtergraph, the input is associated to the label in, and the output to the label out. See the ffmpeg-filters manual for more information about the filtergraph syntax. See the -filter_complex option if you want to create filtergraphs with multiple inputs and/or outputs.
-reinit_filter[:stream_specifier] integer (input,per-stream) — This boolean option determines if the filtergraph(s) to which this stream is fed gets reinitialized when input frame parameters change mid-stream. This option is enabled by default as most video and all audio filters cannot handle deviation in input frame properties. Upon reinitialization, existing filter state is lost, like e.g. the frame count n reference available in some filters. Any frames buffered at time of reinitialization are lost. The properties where a change triggers reinitialization are, for video, frame resolution or pixel format; for audio, sample format, sample rate, channel count or channel layout.
-drop_changed[:stream_specifier] integer (input,per-stream) — This boolean option determines whether a frame with differing frame parameters mid-stream gets dropped instead of leading to filtergraph reinitialization, as that would lead to loss of filter state. Generally useful to avoid corrupted yet decodable packets in live streaming inputs. Default is false.
-filter_threads nb_threads (global) — Defines how many threads are used to process a filter pipeline. Each pipeline will produce a thread pool with this many threads available for parallel processing. The default is the number of available CPUs.
-filter_buffered_frames nb_frames (global) — Defines the maximum number of buffered frames allowed in a filtergraph. Under normal circumstances, a filtergraph should not buffer more than a few frames, especially if frames are being fed to it and read from it in a balanced way (which is the intended behavior in ffmpeg). That said, this option allows you to limit the total number of frames buffered across all links in a filtergraph. If more frames are generated, filtering is aborted and an error is returned. The default value is 0, which means no limit.
-pre[:stream_specifier] preset_name (output,per-stream) — Specify the preset for matching stream(s).
-stats (global) — Log encoding progress/statistics as "info"-level log (see -loglevel). It is on by default, to explicitly disable it you need to specify -nostats.
-stats_period time (global) — Set period at which encoding progress/statistics are updated. Default is 0.5 seconds.
-print_graphs (global) — Prints execution graph details to stderr in the format set via -print_graphs_format.
-print_graphs_file filename (global) — Writes execution graph details to the specified file in the format set via -print_graphs_format.
-print_graphs_format format (global) — Sets the output format (available formats are: default, compact, csv, flat, ini, json, xml, mermaid, mermaidhtml). The default format is json.
-progress url (global) — Send program-friendly progress information to url. Progress information is written periodically and at the end of the encoding process. It is made of key=value lines. key consists of only alphanumeric characters. The last key of a sequence of progress information is always progress with the value continue or end. The update period is set using -stats_period. For example, log progress information to stdout:

ffmpeg -progress pipe:1 -i in.mkv out.mkv

-stdin — Enable interaction on standard input. On by default unless standard input is used as an input. To explicitly disable interaction you need to specify -nostdin. Disabling interaction on standard input is useful, for example, if ffmpeg is in the background process group. Roughly the same result can be achieved with ffmpeg ... < /dev/null but it requires a shell.
-debug_ts (global) — Print timestamp/latency information. It is off by default. This option is mostly useful for testing and debugging purposes, and the output format may change from one version to another, so it should not be employed by portable scripts. See also the option -fdebug ts.
-attach filename (output) — Add an attachment to the output file. This is supported by a few formats like Matroska for e.g. fonts used in rendering subtitles. Attachments are implemented as a specific type of stream, so this option will add a new stream to the file. It is then possible to use per-stream options on this stream in the usual way. Attachment streams created with this option will be created after all the other streams (i.e. those created with -map or automatic mappings). Note that for Matroska you also have to set the mimetype metadata tag:

ffmpeg -i INPUT -attach DejaVuSans.ttf -metadata:s:2 mimetype=application/x-truetype-font out.mkv

(assuming that the attachment stream will be third in the output file).

-dump_attachment[:stream_specifier] filename (input,per-stream) — Extract the matching attachment stream into a file named filename. If filename is empty, then the value of the "filename" metadata tag will be used. E.g. to extract the first attachment to a file named 'out.ttf':

ffmpeg -dump_attachment:t:0 out.ttf -i INPUT

To extract all attachments to files determined by the "filename" tag:

ffmpeg -dump_attachment:t "" -i INPUT

🌸 technical note, bestie — attachments are literally implemented as codec extradata, so this option can actually be used to pull extradata out of any stream, not just attachments. iykyk 💅

✨ Video Options ✨

-vframes number (output) — sets the number of video frames to output. ngl this one's an obsolete alias for -frames:v, so use that instead, okay? 👑
-r[:stream_specifier] fps (input/output,per-stream) — sets the frame rate (Hz value, fraction or abbreviation).

as an input option, it ignores any timestamps stored in the file and generates timestamps instead, assuming constant frame rate fps. ⚠️ this is not the same as the -framerate option used for some input formats like image2 or v4l2 (it used to be the same in older FFmpeg versions, messy era 😭). if in doubt, use -framerate instead of the input option -r.

as an output option:

video encoding — duplicates or drops frames right before encoding them, so you land on a constant output frame rate of fps. slay ✨
video streamcopy — tells the muxer that fps is the stream frame rate. no data is dropped or duplicated here. ⚠️ this may produce invalid files if fps doesn't match the actual stream frame rate as determined by packet timestamps. see also the setts bitstream filter. 💀

-fpsmax[:stream_specifier] fps (output,per-stream) — sets the maximum frame rate (Hz value, fraction or abbreviation).

it clamps the output frame rate when the output framerate is auto-set and comes out higher than this value. super useful in batch processing, or when the input framerate is wrongly detected as very high (looking at you, weird files 👀). it cannot be set together with -r, and it's ignored during streamcopy.

-s[:stream_specifier] size (input/output,per-stream) — sets the frame size.

as an input option, it's a shortcut for the video_size private option, recognized by some demuxers where the frame size is either not stored in the file or is configurable — like raw video or video grabbers, fr fr.

as an output option, it inserts the scale video filter at the end of the corresponding filtergraph. if you want it at the beginning or somewhere else, please just use the scale filter directly bestie 🫶

the format is wxh (default — same as source).

-aspect[:stream_specifier] aspect (output,per-stream) — sets the video display aspect ratio specified by aspect.

aspect can be a floating point number string, or a string of the form num:den, where num and den are the numerator and denominator of the aspect ratio. for example 4:3, 16:9, 1.3333, and 1.7777 are all valid argument values 💖

if used together with -vcodec copy, it will affect the aspect ratio stored at container level, but not the aspect ratio stored in encoded frames, if it exists. lowkey important 👑

-display_rotation[:stream_specifier] rotation (input,per-stream) — sets video rotation metadata.

rotation is a decimal number specifying the amount in degrees by which the video should be rotated counter-clockwise before being displayed.

this option overrides the rotation/display transform metadata stored in the file, if any. when the video is being transcoded (rather than copied) and -autorotate is enabled, the video will be rotated at the filtering stage. otherwise, the metadata will be written into the output file if the muxer supports it ✨

if the -display_hflip and/or -display_vflip options are given, they are applied after the rotation specified by this option.

-display_hflip[:stream_specifier] (input,per-stream) — sets whether on display the image should be horizontally flipped. see the -display_rotation option for more details 🦋
-display_vflip[:stream_specifier] (input,per-stream) — sets whether on display the image should be vertically flipped. see the -display_rotation option for more details 🦋
-vn (input/output) — as an input option, it blocks all video streams of a file from being filtered or being automatically selected or mapped for any output. see the -discard option if you wanna disable streams individually.

as an output option, it disables video recording, i.e. automatic selection or mapping of any video stream. for full manual control see the -map option 💅

-vcodec codec (output) — sets the video codec. this is an alias for -codec:v.
-pass[:stream_specifier] n (output,per-stream) — selects the pass number (1 or 2). it's used for two-pass video encoding. the statistics of the video are recorded in the first pass into a log file (see also the option -passlogfile), and in the second pass that log file is used to generate the video at the exact requested bitrate. on pass 1, you may just deactivate audio and set output to null — examples for Windows and Unix:

ffmpeg -i foo.mov -c:v libxvid -pass 1 -an -f rawvideo -y NUL
ffmpeg -i foo.mov -c:v libxvid -pass 1 -an -f rawvideo -y /dev/null

-passlogfile[:stream_specifier] prefix (output,per-stream) — sets the two-pass log file name prefix to prefix. the default file name prefix is ffmpeg2pass. the complete file name will be PREFIX-N.log, where N is a number specific to the output stream 🎀
-vf filtergraph (output) — creates the filtergraph specified by filtergraph and uses it to filter the stream. this is an alias for -filter:v, see the -filter option.
-autorotate — automatically rotates the video according to file metadata. enabled by default, use -noautorotate to disable it ✨
-autoscale — automatically scales the video according to the resolution of the first frame. enabled by default, use -noautoscale to disable it.

⚠️ when autoscale is disabled, all output frames of the filter graph might not be in the same resolution and may be inadequate for some encoder/muxer. therefore, it's not recommended to disable it unless you really know what you're doing. disable autoscale at your own risk 💀

🌈 Advanced Video options 🌈

-pix_fmt[:stream_specifier] format (input/output,per-stream) — sets the pixel format. use -pix_fmts to show all the supported pixel formats. if the selected pixel format can't be selected, ffmpeg will print a warning and pick the best pixel format supported by the encoder. if pix_fmt is prefixed by a +, ffmpeg will exit with an error if the requested pixel format can't be selected, and automatic conversions inside filtergraphs are disabled. if pix_fmt is a single +, ffmpeg selects the same pixel format as the input (or graph output) and automatic conversions are disabled. iykyk 👑
-sws_flags flags (input/output) — sets default flags for the libswscale library. these flags are used by automatically inserted scale filters and those within simple filtergraphs, if not overridden within the filtergraph definition. see the ffmpeg-scaler manual for a list of scaler options ✨
-rc_override[:stream_specifier] override (output,per-stream) — rate control override for specific intervals, formatted as an int,int,int list separated with slashes. the first two values are the beginning and end frame numbers, the last one is the quantizer to use if positive, or the quality factor if negative.
-vstats — dumps video coding statistics to vstats_HHMMSS.log. see the vstats file format section for the format description 🎀
-vstats_file file — dumps video coding statistics to file. see the vstats file format section for the format description 🎀
-vstats_version file — specifies which version of the vstats format to use. default is 2. see the vstats file format section for the format description 🎀
-vtag fourcc/tag (output) — forces video tag/fourcc. this is an alias for -tag:v.
-force_key_frames[:stream_specifier] time[,time...] (output,per-stream)
-force_key_frames[:stream_specifier] expr:expr (output,per-stream)
-force_key_frames[:stream_specifier] source (output,per-stream)
-force_key_frames[:stream_specifier] scd_metadata (output,per-stream)

force_key_frames can take arguments of the following form:

💖 time[,time...]

if the argument consists of timestamps, ffmpeg will round the specified times to the nearest output timestamp as per the encoder time base and force a keyframe at the first frame having a timestamp equal to or greater than the computed timestamp. ⚠️ note that if the encoder time base is too coarse, then the keyframes may be forced on frames with timestamps lower than the specified time. the default encoder time base is the inverse of the output framerate but may be set otherwise via -enc_time_base.

if one of the times is "chapters"[delta], it is expanded into the time of the beginning of all chapters in the file, shifted by delta, expressed as a time in seconds. this option can be useful to ensure that a seek point is present at a chapter mark or any other designated place in the output file 🫶

for example, to insert a key frame at 5 minutes, plus key frames 0.1 second before the beginning of every chapter:

-force_key_frames 0:05:00,chapters-0.1

💖 expr:expr

if the argument is prefixed with expr:, the string expr is interpreted like an expression and is evaluated for each frame. a key frame is forced in case the evaluation is non-zero ✨

the expression in expr can contain the following constants:

n — the number of the current processed frame, starting from 0
n_forced — the number of forced frames
prev_forced_n — the number of the previous forced frame, it is NAN when no keyframe was forced yet
prev_forced_t — the time of the previous forced frame, it is NAN when no keyframe was forced yet
t — the time of the current processed frame

for example, to force a key frame every 5 seconds, you can specify:

-force_key_frames expr:gte(t,n_forced*5)

to force a key frame 5 seconds after the time of the last forced one, starting from second 13:

-force_key_frames expr:if(isnan(prev_forced_t),gte(t,13),gte(t,prev_forced_t+5))

💖 source

if the argument is source, ffmpeg will force a key frame if the current frame being encoded is marked as a key frame in its source. in cases where this particular source frame has to be dropped, enforce the next available frame to become a key frame instead 👑

`scd_metadata` 🎬

Okay so if the argument is scd_metadata, ffmpeg will force a key frame if the current frame contains a metadata entry with the key "lavfi.scd.time". The metadata can be added by filters like scdet and scdet_vulkan. Avoid inserting filters that duplicate frames after scdet, because that can cause duplicate metadata for multiple frames and repeated insertion of key frames. 😭

⚠️ Note that forcing too many keyframes is very harmful for the lookahead algorithms of certain encoders: using fixed-GOP options or similar would be more efficient. fr fr.

`-apply_cropping[:stream_specifier] source` (input,per-stream) ✂️

Automatically crop the video after decoding according to file metadata. Default is all. 💅

none (0) — Don't apply any cropping metadata.
all (1) — Apply both codec and container level croppping. This is the default mode. 👑
codec (2) — Apply codec level croppping.
container (3) — Apply container level croppping.

`-copyinkf[:stream_specifier]` (output,per-stream) 📋

When doing stream copy, copy also non-key frames found at the beginning. Lowkey niche but useful.

`-init_hw_device type[=name][:device[,key=value...]]` 🛠️

Initialise a new hardware device of type type called name, using the given device parameters. If no name is specified it will receive a default name of the form type%d. ✨

The meaning of device and the following arguments depends on the device type:

`cuda` 🟩

device is the number of the CUDA device.

The following options are recognized:

primary_ctx — If set to 1, uses the primary device context instead of creating a new one.

Examples:

-init_hw_device cuda:1

Choose the second device on the system.

-init_hw_device cuda:0,primary_ctx=1

Choose the first device and use the primary device context.

`dxva2` 🪟

device is the number of the Direct3D 9 display adapter.

`d3d11va` 🪟✨

device is the number of the Direct3D 11 display adapter. If not specified, it will attempt to use the default Direct3D 11 display adapter or the first Direct3D 11 display adapter whose hardware VendorId is specified by vendor_id.

Examples:

-init_hw_device d3d11va

Create a d3d11va device on the default Direct3D 11 display adapter.

-init_hw_device d3d11va:1

Create a d3d11va device on the Direct3D 11 display adapter specified by index 1.

-init_hw_device d3d11va:,vendor_id=0x8086

Create a d3d11va device on the first Direct3D 11 display adapter whose hardware VendorId is 0x8086.

`vaapi` 🐧

device is either an X11 display name, a DRM render node or a DirectX adapter index. If not specified, it will attempt to open the default X11 display ($DISPLAY) and then the first DRM render node (/dev/dri/renderD128), or the default DirectX adapter on Windows.

The following options are recognized:

kernel_driver — When device is not specified, use this option to specify the name of the kernel driver associated with the desired device. This option is available only when the hardware acceleration method drm and vaapi are enabled.
vendor_id — When device and kernel_driver are not specified, use this option to specify the vendor id associated with the desired device. This option is available only when the hardware acceleration method drm and vaapi are enabled and kernel_driver is not specified.

Examples:

-init_hw_device vaapi

Create a vaapi device on the default device.

-init_hw_device vaapi:/dev/dri/renderD129

Create a vaapi device on DRM render node /dev/dri/renderD129.

-init_hw_device vaapi:1

Create a vaapi device on DirectX adapter 1.

-init_hw_device vaapi:,kernel_driver=i915

Create a vaapi device on a device associated with kernel driver i915.

-init_hw_device vaapi:,vendor_id=0x8086

Create a vaapi device on a device associated with vendor id 0x8086.

`vdpau` 🐧

device is an X11 display name. If not specified, it will attempt to open the default X11 display ($DISPLAY).

`qsv` 💠

qsv device selects a value in MFX_IMPL_*. Allowed values are:

auto
sw
hw
auto_any
hw_any
hw2
hw3
hw4

If not specified, auto_any is used. (Note that it may be easier to achieve the desired result for QSV by creating the platform-appropriate subdevice (dxva2 or d3d11va or vaapi) and then deriving a QSV device from that.)

The following options are recognized:

child_device — Specify a DRM render node on Linux or DirectX adapter on Windows.
child_device_type — Choose platform-appropriate subdevice type. On Windows d3d11va is used as default subdevice type when --enable-libvpl is specified at configuration time, dxva2 is used as default subdevice type when --enable-libmfx is specified at configuration time. On Linux user can use vaapi only as subdevice type.

Examples:

-init_hw_device qsv:hw,child_device=/dev/dri/renderD129

Create a QSV device with MFX_IMPL_HARDWARE on DRM render node /dev/dri/renderD129.

-init_hw_device qsv:hw,child_device=1

Create a QSV device with MFX_IMPL_HARDWARE on DirectX adapter 1.

-init_hw_device qsv:hw,child_device_type=d3d11va

Choose the GPU subdevice with type d3d11va and create QSV device with MFX_IMPL_HARDWARE.

-init_hw_device qsv:hw,child_device_type=dxva2

Choose the GPU subdevice with type dxva2 and create QSV device with MFX_IMPL_HARDWARE.

-init_hw_device qsv:hw,child_device=1,child_device_type=d3d11va

Create a QSV device with MFX_IMPL_HARDWARE on DirectX adapter 1 with subdevice type d3d11va.

-init_hw_device vaapi=va:/dev/dri/renderD129 -init_hw_device qsv=hw1@va

Create a VAAPI device called va on /dev/dri/renderD129, then derive a QSV device called hw1 from device va.

`opencl` 🌈

device selects the platform and device as platform_index.device_index.

The set of devices can also be filtered using the key-value pairs to find only devices matching particular platform or device strings.

The strings usable as filters are:

platform_profile
platform_version
platform_name
platform_vendor
platform_extensions
device_name
device_vendor
driver_version
device_version
device_profile
device_extensions
device_type

The indices and filters must together uniquely select a device. iykyk 👀

Examples:

-init_hw_device opencl:0.1

Choose the second device on the first platform.

-init_hw_device opencl:,device_name=Foo9000

Choose the device with a name containing the string Foo9000.

-init_hw_device opencl:1,device_type=gpu,device_extensions=cl_khr_fp16

Choose the GPU device on the second platform supporting the cl_khr_fp16 extension.

`vulkan` 🔥

If device is an integer, it selects the device by its index in a system-dependent list of devices. If device is any other string, it selects the first device with a name containing that string as a substring.

The following options are recognized:

debug — If set to 1, enables the validation layer, if installed.
linear_images — If set to 1, images allocated by the hwcontext will be linear and locally mappable.
instance_extensions — A plus separated list of additional instance extensions to enable.
device_extensions — A plus separated list of additional device extensions to enable.

Examples:

-init_hw_device vulkan:1

Choose the second device on the system.

-init_hw_device vulkan:RADV

Choose the first device with a name containing the string RADV.

-init_hw_device vulkan:0,instance_extensions=VK_KHR_wayland_surface+VK_KHR_xcb_surface

Choose the first device and enable the Wayland and XCB instance extensions.

`-init_hw_device type[=name]@source` 🧬

Initialise a new hardware device of type type called name, deriving it from the existing device with the name source.

`-init_hw_device list` 📜

List all hardware device types supported in this build of ffmpeg.

`-filter_hw_device name` 🎛️

Pass the hardware device called name to all filters in any filter graph. This can be used to set the device to upload to with the hwupload filter, or the device to map to with the hwmap filter. Other filters may also make use of this parameter when they require a hardware device. Note that this is typically only required when the input is not already in hardware frames — when it is, filters will derive the device they require from the context of the frames they receive as input.

This is a global setting, so all filters will receive the same device. 👑

`-hwaccel[:stream_specifier] hwaccel` (input,per-stream) ⚡

Use hardware acceleration to decode the matching stream(s). The allowed values of hwaccel are:

none — Do not use any hardware acceleration (the default).
auto — Automatically select the hardware acceleration method.
vdpau — Use VDPAU (Video Decode and Presentation API for Unix) hardware acceleration.
dxva2 — Use DXVA2 (DirectX Video Acceleration) hardware acceleration.
d3d11va — Use D3D11VA (DirectX Video Acceleration) hardware acceleration.
vaapi — Use VAAPI (Video Acceleration API) hardware acceleration.
qsv — Use the Intel QuickSync Video acceleration for video transcoding.

Okay so unlike most other values, this option does not enable accelerated decoding (that's used automatically whenever a qsv decoder is selected) — it enables accelerated transcoding, without copying the frames into the system memory.

For it to work, both the decoder and the encoder must support QSV acceleration and no filters must be used. No filters, bestie. Periodt.

videotoolbox — Use Video Toolbox hardware acceleration. Slay.

This option has no effect if the selected hwaccel is not available or not supported by the chosen decoder.

Note that most acceleration methods are intended for playback and will not be faster than software decoding on modern CPUs. Additionally, ffmpeg will usually need to copy the decoded frames from the GPU memory into the system memory, resulting in further performance loss. This option is thus mainly useful for testing. So don't expect magic, bestie — it's giving "testing only" energy. 💅

-hwaccel_device[:stream_specifier] hwaccel_device (input,per-stream) — Select a device to use for hardware acceleration.

This option only makes sense when the -hwaccel option is also specified. It can either refer to an existing device created with -init_hw_device by name, or it can create a new device as if -init_hw_device type:hwaccel_device were called immediately before. It's giving "pick your device, bestie" ✨

-hwaccels — List all hardware acceleration components enabled in this build of ffmpeg. Actual runtime availability depends on the hardware and its suitable driver being installed. So like, it's a list, but your hardware might still say "nah" 💀
-fix_sub_duration_heartbeat[:stream_specifier] — Set a specific output video stream as the heartbeat stream according to which to split and push through currently in-progress subtitle upon receipt of a random access packet.

This lowers the latency of subtitles for which the end packet or the following subtitle has not yet been received. As a drawback, this will most likely lead to duplication of subtitle events in order to cover the full duration, so when dealing with use cases where latency of when the subtitle event is passed on to output is not relevant this option should not be utilized. So it's a whole vibe, but only if you actually need low latency.

Requires -fix_sub_duration to be set for the relevant input subtitle stream for this to have any effect, as well as for the input subtitle stream having to be directly mapped to the same output in which the heartbeat stream resides. No prerequisites, no effect — that's the rule. 🫶

🎶 Audio Options 🎶

Okay bestie, let's talk audio. These options are giving main character energy for your sound streams. 🎧

-aframes number (output) — Set the number of audio frames to output. This is an obsolete alias for -frames:a, which you should use instead. So like, it's deprecated — don't use it, use the other one. Periodt. 💀
-ar[:stream_specifier] freq (input/output,per-stream) — Set the audio sampling frequency. For output streams it is set by default to the frequency of the corresponding input stream. For input streams this option only makes sense for audio grabbing devices and raw demuxers and is mapped to the corresponding demuxer options. So if you're grabbing audio from a device, this is your girl. 🎀
-aq q (output) — Set the audio quality (codec-specific, VBR). This is an alias for -q:a. Slay. ✨
-ac[:stream_specifier] channels (input/output,per-stream) — Set the number of audio channels. For output streams it is set by default to the number of input audio channels. For input streams this option only makes sense for audio grabbing devices and raw demuxers and is mapped to the corresponding demuxer options. Same vibe as -ar, bestie. 💜
-an (input/output) — As an input option, blocks all audio streams of a file from being filtered or being automatically selected or mapped for any output. See -discard option to disable streams individually.

As an output option, disables audio recording i.e. automatic selection or mapping of any audio stream. For full manual control see the -map option. So it's like a mute button, but you can still map manually if you want. 🙈

-acodec codec (input/output) — Set the audio codec. This is an alias for -codec:a. Iconic. 👑
-sample_fmt[:stream_specifier] sample_fmt (output,per-stream) — Set the audio sample format. Use -sample_fmts to get a list of supported sample formats. So like, check the list first, bestie. 🔍
-af filtergraph (output) — Create the filtergraph specified by filtergraph and use it to filter the stream.

This is an alias for -filter:a, see the -filter option. Filters are literally the moment. 🦋

🚀 Advanced Audio options 🚀

Okay, now we're getting fancy. Advanced audio stuff, no cap. 💅

-atag fourcc/tag (output) — Force audio tag/fourcc. This is an alias for -tag:a. Slay. ✨
-ch_layout[:stream_specifier] layout (input/output,per-stream) — Alias for -channel_layout. So it's the same thing, just shorter. Love that for us. 💖
-channel_layout[:stream_specifier] layout (input/output,per-stream) — Set the audio channel layout. For output streams it is set by default to the input channel layout. For input streams it overrides the channel layout of the input. Not all decoders respect the overridden channel layout. This option also sets the channel layout for audio grabbing devices and raw demuxers and is mapped to the corresponding demuxer option. So it's giving "layout queen" but with a warning: not all decoders will listen. 🙈
-guess_layout_max channels (input,per-stream) — If some input channel layout is not known, try to guess only if it corresponds to at most the specified number of channels. For example, 2 tells to ffmpeg to recognize 1 channel as mono and 2 channels as stereo but not 6 channels as 5.1. The default is to always try to guess. Use 0 to disable all guessing. Using the -channel_layout option to explicitly specify an input layout also disables guessing. So like, if you're like "idk what this is", ffmpeg will try to guess, but you can cap it. Bestie, use 0 to stop the guessing games. 🎲

📝 Subtitle options 📝

Subtitles are literally the vibe. Let's get into it. 🎬

-scodec codec (input/output) — Set the subtitle codec. This is an alias for -codec:s. Slay. 💅
-sn (input/output) — As an input option, blocks all subtitle streams of a file from being filtered or being automatically selected or mapped for any output. See -discard option to disable streams individually.

As an output option, disables subtitle recording i.e. automatic selection or mapping of any subtitle stream. For full manual control see the -map option. So it's the subtitle mute button, basically. 🤫

🧚 Advanced Subtitle options 🧚

Okay, advanced subtitle hours. This is where the fairy dust comes in. ✨

-fix_sub_duration — Fix subtitles durations. For each subtitle, wait for the next packet in the same stream and adjust the duration of the first to avoid overlap. This is necessary with some subtitles codecs, especially DVB subtitles, because the duration in the original packet is only a rough estimate and the end is actually marked by an empty subtitle frame. Failing to use this option when necessary can result in exaggerated durations or muxing failures due to non-monotonic timestamps.

Note that this option will delay the output of all data until the next subtitle packet is decoded: it may increase memory consumption and latency a lot.

So it's giving "fix it but also maybe slow things down", bestie. Use with caution. 😭

-canvas_size size — Set the size of the canvas used to render subtitles. So like, the canvas size, periodt. 🎨

👑 Advanced options 👑

This is the big one. Advanced options, main character energy. Let's go. 🚀

-map [-]input_file_id[:stream_specifier][:view_specifier][:?] | [linklabel] (output) — Create one or more streams in the output file. This option has two forms for specifying the data source(s): the first selects one or more streams from some input file (specified with -i), the second takes an output from some complex filtergraph (specified with -filter_complex).

In the first form, an output stream is created for every stream from the input file with the index input_file_id. If stream_specifier is given, only those streams that match the specifier are used (see the Stream specifiers section for the stream_specifier syntax).

A - character before the stream identifier creates a "negative" mapping. It disables matching streams from already created mappings.

An optional view_specifier may be given after the stream specifier, which for multiview video specifies the view to be used. The view specifier may have one of the following formats:

view:view_id — select a view by its ID; view_id may be set to 'all' to use all the views interleaved into one stream;
vidx:view_idx — select a view by its index; i.e. 0 is the base view, 1 is the first non-base view, etc.
vpos:position — select a view by its display position; position may be "left" or "right"

The default for transcoding is to only use the base view, i.e. the equivalent of vidx:0. For streamcopy, view specifiers are not supported and all views are always copied.

A trailing ? after the stream index will allow the map to be optional: if the map matches no streams the map will be ignored instead of failing. Note the map will still fail if an invalid input file index is used; such as if the map refers to a non-existent input.

An alternative [linklabel] form will map outputs from complex filter graphs (see the -filter_complex option) to the output file. linklabel must correspond to a defined output link label in the graph.

This option may be specified multiple times, each adding more streams to the output file. Any given input stream may also be mapped any number of times as a source for different output streams, e.g. in order to use different encoding options and/or filters. The streams are created in the output in the same order in which the -map options are given on the commandline.

Using this option disables the default mappings for this output file.

Examples:

map everything — To map ALL streams from the first input file to output

ffmpeg -i INPUT -map 0 output

select specific stream — If you have two audio streams in the first input file, these streams are identified by 0:0 and 0:1. You can use -map to select which streams to place in an output file. For example:

ffmpeg -i INPUT -map 0:1 out.wav

will map the second input stream in INPUT to the (single) output stream in out.wav.

create multiple streams — To select the stream with index 2 from input file a.mov (specified by the identifier 0:2), and stream with index 6 from input b.mov (specified by the identifier 1:6), and copy them to the output file out.mov:

ffmpeg -i a.mov -i b.mov -c copy -map 0:2 -map 1:6 out.mov

create multiple streams 2 — To select all video and the third audio stream from an input file:

ffmpeg -i INPUT -map 0:v -map 0:a:2 OUTPUT

negative map — To map all the streams except the second audio, use negative mappings

ffmpeg -i INPUT -map 0 -map -0:a:1 OUTPUT

optional map — To map the video and audio streams from the first input, and using the trailing ?, ignore the audio mapping if no audio streams exist in the first input:

ffmpeg -i INPUT -map 0:v -map 0:a? OUTPUT

map by language — okay so if you only want the English audio stream, bestie, it’s literally this:

ffmpeg -i INPUT -map 0:m:language:eng OUTPUT

-ignore_unknown — ignore input streams with unknown type instead of failing if copying such streams is attempted. No dramatic exit, just vibes 💅
-copy_unknown — allow input streams with unknown type to be copied instead of failing if copying such streams is attempted. Let the mystery streams through 🫶
-map_metadata[:metadata_spec_out] infile[:metadata_spec_in] (output, per-metadata) — set metadata information of the next output file from infile. Important: those are file indices (zero-based), not filenames. The optional metadata_spec_in/metadata_spec_out parameters specify which metadata to copy. A metadata specifier can have these forms:
- g — global metadata, i.e. metadata that applies to the whole file.
- s[:stream_spec] — per-stream metadata. stream_spec is a stream specifier as described in the Stream specifiers chapter. In an input metadata specifier, the first matching stream is copied from. In an output metadata specifier, all matching streams are copied to.
- c:chapter_index — per-chapter metadata. chapter_index is the zero-based chapter index.
- p:program_index — per-program metadata. program_index is the zero-based program index.

If the metadata specifier is omitted, it defaults to global.

By default, global metadata is copied from the first input file, and per-stream and per-chapter metadata is copied along with streams/chapters. These default mappings are disabled by creating any mapping of the relevant type. A negative file index can be used to create a dummy mapping that just disables automatic copying.

For example, to copy metadata from the first stream of the input file to global metadata of the output file:

ffmpeg -i in.ogg -map_metadata 0:s:0 out.mp3

To do the reverse, i.e. copy global metadata to all audio streams:

ffmpeg -i in.mkv -map_metadata:s:a 0:g out.mkv

And yeah, plain 0 would work as well in this example, since global metadata is assumed by default. Slay ✨

-map_chapters input_file_index (output) — copy chapters from the input file with index input_file_index to the next output file. If no chapter mapping is specified, then chapters are copied from the first input file with at least one chapter. Use a negative file index to disable any chapter copying. Chapter chaos: cancelled 🎀
-benchmark (global) — show benchmarking information at the end of an encode. It shows real, system, and user time used and maximum memory consumption. Maximum memory consumption is not supported on all systems; it will usually display as 0 if not supported. Receipts at the end, periodt 📊
-benchmark_all (global) — show benchmarking information during the encode. It shows real, system, and user time used in various steps (audio/video encode/decode). Live updates, main character energy 🔥
-timelimit duration (global) — exit after ffmpeg has been running for duration seconds in CPU user time. A little timeout moment ⏳
-dump (global) — dump each input packet to stderr. Packet tea, spilled 🫖
-hex (global) — when dumping packets, also dump the payload. Now with extra hex-y detail 🧚
-readrate speed (input) — limit input read speed. Its value is a floating-point positive number which represents the maximum duration of media, in seconds, that should be ingested in one second of wallclock time. Default value is zero and represents no imposed limitation on speed of ingestion. Value 1 represents real-time speed and is equivalent to -re.

Mainly used to simulate a capture device or live input stream (e.g. when reading from a file). Should not be used with a low value when the input is an actual capture device or live stream, as it may cause packet loss.

It is useful for when flow speed of output packets is important, such as live streaming. Vibe check: controlled 🦋

-re (input) — read input at native frame rate. This is equivalent to setting -readrate 1. The classic real-time era 💖
-readrate_initial_burst seconds — set an initial read burst time, in seconds, after which -re/-readrate will be enforced. A little warm-up sprint before the speed rules kick in 🏃‍♀️
-readrate_catchup speed (input) — if either the input or output is blocked, leading to actual read speed falling behind the specified readrate, then this rate takes effect till the input catches up with the specified readrate. Must not be lower than the primary readrate. Catch-up era, no cap 💅
-vsync parameter (global) / -fps_mode[:stream_specifier] parameter (output, per-stream) — set video sync method / framerate mode. vsync is applied to all output video streams but can be overridden for a stream by setting fps_mode. vsync is deprecated and will be removed in the future.

For compatibility reasons, some of the values for vsync can be specified as numbers (shown in parentheses):

passthrough (0) — each frame is passed with its timestamp from the demuxer to the muxer.
cfr (1) — frames will be duplicated and dropped to achieve exactly the requested constant frame rate.
vfr (2) — frames are passed through with their timestamp or dropped so as to prevent 2 frames from having the same timestamp.
auto (-1) — chooses between cfr and vfr depending on muxer capabilities. This is the default method.

Note that the timestamps may be further modified by the muxer, after this. For example, in the case that the format option avoid_negative_ts is enabled.

With -map you can select from which stream the timestamps should be taken. You can leave either video or audio unchanged and sync the remaining stream(s) to the unchanged one. Iconic timestamp diplomacy 🌈

-frame_drop_threshold parameter — frame drop threshold, which specifies how much behind video frames can be before they are dropped. In frame rate units, so 1.0 is one frame. The default is -1.1. One possible usecase is to avoid framedrops in case of noisy timestamps or to increase frame drop precision in case of exact timestamps. Drop it like it’s hot, but precisely 💀
-apad parameters (output, per-stream) — pad the output audio stream(s). This is the same as applying -af apad. Argument is a string of filter parameters composed the same as with the apad filter. -shortest must be set for this output for the option to take effect. Audio padding, but make it fashion 🎧
-copyts — do not process input timestamps, but keep their values without trying to sanitize them. In particular, do not remove the initial start time offset value.

Note that, depending on the vsync option or on specific muxer processing (e.g. in case the format option avoid_negative_ts is enabled), the output timestamps may mismatch with the input timestamps even when this option is selected. Timestamps: preserved, not necessarily matching. It’s giving “trust the process” 😭

-start_at_zero — when used with copyts, shift input timestamps so they start at zero. This means that using e.g. -ss 50 will make output timestamps start at 50 seconds, regardless of what timestamp the input file started at. Zero is the moment ✨
-copytb mode — specify how to set the encoder timebase when stream copying. mode is an integer numeric value, and can assume one of the following values:
- 1 — use the demuxer timebase. The time base is copied to the output encoder from the corresponding input demuxer. This is sometimes required to avoid non monotonically increasing timestamps when copying video streams with variable frame rate.
- 0 — use the decoder timebase. The time base is copied to the output encoder from the corresponding input decoder.
- -1 — try to make the choice automatically, in order to generate a sane output.

Default value is -1. Auto-pilot timebase, bestie 🛟

-enc_time_base[:stream_specifier] timebase (output, per-stream) — set the encoder timebase. timebase can assume one of the following values:
- 0 — assign a default value according to the media type. For video, use 1/framerate; for audio, use 1/samplerate.
- demux — use the timebase from the demuxer.
- filter — use the timebase from the filtergraph.
- a positive number — use the provided number as the timebase. This field can be provided as a ratio of two integers (e.g. 1:24, 1:48000) or as a decimal number (e.g. 0.04166, 2.0833e-5).

Default value is 0. Timebase customization is lowkey a flex 💜

-bitexact (input/output) — enable bitexact mode for (de)muxer and (de/en)coder. Bit-for-bit bestie behavior 🧬
-shortest (output) — finish encoding when the shortest output stream ends. Note that this option may require buffering frames, which introduces extra latency. The maximum amount of this latency may be controlled with the -shortest_buf_duration option. Shortest stream said “I’m done,” and everybody packs up 🎬
-shortest_buf_duration duration (output) — the -shortest option may require buffering potentially large amounts of data when at least one of the streams is “sparse” (i.e. has large gaps between frames — this is typically the case for subtitles). This option controls the maximum duration of buffered frames in seconds. Larger values may allow the -shortest option to produce more accurate results, but increase memory use and latency. The default value is 10 seconds. Buffering, but make it bounded 🍓
-dts_delta_threshold threshold — timestamp discontinuity delta threshold, expressed as a decimal number of seconds. The timestamp discontinuity correction enabled by this option is only applied to input formats accepting timestamp discontinuity (for which the AVFMT_TS_DISCONT flag is enabled), e.g. MPEG-TS and HLS, and is automatically disabled when employing the -copyts option (unless wrapping is detected). If a timestamp discontinuity is detected whose absolute value is greater than threshold, ffmpeg will remove the discontinuity by decreasing/increasing the current DTS and PTS by the corresponding delta value. The default value is 10. Timestamp drama: smoothed, fr fr 🌸
-dts_error_threshold threshold — timestamp error delta threshold, expressed as a decimal number of seconds. The timestamp correction enabled by this option is only applied to input formats not accepting timestamp discontinuity (for which the AVFMT_TS_DISCONT flag is not enabled). If a timestamp discontinuity is detected whose absolute value is greater than threshold, ffmpeg will drop the PTS/DTS timestamp value. The default value is 3600*30 (30 hours), which is arbitrarily picked and quite conservative. We’re not dramatic, we’re just 30 hours patient 😭
-muxdelay seconds (output) — set the maximum demux-decode delay. A tiny bit of mux delay, and we’re out 🫶
-muxpreload seconds (output) — sets the initial demux-decode delay ✨
-streamid output-stream-index:new-value (output) — hands a brand-new stream-id to an output stream, like giving your fave a glow-up name 💅. Okay so important: this option has to come before the output filename it applies to. And if you've got multiple output files going on, you can absolutely reassign a streamid to a different value, no drama.

For example, setting stream 0's PID to 33 and stream 1's PID to 36 for an output mpegts file is literally just:

ffmpeg -i inurl -streamid 0:33 -streamid 1:36 out.ts

-bsf[:stream_specifier] bitstream_filters (input/output,per-stream) — applies bitstream filters to the streams that match, bestie 🎀. The filters get applied to each packet as it's received from the demuxer (when you use it as an input option) or right before it's sent to the muxer (when you use it as an output option).

bitstream_filters is a comma-separated list of bitstream filter specs, and each one looks like this:

<filter>[=<optname0>=<optval0>:<optname1>=<optval1>:...]

Any ,, = or : characters that need to be part of an option value have to be escaped with a backslash ⭐.

Wanna see the whole list of bitstream filters? Use the -bsfs option, iykyk.

E.g.

ffmpeg -bsf:v h264_mp4toannexb -i h264.mp4 -c:v copy -an out.h264

that one applies the h264_mp4toannexb bitstream filter (which converts an MP4-encapsulated H.264 stream to Annex B) to the input video stream, slay.

On the other hand,

ffmpeg -i file.mov -an -vn -bsf:s mov2textsub -c:s copy -f rawvideo sub.txt

applies the mov2textsub bitstream filter (which extracts text from MOV subtitles) to the output subtitle stream. But heads up 🙈 — since both examples use -c copy, it barely matters whether the filters are applied on input or output. That little detail would matter if transcoding was happening, so keep it in mind!

-tag[:stream_specifier] codec_tag (input/output,per-stream) — forces a tag/fourcc for the streams that match 🔥
-timecode hh:mm:ssSEPff — sets the Timecode for writing. SEP is : for non-drop timecode and ; (or .) for drop, periodt.

ffmpeg -i input.mpg -timecode 01:02:03.04 -r 30000/1001 -s ntsc output.mpg

-filter_complex filtergraph (global) — defines a complex filtergraph, aka one with an arbitrary number of inputs and/or outputs 🦋. If you just want a simple graph — one input, one output, same type — go look at the -filter options instead. filtergraph is a description of the filtergraph, exactly as described in the ``Filtergraph syntax'' section of the ffmpeg-filters manual. And you can specify this option multiple times: each use creates a brand-new complex filtergraph, iconic.

Inputs to a complex filtergraph can come from different source types, and you tell them apart by the format of the corresponding link label:

To connect an input stream, use [file_index:stream_specifier] (yep, the same syntax as -map). If stream_specifier matches multiple streams, the first one gets used. For multiview video, the stream specifier may be followed by the view specifier — check the documentation for the -map option for its syntax 👑.
To connect a loopback decoder use [dec:dec_idx], where dec_idx is the index of the loopback decoder to be connected to the given input. For multiview video, the decoder index may be followed by the view specifier, see the documentation for the -map option for its syntax.
To connect an output from another complex filtergraph, use its link label. E.g. the following example:

ffmpeg -i input.mkv \
  -filter_complex '[0:v]scale=size=hd1080,split=outputs=2[for_enc][orig_scaled]' \
  -c:v libx264 -map '[for_enc]' output.mkv \
  -dec 0:0 \
  -filter_complex '[dec:0][orig_scaled]hstack[stacked]' \
  -map '[stacked]' -c:v ffv1 comparison.mkv

reads an input video and

(line 2) uses a complex filtergraph with one input and two outputs to scale the video to 1920x1080 and duplicate the result to both outputs;
(line 3) encodes one scaled output with libx264 and writes the result to output.mkv;
(line 4) decodes this encoded stream with a loopback decoder;
(line 5) places the output of the loopback decoder (i.e. the libx264-encoded video) side by side with the scaled original input;
(line 6) combined video is then losslessly encoded and written into comparison.mkv.

Note that the two filtergraphs cannot be combined into one, because then there would be a cycle in the transcoding pipeline (filtergraph output goes to encoding, from there to decoding, then back to the same graph), and such cycles are not allowed 💀. Rules are rules.

An unlabeled input will be connected to the first unused input stream of the matching type ✨.

Output link labels are referred to with -map. Unlabeled outputs are added to the first output file.

Note that with this option it's possible to use only lavfi sources without normal input files — no cap, main character energy 🌈.

For example, to overlay an image over video:

ffmpeg -i video.mkv -i image.png -filter_complex '[0:v][1:v]overlay[out]' -map
'[out]' out.mkv

Here [0:v] refers to the first video stream in the first input file, which is linked to the first (main) input of the overlay filter. Similarly the first video stream in the second input is linked to the second (overlay) input of overlay 💖.

Assuming there's only one video stream in each input file, we can omit input labels, so the above is equivalent to:

ffmpeg -i video.mkv -i image.png -filter_complex 'overlay[out]' -map
'[out]' out.mkv

Furthermore we can omit the output label and the single output from the filter graph will be added to the output file automatically, so we can simply write:

ffmpeg -i video.mkv -i image.png -filter_complex 'overlay' out.mkv

As a special exception, you can use a bitmap subtitle stream as input: it will be converted into a video with the same size as the largest video in the file, or 720x576 if no video is present. Note that this is an experimental and temporary solution 🫶. It will be removed once libavfilter has proper support for subtitles.

For example, to hardcode subtitles on top of a DVB-T recording stored in MPEG-TS format, delaying the subtitles by 1 second:

ffmpeg -i input.ts -filter_complex \
  '[#0x2ef] setpts=PTS+1/TB [sub] ; [#0x2d0] [sub] overlay' \
  -sn -map '#0x2dc' output.mkv

(0x2d0, 0x2dc and 0x2ef are the MPEG-TS PIDs of respectively the video, audio and subtitles streams; 0:0, 0:3 and 0:7 would have worked too 😌)

To generate 5 seconds of pure red video using the lavfi color source:

ffmpeg -filter_complex 'color=c=red' -t 5 out.mkv

-filter_complex_threads nb_threads (global) — defines how many threads get used to process a filter_complex graph 🔥. Similar to filter_threads but used for -filter_complex graphs only. The default is the number of available CPUs, cute.
-lavfi filtergraph (global) — defines a complex filtergraph, i.e. one with arbitrary number of inputs and/or outputs. Equivalent to -filter_complex 💫.
-accurate_seek (input) — turns accurate seeking in input files on or off when you're using the -ss option. It's enabled by default, so seeking is accurate when transcoding 🎯. Use -noaccurate_seek to disable it, which may be useful e.g. when copying some streams and transcoding the others.
-seek_timestamp (input) — turns seeking by timestamp in input files on or off with the -ss option. Disabled by default. If enabled, the argument to the -ss option is considered an actual timestamp, and is not offset by the start time of the file ⭐. This matters only for files which don't start from timestamp 0, such as transport streams.
-thread_queue_size size (input/output) — for input, this option sets the maximum number of queued packets when reading from the file or device. With low latency / high rate live streams, packets may be discarded if they're not read in a timely manner 😭; setting this value can force ffmpeg to use a separate input thread and read packets as soon as they arrive. By default ffmpeg only does this if multiple inputs are specified.

For output, this option specifies the maximum number of packets that may be queued to each muxing thread.

-sdp_file file (global) — prints sdp information for an output stream to file 🎀. This lets you dump sdp information when at least one output isn't an rtp stream. (Requires at least one of the output formats to be rtp.)
-discard (input) — lets you discard specific streams or frames from streams 🦋. Any input stream can be fully discarded with the value all, whereas selective discarding of frames from a stream happens at the demuxer and isn't supported by all demuxers. The values are:
- none — discard no frame.
- default — the default, which discards no frames.
- noref — discard all non-reference frames.
- bidir — discard all bidirectional frames.
- nokey — discard all frames excepts keyframes.
- all — discard all frames.
-abort_on flags (global) — stop and abort on various conditions 🛑. The available flags are:
- empty_output — no packets were passed to the muxer, so the output is empty.
- empty_output_stream — no packets were passed to the muxer in some of the output streams.
-max_error_rate (global) — sets the fraction of decoding frame failures across all inputs which, when crossed, makes ffmpeg return exit code 69 👀. Crossing this threshold does not terminate processing — it's just a lil heads-up. Range is a floating-point number between 0 to 1. Default is 2/3.
-xerror (global) — stop and exit on error. No notes. 💀
-max_muxing_queue_size packets (output,per-stream) — okay so when transcoding audio and/or video streams, ffmpeg won't begin writing into the output until it has one packet for each such stream 🌸. While it's waiting for that to happen, packets for other streams are buffered. This option sets the size of that buffer, in packets, for the matching output stream.

The default value of this option should be high enough for most uses, so only touch this option if you're sure that you need it, bestie 🙈.

-muxing_queue_data_threshold bytes (output,per-stream) — okay so this one's a minimum threshold, and below it the muxing queue size just isn't counted, periodt. It defaults to 50 megabytes per stream, and it's based on the overall size of the packets passed to the muxer. 💅
-auto_conversion_filters (global) — turns on auto-inserting format conversion filters into all the filter graphs, including the ones you define with -vf, -af, -filter_complex and -lavfi. If filter format negotiation needs a conversion and this is off, the filter init just fails 😭. Conversions can still happen tho — you just insert the relevant conversion filter (scale, aresample) into the graph yourself. It's on by default, so to explicitly disable it you gotta specify -noauto_conversion_filters.
-bits_per_raw_sample[:stream_specifier] value (output,per-stream) — declares that the number of bits per raw sample in the given output stream is value. ⚠️ Heads up bestie: this only sets the info handed to the encoder/muxer, it does not change the stream to conform to that value. Setting values that don't match the stream properties may result in encoding failures or straight-up invalid output files. No cap.
-stats_enc_pre[:stream_specifier] path (output,per-stream)
-stats_enc_post[:stream_specifier] path (output,per-stream)
-stats_mux_pre[:stream_specifier] path (output,per-stream) — write per-frame encoding info about the matching streams into the file given by path. 🎀

-stats_enc_pre writes info about raw video or audio frames right before they're sent off for encoding, while -stats_enc_post writes info about encoded packets as they come back from the encoder. -stats_mux_pre writes info about packets just as they're about to be sent to the muxer. Every frame or packet produces one line in the specified file. The format of that line is controlled by -stats_enc_pre_fmt / -stats_enc_post_fmt / -stats_mux_pre_fmt.

When stats for multiple streams get written into a single file, the lines for different streams will be interleaved. The precise order of that interleaving is not specified and not guaranteed to stay stable between different invocations of the program — even with the same options. 💀

-stats_enc_pre_fmt[:stream_specifier] format_spec (output,per-stream)
-stats_enc_post_fmt[:stream_specifier] format_spec (output,per-stream)
-stats_mux_pre_fmt[:stream_specifier] format_spec (output,per-stream) — specify the format for the lines written with -stats_enc_pre / -stats_enc_post / -stats_mux_pre.

format_spec is a string that may contain directives of the form {fmt}. format_spec is backslash-escaped — use \{, \}, and \\ to write a literal {, }, or \, respectively, into the output.

The directives given with fmt may be one of the following:

fidx — index of the output file.
sidx — index of the output stream in the file.
n — frame number. Pre-encoding: number of frames sent to the encoder so far. Post-encoding: number of packets received from the encoder so far. Muxing: number of packets submitted to the muxer for this stream so far.
ni — input frame number. Index of the input frame (i.e. output by a decoder) that corresponds to this output frame or packet. -1 if unavailable.
tb — timebase in which this frame/packet's timestamps are expressed, as a rational number num/den. Note that encoder and muxer may use different timebases.
tbi — timebase for ptsi, as a rational number num/den. Available when ptsi is available, 0/1 otherwise.
pts — presentation timestamp of the frame or packet, as an integer. Should be multiplied by the timebase to compute presentation time.
ptsi — presentation timestamp of the input frame (see ni), as an integer. Should be multiplied by tbi to compute presentation time. Printed as (2^63 - 1 = 9223372036854775807) when not available.
t — presentation time of the frame or packet, as a decimal number. Equal to pts multiplied by tb.
ti — presentation time of the input frame (see ni), as a decimal number. Equal to ptsi multiplied by tbi. Printed as inf when not available.
dts (packet) — decoding timestamp of the packet, as an integer. Should be multiplied by the timebase to compute presentation time.
dt (packet) — decoding time of the frame or packet, as a decimal number. Equal to dts multiplied by tb.
sn (frame,audio) — number of audio samples sent to the encoder so far.
samp (frame,audio) — number of audio samples in the frame.
size (packet) — size of the encoded packet in bytes.
br (packet) — current bitrate in bits per second.
abr (packet) — average bitrate for the whole stream so far, in bits per second, -1 if it can't be determined at this point.
key (packet) — character 'K' if the packet contains a keyframe, character 'N' otherwise.

Directives tagged with packet may only be used with -stats_enc_post_fmt and -stats_mux_pre_fmt.

Directives tagged with frame may only be used with -stats_enc_pre_fmt.

Directives tagged with audio may only be used with audio streams.

The default format strings are:

pre-encoding

{fidx} {sidx} {n} {t}

post-encoding

{fidx} {sidx} {n} {t}

In the future, new items may be added to the end of the default formatting strings. Users who depend on the format staying exactly the same should prescribe it manually. 🔮

Note that stats for different streams written into the same file may have different formats.

🎀 Preset files 🎀

A preset file contains a sequence of option=value pairs, one for each line, specifying a sequence of options which would be awkward to specify on the command line. Lines starting with the hash ('#') character are ignored and are used to provide comments. Check the presets directory in the FFmpeg source tree for examples. ✨

There are two types of preset files: ffpreset and avpreset files.

ffpreset files

ffpreset files are specified with the vpre, apre, spre, and fpre options. The fpre option takes the filename of the preset instead of a preset name as input and can be used for any kind of codec. For the vpre, apre, and spre options, the options specified in a preset file are applied to the currently selected codec of the same type as the preset option.

The argument passed to the vpre, apre, and spre preset options identifies the preset file to use according to the following rules:

First ffmpeg searches for a file named arg.ffpreset in the directories $FFMPEG_DATADIR (if set), and $HOME/.ffmpeg, and in the datadir defined at configuration time (usually PREFIX/share/ffmpeg) or in a ffpresets folder along the executable on win32, in that order. For example, if the argument is libvpx-1080p, it will search for the file libvpx-1080p.ffpreset. 🦋

If no such file is found, then ffmpeg will search for a file named codec_name-arg.ffpreset in the above-mentioned directories, where codec_name is the name of the codec to which the preset file options will be applied. For example, if you select the video codec with -vcodec libvpx and use -vpre 1080p, then it will search for the file libvpx-1080p.ffpreset.

avpreset files

avpreset files are specified with the pre option. They work similar to ffpreset files, but they only allow encoder-specific options. Therefore, an option=value pair specifying an encoder cannot be used. 🙈

When the pre option is specified, ffmpeg will look for files with the suffix .avpreset in the directories $AVCONV_DATADIR (if set), and $HOME/.avconv, and in the datadir defined at configuration time (usually PREFIX/share/ffmpeg), in that order.

First ffmpeg searches for a file named codec_name-arg.avpreset in the above-mentioned directories, where codec_name is the name of the codec to which the preset file options will be applied. For example, if you select the video codec with -vcodec libvpx and use -pre 1080p, then it will search for the file libvpx-1080p.avpreset.

If no such file is found, then ffmpeg will search for a file named arg.avpreset in the same directories.

🌈 vstats file format 🌈

The -vstats and -vstats_file options enable generation of a file containing statistics about the generated video outputs.

The -vstats_version option controls the format version of the generated file.

With version 1 the format is:

frame= <FRAME> q= <FRAME_QUALITY> PSNR= <PSNR> f_size= <FRAME_SIZE> s_size= <STREAM_SIZE>kB time= <TIMESTAMP> br= <BITRATE>kbits/s avg_br= <AVERAGE_BITRATE>kbits/s

With version 2 the format is:

out= <OUT_FILE_INDEX> st= <OUT_FILE_STREAM_INDEX> frame= <FRAME_NUMBER> q= <FRAME_QUALITY>f PSNR= <PSNR> f_size= <FRAME_SIZE> s_size= <STREAM_SIZE>kB time= <TIMESTAMP> br= <BITRATE>kbits/s avg_br= <AVERAGE_BITRATE>kbits/s

The value corresponding to each key is described below:

avg_br — average bitrate expressed in Kbits/s
br — bitrate expressed in Kbits/s
frame — number of encoded frame
out — out file index
PSNR — Peak Signal to Noise Ratio
q — quality of the frame
f_size — encoded packet size expressed as number of bytes
s_size — stream size expressed in KiB
st — out file stream index
time — time of the packet
type — picture type

See also the -stats_enc options for an alternative way to show encoding statistics. 💖

✨ EXAMPLES ✨

okay bestie, this is the part where the manpage stops explaining and starts showing you the moves 💅

📹 Video and Audio grabbing

If you specify the input format and the device, then ffmpeg can grab video and audio directly — no cap.

ffmpeg -f oss -i /dev/dsp -f video4linux2 -i /dev/video0 /tmp/out.mpg

Or with an ALSA audio source (mono input, card id 1) instead of OSS:

ffmpeg -f alsa -ac 1 -i hw:1 -f video4linux2 -i /dev/video0 /tmp/out.mpg

Note that you must activate the right video source and channel before launching ffmpeg with any TV viewer such as http://linux.bytesex.org/xawtv/ by Gerd Knorr. You also have to set the audio recording levels correctly with a standard mixer. like, actual setup first, bestie — ffmpeg is not a mind reader 💖

🖥️ X11 grabbing

Grab the X11 display with ffmpeg via:

ffmpeg -f x11grab -video_size cif -framerate 25 -i :0.0 /tmp/out.mpg

0.0 is the display.screen number of your X11 server, same as the DISPLAY environment variable.

ffmpeg -f x11grab -video_size cif -framerate 25 -i :0.0+10,20 /tmp/out.mpg

0.0 is the display.screen number of your X11 server, same as the DISPLAY environment variable. 10 is the x-offset and 20 is the y-offset for the grabbing. offsets! we love a little coordinate moment ✨

🎞️ Video and Audio file format conversion

Any supported file format and protocol can serve as input to ffmpeg:

Examples:

You can use YUV files as input:

ffmpeg -i /tmp/test%d.Y /tmp/out.mpg

It will use the files:

/tmp/test0.Y, /tmp/test0.U, /tmp/test0.V,
/tmp/test1.Y, /tmp/test1.U, /tmp/test1.V, etc...

The Y files use twice the resolution of the U and V files. They are raw files, without header. They can be generated by all decent video decoders. You must specify the size of the image with the -s option if ffmpeg cannot guess it.

You can input from a raw YUV420P file:

ffmpeg -i /tmp/test.yuv /tmp/out.avi

test.yuv is a file containing raw YUV planar data. Each frame is composed of the Y plane followed by the U and V planes at half vertical and horizontal resolution.

You can output to a raw YUV420P file:

ffmpeg -i mydivx.avi hugefile.yuv

You can set several input files and output files:

ffmpeg -i /tmp/a.wav -s 640x480 -i /tmp/a.yuv /tmp/a.mpg

Converts the audio file a.wav and the raw YUV video file a.yuv to MPEG file a.mpg.

You can also do audio and video conversions at the same time:

ffmpeg -i /tmp/a.wav -ar 22050 /tmp/a.mp2

Converts a.wav to MPEG audio at 22050 Hz sample rate.

You can encode to several formats at the same time and define a mapping from input stream to output streams:

ffmpeg -i /tmp/a.wav -map 0:a -b:a 64k /tmp/a.mp2 -map 0:a -b:a 128k /tmp/b.mp2

Converts a.wav to a.mp2 at 64 kbits and to b.mp2 at 128 kbits. -map file:index specifies which input stream is used for each output stream, in the order of the definition of output streams. mapping is giving main character energy 🔥

You can transcode decrypted VOBs:

ffmpeg -i snatch_1.vob -f avi -c:v mpeg4 -b:v 800k -g 300 -bf 2 -c:a libmp3lame -b:a 128k snatch.avi

This is a typical DVD ripping example; the input is a VOB file, the output an AVI file with MPEG-4 video and MP3 audio. Note that in this command we use B-frames so the MPEG-4 stream is DivX5 compatible, and GOP size is 300 which means one intra frame every 10 seconds for 29.97fps input video. Furthermore, the audio stream is MP3-encoded so you need to enable LAME support by passing --enable-libmp3lame to configure. The mapping is particularly useful for DVD transcoding to get the desired audio language.

NOTE: To see the supported input formats, use ffmpeg -demuxers.

You can extract images from a video, or create a video from many images:

For extracting images from a video:

ffmpeg -i foo.avi -r 1 -s WxH -f image2 foo-%03d.jpeg

This will extract one video frame per second from the video and will output them in files named foo-001.jpeg, foo-002.jpeg, etc. Images will be rescaled to fit the new WxH values.

If you want to extract just a limited number of frames, you can use the above command in combination with the -frames:v or -t option, or in combination with -ss to start extracting from a certain point in time.

For creating a video from many images:

ffmpeg -f image2 -framerate 12 -i foo-%03d.jpeg -s WxH foo.avi

The syntax foo-%03d.jpeg specifies to use a decimal number composed of three digits padded with zeroes to express the sequence number. It is the same syntax supported by the C printf function, but only formats accepting a normal integer are suitable.

When importing an image sequence, -i also supports expanding shell-like wildcard patterns (globbing) internally, by selecting the image2-specific -pattern_type glob option.

For example, for creating a video from filenames matching the glob pattern foo-*.jpeg:

ffmpeg -f image2 -pattern_type glob -framerate 12 -i 'foo-*.jpeg' -s WxH foo.avi

You can put many streams of the same type in the output:

ffmpeg -i test1.avi -i test2.avi -map 1:1 -map 1:0 -map 0:1 -map 0:0 -c copy -y test12.nut

The resulting output file test12.nut will contain the first four streams from the input files in reverse order. reverse order?? chaotic, but make it fashion 💅

To force CBR video output:

ffmpeg -i myfile.avi -b 4000k -minrate 4000k -maxrate 4000k -bufsize 1835k out.m2v

The four options lmin, lmax, mblmin and mblmax use 'lambda' units, but you may use the QP2LAMBDA constant to easily convert from 'q' units:

ffmpeg -i src.ext -lmax 21*QP2LAMBDA dst.ext

🧠 SYNTAX 🧠

This section documents the syntax and formats employed by the FFmpeg libraries and tools. yes bestie, we're getting into the grammar of it all ✨

🔤 Quoting and escaping

FFmpeg adopts the following quoting and escaping mechanism, unless explicitly specified. The following rules are applied:

' and \ are special characters (respectively used for quoting and escaping). In addition to them, there might be other special characters depending on the specific syntax where the escaping and quoting are employed.
A special character is escaped by prefixing it with a \.
All characters enclosed between '' are included literally in the parsed string. The quote character ' itself cannot be quoted, so you may need to close the quote and escape it.
Leading and trailing whitespaces, unless escaped or quoted, are removed from the parsed string.

Note that you may need to add a second level of escaping when using the command line or a script, which depends on the syntax of the adopted shell language.

The function av_get_token defined in libavutil/avstring.h can be used to parse a token quoted or escaped according to the rules defined above.

The tool tools/ffescape in the FFmpeg source tree can be used to automatically quote or escape a string in a script.

Examples:

Escape the string "Crime d'Amour" containing the ' special character:

Crime d\'Amour

The string above contains a quote, so the ' needs to be escaped when quoting it:

'Crime d'\''Amour'

Include leading or trailing whitespaces using quoting:

'  this string starts and ends with whitespaces  '

Escaping and quoting can be mixed together:

' The string '\'string\'' is a string '

To include a literal \ you can use either escaping or quoting:

'c:\foo' can be written as c:\\foo

📅 Date

The accepted syntax is:

[(YYYY-MM-DD|YYYYMMDD)[T|t| ]]((HH:MM:SS[.m...]]])|(HHMMSS[.m...]]]))[Z]
now

If the value is now it takes the current time.

Time is local time unless Z is appended, in which case it is interpreted as UTC. If the year-month-day part is not specified it takes the current year-month-day. timezone behavior, we love a precise queen 🫶

⏱️ Time duration

There are two accepted syntaxes for expressing time duration.

[-][<HH>:]<MM>:<SS>[.<m>...]

HH expresses the number of hours, MM the number of minutes for a maximum of 2 digits, and SS the number of seconds for a maximum of 2 digits. The m at the end expresses decimal value for SS.

[-]<S>+[.<m>...][s|ms|us]

S expresses the number of seconds, with the optional decimal part m. The optional literal suffixes s, ms or us indicate to interpret the value as seconds, milliseconds or microseconds, respectively.

In both expressions, the optional - indicates negative duration.

Examples:

The following examples are all valid time duration:

55 — 55 seconds
0.2 — 0.2 seconds
200ms — 200 milliseconds, that's 0.2s
200000us — 200000 microseconds, that's 0.2s
12:03:45 — 12 hours, 03 minutes and 45 seconds
23.189 — 23.189 seconds

📐 Video size

Specify the size of the sourced video, it may be a string of the form widthxheight, or the name of a size abbreviation.

The following abbreviations are recognized:

ntsc — 720x480
pal — 720x576
qntsc — 352x240
qpal — 352x288
sntsc — 640x480
spal — 768x576
film — 352x240
ntsc-film — 352x240
sqcif — 128x96
qcif — 176x144
cif — 352x288
4cif — 704x576
16cif — 1408x1152
qqvga — 160x120
qvga — 320x240
vga — 640x480
svga — 800x600
xga — 1024x768
uxga — 1600x1200
qxga — 2048x1536
sxga — 1280x1024
qsxga — 2560x2048
hsxga — 5120x4096
wvga — 852x480
wxga — 1366x768
wsxga — 1600x1024
wuxga — 1920x1200
woxga — 2560x1600
wqsxga — 3200x2048
wquxga — 3840x2400
whsxga — 6400x4096
whuxga — 7680x4800
cga — 320x200
ega — 640x350
hd480 — 852x480
hd720 — 1280x720
hd1080 — 1920x1080
2k — 2048x1080
2kflat — 1998x1080
2kscope — 2048x858
4k — 4096x2160
4kflat — 3996x2160
4kscope — 4096x1716
nhd — 640x360
hqvga — 240x160
wqvga — 400x240
fwqvga — 432x240
hvga — 480x320
qhd — 960x540
2kdci — 2048x1080
4kdci — 4096x2160
uhd2160 — 3840x2160
uhd4320 — 7680x4320

🎚️ Video rate

Specify the frame rate of a video, expressed as the number of frames generated per second. It has to be a string in the format frame_rate_num/frame_rate_den, an integer number, a float number or a valid video frame rate abbreviation. frame rate lore, honestly iconic ✨

okay so these abbreviations are recognized, bestie — write them down, they’re literally the shorthand:

ntsc — 30000/1001
pal — 25/1
qntsc — 30000/1001
qpal — 25/1
sntsc — 30000/1001
spal — 25/1
film — 24/1
ntsc-film — 24000/1001

🎀 Ratio 🎀

A ratio can be expressed as an expression, or in the form numerator:denominator.

⚠️ bestie, a ratio with infinite (1/0) or negative value is considered valid, so you should check on the returned value if you want to exclude those values. Not me making the rules, I’m just reporting them.

The undefined value can be expressed using the "0:0" string.

🎨 Color 🎨

It can be the name of a color as defined below (case insensitive match) or a [0x|#]RRGGBB[AA] sequence, possibly followed by @ and a string representing the alpha component.

The alpha component may be a string composed by 0x followed by an hexadecimal number or a decimal number between 0.0 and 1.0, which represents the opacity value (0x00 or 0.0 means completely transparent, 0xff or 1.0 completely opaque). If the alpha component is not specified then 0xff is assumed.

The string random will result in a random color.

The following names of colors are recognized — yes, all of them, this palette is iconic:

AliceBlue — 0xF0F8FF
AntiqueWhite — 0xFAEBD7
Aqua — 0x00FFFF
Aquamarine — 0x7FFFD4
Azure — 0xF0FFFF
Beige — 0xF5F5DC
Bisque — 0xFFE4C4
Black — 0x000000
BlanchedAlmond — 0xFFEBCD
Blue — 0x0000FF
BlueViolet — 0x8A2BE2
Brown — 0xA52A2A
BurlyWood — 0xDEB887
CadetBlue — 0x5F9EA0
Chartreuse — 0x7FFF00
Chocolate — 0xD2691E
Coral — 0xFF7F50
CornflowerBlue — 0x6495ED
Cornsilk — 0xFFF8DC
Crimson — 0xDC143C
Cyan — 0x00FFFF
DarkBlue — 0x00008B
DarkCyan — 0x008B8B
DarkGoldenRod — 0xB8860B
DarkGray — 0xA9A9A9
DarkGreen — 0x006400
DarkKhaki — 0xBDB76B
DarkMagenta — 0x8B008B
DarkOliveGreen — 0x556B2F
Darkorange — 0xFF8C00
DarkOrchid — 0x9932CC
DarkRed — 0x8B0000
DarkSalmon — 0xE9967A
DarkSeaGreen — 0x8FBC8F
DarkSlateBlue — 0x483D8B
DarkSlateGray — 0x2F4F4F
DarkTurquoise — 0x00CED1
DarkViolet — 0x9400D3
DeepPink — 0xFF1493
DeepSkyBlue — 0x00BFFF
DimGray — 0x696969
DodgerBlue — 0x1E90FF
FireBrick — 0xB22222
FloralWhite — 0xFFFAF0
ForestGreen — 0x228B22
Fuchsia — 0xFF00FF
Gainsboro — 0xDCDCDC
GhostWhite — 0xF8F8FF
Gold — 0xFFD700
GoldenRod — 0xDAA520
Gray — 0x808080
Green — 0x008000
GreenYellow — 0xADFF2F
HoneyDew — 0xF0FFF0
HotPink — 0xFF69B4
IndianRed — 0xCD5C5C
Indigo — 0x4B0082
Ivory — 0xFFFFF0
Khaki — 0xF0E68C
Lavender — 0xE6E6FA
LavenderBlush — 0xFFF0F5
LawnGreen — 0x7CFC00
LemonChiffon — 0xFFFACD
LightBlue — 0xADD8E6
LightCoral — 0xF08080
LightCyan — 0xE0FFFF
LightGoldenRodYellow — 0xFAFAD2
LightGreen — 0x90EE90
LightGrey — 0xD3D3D3
LightPink — 0xFFB6C1
LightSalmon — 0xFFA07A
LightSeaGreen — 0x20B2AA
LightSkyBlue — 0x87CEFA
LightSlateGray — 0x778899
LightSteelBlue — 0xB0C4DE
LightYellow — 0xFFFFE0
Lime — 0x00FF00
LimeGreen — 0x32CD32
Linen — 0xFAF0E6
Magenta — 0xFF00FF
Maroon — 0x800000
MediumAquaMarine — 0x66CDAA
MediumBlue — 0x0000CD
MediumOrchid — 0xBA55D3
MediumPurple — 0x9370D8
MediumSeaGreen — 0x3CB371
MediumSlateBlue — 0x7B68EE
MediumSpringGreen — 0x00FA9A
MediumTurquoise — 0x48D1CC
MediumVioletRed — 0xC71585
MidnightBlue — 0x191970
MintCream — 0xF5FFFA
MistyRose — 0xFFE4E1
Moccasin — 0xFFE4B5
NavajoWhite — 0xFFDEAD
Navy — 0x000080
OldLace — 0xFDF5E6
Olive — 0x808000
OliveDrab — 0x6B8E23
Orange — 0xFFA500
OrangeRed — 0xFF4500
Orchid — 0xDA70D6
PaleGoldenRod — 0xEEE8AA
PaleGreen — 0x98FB98
PaleTurquoise — 0xAFEEEE
PaleVioletRed — 0xD87093
PapayaWhip — 0xFFEFD5
PeachPuff — 0xFFDAB9
Peru — 0xCD853F
Pink — 0xFFC0CB
Plum — 0xDDA0DD
PowderBlue — 0xB0E0E6
Purple — 0x800080
Red — 0xFF0000
RosyBrown — 0xBC8F8F
RoyalBlue — 0x4169E1
SaddleBrown — 0x8B4513
Salmon — 0xFA8072
SandyBrown — 0xF4A460
SeaGreen — 0x2E8B57
SeaShell — 0xFFF5EE
Sienna — 0xA0522D
Silver — 0xC0C0C0
SkyBlue — 0x87CEEB
SlateBlue — 0x6A5ACD
SlateGray — 0x708090
Snow — 0xFFFAFA
SpringGreen — 0x00FF7F
SteelBlue — 0x4682B4
Tan — 0xD2B48C
Teal — 0x008080
Thistle — 0xD8BFD8
Tomato — 0xFF6347
Turquoise — 0x40E0D0
Violet — 0xEE82EE
Wheat — 0xF5DEB3
White — 0xFFFFFF
WhiteSmoke — 0xF5F5F5
Yellow — 0xFFFF00
YellowGreen — 0x9ACD32

🔊 Channel Layout 🔊

A channel layout specifies the spatial disposition of the channels in a multi-channel audio stream. To specify a channel layout, FFmpeg makes use of a special syntax.

Individual channels are identified by an id, as given by the table below:

FL — front left
FR — front right
FC — front center
LFE — low frequency
BL — back left
BR — back right
FLC — front left-of-center
FRC — front right-of-center
BC — back center
SL — side left
SR — side right
TC — top center
TFL — top front left
TFC — top front center
TFR — top front right
TBL — top back left
TBC — top back center
TBR — top back right
DL — downmix left
DR — downmix right
WL — wide left
WR — wide right
SDL — surround direct left
SDR — surround direct right
LFE2 — low frequency 2

Standard channel layout compositions can be specified by using the following identifiers:

mono — FC
stereo — FL+FR
2.1 — FL+FR+LFE
3.0 — FL+FR+FC
3.0(back) — FL+FR+BC
4.0 — FL+FR+FC+BC
quad — FL+FR+BL+BR
quad(side) — FL+FR+SL+SR
3.1 — FL+FR+FC+LFE
5.0 — FL+FR+FC+BL+BR
5.0(side) — FL+FR+FC+SL+SR
4.1 — FL+FR+FC+LFE+BC
5.1 — FL+FR+FC+LFE+BL+BR
5.1(side) — FL+FR+FC+LFE+SL+SR
6.0 — FL+FR+FC+BC+SL+SR
6.0(front) — FL+FR+FLC+FRC+SL+SR
3.1.2 — FL+FR+FC+LFE+TFL+TFR
hexagonal — FL+FR+FC+BL+BR+BC
6.1 — FL+FR+FC+LFE+BC+SL+SR
6.1 — FL+FR+FC+LFE+BL+BR+BC
6.1(front) — FL+FR+LFE+FLC+FRC+SL+SR
7.0 — FL+FR+FC+BL+BR+SL+SR
7.0(front) — FL+FR+FC+FLC+FRC+SL+SR
7.1 — FL+FR+FC+LFE+BL+BR+SL+SR
7.1(wide) — FL+FR+FC+LFE+BL+BR+FLC+FRC
7.1(wide-side) — FL+FR+FC+LFE+FLC+FRC+SL+SR
5.1.2 — FL+FR+FC+LFE+BL+BR+TFL+TFR
octagonal — FL+FR+FC+BL+BR+BC+SL+SR
cube — FL+FR+BL+BR+TFL+TFR+TBL+TBR
5.1.4 — FL+FR+FC+LFE+BL+BR+TFL+TFR+TBL+TBR
7.1.2 — FL+FR+FC+LFE+BL+BR+SL+SR+TFL+TFR
7.1.4 — FL+FR+FC+LFE+BL+BR+SL+SR+TFL+TFR+TBL+TBR
7.2.3 — FL+FR+FC+LFE+BL+BR+SL+SR+TFL+TFR+TBC+LFE2
9.1.4 — FL+FR+FC+LFE+BL+BR+FLC+FRC+SL+SR+TFL+TFR+TBL+TBR
9.1.6 — FL+FR+FC+LFE+BL+BR+FLC+FRC+SL+SR+TFL+TFR+TBL+TBR+TSL+TSR
hexadecagonal — FL+FR+FC+BL+BR+BC+SL+SR+WL+WR+TBL+TBR+TBC+TFC+TFL+TFR
binaural — BIL+BIR
downmix — DL+DR
22.2 — FL+FR+FC+LFE+BL+BR+FLC+FRC+BC+SL+SR+TC+TFL+TFC+TFR+TBL+TBC+TBR+LFE2+TSL+TSR+BFC+BFL+BFR

A custom channel layout can be specified as a sequence of terms, separated by +. Each term can be:

the name of a single channel (e.g. FL, FR, FC, LFE, etc.), each optionally containing a custom name after a @, (e.g. FL@Left, FR@Right, FC@Center, LFE@Low_Frequency, etc.)

A standard channel layout can be specified by the following:

the name of a single channel (e.g. FL, FR, FC, LFE, etc.)
the name of a standard channel layout (e.g. mono, stereo, 4.0, quad, 5.0, etc.)
a number of channels, in decimal, followed by c, yielding the default channel layout for that number of channels (see the function av_channel_layout_default). Note that not all channel counts have a default layout.
a number of channels, in decimal, followed by C, yielding an unknown channel layout with the specified number of channels. Note that not all channel layout specification strings support unknown channel layouts.
a channel layout mask, in hexadecimal starting with 0x (see the AV_CH_* macros in libavutil/channel_layout.h.)

Before libavutil version 53 the trailing character c to specify a number of channels was optional, but now it is required, while a channel layout mask can also be specified as a decimal number (if and only if not followed by c or C).

See also the function av_channel_layout_from_string defined in libavutil/channel_layout.h.

✨ EXPRESSION EVALUATION ✨

Okay bestie, when FFmpeg has to evaluate an arithmetic expression it uses an internal formula evaluator, and that evaluator is implemented through the libavutil/eval.h interface. 💖

An expression may contain unary operators, binary operators, constants, and functions.

Two expressions expr1 and expr2 can be combined into another expression expr1;expr2. They get evaluated one after the other, and the new expression evaluates to the value of expr2.

The binary operators you can use are: +, -, *, /, ^.

The unary operators you can use are: +, -.

Some internal variables can store and load intermediary results. You access them with the ld and st functions, using an index argument from 0 to 9 to pick which internal variable you mean. 🎀

The following functions are available:

abs(x) — compute the absolute value of x.
acos(x) — compute the arccosine of x.
asin(x) — compute the arcsine of x.
atan(x) — compute the arctangent of x.
atan2(y, x) — compute the principal value of the arc tangent of y/x.
between(x, min, max) — return 1 if x is greater than or equal to min and lesser than or equal to max, 0 otherwise.
bitand(x, y), bitor(x, y) — compute the bitwise and/or operation on x and y. The results of evaluating x and y are converted to integers before the bitwise operation. Note that both the conversion to integer and the conversion back to floating point can lose precision. Beware of unexpected results for large numbers, usually 2^53 and larger. 💀
ceil(expr) — round the value of expression expr upwards to the nearest integer. For example, ceil(1.5) is 2.0.
clip(x, min, max) — return the value of x clipped between min and max.
cos(x) — compute the cosine of x.
cosh(x) — compute the hyperbolic cosine of x.
eq(x, y) — return 1 if x and y are equivalent, 0 otherwise.
exp(x) — compute the exponential of x with base e, the Euler's number.
floor(expr) — round the value of expression expr downwards to the nearest integer. For example, floor(-1.5) is -2.0.
gauss(x) — compute the Gauss function of x, corresponding to exp(-x*x/2) / sqrt(2*PI).
gcd(x, y) — return the greatest common divisor of x and y. If both x and y are 0, or either or both are less than zero, then behavior is undefined. 🚨
gt(x, y) — return 1 if x is greater than y, 0 otherwise.
gte(x, y) — return 1 if x is greater than or equal to y, 0 otherwise.
hypot(x, y) — similar to the C function with the same name; it returns sqrt(x*x + y*y), the length of the hypotenuse of a right triangle with sides of length x and y, or the distance of the point (x, y) from the origin.
if(x, y) — evaluate x, and if the result is non-zero return the result of the evaluation of y, return 0 otherwise.
if(x, y, z) — evaluate x, and if the result is non-zero return the evaluation result of y, otherwise the evaluation result of z.
ifnot(x, y) — evaluate x, and if the result is zero return the result of the evaluation of y, return 0 otherwise.
ifnot(x, y, z) — evaluate x, and if the result is zero return the evaluation result of y, otherwise the evaluation result of z.
isinf(x) — return 1.0 if x is +/-INFINITY, 0.0 otherwise.
isnan(x) — return 1.0 if x is NAN, 0.0 otherwise.
ld(idx) — load the value of the internal variable with index idx, which was previously stored with st(idx, expr). The function returns the loaded value.
lerp(x, y, z) — return linear interpolation between x and y by amount of z.
log(x) — compute the natural logarithm of x.
lt(x, y) — return 1 if x is lesser than y, 0 otherwise.
lte(x, y) — return 1 if x is lesser than or equal to y, 0 otherwise.
max(x, y) — return the maximum between x and y.
min(x, y) — return the minimum between x and y.
mod(x, y) — compute the remainder of division of x by y.
not(expr) — return 1.0 if expr is zero, 0.0 otherwise.
pow(x, y) — compute the power of x elevated y; it is equivalent to (x)^(y).
print(t), print(t, l) — print the value of expression t with loglevel l. If l is not specified then a default log level is used. Return the value of the expression printed.
random(idx) — return a pseudo random value between 0.0 and 1.0. idx is the index of the internal variable used to save the seed/state, which can be previously stored with st(idx). To initialize the seed, you need to store the seed value as a 64-bit unsigned integer in the internal variable with index idx. For example, to store the seed with value 42 in the internal variable with index 0 and print a few random values:

st(0,42); print(random(0)); print(random(0)); print(random(0))

randomi(idx, min, max) — return a pseudo random value in the interval between min and max. idx is the index of the internal variable which will be used to save the seed/state, which can be previously stored with st(idx). To initialize the seed, you need to store the seed value as a 64-bit unsigned integer in the internal variable with index idx.
root(expr, max) — find an input value for which the function represented by expr with argument ld(0) is 0 in the interval 0..max. The expression in expr must denote a continuous function or the result is undefined. ld(0) is used to represent the function input value, which means that the given expression will be evaluated multiple times with various input values that the expression can access through ld(0). When the expression evaluates to 0 then the corresponding input value will be returned.
round(expr) — round the value of expression expr to the nearest integer. For example, round(1.5) is 2.0.
sgn(x) — compute the sign of x.
sin(x) — compute the sine of x.
sinh(x) — compute the hyperbolic sine of x.
sqrt(expr) — compute the square root of expr. This is equivalent to (expr)^.5.
squish(x) — compute the expression 1/(1 + exp(4*x)).
st(idx, expr) — store the value of the expression expr in an internal variable. idx specifies the index of the variable where to store the value, and it is a value ranging from 0 to 9. The function returns the value stored in the internal variable. The stored value can be retrieved with ld(var). Note: variables are currently not shared between expressions.
tan(x) — compute the tangent of x.
tanh(x) — compute the hyperbolic tangent of x.
taylor(expr, x), taylor(expr, x, idx) — evaluate a Taylor series at x, given an expression representing the ld(idx)-th derivative of a function at 0. When the series does not converge the result is undefined. ld(idx) is used to represent the derivative order in expr, which means that the given expression will be evaluated multiple times with various input values that the expression can access through ld(idx). If idx is not specified then 0 is assumed. Note, when you have the derivatives at y instead of 0, taylor(expr, x-y) can be used.
time(0) — return the current (wallclock) time in seconds.
trunc(expr) — round the value of expression expr towards zero to the nearest integer. For example, trunc(-1.5) is -1.0.
while(cond, expr) — evaluate expression expr while the expression cond is non-zero, and return the value of the last expr evaluation, or NAN if cond was always false.

The following constants are available:

PI — area of the unit disc, approximately 3.14.
E — exp(1) (Euler's number), approximately 2.718.
PHI — golden ratio (1+sqrt(5))/2, approximately 1.618.

Assuming that an expression is considered "true" if it has a non-zero value, note that:

* works like AND.
+ works like OR.

For example, the construct:

if (A AND B) then C

is equivalent to:

if(A*B, C)

In your C code, you can extend the list of unary and binary functions, and define recognized constants, so that they are available for your expressions.

The evaluator also recognizes the International System unit prefixes. If 'i' is appended after the prefix, binary prefixes are used, which are based on powers of 1024 instead of powers of 1000. The 'B' postfix multiplies the value by 8, and can be appended after a unit prefix or used alone. This allows using, for example, 'KB', 'MiB', 'G' and 'B' as number postfix. 🌈

The list of available International System prefixes follows, with indication of the corresponding powers of 10 and of 2.

y — 10^-24 / 2^-80
z — 10^-21 / 2^-70
a — 10^-18 / 2^-60
f — 10^-15 / 2^-50
p — 10^-12 / 2^-40
n — 10^-9 / 2^-30
u — 10^-6 / 2^-20
m — 10^-3 / 2^-10
c — 10^-2
d — 10^-1
h — 10^2
k — 10^3 / 2^10
K — 10^3 / 2^10
M — 10^6 / 2^20
G — 10^9 / 2^30
T — 10^12 / 2^40
P — 10^15 / 2^50
E — 10^18 / 2^60
Z — 10^21 / 2^70
Y — 10^24 / 2^80

🎛️ CODEC OPTIONS 🎛️

libavcodec provides some generic global options, which can be set on all the encoders and decoders. In addition, each codec may support so-called private options, which are specific for a given codec. 💅

Sometimes a global option may only affect a specific kind of codec, and may be nonsensical or ignored by another, so you need to be aware of the meaning of the specified options. Also some options are meant only for decoding or encoding.

Options may be set by specifying -option value in the FFmpeg tools, or by setting the value explicitly in the AVCodecContext options or using the libavutil/opt.h API for programmatic use.

The list of supported options follows:

b integer (encoding,audio,video) — set bitrate in bits/s. Default value is 200K.
ab integer (encoding,audio) — set audio bitrate in bits/s. Default value is 128K.
bt integer (encoding,video) — set video bitrate tolerance in bits/s. In 1-pass mode, bitrate tolerance specifies how far ratecontrol is willing to deviate from the target average bitrate value. This is not related to min/max bitrate. Lowering tolerance too much has an adverse effect on quality.
flags flags (decoding/encoding,audio,video,subtitles) — set generic flags.

Possible values:

mv4 — use four motion vector by macroblock (mpeg4).
qpel — use 1/4 pel motion compensation.

loop — use the loop filter 🌀
qscale — use fixed qscale, periodt.
pass1 — use internal 2pass ratecontrol in first pass mode.
pass2 — use internal 2pass ratecontrol in second pass mode.
gray — only decode/encode grayscale, so no color for you bestie 🩶
psnr — set error[?] variables during encoding.
truncated — input bitstream might be randomly truncated 💀
drop_changed — don't output frames whose parameters differ from the first decoded frame in the stream. The error AVERROR_INPUT_CHANGED is returned when a frame is dropped.
ildct — use interlaced DCT.
low_delay — force low delay ⚡
global_header — place global headers in extradata instead of every keyframe.
bitexact — only write platform-, build- and time-independent data (except (I)DCT). This ensures that file and data checksums are reproducible and match between platforms. Its primary use is for regression testing, iykyk 🧪
aic — apply H263 advanced intra coding / mpeg4 ac prediction.
ilme — apply interlaced motion estimation.
cgop — use closed gop.
output_corrupt — output even potentially corrupted frames 🙈

time_base (rational number) — set codec time base.

It's the fundamental unit of time (in seconds) that frame timestamps get represented in terms of. For fixed-fps content, timebase should be 1 / frame_rate and timestamp increments should be identically 1.

g (integer, encoding,video) — set the group of picture (GOP) size. Default value is 12.
ar (integer, decoding/encoding,audio) — set audio sampling rate (in Hz).
ac (integer, decoding/encoding,audio) — set number of audio channels.
cutoff (integer, encoding,audio) — set cutoff bandwidth. (Supported only by selected encoders, see their respective documentation sections, fr.)
frame_size (integer, encoding,audio) — set audio frame size.

Every submitted frame except the last one must contain exactly frame_size samples per channel. It may be 0 when the codec has CODEC_CAP_VARIABLE_FRAME_SIZE set — in that case the frame size is not restricted. Some decoders set it to indicate a constant frame size.

frame_number (integer) — set the frame number.
delay (integer) — the delay, bestie.
qcomp (float, encoding,video) — set video quantizer scale compression (VBR). It's used as a constant in the ratecontrol equation. Recommended range for default rc_eq: 0.0-1.0.
qblur (float, encoding,video) — set video quantizer scale blur (VBR).
qmin (integer, encoding,video) — set min video quantizer scale (VBR). Must be included between -1 and 69, default value is 2.
qmax (integer, encoding,video) — set max video quantizer scale (VBR). Must be included between -1 and 1024, default value is 31.
qdiff (integer, encoding,video) — set max difference between the quantizer scale (VBR).
bf (integer, encoding,video) — set max number of B frames between non-B-frames.

Must be an integer between -1 and 16. 0 means B-frames are disabled. If you use -1, it'll pick an automatic value depending on the encoder.

Default value is 0.

b_qfactor (float, encoding,video) — set qp factor between P and B frames.
codec_tag (integer) — the codec tag.
bug (flags, decoding,video) — workaround for encoder bugs that aren't auto detected 🐛

Possible values:

autodetect
xvid_ilace — Xvid interlacing bug (autodetected if fourcc==XVIX)
ump4 — (autodetected if fourcc==UMP4)
no_padding — padding bug (autodetected)
amv
qpel_chroma
std_qpel — old standard qpel (autodetected per fourcc/version)
qpel_chroma2
direct_blocksize — direct-qpel-blocksize bug (autodetected per fourcc/version)
edge — edge padding bug (autodetected per fourcc/version)
hpel_chroma
dc_clip
ms — workaround various bugs in microsoft broken decoders 💀
trunc — trancated frames

strict (integer, decoding/encoding,audio,video) — specify how strictly to follow the standards.

Possible values:

very — strictly conform to an older more strict version of the spec or reference software
strict — strictly conform to all the things in the spec no matter what consequences
normal
unofficial — allow unofficial extensions
experimental — allow non standardized experimental things, experimental (unfinished/work in progress/not well tested) decoders and encoders.

Note: experimental decoders can pose a security risk — do not use this for decoding untrusted input. Like, genuinely bestie, don't 🔥

b_qoffset (float, encoding,video) — set QP offset between P and B frames.
err_detect (flags, decoding,audio,video) — set error detection flags.

Possible values:

crccheck — verify embedded CRCs
bitstream — detect bitstream specification deviations
buffer — detect improper bitstream length
explode — abort decoding on minor error detection 💥
ignore_err — ignore decoding errors, and continue decoding. This is useful if you want to analyze the content of a video and thus want everything to be decoded no matter what. This option will not result in a video that is pleasing to watch in case of errors, ngl.
careful — consider things that violate the spec and have not been seen in the wild as errors
compliant — consider all spec non compliancies as errors
aggressive — consider things that a sane encoder should not do as an error

has_b_frames (integer) — whether it has B frames.
block_align (integer) — the block align.
rc_override_count (integer) — the rc override count.
maxrate (integer, encoding,audio,video) — set max bitrate tolerance (in bits/s). Requires bufsize to be set.
minrate (integer, encoding,audio,video) — set min bitrate tolerance (in bits/s). Most useful in setting up a CBR encode. It's of little use elsewise.
bufsize (integer, encoding,audio,video) — set ratecontrol buffer size (in bits).
i_qfactor (float, encoding,video) — set QP factor between P and I frames.
i_qoffset (float, encoding,video) — set QP offset between P and I frames.
dct (integer, encoding,video) — set DCT algorithm.

Possible values:

auto — autoselect a good one (default)
fastint — fast integer
int — accurate integer
mmx
altivec
faan — floating point AAN DCT

lumi_mask (float, encoding,video) — compress bright areas stronger than medium ones ✨
tcplx_mask (float, encoding,video) — set temporal complexity masking.
scplx_mask (float, encoding,video) — set spatial complexity masking.
p_mask (float, encoding,video) — set inter masking.
dark_mask (float, encoding,video) — compress dark areas stronger than medium ones 🌑
idct (integer, decoding/encoding,video) — select IDCT implementation.

Possible values:

auto
int
simple
simplemmx
simpleauto — automatically pick a IDCT compatible with the simple one
arm
altivec
sh4
simplearm
simplearmv5te
simplearmv6
simpleneon
xvid
faani — floating point AAN IDCT

slice_count (integer) — the slice count.
ec (flags, decoding,video) — set error concealment strategy.

Possible values:

guess_mvs — iterative motion vector (MV) search (slow) 🐌
deblock — use strong deblock filter for damaged MBs
favor_inter — favor predicting from the previous frame instead of the current

bits_per_coded_sample (integer) — the bits per coded sample.
aspect (rational number, encoding,video) — set sample aspect ratio.
sar (rational number, encoding,video) — set sample aspect ratio. Alias to aspect.
debug (flags, decoding/encoding,audio,video,subtitles) — print specific debug info.

Possible values:

pict — picture info
rc — rate control
bitstream
mb_type — macroblock (MB) type
qp — per-block quantization parameter (QP)
dct_coeff
green_metadata — display complexity metadata for the upcoming frame, GoP or for a given duration.
skip
startcode
er — error recognition
mmco — memory management control operations (H.264)
bugs
buffers — picture buffer allocations
thread_ops — threading operations
nomc — skip motion compensation

cmp (integer, encoding,video) — set full pel me compare function.

Possible values:

sad — sum of absolute differences, fast (default)
sse — sum of squared errors
satd — sum of absolute Hadamard transformed differences
dct — sum of absolute DCT transformed differences
psnr — sum of squared quantization errors (avoid, low quality)
bit — number of bits needed for the block
rd — rate distortion optimal, slow
zero — 0
vsad — sum of absolute vertical differences
vsse — sum of squared vertical differences
nsse — noise preserving sum of squared differences
w53 — 5/3 wavelet, only used in snow
w97 — 9/7 wavelet, only used in snow
dctmax
chroma

subcmp (integer, encoding,video) — set sub pel me compare function.

Possible values:

sad — sum of absolute differences, fast (default)
sse — sum of squared errors
satd — sum of absolute Hadamard transformed differences
dct — sum of absolute DCT transformed differences
psnr — sum of squared quantization errors (avoid, low quality)
bit — number of bits needed for the block
rd — rate distortion optimal, slow
zero — 0
vsad — sum of absolute vertical differences
vsse — sum of squared vertical differences
nsse — noise preserving sum of squared differences
w53 — 5/3 wavelet, only used in snow
w97 — 9/7 wavelet, only used in snow
dctmax
chroma

mbcmp (integer, encoding,video) — set macroblock compare function.

Possible values:

okay continuing the compare-function menu, no cap — here's the whole value lineup:

sad — sum of absolute differences, fast (default) ⚡
sse — sum of squared errors
satd — sum of absolute Hadamard transformed differences
dct — sum of absolute DCT transformed differences
psnr — sum of squared quantization errors (avoid, low quality) 😬
bit — number of bits needed for the block
rd — rate distortion optimal, slow 🐢
zero — 0
vsad — sum of absolute vertical differences
vsse — sum of squared vertical differences
nsse — noise preserving sum of squared differences
w53 — 5/3 wavelet, only used in snow
w97 — 9/7 wavelet, only used in snow
dctmax
chroma

ildctcmp integer (encoding,video) — Set interlaced dct compare function. 🧵
- Possible values:
  - sad — sum of absolute differences, fast (default)
  - sse — sum of squared errors
  - satd — sum of absolute Hadamard transformed differences
  - dct — sum of absolute DCT transformed differences
  - psnr — sum of squared quantization errors (avoid, low quality)
  - bit — number of bits needed for the block
  - rd — rate distortion optimal, slow
  - zero — 0
  - vsad — sum of absolute vertical differences
  - vsse — sum of squared vertical differences
  - nsse — noise preserving sum of squared differences
  - w53 — 5/3 wavelet, only used in snow
  - w97 — 9/7 wavelet, only used in snow
  - dctmax
  - chroma
dia_size integer (encoding,video) — Set diamond type & size for motion estimation. 💎
- (1024, INT_MAX) — full motion estimation(slowest)
- (768, 1024] — umh motion estimation
- (512, 768] — hex motion estimation
- (256, 512] — l2s diamond motion estimation
- [2,256] — var diamond motion estimation
- (-1, 2) — small diamond motion estimation
- -1 — funny diamond motion estimation 😭
- (INT_MIN, -1) — sab diamond motion estimation
last_pred integer (encoding,video) — Set amount of motion predictors from the previous frame.
precmp integer (encoding,video) — Set pre motion estimation compare function. 🔍
- Possible values:
  - sad — sum of absolute differences, fast (default)
  - sse — sum of squared errors
  - satd — sum of absolute Hadamard transformed differences
  - dct — sum of absolute DCT transformed differences
  - psnr — sum of squared quantization errors (avoid, low quality)
  - bit — number of bits needed for the block
  - rd — rate distortion optimal, slow
  - zero — 0
  - vsad — sum of absolute vertical differences
  - vsse — sum of squared vertical differences
  - nsse — noise preserving sum of squared differences
  - w53 — 5/3 wavelet, only used in snow
  - w97 — 9/7 wavelet, only used in snow
  - dctmax
  - chroma
pre_dia_size integer (encoding,video) — Set diamond type & size for motion estimation pre-pass.
subq integer (encoding,video) — Set sub pel motion estimation quality.
me_range integer (encoding,video) — Set limit motion vectors range (1023 for DivX player).
global_quality integer (encoding,audio,video)
slice_flags integer
mbd integer (encoding,video) — Set macroblock decision algorithm (high quality mode). 🧠
- Possible values:
  - simple — use mbcmp (default)
  - bits — use fewest bits
  - rd — use best rate distortion
rc_init_occupancy integer (encoding,video) — Set number of bits which should be loaded into the rc buffer before decoding starts.
flags2 flags (decoding/encoding,audio,video,subtitles) — Possible values:
- fast — Allow non spec compliant speedup tricks.
- noout — Skip bitstream encoding.
- ignorecrop — Ignore cropping information from sps.
- local_header — Place global headers at every keyframe instead of in extradata.
- chunks — Frame data might be split into multiple chunks.
- showall — Show all frames before the first keyframe.
- export_mvs — Export motion vectors into frame side-data (see AV_FRAME_DATA_MOTION_VECTORS) for codecs that support it. See also doc/examples/export_mvs.c.
- skip_manual — Do not skip samples and export skip information as frame side data.
- ass_ro_flush_noop — Do not reset ASS ReadOrder field on flush.
- icc_profiles — Generate/parse embedded ICC profiles from/to colorimetry tags.
export_side_data flags (decoding/encoding,audio,video,subtitles) — Possible values:
- mvs — Export motion vectors into frame side-data (see AV_FRAME_DATA_MOTION_VECTORS) for codecs that support it. See also doc/examples/export_mvs.c.
- prft — Export encoder Producer Reference Time into packet side-data (see AV_PKT_DATA_PRFT) for codecs that support it.
- venc_params — Export video encoding parameters through frame side data (see AV_FRAME_DATA_VIDEO_ENC_PARAMS) for codecs that support it. At present, those are H.264 and VP9.
- film_grain — Export film grain parameters through frame side data (see AV_FRAME_DATA_FILM_GRAIN_PARAMS). Supported at present by AV1 decoders.
- enhancements — Export picture enhancement metadata through frame side data, e.g. LCEVC (see AV_FRAME_DATA_LCEVC).
threads integer (decoding/encoding,video) — Set the number of threads to be used, in case the selected codec implementation supports multi-threading. 🧵
- Possible values:
  - auto, 0 — automatically select the number of threads to set
- Default value is auto.
dc integer (encoding,video) — Set intra_dc_precision.
nssew integer (encoding,video) — Set nsse weight.
skip_top integer (decoding,video) — Set number of macroblock rows at the top which are skipped.
skip_bottom integer (decoding,video) — Set number of macroblock rows at the bottom which are skipped.
profile integer (encoding,audio,video) — Set encoder codec profile. Default value is unknown. Encoder specific profiles are documented in the relevant encoder documentation.
level integer (encoding,audio,video) — Set the encoder level. This level depends on the specific codec, and might correspond to the profile level. It is set by default to unknown.
- Possible values:
  - unknown
lowres integer (decoding,audio,video) — Decode at 1= 1/2, 2=1/4, 3=1/8 resolutions.
mblmin integer (encoding,video) — Set min macroblock lagrange factor (VBR).
mblmax integer (encoding,video) — Set max macroblock lagrange factor (VBR).
skip_loop_filter, skip_idct, skip_frame — integer (decoding,video). Make decoder discard processing depending on the frame type selected by the option value.
- skip_loop_filter skips frame loop filtering, skip_idct skips frame IDCT/dequantization, skip_frame skips decoding.
- Possible values:
  - none — Discard no frame.
  - default — Discard useless frames like 0-sized frames.
  - noref — Discard all non-reference frames.
  - bidir — Discard all bidirectional frames.
  - nokey — Discard all frames excepts keyframes.
  - nointra — Discard all frames except I frames.
  - all — Discard all frames.
- Default value is default.
bidir_refine integer (encoding,video) — Refine the two motion vectors used in bidirectional macroblocks.
keyint_min integer (encoding,video) — Set minimum interval between IDR-frames.
refs integer (encoding,video) — Set reference frames to consider for motion compensation.
trellis integer (encoding,audio,video) — Set rate-distortion optimal quantization.
mv0_threshold integer (encoding,video)
compression_level integer (encoding,audio,video)
bits_per_raw_sample integer
channel_layout integer (decoding/encoding,audio) — See the Channel Layout section in the ffmpeg-utils(1) manual for the required syntax.
rc_max_vbv_use float (encoding,video)
rc_min_vbv_use float (encoding,video)
color_primaries integer (decoding/encoding,video) — Possible values:
- bt709 — BT.709
- bt470m — BT.470 M
- bt470bg — BT.470 BG
- smpte170m — SMPTE 170 M
- smpte240m — SMPTE 240 M
- film — Film
- bt2020 — BT.2020
- smpte428, smpte428_1 — SMPTE ST 428-1
- smpte431 — SMPTE 431-2
- smpte432 — SMPTE 432-1
- jedec-p22 — JEDEC P22
color_trc integer (decoding/encoding,video) — Possible values:
- bt709 — BT.709
- gamma22 — BT.470 M
- gamma28 — BT.470 BG
- smpte170m — SMPTE 170 M
- smpte240m — SMPTE 240 M
- linear — Linear
- log, log100 — Log
- log_sqrt, log316 — Log square root
- iec61966_2_4, iec61966-2-4 — IEC 61966-2-4
- bt1361, bt1361e — BT.1361
- iec61966_2_1, iec61966-2-1 — IEC 61966-2-1
- bt2020_10, bt2020_10bit — BT.2020 - 10 bit
- bt2020_12, bt2020_12bit — BT.2020 - 12 bit
- smpte2084 — SMPTE ST 2084
- smpte428, smpte428_1 — SMPTE ST 428-1
- arib-std-b67 — ARIB STD-B67
colorspace integer (decoding/encoding,video) — Possible values:
- rgb — RGB
- bt709 — BT.709
- fcc — FCC
- bt470bg — BT.470 BG
- smpte170m — SMPTE 170 M
- smpte240m — SMPTE 240 M
- ycocg — YCOCG
- bt2020nc, bt2020_ncl — BT.2020 NCL
- bt2020c, bt2020_cl — BT.2020 CL
- smpte2085 — SMPTE 2085
- chroma-derived-nc — Chroma-derived NCL
- chroma-derived-c — Chroma-derived CL
- ictcp — ICtCp
color_range integer (decoding/encoding, video) — when you use this as an input parameter it's basically a lil hint to the decoder like "hey bestie, here's the color_range the input is serving". Possible values:
- tv
- mpeg
- limited — MPEG (219*2^(n-8))
- pc
- jpeg
- full — JPEG (2^n-1)
chroma_sample_location integer (decoding/encoding, video) — where the chroma is at, location-wise. Possible values:
- left
- center
- topleft
- top
- bottomleft
- bottom
alpha_mode integer (decoding/encoding, video) — the alpha vibe. Possible values:
- premultiplied
- straight
log_level_offset integer — sets the log level offset. that's it, that's the whole thing 💅
slices integer (encoding, video) — how many slices you want, used in parallelized encoding. more slices = more parallel besties doing the work.
thread_type flags (decoding/encoding, video) — picks which multithreading methods get used. no cap.

using frame will increase decoding delay by one frame per thread, so clients that literally cannot provide future frames should not use it. read that again, it matters 😭

Possible values:

slice — decode more than one part of a single frame at once. multithreading with slices only works when the video was encoded with slices, iykyk.
frame — decode more than one frame at once.

Default value is slice+frame. iconic default, honestly.

audio_service_type integer (encoding, audio) — sets the audio service type. Possible values:
- ma — Main Audio Service
- ef — Effects
- vi — Visually Impaired
- hi — Hearing Impaired
- di — Dialogue
- co — Commentary
- em — Emergency
- vo — Voice Over
- ka — Karaoke
request_sample_fmt sample_fmt (decoding, audio) — sets the sample format audio decoders should prefer. Default value is "none". manifesting a sample format for u ✨
pkt_timebase rational number — the timebase for packets. look at it. it's a rational number. that's the whole description, she's concise today 💅
sub_charenc encoding (decoding, subtitles) — sets the input subtitles character encoding. so ur subs don't come out looking like mojibake 🙈
field_order field_order (video) — sets/overrides the field order of the video. Possible values:
- progressive — Progressive video
- tt — Interlaced video, top field coded and displayed first
- bb — Interlaced video, bottom field coded and displayed first
- tb — Interlaced video, top coded first, bottom displayed first
- bt — Interlaced video, bottom coded first, top displayed first
skip_alpha bool (decoding, video) — set to 1 to disable processing alpha (transparency). this works like the gray flag in the flags option, which skips chroma information instead of alpha. Default is 0.
codec_whitelist list (input) — a "," separated list of allowed decoders. by default all are allowed. the VIP list of decoders 🎀
dump_separator string (input) — the separator used to separate the fields printed on the command line about the Stream parameters. for example, to separate the fields with newlines and indentation:

ffprobe -dump_separator "
                          "  -i ~/videos/matrixbench_mpeg2.mpg

(yes the separator is literally a newline plus spaces. she contains multitudes.)

max_pixels integer (decoding/encoding, video) — maximum number of pixels per image. you can use this value to avoid out of memory failures due to large images. protect ur RAM bestie 🫶
apply_cropping bool (decoding, video) — enables cropping if the cropping parameters are multiples of the required alignment for the left and top parameters. if the alignment is not met, the cropping will be partially applied to maintain alignment. Default is 1 (enabled).

Note: the required alignment depends on if AV_CODEC_FLAG_UNALIGNED is set and the CPU. AV_CODEC_FLAG_UNALIGNED cannot be changed from the command line. also hardware decoders will not apply left/top Cropping. not them skipping the crop 😭

🎬 DECODERS

Decoders are configured elements in FFmpeg that allow the decoding of multimedia streams. they're the ones turning ur mystery bytes back into actual video. main character energy, honestly.

When you configure your FFmpeg build, all the supported native decoders are enabled by default. decoders requiring an external library must be enabled manually via the corresponding --enable-lib option. you can list all available decoders using the configure option --list-decoders.

You can disable all the decoders with the configure option --disable-decoders and selectively enable / disable single decoders with the options --enable-decoder=DECODER / --disable-decoder=DECODER.

The option -decoders of the ff* tools will display the list of enabled decoders.

🎥 VIDEO DECODERS

A description of some of the currently available video decoders follows. buckle up bestie 💖

av1

AOMedia Video 1 (AV1) decoder. she's everywhere now and she ate.

Options

operating_point — select an operating point of a scalable AV1 bitstream (0 - 31). Default is 0.

hevc

HEVC (AKA ITU-T H.265 or ISO/IEC 23008-2) decoder. yes it has three names, it's that girl.

The decoder supports MV-HEVC multiview streams with at most two views. Views to be output are selected by supplying a list of view IDs to the decoder (the view_ids option). This option may be set either statically before decoder init, or from the get_format() callback — useful for the case when the view count or IDs change dynamically during decoding.

Only the base layer is decoded by default.

Note that if you are using the ffmpeg CLI tool, you should be using view specifiers as documented in its manual, rather than the options documented here. like, read the other manual, bestie, this one's not it for that 🌸

Options

view_ids (MV-HEVC) — specify a list of view IDs that should be output. this option can also be set to a single -1, which will cause all views defined in the VPS to be decoded and output.
view_ids_available (MV-HEVC) — this option may be read by the caller to retrieve an array of view IDs available in the active VPS. the array is empty for single-layer video.

The value of this option is guaranteed to be accurate when read from the get_format() callback. It may also be set at other times (e.g. after opening the decoder), but the value is informational only and may be incorrect (e.g. when the stream contains multiple distinct VPS NALUs). so like... trust it, but only in the callback 💀

view_pos_available (MV-HEVC) — this option may be read by the caller to retrieve an array of view positions (left, right, or unspecified) available in the active VPS, as AVStereo3DView values. When the array is available, its elements apply to the corresponding elements of view_ids_available, i.e. view_pos_available[i] contains the position of view with ID view_ids_available[i].

Same validity restrictions as for view_ids_available apply to this option. same rules, same caveats, no exceptions.

rawvideo

Raw video decoder. it decodes rawvideo streams. what did u expect. she's literally named after her job.

Options

top top_field_first — specify the assumed field type of the input video.
- -1 — the video is assumed to be progressive (default)
- 0 — bottom-field-first is assumed
- 1 — top-field-first is assumed

libdav1d

dav1d AV1 decoder. the fast one. the beloved one.

libdav1d allows libavcodec to decode the AOMedia Video 1 (AV1) codec. Requires the presence of the libdav1d headers and library during configuration. You need to explicitly configure the build with --enable-libdav1d.

Options

The following options are supported by the libdav1d wrapper.

max_frame_delay — set max amount of frames the decoder may buffer internally. the default value is 0 (autodetect).
filmgrain — apply film grain to the decoded video if present in the bitstream. defaults to the internal default of the library.

this option is deprecated and will be removed in the future. see the global option export_side_data to export Film Grain parameters instead of applying it. like pls stop using it bestie, it's on the way out 🙈

oppoint — select an operating point of a scalable AV1 bitstream (0 - 31). defaults to the internal default of the library.
alllayers — output all spatial layers of a scalable AV1 bitstream. the default value is false.

libdavs2

AVS2-P2/IEEE1857.4 video decoder wrapper.

This decoder allows libavcodec to decode AVS2 streams with davs2 library. that's the whole thing, she's a wrapper and she's proud 🦋

libuavs3d

AVS3-P2/IEEE1857.10 video decoder.

libuavs3d allows libavcodec to decode AVS3 streams. Requires the presence of the libuavs3d headers and library during configuration. You need to explicitly configure the build with --enable-libuavs3d.

Options

The following option is supported by the libuavs3d wrapper.

frame_threads — set amount of frame threads to use during decoding. the default value is 0 (autodetect).

libxevd

eXtra-fast Essential Video Decoder (XEVD) MPEG-5 EVC decoder wrapper.

This decoder requires the presence of the libxevd headers and library during configuration. You need to explicitly configure the build with --enable-libxevd.

The xevd project website is at https://github.com/mpeg5/xevd.

Options

The following options are supported by the libxevd wrapper. The xevd-equivalent options or values are listed in parentheses for easy migration. thoughtful queen 💅

To get a more accurate and extensive documentation of the libxevd options, invoke the command xevd_app --help or consult the libxevd documentation.

threads (threads) — force to use a specific number of threads. no autodetect, no vibes, just a number.

QSV Decoders

The family of Intel QuickSync Video decoders (VC1, MPEG-2, H.264, HEVC, JPEG/MJPEG, VP8, VP9, AV1, VVC). the whole squad, all invited to the party ⭐

Common Options

The following options are supported by all qsv decoders.

async_depth — internal parallelization depth, the higher the value the higher the latency. yes that's a tradeoff, no you can't have both, choose ur fighter.
gpu_copy — a GPU-accelerated copy between video and system memory. possible values:
- default
- on
- off

HEVC Options

Extra options for hevc_qsv.

load_plugin — a user plugin to load in an internal session.
- none
- hevc_sw
- hevc_hw
load_plugins — a :-separate list of hexadecimal plugin UIDs to load in an internal session.

v210

Uncompressed 4:2:2 10-bit decoder. no compression, just vibes and bits.

Options

custom_stride — set the line size of the v210 data in bytes. the default value is 0 (autodetect). you can use the special -1 value for a strideless v210 as seen in BOXX files. special little guy 🎀

🎶 AUDIO DECODERS 🎶

okay bestie, it's audio decoder time 💖 here's a description of some of the currently available audio decoders, and yes we're going through every single one like the responsible little Unix gremlin you are ✨

🎧 ac3 🎧

AC-3 audio decoder. iconic behavior honestly.

This decoder implements part of ATSC A/52:2010 and ETSI TS 102 366, as well as the undocumented RealAudio 3 (a.k.a. dnet). no cap, it's doing both the documented and the mysterious 💅

🎚️ AC-3 Decoder Options 🎚️

-drc_scale value — Dynamic Range Scale Factor. The factor to apply to dynamic range values from the AC-3 stream. This factor is applied exponentially. The default value is 1. There are 3 notable scale factor ranges:
- drc_scale == 0 — DRC disabled. Produces full range audio.
- 0 < drc_scale <= 1 — DRC enabled. Applies a fraction of the stream DRC value. Audio reproduction is between full range and full compression.
- drc_scale > 1 — DRC enabled. Applies drc_scale asymmetrically. Loud sounds are fully compressed. Soft sounds are enhanced.

🌸 flac 🌸

FLAC audio decoder. she's giving lossless main character energy 🦋

This decoder aims to implement the complete FLAC specification from Xiph. periodt.

🎀 FLAC Decoder options 🎀

-use_buggy_lpc — The lavc FLAC encoder used to produce buggy streams with high lpc values (like the default value). This option makes it possible to decode such streams correctly by using lavc's old buggy lpc logic for decoding. it's literally a little time machine for buggy streams 😭

🧚 ffwavesynth 🧚

Internal wave synthesizer. very mysterious, very internal, very "don't worry about it" 💜

This decoder generates wave patterns according to predefined sequences. Its use is purely internal and the format of the data it accepts is not publicly documented. so yeah, private vibes only 🙈

🦋 libcelt 🦋

libcelt decoder wrapper.

libcelt allows libavcodec to decode the Xiph CELT ultra-low delay audio codec. Requires the presence of the libcelt headers and library during configuration. You need to explicitly configure the build with --enable-libcelt. slay, but configure it first bestie ✨

🍓 libgsm 🍓

libgsm decoder wrapper.

libgsm allows libavcodec to decode the GSM full rate audio codec. Requires the presence of the libgsm headers and library during configuration. You need to explicitly configure the build with --enable-libgsm.

This decoder supports both the ordinary GSM and the Microsoft variant. two flavors, one wrapper, we love a versatile icon 💅

🎀 libilbc 🎀

libilbc decoder wrapper.

libilbc allows libavcodec to decode the Internet Low Bitrate Codec (iLBC) audio codec. Requires the presence of the libilbc headers and library during configuration. You need to explicitly configure the build with --enable-libilbc.

✨ Options ✨

The following option is supported by the libilbc wrapper.

enhance — Enable the enhancement of the decoded audio when set to 1. The default value is 0 (disabled). so if you want the glow-up, set it to 1 🫶

🌈 libmpeghdec 🌈

libmpeghdec decoder wrapper.

libmpeghdec allows libmpeghdec to decode the MPEG-H 3D audio codec. Requires the presence of the libmpeghdec headers and library during configuration. You need to explicitly configure the build with --enable-libmpeghdec --enable-nonfree. note the --enable-nonfree part, bestie, it's right there in the configure line 👀

🔥 libopencore-amrnb 🔥

libopencore-amrnb decoder wrapper.

libopencore-amrnb allows libavcodec to decode the Adaptive Multi-Rate Narrowband audio codec. Using it requires the presence of the libopencore-amrnb headers and library during configuration. You need to explicitly configure the build with --enable-libopencore-amrnb.

An FFmpeg native decoder for AMR-NB exists, so users can decode AMR-NB without this library. so like, this wrapper is cute but not mandatory 💖

🌟 libopencore-amrwb 🌟

libopencore-amrwb decoder wrapper.

libopencore-amrwb allows libavcodec to decode the Adaptive Multi-Rate Wideband audio codec. Using it requires the presence of the libopencore-amrwb headers and library during configuration. You need to explicitly configure the build with --enable-libopencore-amrwb.

An FFmpeg native decoder for AMR-WB exists, so users can decode AMR-WB without this library. again: native option exists, no stress required 🌸

👑 libopus 👑

libopus decoder wrapper.

libopus allows libavcodec to decode the Opus Interactive Audio Codec. Requires the presence of the libopus headers and library during configuration. You need to explicitly configure the build with --enable-libopus.

An FFmpeg native decoder for Opus exists, so users can decode Opus without this library. iconic of FFmpeg to have a native fallback, honestly 💜

💜 SUBTITLES DECODERS 💜

These are the subtitle decoders, bestie — the little guys that turn squiggly caption data into actual readable words on your screen. Let's go through them one by one, no cap.

🌸 libaribb24

A caption decoder for ARIB STD-B24. It implements profiles A and C of the ARIB STD-B24 standard. That's it, that's the whole vibe.

✨ libaribb24 Decoder Options ✨

-aribb24-base-path path — sets the base path for the libaribb24 library. It's used for reading configuration files (for custom unicode conversions) and for dumping non-text symbols as images under that location. Unset by default — so if you don't set it, nothing's set, periodt.
-aribb24-skip-ruby-text boolean — tells the decoder wrapper to skip text blocks that contain half-height ruby text. It's enabled by default — so she's already skipping that for you, slay.

🦋 libaribcaption

Yet another ARIB STD-B24 caption decoder, this one leaning on the external libaribcaption library. It implements profiles A and C of the Japanese ARIB STD-B24 standard, plus Brazilian ABNT NBR 15606-1, plus the Philippines version of ISDB-T. Versatile icon.

It needs the libaribcaption headers and library (https://github.com/xqq/libaribcaption) present during configuration. You have to explicitly configure the build with "--enable-libaribcaption" — it will not just show up. And listen: if both libaribb24 and libaribcaption are enabled, libaribcaption decoder precedes. She's the main character in that scenario, iykyk.

🎀 libaribcaption Decoder Options 🎀

-sub_type subtitle_type — specifies the format of the decoded subtitles:
- bitmap — graphical image.
- ass — ASS formatted text.
- text — simple text based output without formatting.

The default is ass, same as the libaribb24 decoder. Some present players (e.g., mpv) expect ASS format for ARIB caption. So the default is doing the work for you fr fr.

-caption_encoding encoding_scheme — specifies the encoding scheme of input subtitle text:
- auto — automatically detect text encoding (default).
- jis — 8bit-char JIS encoding defined in ARIB STD B24. This encoding is used in Japan for ISDB captions.
- utf8 — UTF-8 encoding defined in ARIB STD B24. This encoding is used in Philippines for ISDB-T captions.
- latin — Latin character encoding defined in ABNT NBR 15606-1. This encoding is used in South America for SBTVD / ISDB-Tb captions.
-font font_name[,font_name2,...] — specify a comma-separated list of font family names to be used for bitmap or ass type subtitle rendering. Only the first font name is used for ass type subtitle. If not specified, it uses the internally defined default font family. 💅
-ass_single_rect boolean — okay so ARIB STD-B24 specifies that some captions may be displayed at different positions at a time (that's the multi-rectangle subtitle situation). Since some players (e.g., old mpv) can't handle multiple ASS rectangles in a single AVSubtitle, or multiple ASS rectangles of indeterminate duration with the same start timestamp, this option can change the behavior so that all the texts are displayed in a single ASS rectangle. The default is false. If your player cannot handle AVSubtitles with multiple ASS rectangles properly, set this option to true or define ASS_SINGLE_RECT=1 to change default behavior at compilation. Fix it at build time or fix it at runtime, your call bestie.
-force_outline_text boolean — specify whether to always render outline text for all characters, regardless of the indication by character style. The default is false.
-outline_width number (0.0 - 3.0) — specify width for outline text, in dots (relative). The default is 1.5.
-ignore_background boolean — specify whether to ignore background color rendering. The default is false.
-ignore_ruby boolean — specify whether to ignore rendering for ruby-like (furigana) characters. The default is false.
-replace_drcs boolean — specify whether to render replaced DRCS characters as Unicode characters. The default is true.
-replace_msz_ascii boolean — specify whether to replace MSZ (Middle Size; half width) fullwidth alphanumerics with halfwidth alphanumerics. The default is true.
-replace_msz_japanese boolean — specify whether to replace some MSZ (Middle Size; half width) fullwidth japanese special characters with halfwidth ones. The default is true.
-replace_msz_glyph boolean — specify whether to replace MSZ (Middle Size; half width) characters with halfwidth glyphs if the font supports it. This option works under FreeType or DirectWrite renderer with Adobe-Japan1 compliant fonts — e.g., IBM Plex Sans JP, Morisawa BIZ UDGothic, Morisawa BIZ UDMincho, Yu Gothic, Yu Mincho, and Meiryo. The default is true. Font-girlie behavior, honestly iconic.
-canvas_size image_size — specify the resolution of the canvas to render subtitles to; usually, this should be the frame size of the input video. This only applies when "-subtitle_type" is set to bitmap.

The libaribcaption decoder assumes input frame size for bitmap rendering as below:

PROFILE_A : 1440 x 1080 with SAR (PAR) 4:3
PROFILE_C : 320 x 180 with SAR (PAR) 1:1

If the actual frame size of the input video does not match the above assumption, the rendered captions may be distorted. To make the captions undistorted, add the "-canvas_size" option to specify the actual input video size.

Note that the "-canvas_size" option is not required for video with a different size but the same aspect ratio. In such cases, the caption will be stretched or shrunk to the actual video size if the "-canvas_size" option is not specified. If the "-canvas_size" option is specified with a different size, the caption will be stretched or shrunk to the specified size with the calculated SAR. Math girlie eats. 🔥

🌈 libaribcaption decoder usage examples 🌈

Display an MPEG-TS file with ARIB subtitle using the ffplay tool:

ffplay -sub_type bitmap MPEG.TS

Display an MPEG-TS file with input frame size 1920x1080 using the ffplay tool:

ffplay -sub_type bitmap -canvas_size 1920x1080 MPEG.TS

Embed ARIB subtitle in transcoded video:

ffmpeg -sub_type bitmap -i src.m2t -filter_complex "[0:v][0:s]overlay" -vcodec h264 dest.mp4

⭐ dvbsub

Options

compute_clut
- -2 — compute CLUT once if no matching CLUT is in the stream.
- -1 — compute CLUT if no matching CLUT is in the stream.
- 0 — never compute CLUT.
- 1 — always compute CLUT and override the one provided in the stream. She's not asking, she's taking over. 💀
dvb_substream — selects the dvb substream, or all substreams if -1, which is the default.

🍓 dvdsub

This codec decodes the bitmap subtitles used in DVDs — and the same subtitles can also be found in VobSub file pairs and in some Matroska files. Glow-up across containers, we love to see it.

Options

palette — specify the global palette used by the bitmaps. When stored in VobSub, the palette is normally specified in the index file; in Matroska, the palette is stored in the codec extra-data in the same format as in VobSub. In DVDs, the palette is stored in the IFO file, and therefore not available when reading from dumped VOB files. (RIP, she's just not there.)

The format for this option is a string containing 16 24-bits hexadecimal numbers (without 0x prefix) separated by commas, for example "0d00ee, ee450d, 101010, eaeaea, 0ce60b, ec14ed, ebff0b, 0d617a, 7b7b7b, d1d1d1, 7b2a0e, 0d950c, 0f007b, cf0dec, cfa80c, 7c127b".

ifo_palette — specify the IFO file from which the global palette is obtained. (experimental 👀)
forced_subs_only — only decode subtitle entries marked as forced. Some titles have forced and non-forced subtitles in the same track. Setting this flag to 1 will only keep the forced subtitles. Default value is 0.

🧚 libzvbi-teletext

Libzvbi allows libavcodec to decode DVB teletext pages and DVB teletext subtitles. It requires the presence of the libzvbi headers and library during configuration — you need to explicitly configure the build with "--enable-libzvbi". No sneaky auto-enable here, bestie.

Options

txt_page — list of teletext page numbers to decode. Pages that do not match the specified list are dropped. You may use the special "*" string to match all pages, or "subtitle" to match all subtitle pages. Default value is *.
txt_default_region — set the default character set used for decoding, a value between 0 and 87 (see ETS 300 706, Section 15, Table 32). Default value is -1, which does not override the libzvbi default. This option is needed for some legacy level 1.0 transmissions which cannot signal the proper charset. Legacy girlies need extra help sometimes, no shame. 🫶
txt_chop_top — discards the top teletext line. Default value is 1.
txt_format — specifies the format of the decoded subtitles:
- bitmap — the default format; you should use this for teletext pages, because certain graphics and colors cannot be expressed in simple text or even ASS.
- text — simple text based output without formatting.
- ass — formatted ASS output, subtitle pages and teletext pages are returned in different styles; subtitle pages are stripped down to text, but an effort is made to keep the text alignment and the formatting.
txt_left — X offset of generated bitmaps, default is 0.
txt_top — Y offset of generated bitmaps, default is 0.
txt_chop_spaces — chops leading and trailing spaces and removes empty lines from the generated text. This option is useful for teletext based subtitles where empty spaces may be present at the start or at the end of the lines, or empty lines may be present between the subtitle lines because of double-sized teletext characters. Default value is 1. Trim her, she deserves it. ✨
txt_duration — sets the display duration of the decoded teletext pages or subtitles in milliseconds. Default value is -1, which means infinity, or until the next subtitle event comes.
txt_transparent — force a transparent background of the generated teletext bitmaps. Default value is 0, which means an opaque background.
txt_opacity — sets the opacity (0-255) of the teletext background. If txt_transparent is not set, it only affects characters between a start box and an end box, typically subtitles. Default value is 0 if txt_transparent is set, 255 otherwise.

👑 ENCODERS 👑

Encoders are configured elements in FFmpeg which allow the encoding of multimedia streams. The other direction of the pipeline, basically — she takes your frames and squishes them into a stream. 💖

When you configure your FFmpeg build, all the supported native encoders are enabled by default. Encoders requiring an external library must be enabled manually via the corresponding "--enable-lib" option. You can list all available encoders using the configure option "--list-encoders".

You can disable all the encoders with the configure option "--disable-encoders" and selectively enable / disable single encoders with the options "--enable-encoder=ENCODER" / "--disable-encoder=ENCODER".

The option -encoders of the ff* tools will display the list of enabled encoders. Go check her out, she'll tell you exactly who's available. 💅

💖 AUDIO ENCODERS 💖

Okay bestie, here’s the tea on some of the audio encoders you can use right now. 🎧✨

🎀 aac 🎀

Advanced Audio Coding (AAC) encoder. This is the default AAC encoder, natively built right into FFmpeg — no extra plugin energy needed. 💅

🎛️ Options 🎛️

b — Set bit rate in bits/s. Setting this automatically activates constant bit rate (CBR) mode. If you don’t specify it, it’s set to 128kbps.
q — Set quality for variable bit rate (VBR) mode. This option is valid only when using the ffmpeg command-line tool. For library interface users, use global_quality.
cutoff — Set cutoff frequency. If unspecified, the encoder is allowed to dynamically adjust the cutoff to improve clarity on low bitrates.
aac_coder — Set AAC encoder coding method. Possible values:
- twoloop — Two loop searching (TLS) method. This is the default method.

This method first sets quantizers depending on band thresholds and then tries to find an optimal combination by adding or subtracting a specific value from all quantizers and adjusting some individual quantizer a little. It will tune itself based on whether aac_is, aac_ms and aac_pns are enabled.

anmr — Average noise to mask ratio (ANMR) trellis-based solution.

This is an experimental coder which currently produces lower quality, is more unstable and is slower than the default twoloop coder but has potential. It currently has no support for the aac_is or aac_pns options. Not currently recommended.

fast — Constant quantizer method.

Uses a cheaper version of the twoloop algorithm that doesn’t try to do as many clever adjustments. Worse with low bitrates (less than 64kbps), but is better and much faster at higher bitrates.

aac_ms — Sets mid/side coding mode. The default value of auto automatically uses M/S with bands that will benefit from such coding. You can force it for all bands with enable, which is mainly useful for debugging, or disable it with disable.
aac_is — Sets intensity stereo coding tool usage. By default it’s enabled and will automatically toggle IS for similar pairs of stereo bands if that’s beneficial. You can disable it for debugging by setting the value to disable.
aac_pns — Uses perceptual noise substitution to replace low entropy high frequency bands with imperceptible white noise during the decoding process. By default it’s enabled, but you can disable it for debugging with disable.
aac_tns — Enables the use of a multitap FIR filter which spans through the high frequency bands to hide quantization noise during the encoding process and is reverted by the decoder. As well as decreasing unpleasant artifacts in the high range, this also reduces entropy in the high bands and allows more bits to be used by the mid-low bands. By default it’s enabled, but you can disable it for debugging by setting the option to disable.
aac_ltp — Enables the use of the long term prediction extension which increases coding efficiency in very low bandwidth situations, like encoding voice or solo piano music, by extending constant harmonic peaks in bands throughout frames. This option is implied by profile:a aac_low. Use it in conjunction with -ar to decrease the samplerate.
profile — Sets the encoding profile. Possible values:
- aac_low — The default, AAC "Low-complexity" profile. It’s the most compatible and produces decent quality.
- mpeg2_aac_low — Equivalent to -profile:a aac_low -aac_pns 0. PNS was introduced with the MPEG4 specifications.
- aac_ltp — Long term prediction profile, enabled by and will enable the aac_ltp option. Introduced in MPEG4.

If this option is unspecified it is set to aac_low.

🔥 ac3 and ac3_fixed 🔥

AC-3 audio encoders. These encoders implement part of ATSC A/52:2010 and ETSI TS 102 366. 🎬

The ac3 encoder uses floating-point math, while the ac3_fixed encoder only uses fixed-point integer math. This does not mean one is always faster — just that one or the other may be better suited to a particular system. The ac3_fixed encoder is not the default codec for any of the output formats, so you have to specify it explicitly using the option -acodec ac3_fixed to use it.

📼 AC-3 Metadata 📼

The AC-3 metadata options are used to set parameters that describe the audio, but in most cases they do not affect the audio encoding itself. Some of these options directly affect or influence the decoding and playback of the resulting bitstream, while others are just for informational purposes. A few of the options will add bits to the output stream that could otherwise be used for audio data, and will thus affect the quality of the output. Those will be indicated accordingly with a note in the option list below. 👀

These parameters are described in detail in several publicly-available documents:

Metadata Control Options

-per_frame_metadata boolean — Allow Per-Frame Metadata. Specifies if the encoder should check for changing metadata for each frame.
- 0 — The metadata values set at initialization will be used for every frame in the stream. (default)
- 1 — Metadata values can be changed before encoding each frame.

Downmix Levels

-center_mixlev level — Center Mix Level. The amount of gain the decoder should apply to the center channel when downmixing to stereo. This field will only be written to the bitstream if a center channel is present. The value is specified as a scale factor. There are 3 valid values:
- 0.707 — Apply -3dB gain
- 0.595 — Apply -4.5dB gain (default)
- 0.500 — Apply -6dB gain
-surround_mixlev level — Surround Mix Level. The amount of gain the decoder should apply to the surround channel(s) when downmixing to stereo. This field will only be written to the bitstream if one or more surround channels are present. The value is specified as a scale factor. There are 3 valid values:
- 0.707 — Apply -3dB gain
- 0.500 — Apply -6dB gain (default)
- 0.000 — Silence Surround Channel(s)

Audio Production Information

Audio Production Information is optional info describing the mixing environment. Either none or both of the fields are written to the bitstream. ✨

-mixing_level number — Mixing Level. Specifies peak sound pressure level (SPL) in the production environment when the mix was mastered. Valid values are 80 to 111, or -1 for unknown or not indicated. The default value is -1, but that value cannot be used if the Audio Production Information is written to the bitstream. So if the room_type option is not the default value, the mixing_level option must not be -1.
-room_type type — Room Type. Describes the equalization used during the final mixing session at the studio or on the dubbing stage. A large room is a dubbing stage with the industry standard X-curve equalization; a small room has flat equalization. This field will not be written to the bitstream if both the mixing_level option and the room_type option have the default values.
- 0, notindicated — Not Indicated (default)
- 1, large — Large Room
- 2, small — Small Room

Other Metadata Options

-copyright boolean — Copyright Indicator. Specifies whether a copyright exists for this audio.
- 0, off — No Copyright Exists (default)
- 1, on — Copyright Exists
-dialnorm value — Dialogue Normalization. Indicates how far the average dialogue level of the program is below digital 100% full scale (0 dBFS). This parameter determines a level shift during audio reproduction that sets the average volume of the dialogue to a preset level. The goal is to match volume level between program sources. A value of -31dB will result in no volume level change, relative to the source volume, during audio reproduction. Valid values are whole numbers in the range -31 to -1, with -31 being the default.
-dsur_mode mode — Dolby Surround Mode. Specifies whether the stereo signal uses Dolby Surround (Pro Logic). This field will only be written to the bitstream if the audio stream is stereo. Using this option does NOT mean the encoder will actually apply Dolby Surround processing.
- 0, notindicated — Not Indicated (default)
- 1, off — Not Dolby Surround Encoded
- 2, on — Dolby Surround Encoded
-original boolean — Original Bit Stream Indicator. Specifies whether this audio is from the original source and not a copy.
- 0, off — Not Original Source
- 1, on — Original Source (default)

Extended Bitstream Information

The extended bitstream options are part of the Alternate Bit Stream Syntax as specified in Annex D of the A/52:2010 standard. It is grouped into 2 parts. If any one parameter in a group is specified, all values in that group will be written to the bitstream. Default values are used for those that are written but have not been specified. If the mixing levels are written, the decoder will use these values instead of the ones specified in the center_mixlev and surround_mixlev options if it supports the Alternate Bit Stream Syntax. 🧚

Extended Bitstream Information - Part 1

-dmix_mode mode — Preferred Stereo Downmix Mode. Allows the user to select either Lt/Rt (Dolby Surround) or Lo/Ro (normal stereo) as the preferred stereo downmix mode.
- 0, notindicated — Not Indicated (default)
- 1, ltrt — Lt/Rt Downmix Preferred
- 2, loro — Lo/Ro Downmix Preferred
-ltrt_cmixlev level — Lt/Rt Center Mix Level. The amount of gain the decoder should apply to the center channel when downmixing to stereo in Lt/Rt mode.
- 1.414 — Apply +3dB gain
- 1.189 — Apply +1.5dB gain
- 1.000 — Apply 0dB gain
- 0.841 — Apply -1.5dB gain

Okay so the level list wraps up like this, bestie:

0.707 — Apply -3.0dB gain
0.595 — Apply -4.5dB gain (default)
0.500 — Apply -6.0dB gain
0.000 — Silence Center Channel
-ltrt_surmixlev level — Lt/Rt Surround Mix Level. This is how much gain the decoder should slap onto the surround channel(s) when it downmixes to stereo in Lt/Rt mode.
- 0.841 — Apply -1.5dB gain
- 0.707 — Apply -3.0dB gain
- 0.595 — Apply -4.5dB gain
- 0.500 — Apply -6.0dB gain (default)
- 0.000 — Silence Surround Channel(s)
-loro_cmixlev level — Lo/Ro Center Mix Level. This is how much gain the decoder should apply to the center channel when downmixing to stereo in Lo/Ro mode.
- 1.414 — Apply +3dB gain
- 1.189 — Apply +1.5dB gain
- 1.000 — Apply 0dB gain
- 0.841 — Apply -1.5dB gain
- 0.707 — Apply -3.0dB gain
- 0.595 — Apply -4.5dB gain (default)
- 0.500 — Apply -6.0dB gain
- 0.000 — Silence Center Channel
-loro_surmixlev level — Lo/Ro Surround Mix Level. This is how much gain the decoder should apply to the surround channel(s) when downmixing to stereo in Lo/Ro mode.
- 0.841 — Apply -1.5dB gain
- 0.707 — Apply -3.0dB gain
- 0.595 — Apply -4.5dB gain
- 0.500 — Apply -6.0dB gain (default)
- 0.000 — Silence Surround Channel(s)

🫶 Extended Bitstream Information - Part 2

-dsurex_mode mode — Dolby Surround EX Mode. Tells you whether the stream uses Dolby Surround EX (7.1 matrixed to 5.1). Using this option does NOT mean the encoder will actually apply Dolby Surround EX processing.
- 0 / notindicated — Not Indicated (default)
- 1 / on — Dolby Surround EX Off
- 2 / off — Dolby Surround EX On
-dheadphone_mode mode — Dolby Headphone Mode. Tells you whether the stream uses Dolby Headphone encoding (multi-channel matrixed to 2.0 for headphones). Using this option does NOT mean the encoder will actually apply Dolby Headphone processing.
- 0 / notindicated — Not Indicated (default)
- 1 / on — Dolby Headphone Off
- 2 / off — Dolby Headphone On
-ad_conv_type type — A/D Converter Type. Tells you whether the audio has passed through HDCD A/D conversion.
- 0 / standard — Standard A/D Converter (default)
- 1 / hdcd — HDCD A/D Converter

🌈 Other AC-3 Encoding Options

-stereo_rematrixing boolean — Stereo Rematrixing. Turns rematrixing on/off for stereo input. This is an optional AC-3 feature that bumps quality by selectively encoding the left/right channels as mid/side. It’s enabled by default, and it’s highly recommended that you leave it enabled except for testing purposes.
cutoff frequency — Sets the lowpass cutoff frequency. If you don’t specify it, the encoder picks a default determined by various other encoding parameters.

💜 Floating-Point-Only AC-3 Encoding Options

Heads up: these options are only valid for the floating-point encoder, and they do not exist for the fixed-point encoder because the corresponding features aren’t implemented in fixed-point. No cap.

-channel_coupling boolean — Enables/Disables channel coupling, which is an optional AC-3 feature that increases quality by combining high frequency information from multiple channels into a single channel. The per-channel high frequency info gets sent with less accuracy in both the frequency and time domains. This lets more bits go to lower frequencies while preserving enough information to reconstruct the high frequencies. This option is enabled by default for the floating-point encoder and should generally stay enabled except for testing purposes or to increase encoding speed.
- -1 / auto — Selected by Encoder (default)
- 0 / off — Disable Channel Coupling
- 1 / on — Enable Channel Coupling
-cpl_start_band number — Coupling Start Band. Sets the channel coupling start band, from 1 to 15. If you use a value higher than the bandwidth, it gets reduced to 1 less than the coupling end band. If auto is used, the start band is determined by the encoder based on the bit rate, sample rate, and channel layout. This option has no effect if channel coupling is disabled.
- -1 / auto — Selected by Encoder (default)

🌸 flac

FLAC (Free Lossless Audio Codec) Encoder

✨ Options

The following options are supported by FFmpeg’s flac encoder. Slay.

compression_level — Sets the compression level. This chooses defaults for many other options if they aren’t set explicitly. Valid values are from 0 to 12, and 5 is the default.
frame_size — Sets the size of the frames in samples per channel.
lpc_coeff_precision — Sets the LPC coefficient precision. Valid values are from 1 to 15, and 15 is the default.
lpc_type — Sets the first stage LPC algorithm.
- none — LPC is not used
- fixed — fixed LPC coefficients
- levinson
- cholesky
lpc_passes — Number of passes to use for Cholesky factorization during LPC analysis.
min_partition_order — The minimum partition order.
max_partition_order — The maximum partition order.
prediction_order_method
- estimation
- 2level
- 4level
- 8level
- search — Bruteforce search
- log
ch_mode — Channel mode.
- auto — The mode is chosen automatically for each frame
- indep — Channels are independently coded
- left_side
- right_side
- mid_side
exact_rice_parameters — Chooses if rice parameters are calculated exactly or approximately. If set to 1, they’re chosen exactly — which slows the code down slightly and improves compression slightly. Cute little tradeoff.
multi_dim_quant — Multi Dimensional Quantization. If set to 1, a 2nd stage LPC algorithm is applied after the first stage to finetune the coefficients. This is quite slow and slightly improves compression.

🦋 opus

Opus encoder.

This is a native FFmpeg encoder for the Opus format. Currently, it’s in development and only implements the CELT part of the codec. Its quality is usually worse and at best is equal to the libopus encoder.

✨ Options

b — Set bit rate in bits/s. If unspecified, it uses the number of channels and the layout to make a good guess.
opus_delay — Sets the maximum delay in milliseconds. Lower delays than 20ms will very quickly decrease quality.

🎀 libfdk_aac

libfdk-aac AAC (Advanced Audio Coding) encoder wrapper.

The libfdk-aac library is based on the Fraunhofer FDK AAC code from the Android project.

It requires the presence of the libfdk-aac headers and library during configuration. You need to explicitly configure the build with --enable-libfdk-aac. The library is also incompatible with GPL, so if you allow the use of GPL, you should configure with --enable-gpl --enable-nonfree --enable-libfdk-aac.

This encoder has support for the AAC-HE profiles.

VBR encoding, enabled through the vbr or flags +qscale options, is experimental and only works with some combinations of parameters.

Support for encoding 7.1 audio is only available with libfdk-aac 0.1.3 or higher.

For more information see the fdk-aac project at <http://sourceforge.net/p/opencore-amr/fdk-aac/>.

✨ Options

The following options are mapped on the shared FFmpeg codec options.

b — Set bit rate in bits/s. If the bitrate is not explicitly specified, it is automatically set to a suitable value depending on the selected profile. In case VBR mode is enabled the option is ignored.
ar — Set audio sampling rate (in Hz).
channels — Set the number of audio channels.
flags +qscale — Enable fixed quality, VBR (Variable Bit Rate) mode. Note that VBR is implicitly enabled when the vbr value is positive.
cutoff — Set cutoff frequency. If not specified (or explicitly set to 0) it will use a value automatically computed by the library. Default value is 0.
profile — Set audio profile.

The following profiles are recognized:

aac_low — Low Complexity AAC (LC)
aac_he — High Efficiency AAC (HE-AAC)
aac_he_v2 — High Efficiency AAC version 2 (HE-AACv2)
aac_ld — Low Delay AAC (LD)
aac_eld — Enhanced Low Delay AAC (ELD)

If not specified it is set to aac_low.

The following are private options of the libfdk_aac encoder. Secret menu energy, honestly.

afterburner — Enable afterburner feature if set to 1, disabled if set to 0. This improves the quality but also the required processing power.

Default value is 1.

eld_sbr — Enable SBR (Spectral Band Replication) for ELD if set to 1, disabled if set to 0.

Default value is 0.

eld_v2 — Enable ELDv2 (LD-MPS extension for ELD stereo signals) for ELDv2 if set to 1, disabled if set to 0.

Note that option is available when fdk-aac version (AACENCODER_LIB_VL0.AACENCODER_LIB_VL1.AACENCODER_LIB_VL2) > (4.0.0).

Default value is 0.

signaling — Set SBR/PS signaling style.

It can assume one of the following values:

default — choose signaling implicitly (explicit hierarchical by default, implicit if global header is disabled)
implicit — implicit backwards compatible signaling
explicit_sbr — explicit SBR, implicit PS signaling
explicit_hierarchical — explicit hierarchical signaling

Default value is default.

latm — Output LATM/LOAS encapsulated data if set to 1, disabled if set to 0.

Default value is 0.

header_period — Set StreamMuxConfig and PCE repetition period (in frames) for sending in-band configuration buffers within LATM/LOAS transport layer.

Must be a 16-bits non-negative integer.

Default value is 0.

vbr — Set VBR mode, from 1 to 5. 1 is lowest quality (though still pretty good) and 5 is highest quality. A value of 0 will disable VBR, and CBR (Constant Bit Rate) is enabled.

Currently only the aac_low profile supports VBR encoding.

VBR modes 1–5 are basically a vibe check for bitrate — they line up roughly like this:

32 kbps/channel
40 kbps/channel
48-56 kbps/channel
64 kbps/channel
about 80-96 kbps/channel

Default value is 0 — aka "no thoughts, just defaults" ✨

frame_length — sets the audio frame length in samples, bestie. The default is whatever the library's internal default is (iykyk), so go read the library's docs for which values are actually supported 🫶

Examples 🎀

Wanna convert an audio file to VBR AAC in an M4A (MP4) container using ffmpeg? Slay, here's the line:

ffmpeg -i input.wav -codec:a libfdk_aac -vbr 3 output.m4a

Or convert to CBR 64k kbps AAC using the High-Efficiency AAC profile — iconic:

ffmpeg -i input.wav -c:a libfdk_aac -profile:a aac_he -b:a 64k output.m4a

🌸 liblc3 🌸

The liblc3 LC3 (Low Complexity Communication Codec) encoder wrapper. Literally the moment 💅

You need the liblc3 headers and library present during configuration, and you have to explicitly configure the build with --enable-liblc3.

This encoder supports the Bluetooth SIG LC3 codec for the LE Audio protocol, plus these LC3plus features:

Frame duration of 2.5 and 5ms.
High-Resolution mode, 48 KHz, and 96 kHz sampling rates.

For more info, the liblc3 project lives at <https://github.com/google/liblc3> — go be nosy ✨

Options 🦋

These options are mapped onto the shared FFmpeg codec options.

b bitrate — set the bit rate in bits/s. This decides the fixed size of the encoded frames for whichever frame duration you picked.
ar frequency — set the audio sampling rate (in Hz).
channels — set the number of audio channels.
frame_duration — set the audio frame duration in milliseconds. Default is 10ms. Allowed frame durations are 2.5ms, 5ms, 7.5ms and 10ms. LC3 (Bluetooth LE Audio) allows 7.5ms and 10ms; LC3plus allows 2.5ms, 5ms and 10ms.

The 10ms frame duration is available in both LC3 and LC3 plus standard, no cap. In this mode the produced bitstream can be referenced either as LC3 or LC3plus 👑

high_resolution boolean — enable the high-resolution mode by setting it to 1. High-resolution mode is available with all LC3plus frame durations and for sampling rates of 48 KHz and 96 KHz.

The encoder automatically turns this mode off at lower sampling rates and activates it at 96 KHz.

This mode should be preferred at high bitrates — in it, the audio bandwidth always goes up to the Nyquist frequency, compared to LC3 at 48 KHz which limits the bandwidth to 20 KHz. The range! 💖

🎀 libmp3lame 🎀

The LAME (Lame Ain't an MP3 Encoder) MP3 encoder wrapper. A classic, honestly.

You need the libmp3lame headers and library present during configuration, and you must explicitly configure the build with --enable-libmp3lame.

Check out libshine if you want a fixed-point MP3 encoder — although it's lower quality, just so you know 👀

Options ✨

These are supported by the libmp3lame wrapper. The lame-equivalent names are in parentheses.

b (-b) — set the bitrate expressed in bits/s for CBR or ABR. LAME's "bitrate" is expressed in kilobits/s.
q (-V) — set the constant quality setting for VBR. This option is only valid when you're using the ffmpeg command-line tool. Library interface users, use global_quality instead 🫶
compression_level (-q) — set the algorithm quality. Valid arguments are integers in the 0-9 range, where 0 means highest quality but slowest, and 9 means fastest while producing the worst quality. Pick your fighter 💀
cutoff (--lowpass) — set the lowpass cutoff frequency. If you don't specify it, the encoder dynamically adjusts the cutoff.
reservoir — enable use of the bit reservoir when set to 1. Default is 1. LAME has this on by default but it can be overridden by using the --nores option.
joint_stereo (-m j) — lets the encoder use, frame by frame, either L/R stereo or mid/side stereo. Default is 1.
abr (--abr) — enable ABR when set to 1. The lame --abr sets the target bitrate, while this option only tells FFmpeg to use ABR — it still relies on b to set the bitrate. Teamwork, I guess 🤝
copyright (-c) — set the MPEG audio copyright flag when set to 1. Default is 0 (disabled).
original (-o) — set the MPEG audio original flag when set to 1. Default is 1 (enabled).

🦋 libopencore-amrnb 🦋

The OpenCORE Adaptive Multi-Rate Narrowband encoder.

You need the libopencore-amrnb headers and library present during configuration, and you must explicitly configure the build with --enable-libopencore-amrnb --enable-version3.

This one is a mono-only encoder, periodt. Officially it only supports the 8000Hz sample rate, but you can override that by setting strict to unofficial or lower.

Options 💅

b — set the bitrate in bits per second. Only these bitrates are supported — otherwise libavcodec will round to the nearest valid bitrate:

dtx — allow discontinuous transmission (generate comfort noise) when set to 1. Default is 0 (disabled).

👑 libopus 👑

The libopus Opus Interactive Audio Codec encoder wrapper. Main character energy, fr fr.

You need the libopus headers and library present during configuration, and you must explicitly configure the build with --enable-libopus.

Option Mapping 🌈

Most libopus options are modelled after the opusenc utility from opus-tools. Here's the mapping chart describing options supported by the libopus wrapper, with their opusenc-equivalent in parentheses.

b (bitrate) — set the bit rate in bits/s. FFmpeg's b option is expressed in bits/s, while opusenc's bitrate is in kilobits/s. Watch those units bestie 😭
vbr (vbr, hard-cbr, and cvbr) — set the VBR mode. The FFmpeg vbr option has these valid arguments, with the opusenc equivalents in parentheses:
- off (hard-cbr) — use constant bit rate encoding.
- on (vbr) — use variable bit rate encoding (the default).
- constrained (cvbr) — use constrained variable bit rate encoding.
compression_level (comp) — set the encoding algorithm complexity. Valid options are integers in the 0-10 range. 0 gives the fastest encodes but lower quality, while 10 gives the highest quality but slowest encoding. The default is 10.
frame_duration (framesize) — set the maximum frame size, or duration of a frame, in milliseconds. The argument must be exactly one of: 2.5, 5, 10, 20, 40, 60. Smaller frame sizes get you lower latency but less quality at a given bitrate. Sizes greater than 20ms are only interesting at fairly low bitrates. The default is 20ms.
packet_loss (expect-loss) — set the expected packet loss percentage. The default is 0.
fec (n/a) — enable inband forward error correction. packet_loss must be non-zero to take advantage of it — the frequency of FEC 'side-data' is proportional to the expected packet loss. Default is disabled.
application (N.A.) — set the intended application type. Valid options:
- voip — favor improved speech intelligibility.
- audio — favor faithfulness to the input (the default).
- lowdelay — restrict to only the lowest delay modes by disabling voice-optimized modes.
cutoff (N.A.) — set the cutoff bandwidth in Hz. The argument must be exactly one of: 4000, 6000, 8000, 12000, or 20000, corresponding to narrowband, mediumband, wideband, super wideband, and fullband respectively. The default is 0 (cutoff disabled). Note that libopus forces a wideband cutoff for bitrates < 15 kbps, unless CELT-only (application set to lowdelay) mode is used. Sneaky but valid 💀
mapping_family (mapping_family) — set the channel mapping family the encoder should use. The default value of -1 uses mapping family 0 for mono and stereo inputs, and mapping family 1 otherwise. The default also disables the surround masking and LFE bandwidth optimizations in libopus, and requires that the input contains 8 channels or fewer.

Other values include 0 for mono and stereo, 1 for surround sound with masking and LFE bandwidth optimizations, and 255 for independent streams with an unspecified channel layout.

apply_phase_inv (N.A.) (requires libopus >= 1.2) — if set to 0, disables the use of phase inversion for intensity stereo, improving the quality of mono downmixes but slightly reducing normal stereo quality. The default is 1 (phase inversion enabled).

🍓 libshine 🍓

The Shine Fixed-Point MP3 encoder wrapper.

Shine is a fixed-point MP3 encoder. It performs far better on platforms without an FPU — think armel CPUs, and some phones and tablets. But because it's more targeted at performance than quality, it's not on par with LAME and other production-grade encoders quality-wise. Also, according to the project's homepage, this encoder may not be free of bugs, since the code was written a long time ago and the project was dead for at least 5 years. Not me trusting it anyway 🙈

This encoder only supports stereo and mono input, and it's also CBR-only.

The original project (last updated in early 2007) is at <http://sourceforge.net/projects/libshine-fxp/>. We only support the updated fork by the Savonet/Liquidsoap project at <https://github.com/savonet/shine>.

You need the libshine headers and library present during configuration, and you must explicitly configure the build with --enable-libshine.

💜 libtwolame 💜

The TwoLAME MP2 encoder wrapper.

You need the libtwolame headers and library present during configuration, and you must explicitly configure the build with --enable-libtwolame.

Options ⭐

These are supported by the libtwolame wrapper. The twolame-equivalent options follow the FFmpeg ones and are in parentheses.

b (-b) — set the bitrate expressed in bits/s for CBR. The twolame b option is expressed in kilobits/s. Default value is 128k.
q (-V) — set the quality for experimental VBR support. The maximum value range is from -50 to 50, and the useful range is from -10 to 10. The higher the value, the better the quality. This option is only valid when using the ffmpeg command-line tool. Library interface users, use global_quality instead 🫶

okay so picking up right where we left off — the next encoder knobs are giving very much “tiny details, big vibe checks” 💅

mode (--mode) — sets the mode of the resulting audio. Possible values are:
- auto — choose mode automatically based on the input. This is the default.
- stereo — stereo.
- joint_stereo — joint stereo.
- dual_channel — dual channel.
- mono — mono.
psymodel (--psyc-mode) — sets the psychoacoustic model to use in encoding. The argument must be an integer between -1 and 4, inclusive. The higher the value, the better the quality. The default value is 3.
energy_levels (--energy) — enables energy levels extensions when set to 1. The default value is 0 (disabled).
error_protection (--protect) — enables CRC error protection when set to 1. The default value is 0 (disabled).
copyright (--copyright) — sets the MPEG audio copyright flag when set to 1. The default value is 0 (disabled).
original (--original) — sets the MPEG audio original flag when set to 1. The default value is 0 (disabled).

🎧 libvo-amrwbenc 🎧

VisualOn Adaptive Multi-Rate Wideband encoder — say that five times fast 😭

Requires the presence of the libvo-amrwbenc headers and library during configuration. You need to explicitly configure the build with --enable-libvo-amrwbenc --enable-version3.

This is a mono-only encoder. Officially it only supports 16000Hz sample rate, but you can override it by setting strict to unofficial or lower.

✨ Options ✨

b — set bitrate in bits/s. Only the following bitrates are supported, otherwise libavcodec will round to the nearest valid bitrate:
- 6600
- 8850
- 12650
- 14250
- 15850
- 18250
- 19850
- 23050
- 23850
dtx — allow discontinuous transmission (generate comfort noise) when set to 1. The default value is 0 (disabled).

🦋 libvorbis 🦋

libvorbis encoder wrapper — literally the classic Vorbis vibe ✨

Requires the presence of the libvorbisenc headers and library during configuration. You need to explicitly configure the build with --enable-libvorbis.

✨ Options ✨

The following options are supported by the libvorbis wrapper. The oggenc-equivalent of the options are listed in parentheses.

To get a more accurate and extensive documentation of the libvorbis options, consult the libvorbisenc's and oggenc's documentations. See http://xiph.org/vorbis/, http://wiki.xiph.org/Vorbis-tools, and oggenc(1).

b (-b) — set bitrate expressed in bits/s for ABR. oggenc -b is expressed in kilobits/s.
q (-q) — set constant quality setting for VBR. The value should be a float number in the range of -1.0 to 10.0. The higher the value, the better the quality. The default value is 3.0.

This option is valid only using the ffmpeg command-line tool. For library interface users, use global_quality.

cutoff (--advanced-encode-option lowpass_frequency=N) — set cutoff bandwidth in Hz, a value of 0 disables cutoff. oggenc's related option is expressed in kHz. The default value is 0 (cutoff disabled).
minrate (-m) — set minimum bitrate expressed in bits/s. oggenc -m is expressed in kilobits/s.
maxrate (-M) — set maximum bitrate expressed in bits/s. oggenc -M is expressed in kilobits/s. This only has effect on ABR mode.
iblock (--advanced-encode-option impulse_noisetune=N) — set noise floor bias for impulse blocks. The value is a float number from -15.0 to 0.0. A negative bias instructs the encoder to pay special attention to the crispness of transients in the encoded audio. The tradeoff for better transient response is a higher bitrate.

🌈 mjpeg 🌈

Motion JPEG encoder.

✨ Options ✨

huffman — set the huffman encoding strategy. Possible values:
- default — use the default huffman tables. This is the default strategy.
- optimal — compute and use optimal huffman tables.

🐇 wavpack 🐇

WavPack lossless audio encoder.

✨ Options ✨

The equivalent options for wavpack command line utility are listed in parentheses.

Shared options

The following shared options are effective for this encoder. Only special notes about this particular encoder will be documented here. For the general meaning of the options, see the Codec Options chapter.

frame_size (--blocksize) — for this encoder, the range for this option is between 128 and 131072. Default is automatically decided based on sample rate and number of channel.

For the complete formula of calculating default, see libavcodec/wavpackenc.c.

compression_level (-f, -h, -hh, and -x) — yes, that's literally the whole line in the manual. No extra lore here 🫶

Private options

joint_stereo (-j) — set whether to enable joint stereo. Valid values are:
- on (1) — force mid/side audio encoding.
- off (0) — force left/right audio encoding.
- auto — let the encoder decide automatically.
optimize_mono — set whether to enable optimization for mono. This option is only effective for non-mono streams. Available values:
- on — enabled
- off — disabled

🎬 VIDEO ENCODERS 🎬

A description of some of the currently available video encoders follows.

🕹️ a64_multi, a64_multi5 🕹️

A64 / Commodore 64 multicolor charset encoder. a64_multi5 is extended with 5th color (colram).

🎞️ Cinepak 🎞️

Cinepak aka CVID encoder. Compatible with Windows 3.1 and vintage MacOS.

✨ Options ✨

g integer — keyframe interval. A keyframe is inserted at least every -g frames, sometimes sooner.
q:v integer — quality factor. Lower is better. Higher gives lower bitrate. The following table lists bitrates when encoding akiyo_cif.y4m for various values of -q:v with -g 100:
- -q:v 1 1918 kb/s
- -q:v 2 1735 kb/s
- -q:v 4 1500 kb/s
- -q:v 10 1041 kb/s
- -q:v 20 826 kb/s
- -q:v 40 553 kb/s
- -q:v 100 394 kb/s
- -q:v 200 312 kb/s
- -q:v 400 266 kb/s
- -q:v 1000 237 kb/s
max_extra_cb_iterations integer — max extra codebook recalculation passes, more is better and slower.
skip_empty_cb boolean — avoid wasting bytes, ignore vintage MacOS decoder.
max_strips integer — the maximum number of strips to use. Combined note for both strip options: wider range sometimes improves quality. More strips is generally better quality but costs more bits. Fewer strips tend to yield more keyframes. Vintage compatible is 1..3.
min_strips integer — the minimum number of strips to use. Same combined note as above: wider range sometimes improves quality. More strips is generally better quality but costs more bits. Fewer strips tend to yield more keyframes. Vintage compatible is 1..3.
strip_number_adaptivity integer — how much number of strips is allowed to change between frames. Higher is better but slower.

🧚 ffv1 🧚

FFv1 Encoder.

✨ Options ✨

The following options are supported by FFmpeg's FFv1 encoder.

context — sets the context size, 0 (default) is small, 1 is big.
coder — set the coder:
- rice — Golomb rice coder
- range_def — range coder with default table
- range_tab — range coder with custom table
slicecrc — -1 (default, automatic), 1 use crc with zero initial and final state, 2 use crc with non zero initial and final state.
qtable — possible values:
- default — default, automatic
- 8bit — use 8bit default
- greater8bit — use >8bit default
remap_optimizer — 0 - 5, default 3, how much effort the encoder puts into optimizing the remap table.

🎀 GIF 🎀

GIF image/animation encoder.

✨ Options ✨

gifflags integer — sets the flags used for GIF encoding.
- offsetting — enables picture offsetting. Default is enabled.
- transdiff — enables transparency detection between frames. Default is enabled.
gifimage integer — enables encoding one full GIF image per frame, rather than an animated GIF. Default value is 0.
global_palette integer — writes a palette to the global GIF header where feasible.

If disabled, every frame will always have a palette written, even if there is a global palette supplied.

Default value is 1.

🦋 Hap 🦋

Vidvox Hap video encoder.

✨ Options ✨

format integer — specifies the Hap format to encode.
- hap
- hap_alpha
- hap_q

Default value is hap.

chunks integer — specifies the number of chunks to split frames into, between 1 and 64. This permits multithreaded decoding of large frames, potentially at the cost of data-rate. The encoder may modify this value to divide frames evenly.

Default value is 1.

compressor integer — specifies the second-stage compressor to use. If set to none, chunks will be limited to 1, as chunked uncompressed frames offer no benefit.
- none
- snappy

Default value is snappy.

🖼️ jpeg2000 🖼️

The native jpeg 2000 encoder is lossy by default, the -q:v option can be used to set the encoding quality. Lossless encoding can be selected with -pred 1.

✨ Options ✨

format integer — can be set to either j2k or jp2 (the default) that makes it possible to store non-rgb pix_fmts.
tile_width integer — sets tile width. Range is 1 to 1073741824. Default is 256.
tile_height integer — sets tile height. Range is 1 to 1073741824. Default is 256.
pred integer — allows setting the discrete wavelet transform (DWT) type:
- dwt97int (Lossy)
- dwt53 (Lossless)

Default is dwt97int.

sop boolean — enable this to add SOP marker at the start of each packet. Disabled by default.
eph boolean — enable this to add EPH marker at the end of each packet header. Disabled by default.
prog integer — sets the progression order to be used by the encoder. Possible values are:
- lrcp
- rlcp
- rpcl
- pcrl
- cprl

Set to lrcp by default.

layer_rates string — by default, when this option is not used, compression is done using the quality metric. This option allows for compression using compression ratio. The compression ratio for each level could be specified. The compression ratio of a layer l species the what ratio of total file size is contained in the first l layers.

Example usage:

ffmpeg -i input.bmp -c:v jpeg2000 -layer_rates "100,10,1" output.j2k

This would compress the image to contain 3 layers, where the data contained in the first layer would be compressed by 1000 times, compressed by 100 in the first two layers, and shall contain all data while using all 3 layers.

🦀 librav1e 🦀

rav1e AV1 encoder wrapper.

okay so bestie, the rav1e wrapper needs the rav1e headers and library present when you configure — you gotta explicitly configure the build with --enable-librav1e. no drive-by installs, she needs a formal invite 💅

Options

qmax — sets the max quantizer to use when you're in bitrate mode.
qmin — sets the min quantizer to use when you're in bitrate mode.
qp — uses quantizer mode to encode at the given quantizer (0-255).
speed — picks the speed preset (0-10) to encode with.
tiles — picks how many tiles to encode with.
tile-rows — picks how many rows of tiles to encode with.
tile-columns — picks how many columns of tiles to encode with.
rav1e-params — sets rav1e options using a list of key=value pairs separated by ":". Want the full menu? peep rav1e --help for the option list, iykyk.

Like, say you wanna pass librav1e encoding options with -rav1e-params, it's giving:

ffmpeg -i input -c:v librav1e -b:v 500K -rav1e-params speed=5:low_latency=true output.mp4

🦋 libaom-av1 🦋

the libaom AV1 encoder wrapper, she's the moment.

needs the libaom headers and library hanging around during configuration — you have to explicitly configure the build with --enable-libaom.

Options

this wrapper supports the following standard libavcodec options, periodt:

b — set the bitrate target in bits/second. by default this uses variable-bitrate mode. if maxrate and minrate are also set to the same value, then it uses constant-bitrate mode instead — otherwise, if crf is set too, it uses constrained-quality mode. three modes, one flag, main character energy fr.
g, keyint_min — set key frame placement. the GOP size sets the maximum distance between key frames; if it's zero the output stream will be intra-only. the minimum distance is ignored unless it's the same as the GOP size, in which case key frames always show up at a fixed interval. not set by default, so without this option the library has completely free choice about where to place key frames — she's improvising 💀
qmin, qmax — set min/max quantisation values. valid range is 0 to 63 (warning: this does not match the quantiser values actually used by AV1 — divide by four to map real quantiser values to this range, no cap). defaults to min/max, i.e. no constraint.
minrate, maxrate, bufsize, rc_init_occupancy — set rate control buffering parameters. not used if not set — defaults to unconstrained variable bitrate.
threads — set the number of threads to use while encoding. heads up: this may require the tiles or row-mt options to also be set for the specified thread count to actually get used fully. defaults to the number of hardware threads supported by the host machine. 🧚
profile — set the encoding profile. defaults to using the profile that matches the bit depth and chroma subsampling of the input.

the wrapper also has some specific options of her own:

cpu-used — set the quality/encoding speed tradeoff. valid range is 0 to 8, higher numbers meaning greater speed and lower quality. the default value is 1, which will be slow and high quality. she's thorough, okay 😭
auto-alt-ref — enable use of alternate reference frames. defaults to the internal default of the library.
arnr-max-frames (frames) — set altref noise reduction max frame count. default is -1.
arnr-strength (strength) — set altref noise reduction filter strength. range is -1 to 6. default is -1.
aq-mode (aq-mode) — set adaptive quantization mode. possible values:
- none (0) — disabled.
- variance (1) — variance-based.
- complexity (2) — complexity-based.
- cyclic (3) — cyclic refresh.
tune (tune) — set the distortion metric the encoder is tuned with. default is "psnr".
- psnr (0)
- ssim (1)
lag-in-frames — set the maximum number of frames the encoder may keep in flight at any one time for lookahead purposes. defaults to the internal default of the library.
error-resilience — enable error resilience features:
- default — improve resilience against losses of whole frames.

not enabled by default.

crf — set the quality/size tradeoff for constant-quality (no bitrate target) and constrained-quality (with maximum bitrate target) modes. valid range is 0 to 63, higher numbers meaning lower quality and smaller output size. only used if set; by default only the bitrate target is used.
static-thresh — set a change threshold on blocks below which they'll be skipped by the encoder. defined in arbitrary units as a nonnegative integer, defaulting to zero (no blocks are skipped).
drop-threshold — set a threshold for dropping frames when close to rate control bounds. defined as a percentage of the target buffer — when the rate control buffer falls below this percentage, frames get dropped until it's refilled above the threshold. defaults to zero (no frames are dropped). 🫶
denoise-noise-level (level) — amount of noise to be removed for grain synthesis. grain synthesis is disabled if this option isn't set or is set to 0.
denoise-block-size (pixels) — block size used for denoising for grain synthesis. if not set, the AV1 codec uses the default value of 32.
undershoot-pct (pct) — set datarate undershoot (min) percentage of the target bitrate. range is -1 to 100. default is -1.
overshoot-pct (pct) — set datarate overshoot (max) percentage of the target bitrate. range is -1 to 1000. default is -1.
minsection-pct (pct) — minimum percentage variation of the GOP bitrate from the target bitrate. if minsection-pct isn't set, the libaomenc wrapper computes it like so: "(minrate * 100 / bitrate)". range is -1 to 100. default is -1 (unset).
maxsection-pct (pct) — maximum percentage variation of the GOP bitrate from the target bitrate. if maxsection-pct isn't set, the libaomenc wrapper computes it like so: "(maxrate * 100 / bitrate)". range is -1 to 5000. default is -1 (unset).
frame-parallel (boolean) — enable frame parallel decodability features. default is true.
tiles — set the number of tiles to encode the input video with, as columns x rows. larger numbers allow greater parallelism in both encoding and decoding, but may decrease coding efficiency. defaults to the minimum number of tiles required by the size of the input video (which is 1x1 — that is, a single tile — for sizes up to and including 4K). 🌈
tile-columns, tile-rows — set the number of tiles as log2 of the number of tile rows and columns. provided for compatibility with libvpx/VP9.
row-mt (Requires libaom >= 1.0.0-759-g90a15f4f2) — enable row based multi-threading. disabled by default.
enable-cdef (boolean) — enable Constrained Directional Enhancement Filter. the libaom-av1 encoder enables CDEF by default.
enable-restoration (boolean) — enable Loop Restoration Filter. default is true for libaom-av1.
enable-global-motion (boolean) — enable the use of global motion for block prediction. default is true.
enable-intrabc (boolean) — enable block copy mode for intra block prediction. this mode is useful for screen content. default is true.
enable-rect-partitions (boolean) (Requires libaom >= v2.0.0) — enable rectangular partitions. default is true.
enable-1to4-partitions (boolean) (Requires libaom >= v2.0.0) — enable 1:4/4:1 partitions. default is true.
enable-ab-partitions (boolean) (Requires libaom >= v2.0.0) — enable AB shape partitions. default is true.
enable-angle-delta (boolean) (Requires libaom >= v2.0.0) — enable angle delta intra prediction. default is true.
enable-cfl-intra (boolean) (Requires libaom >= v2.0.0) — enable chroma predicted from luma intra prediction. default is true.
enable-filter-intra (boolean) (Requires libaom >= v2.0.0) — enable filter intra predictor. default is true.
enable-intra-edge-filter (boolean) (Requires libaom >= v2.0.0) — enable intra edge filter. default is true.
enable-smooth-intra (boolean) (Requires libaom >= v2.0.0) — enable smooth intra prediction mode. default is true.
enable-paeth-intra (boolean) (Requires libaom >= v2.0.0) — enable paeth predictor in intra prediction. default is true.
enable-palette (boolean) (Requires libaom >= v2.0.0) — enable palette prediction mode. default is true.
enable-flip-idtx (boolean) (Requires libaom >= v2.0.0) — enable extended transform type, including FLIPADST_DCT, DCT_FLIPADST, FLIPADST_FLIPADST, ADST_FLIPADST, FLIPADST_ADST, IDTX, V_DCT, H_DCT, V_ADST, H_ADST, V_FLIPADST, H_FLIPADST. default is true. that's a lotta types bestie, she ate 💅
enable-tx64 (boolean) (Requires libaom >= v2.0.0) — enable 64-pt transform. default is true.
reduced-tx-type-set (boolean) (Requires libaom >= v2.0.0) — use reduced set of transform types. default is false.
use-intra-dct-only (boolean) (Requires libaom >= v2.0.0) — use DCT only for INTRA modes. default is false.
use-inter-dct-only (boolean) (Requires libaom >= v2.0.0) — use DCT only for INTER modes. default is false.
use-intra-default-tx-only (boolean) (Requires libaom >= v2.0.0) — use Default-transform only for INTRA modes. default is false.
enable-ref-frame-mvs (boolean) (Requires libaom >= v2.0.0) — enable temporal mv prediction. default is true.
enable-reduced-reference-set (boolean) (Requires libaom >= v2.0.0) — use reduced set of single and compound references. default is false.
enable-obmc (boolean) (Requires libaom >= v2.0.0) — enable obmc. default is true.
enable-dual-filter (boolean) (Requires libaom >= v2.0.0) — enable dual filter. default is true.
enable-diff-wtd-comp (boolean) (Requires libaom >= v2.0.0) — enable difference-weighted compound. default is true.
enable-dist-wtd-comp (boolean) (Requires libaom >= v2.0.0) — enable distance-weighted compound. default is true.
enable-onesided-comp (boolean) (Requires libaom >= v2.0.0) — enable one sided compound. default is true.
enable-interinter-wedge (boolean) (Requires libaom >= v2.0.0) — enable interinter wedge compound. default is true.
enable-interintra-wedge (boolean) (Requires libaom >= v2.0.0) — enable interintra wedge compound. default is true.
enable-masked-comp (boolean) (Requires libaom >= v2.0.0) — enable masked compound. default is true.
enable-interintra-comp (boolean) (Requires libaom >= v2.0.0) — enable interintra compound. default is true.
enable-smooth-interintra (boolean) (Requires libaom >= v2.0.0) — enable smooth interintra mode. default is true. ✨

aom-params

okay so this is where you pass libaom its little extras — you give it a list of key=value pairs separated by :. Wanna know which keys exist? Go run aomenc --help and look at the section literally called "AV1 Specific Options" 💅

Like, for example, to specify libaom encoding options with -aom-params:

ffmpeg -i input -c:v libaom-av1 -b:v 500K -aom-params tune=psnr:enable-tpl-model=1 output.mp4

🌈 liboapv 🌈

The Advanced Professional Video codec encoder wrapper, bestie ✨

Fair warning: this encoder needs the liboapv headers and library chillin' there at configuration time, so you gotta explicitly configure your build with --enable-liboapv.

Many liboapv encoder options are mapped to FFmpeg global codec options, while the unique encoder options get handed to you through private options. iykyk 🫶

The apv project website lives at https://github.com/AcademySoftwareFoundation/openapv.

✨ Options ✨

These are the ones the liboapv wrapper supports:

Want the extensive documentation of the liboapv options? Go consult the liboapv docs, queen 👑

preset — set the quality-speed tradeoff, one of fastest, fast, medium, slow, placebo, default.
qp — set the quantization parameter value for CQP rate control mode.
oapv-params (parse_apv_params) — set liboapvenc options using a list of key=value pairs separated by :. Peep the liboapv encoder user guide for the list of accepted parameters.

🦋 libsvtav1 🦋

SVT-AV1 encoder wrapper, no cap 💖

Needs the SVT-AV1 headers and library present during configuration. You gotta explicitly configure the build with --enable-libsvtav1.

🌸 Options 🌸

profile — set the encoding profile.
- main
- high
- professional
level — set the operating point level. For example: '4.0'
hielevel — set the Hierarchical prediction levels.
- 3level
- 4level — this one is the default, fyi.
tier — set the operating point tier.
- main — this is the default ✨
- high
qmax — set the maximum quantizer to use when using a bitrate mode.
qmin — set the minimum quantizer to use when using a bitrate mode.
crf — constant rate factor value used in crf rate control mode (0-63).
qp — set the quantizer used in cqp rate control mode (0-63).
sc_detection — enable scene change detection.
la_depth — set number of frames to look ahead (0-120).
preset — set the quality-speed tradeoff, in the range 0 to 13. Higher values are faster but lower quality (you get what you pay for, literally 💀).
tile_rows — set log2 of the number of rows of tiles to use (0-6).
tile_columns — set log2 of the number of columns of tiles to use (0-4).
svtav1-params — set SVT-AV1 options using a list of key=value pairs separated by :. See the SVT-AV1 encoder user guide for a list of accepted parameters.

🎀 libsvtjpegxs 🎀

SVT-JPEG-XS encoder wrapper, she's niche and she's iconic 💅

Needs the SVT-JPEG-XS headers and library during configuration. You gotta explicitly configure the build with --enable-libsvtjpegxs.

⭐ Options ⭐

decomp_v — set vertical decomposition level.
decomp_h — set horizontal decomposition level.
quantization — set the quantization algorithm.
- deadzone
- uniform
coding-signs — enable Signs handling strategy.
- disable
- fast
- full
coding-sigf — enable Significance coding.
coding-vpred — enable Vertical Prediction coding.
- disable
- no_residuals
- no_coeffs

🧚 libjxl 🧚

The libjxl JPEG XL encoder wrapper, obsessed with her honestly 💜

Needs the libjxl headers and library during configuration. You gotta explicitly configure the build with --enable-libjxl.

🔥 Options 🔥

The libjxl wrapper supports these:

distance — set the target Butteraugli distance. This is a quality setting, bestie: lower distance = higher quality, with distance=1.0 roughly comparable to libjpeg Quality 90 for photographic content. Setting distance=0.0 yields true lossless encoding. Valid values range between 0.0 and 15.0, and sane values rarely exceed 5.0. Setting distance=0.1 usually attains transparency for most input. The default is 1.0.
effort — set the encoding effort used. Higher effort values produce more consistent quality and usually produce a better quality/bpp curve, at the cost of more CPU time required (your fans will let you know 😭). Valid values range from 1 to 9, and the default is 7.
modular — force the encoder to use Modular mode instead of choosing automatically. The default is to use VarDCT for lossy encoding and Modular for lossless. VarDCT is generally superior to Modular for lossy encoding but does not support lossless encoding. Pick your fighter 🥊

💖 libkvazaar 💖

Kvazaar H.265/HEVC encoder, slay 🦋

Needs the libkvazaar headers and library during configuration. You gotta explicitly configure the build with --enable-libkvazaar.

✨ Options ✨

b — set target video bitrate in bit/s and enable rate control.
kvazaar-params — set kvazaar parameters as a list of name=value pairs separated by commas (,). See the kvazaar documentation for a list of options.

👑 libopenh264 👑

Cisco libopenh264 H.264/MPEG-4 AVC encoder wrapper — a certified classic 🎀

This encoder needs the libopenh264 headers and library during configuration. You gotta explicitly configure the build with --enable-libopenh264. The library is detected using pkg-config.

For more info about the library see http://www.openh264.org.

🌸 Options 🌸

These FFmpeg global options affect the configurations of the libopenh264 encoder:

b — set the bitrate (as a number of bits per second).
g — set the GOP size.
maxrate — set the max bitrate (as a number of bits per second).
flags +global_header — set global header in the bitstream.
slices — set the number of slices, used in parallelized encoding. Default value is 0. This is only used when slice_mode is set to fixed.
loopfilter — enable loop filter, if set to 1 (automatically enabled). To disable, set a value of 0.
profile — set profile restrictions. If set to the value of main, enable CABAC (set the "SEncParamExt.iEntropyCodingModeFlag" flag to 1).
max_nal_size — set maximum NAL size in bytes.
allow_skip_frames — allow skipping frames to hit the target bitrate if set to 1.

🌈 libtheora 🌈

The libtheora Theora encoder wrapper, vintage but make it cute 💜

Needs the libtheora headers and library during configuration. You gotta explicitly configure the build with --enable-libtheora.

For more info about the libtheora project see http://www.theora.org/.

⭐ Options ⭐

These global options get mapped to internal libtheora options, and they affect the quality and the bitrate of the encoded stream:

b — set the video bitrate in bit/s for CBR (Constant Bit Rate) mode. In case VBR (Variable Bit Rate) mode is enabled, this option is ignored. rude but okay 🙈
flags — used to enable constant quality mode (VBR) encoding through the qscale flag, and to enable the "pass1" and "pass2" modes.
g — set the GOP size.
global_quality — set the global quality as an integer in lambda units.

Only relevant when VBR mode is enabled with "flags +qscale". The value is converted to QP units by dividing it by FF_QP2LAMBDA, clipped in the [0 - 10] range, and then multiplied by 6.3 to get a value in the native libtheora range [0-63]. A higher value corresponds to a higher quality.

q — enable VBR mode when set to a non-negative value, and set constant quality value as a double floating point value in QP units.

The value is clipped in the [0-10] range, and then multiplied by 6.3 to get a value in the native libtheora range [0-63].

This option is valid only using the ffmpeg command-line tool. For library interface users, use global_quality. (iykyk ✨)

🎀 Examples 🎀

Set maximum constant quality (VBR) encoding with ffmpeg:

ffmpeg -i INPUT -codec:v libtheora -q:v 10 OUTPUT.ogg

Use ffmpeg to convert a CBR 1000 kbps Theora video stream:

ffmpeg -i INPUT -codec:v libtheora -b:v 1000k OUTPUT.ogg

💅 libvpx 💅

VP8/VP9 format supported through libvpx, the main character of web video fr fr 🔥

Needs the libvpx headers and library during configuration. You gotta explicitly configure the build with --enable-libvpx.

🦋 Options 🦋

The following options are supported by the libvpx wrapper. The vpxenc-equivalent options or values are listed in parentheses for easy migration (we love a smooth transition 💖).

To reduce the duplication of documentation, only the private options and some others requiring special attention are documented here. For the documentation of the undocumented generic options, see the Codec Options chapter.

To get more documentation of the libvpx options, invoke the command ffmpeg -h encoder=libvpx, ffmpeg -h encoder=libvpx-vp9 or vpxenc --help. Further information is available in the libvpx API documentation.

b (target-bitrate) — set bitrate in bits/s. Note that FFmpeg's b option is expressed in bits/s, while vpxenc's target-bitrate is in kilobits/s. watch that unit slippage bestie 😭
g (kf-max-dist)
keyint_min (kf-min-dist)
qmin (min-q) — Minimum (Best Quality) Quantizer.
qmax (max-q) — Maximum (Worst Quality) Quantizer. Can be changed per-frame.
bufsize (buf-sz, buf-optimal-sz) — set ratecontrol buffer size (in bits). Note vpxenc's options are specified in milliseconds, and the libvpx wrapper converts this value as follows: "buf-sz = bufsize * 1000 / bitrate", "buf-optimal-sz = bufsize * 1000 / bitrate * 5 / 6".
rc_init_occupancy (buf-initial-sz) — set number of bits which should be loaded into the rc buffer before decoding starts. Note vpxenc's option is specified in milliseconds, and the libvpx wrapper converts this value as follows: "rc_init_occupancy * 1000 / bitrate".
undershoot-pct — set datarate undershoot (min) percentage of the target bitrate.
overshoot-pct — set datarate overshoot (max) percentage of the target bitrate.
skip_threshold (drop-frame)
qcomp (bias-pct)
maxrate (maxsection-pct) — set GOP max bitrate in bits/s. Note vpxenc's option is specified as a percentage of the target bitrate, and the libvpx wrapper converts this value as follows: "(maxrate * 100 / bitrate)".
minrate (minsection-pct) — set GOP min bitrate in bits/s. Note vpxenc's option is specified as a percentage of the target bitrate, and the libvpx wrapper converts this value as follows: "(minrate * 100 / bitrate)".
minrate, maxrate, b end-usage=cbr — "(minrate == maxrate == bitrate)".
crf (end-usage=cq, cq-level) — the constant-rate-factor knob, it lowkey lets you dial in quality directly 💅
tune (tune) — tells the encoder what to optimize for, and these are your choices:
- psnr (psnr) — optimize for PSNR numbers
- ssim (ssim) — optimize for SSIM numbers
quality, deadline (deadline) — the effort dial ✨
- best — use the best quality deadline. Ngl the naming is doing too much here: it's poorly named and quite slow, and you should honestly avoid it because it may give you worse quality output than good. Yeah. Iconic behavior honestly 😭
- good — use the good quality deadline. This is a good trade-off between speed and quality when used with the cpu-used option. This is the one, bestie 🫶
- realtime — use the realtime quality deadline.
speed, cpu-used (cpu-used) — sets the quality/speed ratio modifier. Higher values speed up the encode at the cost of quality. No cap, you're trading one for the other.
nr (noise-sensitivity) — noise sensitivity setting.
static-thresh — sets a change threshold on blocks, below which they will be skipped by the encoder.
slices (token-parts) — ⚠️ heads up: FFmpeg's slices option gives the total number of partitions, while vpxenc's token-parts is given as log2(partitions). Not the same thing, don't get got.
max-intra-rate — sets the maximum I-frame bitrate as a percentage of the target bitrate. A value of 0 means unlimited.
force_key_frames — literally just "VPX_EFLAG_FORCE_KF".
Alternate reference frame related 🦋 — the whole alt-ref squad:
- auto-alt-ref — enable use of alternate reference frames (2-pass only). Values greater than 1 enable multi-layer alternate reference frames (VP9 only).
- arnr-maxframes — set altref noise reduction max frame count.
- arnr-type — set altref noise reduction filter type: backward, forward, centered.
- arnr-strength — set altref noise reduction filter strength.
- rc-lookahead, lag-in-frames (lag-in-frames) — set number of frames to look ahead for frametype and ratecontrol.
- min-gf-interval — set minimum golden/alternate reference frame interval (VP9 only).
error-resilient — enable error resiliency features.
sharpness integer — increase sharpness at the expense of lower PSNR. The valid range is [0, 7].
ts-parameters — sets the temporal scalability configuration using a :-separated list of key=value pairs. For example, to specify temporal scalability parameters with ffmpeg:

ffmpeg -i INPUT -c:v libvpx -ts-parameters ts_number_layers=3:\
ts_target_bitrate=250,500,1000:ts_rate_decimator=4,2,1:\
ts_periodicity=4:ts_layer_id=0,2,1,2:ts_layering_mode=3 OUTPUT

Below is a brief explanation of each of the parameters, please refer to struct vpx_codec_enc_cfg in vpx/vpx_encoder.h for more details 💖

ts_number_layers — number of temporal coding layers.
ts_target_bitrate — target bitrate for each temporal layer (in kbps). (bitrate should be inclusive of the lower temporal layer).
ts_rate_decimator — frame rate decimation factor for each temporal layer.
ts_periodicity — length of the sequence defining frame temporal layer membership.
ts_layer_id — template defining the membership of frames to temporal layers.
ts_layering_mode — (optional) selecting the temporal structure from a set of pre-defined temporal layering modes. Currently supports the following options.
- 0 — no temporal layering flags are provided internally, relies on flags being passed in using "metadata" field in "AVFrame" with following keys:
  - vp8-flags — sets the flags passed into the encoder to indicate the referencing scheme for the current frame. Refer to function vpx_codec_encode in vpx/vpx_encoder.h for more details.
  - temporal_id — explicitly sets the temporal id of the current frame to encode.
- 2 — two temporal layers. 0-1...
- 3 — three temporal layers. 0-2-1-2...; with single reference frame.
- 4 — same as option "3", except there is a dependency between the two temporal layer 2 frames within the temporal period.

VP8-specific options 🍓
- screen-content-mode — screen content mode, one of: 0 (off), 1 (screen), 2 (screen with more aggressive rate control).
VP9-specific options 👑
- lossless — enable lossless mode.
- tile-columns — set number of tile columns to use. Note this is given as log2(tile_columns). For example, 8 tile columns would be requested by setting the tile-columns option to 3.
- tile-rows — set number of tile rows to use. Note this is given as log2(tile_rows). For example, 4 tile rows would be requested by setting the tile-rows option to 2.
- frame-parallel — enable frame parallel decodability features.
- aq-mode — set adaptive quantization mode (0: off (default), 1: variance 2: complexity, 3: cyclic refresh, 4: equator360). Yes that 2: is written exactly like that, don't @ me 😭
- colorspace color-space — set input color space. The VP9 bitstream supports signaling the following colorspaces:
  - rgb → sRGB
  - bt709 → bt709
  - unspecified → unknown
  - bt470bg → bt601
  - smpte170m → smpte170
  - smpte240m → smpte240
  - bt2020_ncl → bt2020
- row-mt boolean — enable row based multi-threading.
- tune-content — set content type: default (0), screen (1), film (2).
- corpus-complexity — Corpus VBR mode is a variant of standard VBR where the complexity distribution midpoint is passed in rather than calculated for a specific clip or chunk. The valid range is [0, 10000]. 0 (default) uses standard VBR.
- enable-tpl boolean — enable temporal dependency model.
- ref-frame-config — using per-frame metadata, set members of the structure vpx_svc_ref_frame_config_t in vpx/vp8cx.h to fine-control referencing schemes and frame buffer management. Use a :-separated list of key=value pairs. For example,

av_dict_set(&av_frame->metadata, "ref-frame-config", \
"rfc_update_buffer_slot=7:rfc_lst_fb_idx=0:rfc_gld_fb_idx=1:rfc_alt_fb_idx=2:rfc_reference_last=0:rfc_reference_golden=0:rfc_reference_alt_ref=0");

rfc_update_buffer_slot — indicates the buffer slot number to update
rfc_update_last — indicates whether to update the LAST frame
rfc_update_golden — indicates whether to update GOLDEN frame
rfc_update_alt_ref — indicates whether to update ALT_REF frame
rfc_lst_fb_idx — LAST frame buffer index
rfc_gld_fb_idx — GOLDEN frame buffer index
rfc_alt_fb_idx — ALT_REF frame buffer index
rfc_reference_last — indicates whether to reference LAST frame
rfc_reference_golden — indicates whether to reference GOLDEN frame
rfc_reference_alt_ref — indicates whether to reference ALT_REF frame
rfc_reference_duration — indicates frame duration

For more information about libvpx see: <http://www.webmproject.org/>

libvvenc 🌈

VVenC H.266/VVC encoder wrapper.

This encoder requires the presence of the libvvenc headers and library during configuration. You need to explicitly configure the build with --enable-libvvenc.

The VVenC project website is at <https://github.com/fraunhoferhhi/vvenc>.

Supported Pixel Formats

VVenC supports only 10-bit color spaces as input. But the internal (encoded) bit depth can be set to 8-bit or 10-bit at runtime. Wild, I know 💀

Options

b — sets target video bitrate.
g — set the GOP size. Currently support for g=1 (Intra only) or default.
preset — set the VVenC preset.
levelidc — set level idc.
tier — set vvc tier.
qp — set constant quantization parameter.
subopt boolean — set subjective (perceptually motivated) optimization. Default is 1 (on).
bitdepth8 boolean — set 8bit coding mode instead of using 10bit. Default is 0 (off).
period — set (intra) refresh period in seconds.
vvenc-params — set vvenc options using a list of key=value couples separated by ":". See vvencapp --fullhelp or vvencFFapp --fullhelp for a list of options.

For example, the options might be provided as:

intraperiod=64:decodingrefreshtype=idr:poc0idr=1:internalbitdepth=8

For example the encoding options might be provided with -vvenc-params:

ffmpeg -i input -c:v libvvenc -b 1M -vvenc-params intraperiod=64:decodingrefreshtype=idr:poc0idr=1:internalbitdepth=8 output.mp4

libwebp 🎀

libwebp WebP Image encoder wrapper.

libwebp is Google's official encoder for WebP images. It can encode in either lossy or lossless mode. Lossy images are essentially a wrapper around a VP8 frame. Lossless images are a separate codec developed by Google.

Pixel Format

Currently, libwebp only supports YUV420 for lossy and RGB for lossless due to limitations of the format and libwebp. Alpha is supported for either mode. Because of API limitations, if RGB is passed in when encoding lossy or YUV is passed in for encoding lossless, the pixel format will automatically be converted using functions from libwebp. This is not ideal and is done only for convenience. Not me side-eyeing the auto-conversion 🙈

Options

-lossless boolean — enables/disables use of lossless mode. Default is 0.
-compression_level integer — for lossy, this is a quality/speed tradeoff. Higher values give better quality for a given size at the cost of increased encoding time. For lossless, this is a size/speed tradeoff. Higher values give smaller size at the cost of increased encoding time. More specifically, it controls the number of extra algorithms and compression tools used, and varies the combination of these tools. This maps to the method option in libwebp. The valid range is 0 to 6. Default is 4.
-quality float — for lossy encoding, this controls image quality. For lossless encoding, this controls the effort and time spent in compression. Range is 0 to 100. Default is 75.
-preset type — configuration preset. This does some automatic settings based on the general type of the image.
- none — do not use a preset.
- default — use the encoder default.
- picture — digital picture, like portrait, inner shot
- photo — outdoor photograph, with natural lighting
- drawing — hand or line drawing, with high-contrast details
icon — small-sized colorful images 🎨
text — text-like stuff 📝

🌈 `libx264`, `libx264rgb` 🌈

okay bestie, this is the x264 wrapper — the H.264/MPEG-4 AVC encoder glow-up 💅✨

Heads up though: this encoder needs the libx264 headers AND the library present at configure time, no cap. You have to explicitly configure the build with --enable-libx264. Miss that and it's giving "where is she" 😭

libx264 supports an impressive number of features, including 8x8 and 4x4 adaptive spatial transform, adaptive B-frame placement, CAVLC/CABAC entropy coding, interlacing (MBAFF), lossless mode, and psy optimizations for detail retention (adaptive quantization, psy-RD, psy-trellis). like, she ate 🔥

Many libx264 encoder options are mapped onto FFmpeg global codec options, while the unique encoder options come through as private options. Also, the x264opts and x264-params private options let you pass a list of key=value tuples exactly as accepted by the libx264 x264_param_parse function. iykyk 🫶

The x264 project website is at <http://www.videolan.org/developers/x264.html>. go say hi 🦋

The libx264rgb encoder is literally the same as libx264, except it accepts packed RGB pixel formats as input instead of YUV. that's the whole vibe check 🌸

🎀 Supported Pixel Formats 🎀

x264 supports 8- to 10-bit color spaces. The exact bit depth is controlled at x264's configure time — so that's a build-time decision, not a runtime one, periodt ⭐

✨ Options ✨

The following options are supported by the libx264 wrapper. The x264-equivalent options or values are listed in parentheses for easy migration 💖

To reduce the duplication of documentation, only the private options and some others requiring special attention are documented here. For the documentation of the undocumented generic options, see the Codec Options chapter. (yes, we're delegating, she's busy 💅)

To get a more accurate and extensive documentation of the libx264 options, invoke the command x264 --fullhelp or consult the libx264 documentation. go straight to the source, bestie 👑

In the list below, note that the x264 option name is shown in parentheses after the libavcodec corresponding name, in case there is a direct mapping. 🔍

b (bitrate) — Set bitrate in bits/s. Important lore: FFmpeg's b option is expressed in bits/s, while x264's bitrate is in kilobits/s. do NOT mix those up, it's giving disaster 😭
bf (bframes) — Number of B-frames between I and P-frames.
g (keyint) — Maximum GOP size.
qmin (qpmin) — Minimum quantizer scale.
qmax (qpmax) — Maximum quantizer scale.
qdiff (qpstep) — Maximum difference between quantizer scales.
qblur (qblur) — Quantizer curve blur.
qcomp (qcomp) — Quantizer curve compression factor.
refs (ref) — Number of reference frames each P-frame can use. The range is 0-16. 🎯
level (level) — Set the x264_param_t.i_level_idc value in case the value is positive, it is ignored otherwise.

This value can be set using the AVCodecContext API (e.g. by setting the AVCodecContext value directly), and is specified as an integer mapped on a corresponding level (e.g. the value 31 maps to H.264 level IDC 3.1, as defined in the x264_levels table). It is ignored when set to a non positive value.

Alternatively it can be set as a private option, overriding the value set in AVCodecContext, and in this case must be specified as the level IDC identifier (e.g. 3.1), as defined by H.264 Annex A. two ways to say the same thing, pick your fighter 🥊

sc_threshold (scenecut) — Sets the threshold for the scene change detection.
trellis (trellis) — Performs Trellis quantization to increase efficiency. Enabled by default. ✅
nr (nr) — Noise reduction.
me_range (merange) — Maximum range of the motion search in pixels.
me_method (me) — Set motion estimation method. Possible values in the decreasing order of speed:
- dia (dia) / epzs (dia) — Diamond search with radius 1 (fastest). epzs is an alias for dia.
- hex (hex) — Hexagonal search with radius 2.
- umh (umh) — Uneven multi-hexagon search.
- esa (esa) — Exhaustive search.
- tesa (tesa) — Hadamard exhaustive search (slowest). she takes her time 💀
forced-idr — Normally, when forcing a I-frame type, the encoder can select any type of I-frame. This option forces it to choose an IDR-frame. no negotiating 🔒
subq (subme) — Sub-pixel motion estimation method.
b_strategy (b-adapt) — Adaptive B-frame placement decision algorithm. Use only on first-pass. ⚠️
keyint_min (min-keyint) — Minimum GOP size.
coder — Set entropy encoder. Possible values:
- ac — Enable CABAC.
- vlc — Enable CAVLC and disable CABAC. It generates the same effect as x264's --no-cabac option.
cmp — Set full pixel motion estimation comparison algorithm. Possible values:
- chroma — Enable chroma in motion estimation.
- sad — Ignore chroma in motion estimation. It generates the same effect as x264's --no-chroma-me option.
threads (threads) — Number of encoding threads. 🧵
thread_type — Set multithreading technique. Possible values:
- slice — Slice-based multithreading. It generates the same effect as x264's --sliced-threads option.
- frame — Frame-based multithreading.
flags — Set encoding flags. It can be used to disable closed GOP and enable open GOP by setting it to -cgop. The result is similar to the behavior of x264's --open-gop option. wild syntax, but ok 💫
rc_init_occupancy (vbv-init) — Initial VBV buffer occupancy.
preset (preset) — Set the encoding preset.
tune (tune) — Set tuning of the encoding params.
profile (profile) — Set profile restrictions.
fastfirstpass — Enable fast settings when encoding first pass, when set to 1. When set to 0, it has the same effect of x264's --slow-firstpass option.
crf (crf) — Set the quality for constant quality mode.
crf_max (crf-max) — In CRF mode, prevents VBV from lowering quality beyond this point. a little boundary, as a treat 🛑
qp (qp) — Set constant quantization rate control method parameter.
aq-mode (aq-mode) — Set AQ method. Possible values:
- none (0) — Disabled.
- variance (1) — Variance AQ (complexity mask).
- autovariance (2) — Auto-variance AQ (experimental). ⚗️
aq-strength (aq-strength) — Set AQ strength, reduce blocking and blurring in flat and textured areas.
psy — Use psychovisual optimizations when set to 1. When set to 0, it has the same effect as x264's --no-psy option.
psy-rd (psy-rd) — Set strength of psychovisual optimization, in psy-rd:psy-trellis format.
rc-lookahead (rc-lookahead) — Set number of frames to look ahead for frametype and ratecontrol. 👀
weightb — Enable weighted prediction for B-frames when set to 1. When set to 0, it has the same effect as x264's --no-weightb option.
weightp (weightp) — Set weighted prediction method for P-frames. Possible values:
- none (0) — Disabled.
- simple (1) — Enable only weighted refs.
- smart (2) — Enable both weighted refs and duplicates. 🧠
ssim (ssim) — Enable calculation and printing SSIM stats after the encoding.
intra-refresh (intra-refresh) — Enable the use of Periodic Intra Refresh instead of IDR frames when set to 1.
avcintra-class (class) — Configure the encoder to generate AVC-Intra. Valid values are 50, 100 and 200. 5️⃣0️⃣
bluray-compat (bluray-compat) — Configure the encoder to be compatible with the bluray standard. It is a shorthand for setting bluray-compat=1 force-cfr=1. shorthand queen 💿
b-bias (b-bias) — Set the influence on how often B-frames are used.
b-pyramid (b-pyramid) — Set method for keeping of some B-frames as references. Possible values:
- none (none) — Disabled.
- strict (strict) — Strictly hierarchical pyramid.
- normal (normal) — Non-strict (not Blu-ray compatible). ⚠️
mixed-refs — Enable the use of one reference per partition, as opposed to one reference per macroblock when set to 1. When set to 0, it has the same effect as x264's --no-mixed-refs option.
8x8dct — Enable adaptive spatial transform (high profile 8x8 transform) when set to 1. When set to 0, it has the same effect as x264's --no-8x8dct option.
fast-pskip — Enable early SKIP detection on P-frames when set to 1. When set to 0, it has the same effect as x264's --no-fast-pskip option.
aud (aud) — Enable use of access unit delimiters when set to 1.
mbtree — Enable use macroblock tree ratecontrol when set to 1. When set to 0, it has the same effect as x264's --no-mbtree option. 🌳
deblock (deblock) — Set loop filter parameters, in alpha:beta form.
cplxblur (cplxblur) — Set fluctuations reduction in QP (before curve compression).
partitions (partitions) — Set partitions to consider as a comma-separated list of values. Possible values in the list:
- p8x8 — 8x8 P-frame partition.
- p4x4 — 4x4 P-frame partition.
- b8x8 — 4x4 B-frame partition.
- i8x8 — 8x8 I-frame partition.
- i4x4 — 4x4 I-frame partition. (Enabling p4x4 requires p8x8 to be enabled. Enabling i8x8 requires adaptive spatial transform (8x8dct option) to be enabled.) ⚠️ read that twice bestie
- none (none) — Do not consider any partitions.
- all (all) — Consider every partition. all of them, greedy 💅
direct-pred (direct) — Set direct MV prediction mode. Possible values:
- none (none) — Disable MV prediction.
- spatial (spatial) — Enable spatial predicting.
- temporal (temporal) — Enable temporal predicting.
- auto (auto) — Automatically decided. she'll figure it out 🧚
slice-max-size (slice-max-size) — Set the limit of the size of each slice in bytes. If not specified but RTP payload size (ps) is specified, that is used. 🔗
stats (stats) — Set the file name for multi-pass stats.
nal-hrd (nal-hrd) — Set signal HRD information (requires vbv-bufsize to be set). Possible values:
- none (none) — Disable HRD information signaling.
- vbr (vbr) — Variable bit rate.
- cbr (cbr) — Constant bit rate (not allowed in MP4 container). 🚫
x264opts opts / x264-params opts — Override the x264 configuration using a :-separated list of key=value options.

The argument for both options is a list of key=value couples separated by ":". With x264opts the value can be omitted, and the value 1 is assumed in that case. little shortcut, love that for us 🎀

Okay so real talk about separators for a sec 💅 — for filter and psy-rd option values that use ":" as a separator themselves, use "," instead. They've accepted that too since forever ago, but it's kept undocumented for some reason 🙈 (iykyk).

So like, the options might be provided as:

level=30:bframes=0:weightp=0:cabac=0:ref=1:vbv-maxrate=768:vbv-bufsize=2000:analyse=all:me=umh:no-fast-pskip=1:subq=6:8x8dct=0:trellis=0

And say you wanna specify libx264 encoding options with ffmpeg, bestie, you do:

ffmpeg -i foo.mpg -c:v libx264 -x264opts keyint=123:min-keyint=20 -an out.mkv

To get the complete list of the libx264 options, invoke the command x264 --fullhelp or consult the libx264 documentation ✨ (yes, go read the docs, she's worth it).

a53cc boolean — Import closed captions (which must be ATSC compatible format) into output. Only the mpeg2 and h264 decoders provide these. Default is 1 (on). 🎀
udu_sei boolean — Import user data unregistered SEI if available into output. Default is 0 (off).
mb_info boolean — Set mb_info data through AVFrameSideData, only useful when used from the API. Default is 0 (off). (API girlies only, lowkey.)

Encoding ffpresets for common usages are provided so they can be used with the general presets system (e.g. passing the pre option) 🦋.

✨ libx265 ✨

x265 H.265/HEVC encoder wrapper. She's iconic, she's efficient 👑

This encoder requires the presence of the libx265 headers and library during configuration. You need to explicitly configure the build with --enable-libx265.

Options

b — Sets target video bitrate.
bf — listed all on her own with no description, just vibing 🫶
g — Set the GOP size.
keyint_min — Minimum GOP size.
refs — Number of reference frames each P-frame can use. The range is from 1-16.
preset — Set the x265 preset.
tune — Set the x265 tune parameter.
profile — Set profile restrictions.
crf — Set the quality for constant quality mode.
qp — Set constant quantization rate control method parameter.
qmin — Minimum quantizer scale.
qmax — Maximum quantizer scale.
qdiff — Maximum difference between quantizer scales.
qblur — Quantizer curve blur.
qcomp — Quantizer curve compression factor.
i_qfactor — (no description, she's a mystery, respect it 💀)
b_qfactor — (same, no description)
forced-idr — Normally, when forcing a I-frame type, the encoder can select any type of I-frame. This option forces it to choose an IDR-frame. 🔥
x265-stats — Specify the file name for 2-pass stats. This is set automatically when using the -passlogfile option.
udu_sei boolean — Import user data unregistered SEI if available into output. Default is 0 (off).
x265-params — Set x265 options using a list of key=value couples separated by ":". See x265 --help for a list of options.

So like, to specify libx265 encoding options with -x265-params:

ffmpeg -i input -c:v libx265 -x265-params crf=26:psy-rd=1 output.mp4

🌈 libxavs2 🌈

xavs2 AVS2-P2/IEEE1857.4 encoder wrapper. Niche? Maybe. Slay? Absolutely 💜

This encoder requires the presence of the libxavs2 headers and library during configuration. You need to explicitly configure the build with --enable-libxavs2.

The following standard libavcodec options are used:

b / bit_rate
g / gop_size
bf / max_b_frames

The encoder also has its own specific options:

Options

lcu_row_threads — Set the number of parallel threads for rows from 1 to 8 (default 5).
initial_qp — Set the xavs2 quantization parameter from 1 to 63 (default 34). This is used to set the initial qp for the first frame.
qp — Set the xavs2 quantization parameter from 1 to 63 (default 34). This is used to set the qp value under constant-QP mode.
max_qp — Set the max qp for rate control from 1 to 63 (default 55).
min_qp — Set the min qp for rate control from 1 to 63 (default 20).
speed_level — Set the Speed level from 0 to 9 (default 0). Higher is better but slower. 😭 (the eternal tradeoff, bestie)
log_level — Set the log level from -1 to 3 (default 0). -1: none, 0: error, 1: warning, 2: info, 3: debug.
xavs2-params — Set xavs2 options using a list of key=value couples separated by ":".

Example time, to specify libxavs2 encoding options with -xavs2-params:

ffmpeg -i input -c:v libxavs2 -xavs2-params RdoqLevel=0 output.avs2

🦋 libxeve 🦋

eXtra-fast Essential Video Encoder (XEVE) MPEG-5 EVC encoder wrapper. The xeve-equivalent options or values are listed in parentheses for easy migration ✨ (thoughtful queen behavior).

This encoder requires the presence of the libxeve headers and library during configuration. You need to explicitly configure the build with --enable-libxeve.

Many libxeve encoder options are mapped to FFmpeg global codec options, while unique encoder options are provided through private options. Additionally the xeve-params private options allows one to pass a list of key=value tuples as accepted by the libxeve parse_xeve_params function.

The xeve project website is at <https://github.com/mpeg5/xeve>.

Options

The following options are supported by the libxeve wrapper. The xeve-equivalent options or values are listed in parentheses for easy migration.

To reduce the duplication of documentation, only the private options and some others requiring special attention are documented here. For the documentation of the undocumented generic options, see the Codec Options chapter.

To get a more accurate and extensive documentation of the libxeve options, invoke the command xeve_app --help or consult the libxeve documentation.

b (bitrate) — Set target video bitrate in bits/s. Note that FFmpeg's b option is expressed in bits/s, while xeve's bitrate is in kilobits/s. ⚠️ (units will get you, fr fr)
bf (bframes) — Set the maximum number of B frames (1,3,7,15).
g (keyint) — Set the GOP size (I-picture period).
preset (preset) — Set the xeve preset. Set the encoder preset value to determine encoding speed [fast, medium, slow, placebo].
tune (tune) — Set the encoder tune parameter [psnr, zerolatency].
profile (profile) — Set the encoder profile [0: baseline; 1: main].
crf (crf) — Set the quality for constant quality mode. Constant rate factor <10..49> [default: 32].
qp (qp) — Set constant quantization rate control method parameter. Quantization parameter qp <0..51> [default: 32].
threads (threads) — Force to use a specific number of threads.

🎀 libxvid 🎀

Xvid MPEG-4 Part 2 encoder wrapper. Old school but she still eats 💅

This encoder requires the presence of the libxvidcore headers and library during configuration. You need to explicitly configure the build with --enable-libxvid --enable-gpl.

The native mpeg4 encoder supports the MPEG-4 Part 2 format, so users can encode to this format without this library. (No cap — you might not even need her!)

Options

The following options are supported by the libxvid wrapper. Some of the following options are listed but are not documented, and correspond to shared codec options. See the Codec Options chapter for their documentation. The other shared options which are not listed have no effect for the libxvid encoder.

b — listed, shared codec option, not documented here.
g — same deal.
qmin — same deal.
qmax — same deal.
mpeg_quant — same deal.
threads — same deal.
bf — same deal.
b_qfactor — same deal.
b_qoffset — same deal.
flags — Set specific encoding flags. Possible values:
- mv4 — Use four motion vector by macroblock.
- aic — Enable high quality AC prediction.
- gray — Only encode grayscale.
- qpel — Enable quarter-pixel motion compensation.
- cgop — Enable closed GOP.
- global_header — Place global headers in extradata instead of every keyframe.
gmc — Enable the use of global motion compensation (GMC). Default is 0 (disabled).
me_quality — Set motion estimation quality level. Possible values in decreasing order of speed and increasing order of quality:
- 0 — Use no motion estimation (default).
- 1, 2 — Enable advanced diamond zonal search for 16x16 blocks and half-pixel refinement for 16x16 blocks.
- 3, 4 — Enable all of the things described above, plus advanced diamond zonal search for 8x8 blocks and half-pixel refinement for 8x8 blocks, also enable motion estimation on chroma planes for P and B-frames.
- 5, 6 — Enable all of the things described above, plus extended 16x16 and 8x8 blocks search.
mbd — Set macroblock decision algorithm. Possible values in the increasing order of quality:
- simple — Use macroblock comparing function algorithm (default).
- bits — Enable rate distortion-based half pixel and quarter pixel refinement for 16x16 blocks.
- rd — Enable all of the things described above, plus rate distortion-based half pixel and quarter pixel refinement for 8x8 blocks, and rate distortion-based search using square pattern.
lumi_aq — Enable lumi masking adaptive quantization when set to 1. Default is 0 (disabled).
variance_aq — Enable variance adaptive quantization when set to 1. Default is 0 (disabled).

When combined with lumi_aq, the resulting quality will not be better than any of the two specified individually. In other words, the resulting quality will be the worse one of the two effects. 😭💀 (two great tastes that do NOT taste great together)

trellis — Set rate-distortion optimal quantization.
ssim — Set structural similarity (SSIM) displaying method. Possible values:
- off — Disable displaying of SSIM information.
- avg — Output average SSIM at the end of encoding to stdout. The format of showing the average SSIM is:

Average SSIM: %f

For users who are not familiar with C, %f means a float number, or a decimal (e.g. 0.939232).

frame — Output both per-frame SSIM data during encoding and average SSIM at the end of encoding to stdout. The format of per-frame information is:

SSIM: avg: %1.3f min: %1.3f max: %1.3f

For users who are not familiar with C, %1.3f means a float number rounded to 3 digits after the dot (e.g. 0.932).

ssim_acc — Set SSIM accuracy. Valid options are integers within the range of 0-4, while 0 gives the most accurate result and 4 computes the fastest.

⭐ MediaCodec ⭐

MediaCodec encoder wrapper enables hardware-accelerated video encoding on Android device 🫶📱 It supports H.264, H.265 (HEVC), VP8, VP9, MPEG-4, and AV1 encoding (whether works or not is device dependent). Yeah — device dependent — so your mileage may absolutely vary, bestie 🙈

Okay so bestie, Android gives you two sets of APIs: Java MediaCodec and NDK MediaCodec. The MediaCodec encoder wrapper supports both. 🫶 Note that the NDK MediaCodec API operates without requiring a JVM, but may fail to function outside the JVM environment due to dependencies on system framework services, particularly after Android 15. 💀

Here are the knobs 💅:

ndk_codec boolean — Use the NDK-based MediaCodec API instead of the Java API. Enabled by default if av_jni_get_java_vm() return NULL.
ndk_async boolean — Use NDK MediaCodec in async mode. Async mode has less overhead than poll in a loop in sync mode. The drawback of async mode is AV_CODEC_FLAG_GLOBAL_HEADER doesn't work (use extract_extradata bsf when necessary). It doesn't work and will be disabled automatically on devices below Android 8.0.
codec_name string — A codec type can have multiple implementations on a single device, this option specify which backend to use (via MediaCodec createCodecByName API). It's NULL by default, and encoder is created by createEncoderByType.
bitrate_mode integer — Possible values:
- cq — Constant quality mode
- vbr — Variable bitrate mode
- cbr — Constant bitrate mode
- cbr_fd — Constant bitrate mode with frame drops
pts_as_dts boolean — Use PTS as DTS. This is a workaround since MediaCodec API doesn't provide decoding timestamp. It is enabled automatically if B frame is 0.
operating_rate integer — The desired operating rate that the codec will need to operate at, zero for unspecified. This is used for cases like high-speed/slow-motion video capture, where the video encoder format contains the target playback rate (e.g. 30fps), but the component must be able to handle the high operating capture rate (e.g. 240fps). This rate will be used by codec for resource planning and setting the operating points.
qp_i_min integer — Minimum quantization parameter for I frame.
qp_p_min integer — Minimum quantization parameter for P frame.
qp_b_min integer — Minimum quantization parameter for B frame.
qp_i_max integer — Maximum quantization parameter for I frame.
qp_p_max integer — Maximum quantization parameter for P frame.
qp_b_max integer — Maximum quantization parameter for B frame.

🎬 MediaFoundation

The following wrappers for encoders in the MediaFoundation framework are available — and yes, the original list is giving exactly this:

*<h264_mf>
*<hevc_mf>
*<av1_mf>

These support both software and hardware encoding. ✨

Video encoders can take input in either of nv12 or yuv420p form (some encoders support both, some support only either - in practice, nv12 is the safer choice, especially among HW encoders).

Hardware-accelerated encoding requires D3D11, including hardware scaling capabilities through the scale_d3d11 filter. 🔥

To list all available options for the MediaFoundation encoders, use:

ffmpeg -h encoder=<encoder> e.g. ffmpeg -h encoder=h264_mf

Options

rate_control — Select rate control mode. Available modes:
- default — Default mode
- cbr — CBR mode
- pc_vbr — Peak constrained VBR mode
- u_vbr — Unconstrained VBR mode
- quality — Quality mode
- ld_vbr — Low delay VBR mode (requires Windows 8+)
- g_vbr — Global VBR mode (requires Windows 8+)
- gld_vbr — Global low delay VBR mode (requires Windows 8+)
scenario — Select usage scenario. Available scenarios:
- default — Default scenario
- display_remoting — Display remoting scenario
- video_conference — Video conference scenario
- archive — Archive scenario
- live_streaming — Live streaming scenario
- camera_record — Camera record scenario
- display_remoting_with_feature_map — Display remoting with feature map scenario
quality — Set encoding quality (0-100). -1 means default quality.
hw_encoding — Force hardware encoding (0-1). Default is 0 (disabled).

Examples

Hardware encoding:

ffmpeg -i input.mp4 -c:v h264_mf -hw_encoding 1 output.mp4

Hardware-accelerated decoding with hardware encoding:

ffmpeg -hwaccel d3d11va -i input.mp4 -c:v h264_mf -hw_encoding 1 output.mp4

Hardware-accelerated decoding, HW scaling and encoding with quality setting:

ffmpeg -hwaccel d3d11va -hwaccel_output_format d3d11 -i input.mp4 -vf scale_d3d11=1920:1080 -c:v hevc_mf -hw_encoding 1 -quality 80 output.mp4

🎨 Microsoft RLE

Microsoft RLE aka MSRLE encoder. Only 8-bit palette mode supported. Compatible with Windows 3.1 and Windows 95. 🖥️

Options

g integer — Keyframe interval. A keyframe is inserted at least every "-g" frames, sometimes sooner.

📼 mpeg2

MPEG-2 video encoder.

Options

profile — Select the mpeg2 profile to encode:
- 422
- high
- ss — Spatially Scalable
- snr — SNR Scalable
- main
- simple
level — Select the mpeg2 level to encode:
- high
- high1440
- main
- low
seq_disp_ext integer — Specifies if the encoder should write a sequence_display_extension to the output.
- -1, auto — Decide automatically to write it or not (this is the default) by checking if the data to be written is different from the default or unspecified values.
- 0, never — Never write it.
- 1, always — Always write it.
video_format integer — Specifies the video_format written into the sequence display extension indicating the source of the video pictures. The default is unspecified, can be component, pal, ntsc, secam or mac. For maximum compatibility, use component.
a53cc boolean — Import closed captions (which must be ATSC compatible format) into output. Default is 1 (on).

🖼️ png

PNG image encoder.

Options

compression_level — Sets the compression level, from 0 to 9(default)

Private options

dpi integer — Set physical density of pixels, in dots per inch, unset by default
dpm integer — Set physical density of pixels, in dots per meter, unset by default
pred method — Set prediction method (none, sub, up, avg, paeth, mixed), default is paeth

🍎 ProRes

Apple ProRes encoder.

FFmpeg contains 2 ProRes encoders, the prores-aw and prores-ks encoder. The used encoder can be chosen with the "-vcodec" option.

Private Options for prores-ks

profile integer — Select the ProRes profile to encode
- proxy
- lt
- standard
- hq
- 4444
- 4444xq
quant_mat integer — Select quantization matrix.
- auto
- default
- proxy
- lt
- standard
- hq

If set to auto, the matrix matching the profile will be picked. If not set, the matrix providing the highest quality, default, will be picked.

bits_per_mb integer — How many bits to allot for coding one macroblock. Different profiles use between 200 and 2400 bits per macroblock, the maximum is 8000.
mbs_per_slice integer — Number of macroblocks in each slice (1-8); the default value (8) should be good in almost all situations.
vendor string — Override the 4-byte vendor ID. A custom vendor ID like apl0 would claim the stream was produced by the Apple encoder.
alpha_bits integer — Specify number of bits for alpha component. Possible values are 0, 8 and 16. Use 0 to disable alpha plane coding.

Speed considerations

In the default mode of operation the encoder has to honor frame constraints (i.e. not produce frames with size bigger than requested) while still making output picture as good as possible. A frame containing a lot of small details is harder to compress and the encoder would spend more time searching for appropriate quantizers for each slice. 😭

Setting a higher bits_per_mb limit will improve the speed.

For the fastest encoding speed set the qscale parameter (4 is the recommended value) and do not set a size constraint.

⚡ QSV Encoders

The family of Intel QuickSync Video encoders (MPEG-2, H.264, HEVC, JPEG/MJPEG, VP9, AV1)

Ratecontrol Method

The ratecontrol method is selected as follows:

When global_quality is specified, a quality-based mode is used. Specifically this means either
- CQP - constant quantizer scale, when the qscale codec flag is also set (the -qscale ffmpeg option).
- LA_ICQ - intelligent constant quality with lookahead, when the look_ahead option is also set.
- ICQ -- intelligent constant quality otherwise. For the ICQ modes, global quality range is 1 to 51, with 1 being the best quality.
Otherwise when the desired average bitrate is specified with the b option, a bitrate-based mode is used.
- LA - VBR with lookahead, when the look_ahead option is specified.
- VCM - video conferencing mode, when the vcm option is set.
- CBR - constant bitrate, when maxrate is specified and equal to the average bitrate.
- VBR - variable bitrate, when maxrate is specified, but is higher than the average bitrate.
- AVBR - average VBR mode, when maxrate is not specified, both avbr_accuracy and avbr_convergence are set to non-zero. This mode is available for H264 and HEVC on Windows.
Otherwise the default ratecontrol method CQP is used.

Note that depending on your system, a different mode than the one you specified may be selected by the encoder. Set the verbosity level to verbose or higher to see the actual settings used by the QSV runtime. 👑

Global Options -> MSDK Options

Additional libavcodec global options are mapped to MSDK options as follows:

g/gop_size -> GopPicSize
bf/max_b_frames+1 -> GopRefDist
rc_init_occupancy/rc_initial_buffer_occupancy -> InitialDelayInKB
slices -> NumSlice
refs -> NumRefFrame
b_strategy/b_frame_strategy -> BRefType
cgop/CLOSED_GOP codec flag -> GopOptFlag
For the CQP mode, the i_qfactor/i_qoffset and b_qfactor/b_qoffset set the difference between QPP and QPI, and QPP and QPB respectively.
Setting the coder option to the value vlc will make the H.264 encoder use CAVLC instead of CABAC.

Common Options

Following options are used by all qsv encoders.

async_depth — tells the app how many async operations it can fire off before it actually has to stop and sync the result. Set it to zero and the value just… isn't specified. default vibes only ✨
preset — the speed-vs-quality dial, listing out a whole range of choices from veryfast (= best speed) all the way to veryslow (= best quality). Pick your fighter:
- veryfast
- faster
- fast
- medium
- slow
- slower
- veryslow
forced_idr — forces I frames to be IDR frames. no negotiation 💅
low_power — for encoders, flip this flag to ON to reduce power consumption and GPU usage. your laptop's fans will literally thank you 🫶

⚡ Runtime Options ⚡

okay so these are the ones you can actually use during qsv encoding, mid-vibe, no restart needed:

global_quality, i_quant_factor, i_quant_offset, b_quant_factor, b_quant_offset — supported in h264_qsv and hevc_qsv. change these values to reset the qsv codec's qp configuration. 🔥
max_frame_size — supported in h264_qsv and hevc_qsv. change this value to reset the qsv codec's MaxFrameSize configuration.
gop_size — change this value to reset the qsv codec's gop configuration.
int_ref_type, int_ref_cycle_size, int_ref_qp_delta, int_ref_cycle_dist — supported in h264_qsv and hevc_qsv. change these values to reset the qsv codec's Intra Refresh configuration.
qmax, qmin, max_qp_i, min_qp_i, max_qp_p, min_qp_p, max_qp_b, min_qp_b — supported in h264_qsv. change these values to reset the qsv codec's max/min qp configuration.
low_delay_brc — supported in h264_qsv, hevc_qsv and av1_qsv. change this value to reset the qsv codec's low_delay_brc configuration.
framerate — change this value to reset the qsv codec's framerate configuration.
bit_rate, rc_buffer_size, rc_initial_buffer_occupancy, rc_max_rate — change these values to reset the qsv codec's bitrate control configuration.
pic_timing_sei — supported in h264_qsv and hevc_qsv. change this value to reset the qsv codec's pic_timing_sei configuration.
qsv_params — sets QSV encoder parameters as a colon-separated list of key-value pairs. 💅

the qsv_params should be formatted as "key1=value1:key2=value2:...".

these parameters get passed directly to the underlying Intel Quick Sync Video (QSV) encoder using the MFXSetParameter function. no middleman, no cap.

example:

ffmpeg -i input.mp4 -c:v h264_qsv -qsv_params "CodingOption1=1:CodingOption2=2" output.mp4

this option gives you fine-grained control over all the various encoder-specific settings the QSV encoder provides. you're basically in the driver's seat now 👑

🎀 H264 options 🎀

these are the ones used by h264_qsv, and there are so many of them, buckle up bestie:

extbrc — extended bitrate control.
recovery_point_sei — set this flag to insert the recovery point SEI message at the beginning of every intra refresh cycle.
rdo — enable rate distortion optimization.
max_frame_size — maximum encoded frame size in bytes.
max_frame_size_i — maximum encoded frame size for I frames in bytes. if this value is set as larger than zero, then for I frames the value set by max_frame_size is ignored. it said what it said 💀
max_frame_size_p — maximum encoded frame size for P frames in bytes. if this value is set as larger than zero, then for P frames the value set by max_frame_size is ignored.
max_slice_size — maximum encoded slice size in bytes.
bitrate_limit — toggle bitrate limitations. modifies bitrate to be in the range imposed by the QSV encoder. setting this flag off may lead to violation of HRD conformance. mind that specifying bitrate below the QSV encoder range might significantly affect quality. if on, this option takes effect in non-CQP modes: if bitrate is not in the range imposed by the QSV encoder, it will be changed to be in the range.

⚠️ fr fr, turning this off can break HRD conformance. don't say nobody warned you 🙈

mbbrc — setting this flag enables macroblock level bitrate control that generally improves subjective visual quality. enabling this flag may have negative impact on performance and objective visual quality metric. trade-offs, bestie, trade-offs 😭
low_delay_brc — setting this flag turns on or off the LowDelayBRC feature in the qsv plugin, which provides more accurate bitrate control to minimize the variance of bitstream size frame by frame. value: -1-default 0-off 1-on
adaptive_i — this flag controls insertion of I frames by the QSV encoder. turn ON this flag to allow changing of frame type from P and B to I.
adaptive_b — this flag controls changing of frame type from B to P.
p_strategy — enable P-pyramid: 0-default 1-simple 2-pyramid (bf need to be set to 0).
b_strategy — this option controls usage of B frames as reference.
dblk_idc — this option disables deblocking. it has value in range 0~2.
cavlc — if set, CAVLC is used; if unset, CABAC is used for encoding.
vcm — video conferencing mode, please see ratecontrol method.
idr_interval — distance (in I-frames) between IDR frames.
pic_timing_sei — insert picture timing SEI with pic_struct_syntax element.
single_sei_nal_unit — put all the SEI messages into one NALU.
max_dec_frame_buffering — maximum number of frames buffered in the DPB.
look_ahead — use VBR algorithm with look ahead.
look_ahead_depth — depth of look ahead in number frames.
look_ahead_downsampling — downscaling factor for the frames saved for the lookahead analysis:
- unknown
- auto
- off
- 2x
- 4x
int_ref_type — specifies intra refresh type. the major goal of intra refresh is improvement of error resilience without significant impact on encoded bitstream size caused by I frames. the SDK encoder achieves this by encoding part of each frame in the refresh cycle using intra MBs. none means no refresh. vertical means vertical refresh, by column of MBs. horizontal means horizontal refresh, by rows of MBs. slice means horizontal refresh by slices without overlapping. in case of slice, in_ref_cycle_size is ignored. to enable intra refresh, B frame should be set to 0. the way this option has lore is unreal 📖
int_ref_cycle_size — specifies number of pictures within the refresh cycle starting from 2. 0 and 1 are invalid values. periodt.
int_ref_qp_delta — specifies QP difference for inserted intra MBs. this is a signed value in [-51, 51] range if target encoding bit-depth for luma samples is 8, and this range is [-63, 63] for 10 bit-depth or [-75, 75] for 12 bit-depth respectively.
int_ref_cycle_dist — distance between the beginnings of the intra-refresh cycles in frames.
profile — pick one:
- unknown
- baseline
- main
- high
a53cc — use A53 Closed Captions (if available).
aud — insert the Access Unit Delimiter NAL.
mfmode — Multi-Frame Mode:
- off
- auto
repeat_pps — repeat pps for every frame.
max_qp_i — maximum video quantizer scale for I frame.
min_qp_i — minimum video quantizer scale for I frame.
max_qp_p — maximum video quantizer scale for P frame.
min_qp_p — minimum video quantizer scale for P frame.
max_qp_b — maximum video quantizer scale for B frame.
min_qp_b — minimum video quantizer scale for B frame.
scenario — provides a hint to the encoder about the scenario for the encoding session:
- unknown
- displayremoting
- videoconference
- archive
- livestreaming
- cameracapture
- videosurveillance
- gamestreaming
- remotegaming
avbr_accuracy — accuracy of the AVBR ratecontrol (unit of tenth of percent).
avbr_convergence — convergence of the AVBR ratecontrol (unit of 100 frames).

the parameters avbr_accuracy and avbr_convergence are for the average variable bitrate control (AVBR) algorithm. the algorithm focuses on overall encoding quality while meeting the specified bitrate, target_bitrate, within the accuracy range avbr_accuracy, after a avbr_Convergence period. this method does not follow HRD and the instant bitrate is not capped or padded. iykyk 🌸

skip_frame — use per-frame metadata "qsv_skip_frame" to skip a frame when encoding. this option defines the usage of that metadata:
- no_skip — frame skipping is disabled.
- insert_dummy — encoder inserts into the bitstream a frame where all macroblocks are encoded as skipped.
- insert_nothing — similar to insert_dummy, but the encoder inserts nothing into the bitstream. the skipped frames are still used in brc. for example, gop still include skipped frames, and the frames after skipped frames will be larger in size.
- brc_only — skip_frame metadata indicates the number of missed frames before the current frame.

👑 HEVC Options 👑

and these are the ones used by hevc_qsv — yes there's overlap, yes they're still all listed, we don't skip 💜

extbrc — extended bitrate control.
recovery_point_sei — set this flag to insert the recovery point SEI message at the beginning of every intra refresh cycle.
rdo — enable rate distortion optimization.
max_frame_size — maximum encoded frame size in bytes.
max_frame_size_i — maximum encoded frame size for I frames in bytes. if this value is set as larger than zero, then for I frames the value set by max_frame_size is ignored.
max_frame_size_p — maximum encoded frame size for P frames in bytes. if this value is set as larger than zero, then for P frames the value set by max_frame_size is ignored.
max_slice_size — maximum encoded slice size in bytes.
mbbrc — setting this flag enables macroblock level bitrate control that generally improves subjective visual quality. enabling this flag may have negative impact on performance and objective visual quality metric.
low_delay_brc — setting this flag turns on or off the LowDelayBRC feature in the qsv plugin, which provides more accurate bitrate control to minimize the variance of bitstream size frame by frame. value: -1-default 0-off 1-on
adaptive_i — this flag controls insertion of I frames by the QSV encoder. turn ON this flag to allow changing of frame type from P and B to I.
adaptive_b — this flag controls changing of frame type from B to P.
p_strategy — enable P-pyramid: 0-default 1-simple 2-pyramid (bf need to be set to 0).
b_strategy — this option controls usage of B frames as reference.
dblk_idc — this option disables deblocking. it has value in range 0~2.
idr_interval — distance (in I-frames) between IDR frames:
- begin_only — output an IDR-frame only at the beginning of the stream.
load_plugin — a user plugin you wanna load up in an internal session ✨
- none
- hevc_sw
- hevc_hw
load_plugins — a :-separated list of hexadecimal plugin UIDs to load in an internal session. lowkey a whole plugin haul 💅
look_ahead_depth — depth of look ahead, in number of frames. only available when the extbrc option is enabled (so like, turn that on first bestie).
profile — sets the encoding profile (scc requires libmfx >= 1.32).
- unknown
- main
- main10
- mainsp
- rext
- scc
tier — sets the encoding tier (only level >= 4 can support high tier, no cap). this option only takes effect when the level option is specified, so don't skip it and expect magic 🙈
- main
- high
gpb — 1: GPB (generalized P/B frame)
- 0: regular P frame.
tile_cols — number of columns for tiled encoding.
tile_rows — number of rows for tiled encoding.
aud — insert the Access Unit Delimiter NAL. little punctuation mark for your stream 🎀
pic_timing_sei — insert picture timing SEI with pic_struct_syntax element.
transform_skip — turn this option ON to enable transformskip. it's supported on platforms equal or newer than ICL.
int_ref_type — specifies intra refresh type. the major goal of intra refresh is improvement of error resilience without significant impact on encoded bitstream size caused by I frames. the SDK encoder achieves this by encoding part of each frame in the refresh cycle using intra MBs. none means no refresh. vertical means vertical refresh, by column of MBs. horizontal means horizontal refresh, by rows of MBs. slice means horizontal refresh by slices without overlapping. in the case of slice, in_ref_cycle_size is ignored. to enable intra refresh, B frame should be set to 0 ⚠️
int_ref_cycle_size — specifies number of pictures within a refresh cycle, starting from 2. 0 and 1 are invalid values, periodt.
int_ref_qp_delta — specifies QP difference for inserted intra MBs. this is a signed value in [-51, 51] range if the target encoding bit-depth for luma samples is 8, and this range is [-63, 63] for 10 bit-depth or [-75, 75] for 12 bit-depth respectively. precision queen 💅
int_ref_cycle_dist — distance between the beginnings of the intra-refresh cycles, in frames.
max_qp_i — maximum video quantizer scale for I frame.
min_qp_i — minimum video quantizer scale for I frame.
max_qp_p — maximum video quantizer scale for P frame.
min_qp_p — minimum video quantizer scale for P frame.
max_qp_b — maximum video quantizer scale for B frame.
min_qp_b — minimum video quantizer scale for B frame.
scenario — gives the encoder a lil hint about the scenario for the encoding session 🫶
- unknown
- displayremoting
- videoconference
- archive
- livestreaming
- cameracapture
- videosurveillance
- gamestreaming
- remotegaming
avbr_accuracy — accuracy of the AVBR rate control (unit of tenth of percent).
avbr_convergence — convergence of the AVBR rate control (unit of 100 frames).

okay so the parameters avbr_accuracy and avbr_convergence are for the average variable bitrate control (AVBR) algorithm. the algorithm focuses on overall encoding quality while meeting the specified bitrate, target_bitrate, within the accuracy range avbr_accuracy, after an avbr_Convergence period. this method does not follow HRD and the instant bitrate is not capped or padded. no cap, literally 🧢

skip_frame — use per-frame metadata "qsv_skip_frame" to skip a frame when encoding. this option defines the usage of this metadata.
- no_skip — frame skipping is disabled.
- insert_dummy — encoder inserts into the bitstream a frame where all macroblocks are encoded as skipped.
- insert_nothing — similar to insert_dummy, but the encoder inserts nothing into the bitstream. the skipped frames are still used in brc. for example, gop still includes skipped frames, and the frames after skipped frames will be larger in size 👀
- brc_only — skip_frame metadata indicates the number of missed frames before the current frame.

MPEG2 Options

these options are used by mpeg2_qsv 🎬

profile
- unknown
- simple
- main
- high

VP9 Options

these options are used by vp9_qsv 🦋

profile
- unknown
- profile0
- profile1
- profile2
- profile3
tile_cols — number of columns for tiled encoding (requires libmfx >= 1.29).
tile_rows — number of rows for tiled encoding (requires libmfx >= 1.29).

AV1 Options

these options are used by av1_qsv (requires libvpl) 👑

profile
- unknown
- main
tile_cols — number of columns for tiled encoding.
tile_rows — number of rows for tiled encoding.
adaptive_i — this flag controls insertion of I frames by the QSV encoder. turn ON this flag to allow changing of frame type from P and B to I.
adaptive_b — this flag controls changing of frame type from B to P.
b_strategy — this option controls usage of B frames as reference.
extbrc — extended bitrate control.
look_ahead_depth — depth of look ahead, in number of frames, available when the extbrc option is enabled.
low_delay_brc — setting this flag turns on or off the LowDelayBRC feature in the qsv plugin, which provides more accurate bitrate control to minimize the variance of bitstream size frame by frame. value: -1-default, 0-off, 1-on.
max_frame_size — set the allowed max size in bytes for each frame. if the frame size exceeds the limitation, the encoder will adjust the QP value to control the frame size. invalid in CQP rate control mode ⚠️
max_frame_size_i — maximum encoded frame size for I frames in bytes. if this value is set as larger than zero, then for I frames the value set by max_frame_size is ignored.
max_frame_size_p — maximum encoded frame size for P frames in bytes. if this value is set as larger than zero, then for P frames the value set by max_frame_size is ignored.

snow

Options

iterative_dia_size — dia size for the iterative motion estimation. snow is such a vibe honestly ❄️

VAAPI encoders

wrappers for hardware encoders accessible via VAAPI 🔥

these encoders only accept input in VAAPI hardware surfaces. if you have input in software frames, use the hwupload filter to upload them to the GPU. don't fight it, just upload 🫶

the following standard libavcodec options are used:

g / gop_size
bf / max_b_frames
profile

if not set, this will be determined automatically from the format of the input frames and the profiles supported by the driver. she'll figure it out herself 💅

level
b / bit_rate
maxrate / rc_max_rate
bufsize / rc_buffer_size
rc_init_occupancy / rc_initial_buffer_occupancy
compression_level

speed / quality tradeoff: higher values are faster / worse quality.

q / global_quality

size / quality tradeoff: higher values are smaller / worse quality.

qmin
qmax
i_qfactor / i_quant_factor
i_qoffset / i_quant_offset
b_qfactor / b_quant_factor
b_qoffset / b_quant_offset
slices

all encoders support the following options:

low_power — some drivers/platforms offer a second encoder for some codecs intended to use less power than the default encoder; setting this option will attempt to use that encoder. note that it may support a reduced feature set, so some other options may not be available in this mode 🙈
idr_interval — set the number of normal intra frames between full-refresh (IDR) frames in open-GOP mode. the intra frames are still IRAPs, but will not include global headers and may have non-decodable leading pictures.
b_depth — set the B-frame reference depth. when set to one (the default), all B-frames will refer only to P- or I-frames. when set to greater values multiple layers of B-frames will be present, frames in each layer only referring to frames in higher layers. layers on layers, it's giving onion 🧅
async_depth — maximum processing parallelism. increase this to improve single channel performance. this option doesn't work if the driver doesn't implement the vaSyncBuffer function. please make sure there are enough hw_frames allocated if a large number of async_depth is used. be responsible bestie ⚠️
max_frame_size — set the allowed max size in bytes for each frame. if the frame size exceeds the limitation, the encoder will adjust the QP value to control the frame size. invalid in CQP rate control mode.
rc_mode — set the rate control mode to use. a given driver may only support a subset of modes.

possible modes:

auto — choose the mode automatically based on driver support and the other options. this is the default.
CQP — constant-quality.
CBR — constant-bitrate.
VBR — variable-bitrate.
ICQ — intelligent constant-quality.
QVBR — quality-defined variable-bitrate.
AVBR — average variable bitrate.

blbrc — enable block level rate control, which assigns a different bitrate block by block. invalid for CQP mode.

each encoder also has its own specific options, naturally:

av1_vaapi

profile sets the value of seq_profile. tier sets the value of seq_tier. level sets the value of seq_level_idx.

tiles — set the number of tiles to encode the input video with, as columns x rows. (default is auto, which means use minimal tile column/row number).
tile_groups — set tile groups number. all the tiles will be distributed as evenly as possible to each tile group. (default is 1).

h264_vaapi

profile sets the value of profile_idc and the constraint_set*_flags. level sets the value of level_idc.

coder — set entropy encoder (default is cabac). possible values:
- ac
- cabac — use CABAC.
- vlc
- cavlc — use CAVLC.
aud — include access unit delimiters in the stream (not included by default).
sei — set SEI message types to include. some combination of the following values:
- identifier — include a user_data_unregistered message containing information about the encoder.
- timing — include picture timing parameters (buffering_period and pic_timing messages).
recovery_point — include recovery points where it’s appropriate (the recovery_point messages). 🫶

💜 `hevc_vaapi`

profile and level set the values of general_profile_idc and general_level_idc respectively.

aud — include access unit delimiters in the stream (not included by default).
tier — set general_tier_flag. This may affect the level chosen for the stream if it is not explicitly specified.
sei — set SEI message types to include. Some combination of the following values:
- hdr — include HDR metadata if the input frames have it (mastering_display_colour_volume and content_light_level messages).
tiles — set the number of tiles to encode the input video with, as columns x rows. Larger numbers allow greater parallelism in both encoding and decoding, but may decrease coding efficiency.

🦋 `mjpeg_vaapi`

Only baseline DCT encoding is supported. The encoder always uses the standard quantisation and huffman tables — global_quality scales the standard quantisation table (range 1-100).

For YUV, 4:2:0, 4:2:2 and 4:4:4 subsampling modes are supported. RGB is also supported, and will create an RGB JPEG.

jfif — include JFIF header in each frame (not included by default).
huffman — include standard huffman tables (on by default). Turning this off will save a few hundred bytes in each output frame, but may lose compatibility with some JPEG decoders which don’t fully handle MJPEG.

🌸 `mpeg2_vaapi`

profile and level set the value of profile_and_level_indication.

🎀 `vp8_vaapi`

B-frames are not supported.

global_quality sets the q_idx used for non-key frames (range 0-127).

loop_filter_level — manually set the loop filter parameters.
loop_filter_sharpness — manually set the loop filter parameters.

⭐ `vp9_vaapi`

global_quality sets the q_idx used for P-frames (range 0-255).

loop_filter_level — manually set the loop filter parameters.
loop_filter_sharpness — manually set the loop filter parameters.

B-frames are supported, but the output stream is always in encode order rather than display order. If B-frames are enabled, it may be necessary to use the vp9_raw_reorder bitstream filter to modify the output stream to display frames in the correct order.

Only normal frames are produced — the vp9_superframe bitstream filter may be required to produce a stream usable with all decoders.

💖 `vbn`

Vizrt Binary Image encoder.

This format is used by the broadcast vendor Vizrt for quick texture streaming. Advanced features of the format such as LZW compression of texture data or generation of mipmaps are not supported.

Options

format string — sets the texture compression used by the VBN file. Can be dxt1, dxt5 or raw. Default is dxt5.

🌈 `vc2`

SMPTE VC-2 (previously BBC Dirac Pro). This codec was primarily aimed at professional broadcasting but since it supports yuv420, yuv422 and yuv444 at 8 (limited range or full range), 10 or 12 bits, this makes it suitable for other tasks which require low overhead and low compression (like screen recording).

Options

b — sets target video bitrate. Usually that’s around 1:6 of the uncompressed video bitrate (e.g. for 1920x1080 50fps yuv422p10 that’s around 400Mbps). Higher values (close to the uncompressed bitrate) turn on lossless compression mode.
field_order — enables field coding when set (e.g. to tt — top field first) for interlaced inputs. Should increase compression with interlaced content as it splits the fields and encodes each separately.
wavelet_depth — sets the total amount of wavelet transforms to apply, between 1 and 5 (default). Lower values reduce compression and quality. Less capable decoders may not be able to handle values of wavelet_depth over 3.
wavelet_type — sets the transform type. Currently only 5_3 (LeGall) and 9_7 (Deslauriers-Dubuc) are implemented, with 9_7 being the one with better compression and thus is the default.
slice_width — sets the slice size for each slice. Larger values result in better compression. For compatibility with other more limited decoders use slice_width of 32.
slice_height — sets the slice size for each slice. Larger values result in better compression. For compatibility with other more limited decoders use slice_height of 8.
tolerance — sets the undershoot tolerance of the rate control system in percent. This is to prevent an expensive search from being run.
qm — sets the quantization matrix preset to use by default or when wavelet_depth is set to 5.
- default — uses the default quantization matrix from the specifications, extended with values for the fifth level. This provides a good balance between keeping detail and omitting artifacts.
- flat — use a completely zeroed out quantization matrix. This increases PSNR but might reduce perception. Use in bogus benchmarks.
- color — reduces detail but attempts to preserve color at extremely low bitrates.

🎀 SUBTITLES ENCODERS 🎀

💅 `dvbsub`

This codec encodes the bitmap subtitle format that is used in DVB broadcasts and recordings. The bitmaps are typically embedded in a container such as MPEG-TS as a separate stream.

Options

min_bpp integer (2, 4, or 8) — set a minimum bits-per-pixel value for the subtitle color lookup tables.

DVB supports 2, 4, and 8 bits-per-pixel color lookup tables. This option enables forcing a particular bits-per-pixel value regardless of the number of colors. Since not all players support or properly support 2 bits-per-pixel, this value defaults to 4.

🍓 `dvdsub`

This codec encodes the bitmap subtitle format that is used in DVDs. Typically they are stored in VOBSUB file pairs (*.idx + *.sub), and they can also be used in Matroska files.

Options

palette — specify the global palette used by the bitmaps.

The format for this option is a string containing 16 24-bits hexadecimal numbers (without 0x prefix) separated by commas, for example "0d00ee, ee450d, 101010, eaeaea, 0ce60b, ec14ed, ebff0b, 0d617a, 7b7b7b, d1d1d1, 7b2a0e, 0d950c, 0f007b, cf0dec, cfa80c, 7c127b".

even_rows_fix — when set to 1, enable a work-around that makes the number of pixel rows even in all subtitles. This fixes a problem with some players that cut off the bottom row if the number is odd. The work-around just adds a fully transparent row if needed. The overhead is low, typically one byte per subtitle on average.

By default, this work-around is disabled.

🧚 `lrc`

This codec encodes the LRC lyrics format.

Options

precision — specify the precision of the fractional part of the timestamp. Time base is determined based on this value.

Defaults to 2 for centiseconds.

🌈 BITSTREAM FILTERS 🌈

When you configure your FFmpeg build, all the supported bitstream filters are enabled by default. You can list all available ones using the configure option "--list-bsfs".

You can disable all the bitstream filters using the configure option "--disable-bsfs", and selectively enable any bitstream filter using the option "--enable-bsf=BSF", or you can disable a particular bitstream filter using the option "--disable-bsf=BSF".

The option "-bsfs" of the ff* tools will display the list of all the supported bitstream filters included in your build.

The ff* tools have a -bsf option applied per stream, taking a comma-separated list of filters, whose parameters follow the filter name after a '='.

ffmpeg -i INPUT -c:v copy -bsf:v filter1[=opt1=str1:opt2=str2][,filter2] OUTPUT

Below is a description of the currently available bitstream filters, with their parameters, if any.

💜 `aac_adtstoasc`

Convert MPEG-2/4 AAC ADTS to an MPEG-4 Audio Specific Configuration bitstream.

This filter creates an MPEG-4 AudioSpecificConfig from an MPEG-2/4 ADTS header and removes the ADTS header.

This filter is required for example when copying an AAC stream from a raw ADTS AAC or an MPEG-TS container to MP4A-LATM, to an FLV file, or to MOV/MP4 files and related formats such as 3GP or M4A. Please note that it is auto-inserted for MP4A-LATM and MOV/MP4 and related formats.

🔥 `av1_metadata`

Modify metadata embedded in an AV1 stream.

td — insert or remove temporal delimiter OBUs in all temporal units of the stream.
- insert — insert a TD at the beginning of every TU which does not already have one.
- remove — remove the TD from the beginning of every TU which has one.
color_primaries — set the color description fields in the stream (see AV1 section 6.4.2).
transfer_characteristics — set the color description fields in the stream (see AV1 section 6.4.2).
matrix_coefficients — set the color description fields in the stream (see AV1 section 6.4.2).
color_range — set the color range in the stream (see AV1 section 6.4.2; note that this cannot be set for streams using BT.709 primaries, sRGB transfer characteristic and identity (RGB) matrix coefficients).
- tv — limited range.
- pc — full range.
chroma_sample_position — set the chroma sample location in the stream (see AV1 section 6.4.2). This can only be set for 4:2:0 streams.
- vertical — left position (matching the default in MPEG-2 and H.264).
- colocated — top-left position.
tick_rate — set the tick rate (time_scale / num_units_in_display_tick) in the timing info in the sequence header.
num_ticks_per_picture — set the number of ticks in each picture, to indicate that the stream has a fixed framerate. Ignored if tick_rate is not also set.
delete_padding — deletes Padding OBUs.

🫶 `chomp`

Remove zero padding at the end of a packet.

💀 `dca_core`

Extract the core from a DCA/DTS stream, dropping extensions such as DTS-HD.

👑 `dovi_rpu`

Manipulate Dolby Vision metadata in a HEVC/AV1 bitstream, optionally enabling metadata compression.

strip — if enabled, strip all Dolby Vision metadata (configuration record + RPU data blocks) from the stream.
compression — which compression level to enable.
- none — no metadata compression.
- limited — limited metadata compression scheme. Should be compatible with most devices. This is the default.
- extended — extended metadata compression. Devices are not required to support this. Note that this level currently behaves the same as limited in libavcodec.

✨ `dump_extra`

Add extradata to the beginning of the filtered packets except when said packets already exactly begin with the extradata that is intended to be added.

freq — okay so this is the lil extra arg that tells the filter which packets get the extradata glow-up. It takes these values:
- k, keyframe — adds extradata to all key packets 🔑
- e, all — adds extradata to literally every packet, no cap

If you don't specify it, it just assumes k. Default behavior, we love her.

So like, for example, this ffmpeg command forces a global header (which, yeah, disables individual packet headers) in the H.264 packets that the libx264 encoder makes, but then corrects them by adding the header stored in extradata to the key packets:

ffmpeg -i INPUT -map 0 -flags:v +global_header -c:v libx264 -bsf:v dump_extra out.ts

🩹 `dv_error_marker`

Blocks in DV that are marked as damaged get swapped out for blocks of whatever color you pick. It's giving "cover it up with a cute sticker" energy but for corrupted video data.

color — the color used to replace the damaged blocks 💅
sta — a 16 bit mask that says which of the 16 possible error status values get replaced by colored blocks. 0xFFFE is the default, and it replaces all the non-0 error status values.
- ok — no error, no concealment ✨
- err — error, no concealment 😭
- res — reserved
- notok — error or concealment
- notres — not reserved
- Aa, Ba, Ca, Ab, Bb, Cb, A, B, C, a, b, erri, erru — the specific error status code

see page 44-46 or section 5.5 of <http://web.archive.org/web/20060927044735/http://www.smpte.org/smpte_store/standards/pdf/s314m.pdf> — yes bestie, it's a real link, go read the spec, I believe in you 🫶

🧲 `eac3_core`

Extracts the core out of an E-AC-3 stream, dropping the extra channels. Basically stripping it down to the basics, minimalism era.

🔤 `eia608_to_smpte436m`

Converts from a EIA_608 stream to a SMPTE_436M_ANC data stream — and the way it does it is wrapping the closed captions in CTA-708 CDP VANC packets. Iconic little pipeline.

line_number — picks which line number the generated VANC packets go on. You generally want either line 9 (the default) or 11.
wrapping_type — picks the SMPTE 436M wrapping type, defaults to vanc_frame. It accepts:
- vanc_frame — VANC frame (interlaced or segmented progressive frame)
- vanc_field_1
- vanc_field_2
- vanc_progressive_frame
sample_coding — picks the SMPTE 436M sample coding, defaults to 8bit_luma. It accepts:
- 8bit_luma — 8-bit component luma samples
- 8bit_color_diff — 8-bit component color difference samples
- 8bit_luma_and_color_diff — 8-bit component luma and color difference samples
- 10bit_luma — 10-bit component luma samples
- 10bit_color_diff — 10-bit component color difference samples
- 10bit_luma_and_color_diff — 10-bit component luma and color difference samples
- 8bit_luma_parity_error — 8-bit component luma samples with parity error
- 8bit_color_diff_parity_error — 8-bit component color difference samples with parity error
- 8bit_luma_and_color_diff_parity_error — 8-bit component luma and color difference samples with parity error
initial_cdp_sequence_cntr — the initial value of the CDP's 16-bit unsigned integer cdp_hdr_sequence_cntr and cdp_ftr_sequence_cntr fields. Defaults to 0.
cdp_frame_rate — sets the CDP's cdp_frame_rate field. And like, heads up: this doesn't actually change the timing of the data stream, it just changes the values inserted in that field in the generated CDP packets. Defaults to 30000/1001.

📤 `extract_extradata`

Extracts the in-band extradata. Let me explain, because this one's actually kind of cool:

Certain codecs let the long-term headers (think MPEG-2 sequence headers, or H.264/HEVC (VPS/)SPS/PPS) be transmitted either "in-band" (like, as a part of the bitstream that contains the coded frames) or "out of band" (like, on the container level). That second form? FFmpeg calls it "extradata" 💅

This bitstream filter detects the in-band headers and makes them available as extradata. Free the headers, honestly.

remove — when this option is on, the long-term headers get removed from the bitstream after extraction. Gone, deleted, no notes.

🚦 `filter_units`

Removes units whose types are in — or not in — a given set, straight outta the stream.

pass_types — a list of unit types (or ranges of unit types) to pass through while removing all the others. You write it as a '|'-separated list of unit type values, or ranges of values with -.
remove_types — same exact thing as pass_types, except the units in the given set get removed and everything else passes through. Inverse queen 👑

The types used by pass_types and remove_types correspond to NAL unit types (nal_unit_type) in H.264, HEVC and H.266 (see Table 7-1 in the H.264 and HEVC specifications or Table 5 in the H.266 specification), to marker values for JPEG (without 0xFF prefix) and to start codes without start code prefix (i.e. the byte following the 0x000001) for MPEG-2. For VP8 and VP9, every unit has type zero.

Extradata is unchanged by this transformation, but bestie, note this: if the stream contains inline parameter sets then the output may be unusable if they're removed. Like, actually unusable. Don't do that and then come crying to me 😭

For example, to remove all non-VCL NAL units from an H.264 stream:

ffmpeg -i INPUT -c:v copy -bsf:v 'filter_units=pass_types=1-5' OUTPUT

To remove all AUDs, SEI and filler from an H.265 stream:

ffmpeg -i INPUT -c:v copy -bsf:v 'filter_units=remove_types=35|38-40' OUTPUT

To remove all user data from a MPEG-2 stream, including Closed Captions:

ffmpeg -i INPUT -c:v copy -bsf:v 'filter_units=remove_types=178' OUTPUT

To remove all SEI from a H264 stream, including Closed Captions:

ffmpeg -i INPUT -c:v copy -bsf:v 'filter_units=remove_types=6' OUTPUT

To remove all prefix and suffix SEI from a HEVC stream, including Closed Captions and dynamic HDR:

ffmpeg -i INPUT -c:v copy -bsf:v 'filter_units=remove_types=39|40' OUTPUT

🎨 `hapqa_extract`

Extracts the Rgb or Alpha part of an HAPQA file — without recompression, because we're not out here degrading quality — so you can make an HAPQ or an HAPAlphaOnly file. Crafty.

texture — specifies which texture to keep.
- color
- alpha

Convert HAPQA to HAPQ:

ffmpeg -i hapqa_inputfile.mov -c copy -bsf:v hapqa_extract=texture=color -tag:v HapY -metadata:s:v:0 encoder="HAPQ" hapq_file.mov

Convert HAPQA to HAPAlphaOnly:

ffmpeg -i hapqa_inputfile.mov -c copy -bsf:v hapqa_extract=texture=alpha -tag:v HapA -metadata:s:v:0 encoder="HAPAlpha Only" hapalphaonly_file.mov

🔧 `h264_metadata`

Modifies metadata that's embedded in an H.264 stream. She's a little surgeon, fr.

aud — inserts or removes AUD NAL units in all access units of the stream. Options:
- pass
- insert
- remove

Default is pass.

sample_aspect_ratio — sets the sample aspect ratio of the stream in the VUI parameters. See H.264 table E-1.
overscan_appropriate_flag — sets whether the stream is suitable for display using overscan or not (see H.264 section E.2.1).
video_format, video_full_range_flag — sets the video format in the stream (see H.264 section E.2.1 and table E-2).
colour_primaries, transfer_characteristics, matrix_coefficients — sets the colour description in the stream (see H.264 section E.2.1 and tables E-3, E-4 and E-5).
chroma_sample_loc_type — sets the chroma sample location in the stream (see H.264 section E.2.1 and figure E-1).
tick_rate — sets the tick rate (time_scale / num_units_in_tick) in the VUI parameters. This is the smallest time unit representable in the stream, and in many cases it represents the field rate of the stream (double the frame rate). Math! ✨
fixed_frame_rate_flag — sets whether the stream has a fixed framerate — typically this indicates that the framerate is exactly half the tick rate, but the exact meaning depends on interlacing and the picture structure (see H.264 section E.2.1 and table E-6).
zero_new_constraint_set_flags — zeroes constraint_set4_flag and constraint_set5_flag in the SPS. These bits were reserved in a previous version of the H.264 spec, and thus some hardware decoders require these to be zero. And good news: the result of zeroing this is still a valid bitstream 👑
crop_left, crop_right, crop_top, crop_bottom — sets the frame cropping offsets in the SPS. These values will replace the current ones if the stream is already cropped.

These fields are set in pixels. Note that some sizes may not be representable if the chroma is subsampled or the stream is interlaced (see H.264 section 7.4.2.1.1). So like, don't come for me if your weird number doesn't fit 🙈

sei_user_data — inserts a string as SEI unregistered user data. The argument must be of the form UUID+string, where the UUID is as hex digits possibly separated by hyphens, and the string can be anything.

For example, 086f3693-b7b3-4f2c-9653-21492feee5b8+hello will insert the string ``hello'' associated with the given UUID. Cute little hidden message, iykyk.

delete_filler — deletes both filler NAL units and filler SEI messages. Bye, filler, you were never that important.
display_orientation — inserts, extracts or removes Display orientation SEI messages. See H.264 section D.1.27 and D.2.27 for syntax and semantics. Options:
- pass
- insert
- remove
- extract

Default is pass.

Insert mode works in conjunction with rotate and flip options. Any pre-existing Display orientation messages will be removed in insert or remove mode. Extract mode attaches the display matrix to the packet as side data.

rotate — sets rotation in display orientation SEI (anticlockwise angle in degrees). Range is -360 to +360. Default is NaN. Spin her around 💫
flip — sets flip in display orientation SEI. Options:
- horizontal
- vertical

Default is unset.

level — sets the level in the SPS. Refer to H.264 section A.3 and tables A-1 to A-5.

The argument must be the name of a level (for example, 4.2), a level_idc value (for example, 42), or the special name auto indicating that the filter should attempt to guess the level from the input stream properties. Let her guess, she's smart.

🎬 `h264_mp4toannexb`

Converts an H.264 bitstream from length prefixed mode to start code prefixed mode (as defined in the Annex B of the ITU-T H.264 specification).

This is required by some streaming formats, typically the MPEG-2 transport stream format (muxer mpegts). So if you're going to mpegts, you literally need this, no cap.

For example, to remux an MP4 file containing an H.264 stream to mpegts format with ffmpeg, you can use the command:

ffmpeg -i INPUT.mp4 -codec copy -bsf:v h264_mp4toannexb OUTPUT.ts

Please note that this filter is auto-inserted for MPEG-TS (muxer mpegts) and raw H.264 (muxer h264) output formats. Like, it just does it for you. You don't have to ask 💅

🔁 `h264_redundant_pps`

This applies a specific fixup to some Blu-ray BDMV H264 streams that contain redundant PPSs. The PPSs modify irrelevant parameters of the stream, confusing other transformations which require the correct extradata. So she's basically the friend who fixes your outfit before you go out — the PPSs were messing up the vibe for everyone else 🫶

okay so the encoder on those impacted streams? it keeps adding extra PPSs all the way through the stream, and it also switches up the initial QP and whether weighted prediction was on 😭. that’s a whole mess when you copy the stream into a global header container, because the starting PPS is not it for the rest of the stream. like, one side effect? seeking will return garbled output until a new PPS shows up. not cute.

This BSF removes those extra PPSs and rewrites the slice headers so the stream uses a single leading PPS in the global header, which resolves the issue. slay 💅.

`hevc_metadata`

Modify metadata embedded in an HEVC stream ✨.

aud — Insert or remove AUD NAL units in all access units of the stream.
- insert
- remove
sample_aspect_ratio — Set the sample aspect ratio in the stream in the VUI parameters.
video_format, video_full_range_flag — Set the video format in the stream (see H.265 section E.3.1 and table E.2).
colour_primaries, transfer_characteristics, matrix_coefficients — Set the colour description in the stream (see H.265 section E.3.1 and tables E.3, E.4 and E.5).
chroma_sample_loc_type — Set the chroma sample location in the stream (see H.265 section E.3.1 and figure E.1).
tick_rate — Set the tick rate in the VPS and VUI parameters (time_scale / num_units_in_tick). Combined with num_ticks_poc_diff_one, this can set a constant framerate in the stream. Note that it is likely to be overridden by container parameters when the stream is in a container.
num_ticks_poc_diff_one — Set poc_proportional_to_timing_flag in VPS and VUI and use this value to set num_ticks_poc_diff_one_minus1 (see H.265 sections 7.4.3.1 and E.3.1). Ignored if tick_rate is not also set.
crop_left, crop_right, crop_top, crop_bottom — Set the conformance window cropping offsets in the SPS. These values will replace the current ones if the stream is already cropped. These fields are set in pixels. Note that some sizes may not be representable if the chroma is subsampled (H.265 section 7.4.3.2.1).
width, height — Set width and height after crop.
level — Set the level in the VPS and SPS. See H.265 section A.4 and tables A.6 and A.7. The argument must be the name of a level (for example, 5.1), a general_level_idc value (for example, 153 for level 5.1), or the special name auto indicating that the filter should attempt to guess the level from the input stream properties.

`hevc_mp4toannexb`

Convert an HEVC/H.265 bitstream from length prefixed mode to start code prefixed mode (as defined in the Annex B of the ITU-T H.265 specification).

This is required by some streaming formats, typically the MPEG-2 transport stream format (muxer mpegts).

For example to remux an MP4 file containing an HEVC stream to mpegts format with ffmpeg, you can use the command:

ffmpeg -i INPUT.mp4 -codec copy -bsf:v hevc_mp4toannexb OUTPUT.ts

Please note that this filter is auto-inserted for MPEG-TS (muxer mpegts) and raw HEVC/H.265 (muxer h265 or hevc) output formats.

`imxdump`

Modifies the bitstream to fit in MOV and to be usable by the Final Cut Pro decoder. This filter only applies to the mpeg2video codec, and is likely not needed for Final Cut Pro 7 and newer with the appropriate -tag:v.

For example, to remux 30 MB/sec NTSC IMX to MOV:

ffmpeg -i input.mxf -c copy -bsf:v imxdump -tag:v mx3n output.mov

`mjpeg2jpeg`

Convert MJPEG/AVI1 packets to full JPEG/JFIF packets.

MJPEG is a video codec wherein each video frame is essentially a JPEG image. The individual frames can be extracted without loss, e.g. by

ffmpeg -i ../some_mjpeg.avi -c:v copy frames_%d.jpg

Unfortunately, these chunks are incomplete JPEG images, because they lack the DHT segment required for decoding. Quoting from http://www.digitalpreservation.gov/formats/fdd/fdd000063.shtml:

Avery Lee, writing in the rec.video.desktop newsgroup in 2001, commented that "MJPEG, or at least the MJPEG in AVIs having the MJPG fourcc, is restricted JPEG with a fixed -- and omitted -- Huffman table. The JPEG must be YCbCr colorspace, it must be 4:2:2, and it must use basic Huffman encoding, not arithmetic or progressive. . . . You can indeed extract the MJPEG frames and decode them with a regular JPEG decoder, but you have to prepend the DHT segment to them, or else the decoder won't have any idea how to decompress the data. The exact table necessary is given in the OpenDML spec."

This bitstream filter patches the header of frames extracted from an MJPEG stream (carrying the AVI1 header ID and lacking a DHT segment) to produce fully qualified JPEG images.

ffmpeg -i mjpeg-movie.avi -c:v copy -bsf:v mjpeg2jpeg frame_%d.jpg
exiftran -i -9 frame*.jpg
ffmpeg -i frame_%d.jpg -c:v copy rotated.avi

`mjpegadump`

Add an MJPEG A header to the bitstream, to enable decoding by Quicktime.

`mov2textsub`

Extract a representable text file from MOV subtitles, stripping the metadata header from each subtitle packet.

`mpeg2_metadata`

Modify metadata embedded in an MPEG-2 stream ✨.

display_aspect_ratio — Set the display aspect ratio in the stream. The following fixed values are supported:
- 4/3
- 16/9
- 221/100

Any other value will result in square pixels being signalled instead (see H.262 section 6.3.3 and table 6-3).

frame_rate — Set the frame rate in the stream. This is constructed from a table of known values combined with a small multiplier and divisor - if the supplied value is not exactly representable, the nearest representable value will be used instead (see H.262 section 6.3.3 and table 6-4).
video_format — Set the video format in the stream (see H.262 section 6.3.6 and table 6-6).
colour_primaries, transfer_characteristics, matrix_coefficients — Set the colour description in the stream (see H.262 section 6.3.6 and tables 6-7, 6-8 and 6-9).

`mpeg4_unpack_bframes`

Unpack DivX-style packed B-frames.

DivX-style packed B-frames are not valid MPEG-4 and were only a workaround for the broken Video for Windows subsystem. They use more space, can cause minor AV sync issues, require more CPU power to decode (unless the player has some decoded picture queue to compensate the 2,0,2,0 frame per packet style) and cause trouble if copied into a standard container like MP4 or mpeg-ps/ts, because MPEG-4 decoders may not be able to decode them, since they are not valid MPEG-4.

For example to fix an AVI file containing an MPEG-4 stream with DivX-style packed B-frames using ffmpeg, you can use the command:

ffmpeg -i INPUT.avi -codec copy -bsf:v mpeg4_unpack_bframes OUTPUT.avi

`noise`

Damages the contents of packets or simply drops them without damaging the container. Can be used for fuzzing or testing error resilience/concealment.

Parameters:

amount — Accepts an expression whose evaluation per-packet determines how often bytes in that packet will be modified. A value below 0 will result in a variable frequency. Default is 0 which results in no modification. However, if neither amount nor drop is specified, amount will be set to -1. See below for accepted variables.
drop — Accepts an expression evaluated per-packet whose value determines whether that packet is dropped. Evaluation to a positive value results in the packet being dropped. Evaluation to a negative value results in a variable chance of it being dropped, roughly inverse in proportion to the magnitude of the value. Default is 0 which results in no drops. See below for accepted variables.
dropamount — Accepts a non-negative integer, which assigns a variable chance of it being dropped, roughly inverse in proportion to the value. Default is 0 which results in no drops. This option is kept for backwards compatibility and is equivalent to setting drop to a negative value with the same magnitude i.e. dropamount=4 is the same as drop=-4. Ignored if drop is also specified.

Both amount and drop accept expressions containing the following variables:

n — The index of the packet, starting from zero.
tb — The timebase for packet timestamps.
pts — Packet presentation timestamp.
dts — Packet decoding timestamp.
nopts — Constant representing AV_NOPTS_VALUE.
startpts — First non-AV_NOPTS_VALUE PTS seen in the stream.
startdts — First non-AV_NOPTS_VALUE DTS seen in the stream.
duration, d — Packet duration, in timebase units.
pos — Packet position in input; may be -1 when unknown or not set.
size — Packet size, in bytes.
key — Whether packet is marked as a keyframe.
state — A pseudo random integer, primarily derived from the content of packet payload.

Examples

Apply modification to every byte but don't drop any packets.

ffmpeg -i INPUT -c copy -bsf noise=1 output.mkv

Drop every video packet not marked as a keyframe after timestamp 30s but do not modify any of the remaining packets.

ffmpeg -i INPUT -c copy -bsf:v noise=drop='gt(pts*tb\,30)*not(key)' output.mkv

Drop one second of audio every 10 seconds and add some random noise to the rest.

ffmpeg -i INPUT -c copy -bsf:a noise=amount=-1:drop='between(mod(pts*tb\,10)\,9\,10)' output.mkv

`null`

This bitstream filter passes the packets through unchanged. iconic, no notes 💅.

`pcm_rechunk`

Repacketize PCM audio to a fixed number of samples per packet or a fixed packet rate per second. This is similar to the asetnsamples audio filter but works on audio packets instead of audio frames.

nb_out_samples, n — Set the number of samples per each output audio packet. The number is intended as the number of samples per each channel. Default value is 1024.
pad, p — If set to 1, the filter will pad the last audio packet with silence, so that it will contain the same number of samples (or roughly the same number of samples, see frame_rate) as the previous ones. Default value is 1.
frame_rate, r — This option makes the filter output a fixed number of packets per second instead of a fixed number of samples per packet. If the audio sample rate is not divisible by the frame rate then the number of samples will not be constant but will vary slightly so that each packet will start as close to the frame boundary as possible. Using this option has precedence over nb_out_samples.

You can generate the well known 1602-1601-1602-1601-1602 pattern of 48kHz audio for NTSC frame rate using the frame_rate option 🎀.

ffmpeg -f lavfi -i sine=r=48000:d=1 -c pcm_s16le -bsf pcm_rechunk=r=30000/1001 -f framecrc -

pgs_frame_merge 🎀

Bestie, this one merges a sequence of PGS Subtitle segments that ends with an “end of display set” segment into a single packet. It’s literally required by some containers that support PGS subtitles (muxer matroska). No cap. ✨

prores_metadata 🦋

Okay so this filter modifies color property metadata embedded in a prores stream. Obsessed.

color_primaries — set the color primaries. Available values are:
- auto — keep the same color primaries property (default).
- unknown
- bt709
- bt470bg — BT601 625
- smpte170m — BT601 525
- bt2020
- smpte431 — DCI P3
- smpte432 — P3 D65
transfer_characteristics — set the color transfer. Available values are:
- auto — keep the same transfer characteristics property (default).
- unknown
- bt709 — BT 601, BT 709, BT 2020
- smpte2084 — SMPTE ST 2084
- arib-std-b67 — ARIB STD-B67
matrix_coefficients — set the matrix coefficient. Available values are:
- auto — keep the same colorspace property (default).
- unknown
- bt709
- smpte170m — BT 601
- bt2020nc

Set Rec709 colorspace for each frame of the file 💅

ffmpeg -i INPUT -c copy -bsf:v prores_metadata=color_primaries=bt709:color_trc=bt709:colorspace=bt709 output.mov

Set Hybrid Log-Gamma parameters for each frame of the file 🌈

ffmpeg -i INPUT -c copy -bsf:v prores_metadata=color_primaries=bt2020:color_trc=arib-std-b67:colorspace=bt2020nc output.mov

remove_extra 🧚

Remove extradata from packets. She’s doing the most but in a minimal way.

It accepts the following parameter:

freq — set which frame types to remove extradata from.
- k — remove extradata from non-keyframes only.
- keyframe — remove extradata from keyframes only.
- e, all — remove extradata from all frames.

setts 👑

Set PTS and DTS in packets. Main character energy fr fr.

It accepts the following parameters:

ts, pts, dts — set expressions for PTS, DTS or both.
duration — set expression for duration.
time_base — set output time base.
prescale — set whether to convert time fields to user-set output time base before evaluation of expressions. Defaults to 0.

The expressions are evaluated through the eval API and can contain the following constants:

N — the count of the input packet. Starting from 0.
TS — the demux timestamp in input in case of ts or dts option or presentation timestamp in case of pts option.
POS — the original position in the file of the packet, or undefined if undefined for the current packet.
DTS — the demux timestamp in input.
PTS — the presentation timestamp in input.
DURATION — the duration in input.
STARTDTS — the DTS of the first packet.
STARTPTS — the PTS of the first packet.
PREV_INDTS — the previous input DTS.
PREV_INPTS — the previous input PTS.
PREV_INDURATION — the previous input duration.
PREV_OUTDTS — the previous output DTS.
PREV_OUTPTS — the previous output PTS.
PREV_OUTDURATION — the previous output duration.
NEXT_DTS — the next input DTS.
NEXT_PTS — the next input PTS.
NEXT_DURATION — the next input duration.
TB — the timebase in which time fields are denominated. The user-set output timebase if prescale is enabled, else the input timebase.
TB_OUT — the output timebase.
SR — the sample rate of stream packet belongs.
NOPTS — the AV_NOPTS_VALUE constant.

For example, to set PTS equal to DTS (not recommended if B-frames are involved):

ffmpeg -i INPUT -c:a copy -bsf:a setts=pts=DTS out.mkv

showinfo 💖

Log basic packet information. Mainly useful for testing, debugging, and development. Like a little vibe check for your packets, iykyk.

smpte436m_to_eia608 🌸

Convert from a SMPTE_436M_ANC data stream to a EIA_608 stream, extracting the closed captions from CTA-708 CDP VANC packets, and ignoring all other data. She’s selective, and honestly? Same.

text2movsub 🎀

Convert text subtitles to MOV subtitles (as used by the mov_text codec) with metadata headers.

trace_headers 🔥

Log trace output containing all syntax elements in the coded stream headers (everything above the level of individual coded blocks). This can be useful for debugging low-level stream issues.

Supports AV1, H.264, H.265, (M)JPEG, MPEG-2 and VP9, but depending on the build only a subset of these may be available. So like… check your build bestie, no cap.

truehd_core 💜

Extract the core from a TrueHD stream, dropping ATMOS data. She’s stripping it down to the essentials, periodt.

vp9_metadata 🦋

Modify metadata embedded in a VP9 stream. Little header glow-up.

color_space — set the color space value in the frame header. Note that any frame set to RGB will be implicitly set to PC range and that RGB is incompatible with profiles 0 and 2.
- unknown
- bt601
- bt709
- smpte170
- smpte240
- bt2020
- rgb
color_range — set the color range value in the frame header. Note that any value imposed by the color space will take precedence over this value.
- tv
- pc

vp9_superframe ⭐

Merge VP9 invisible (alt-ref) frames back into VP9 superframes. This fixes merging of split/segmented VP9 streams where the alt-ref frame was split from its visible counterpart. Reunited and it feels so good 😭

vp9_superframe_split ✨

Split VP9 superframes into single frames. Simple. Iconic. No notes.

vp9_raw_reorder 🌈

Given a VP9 stream with correct timestamps but possibly out of order, insert additional show-existing-frame packets to correct the ordering. She’s putting the timeline back in line, bestie. 🫶

✨ FORMAT OPTIONS ✨

Okay bestie, so the libavformat library ships some generic global options that you can set on all the muxers and demuxers. And on top of that? Each muxer or demuxer can have its own lil private options that are specific to that one component. iykyk 💅

You set these by passing -option value in the FFmpeg tools, or by setting the value explicitly in the AVFormatContext options, or via the libavutil/opt.h API if you're doing the programmatic thing.

Here's the full list, no skips, we're reading the whole menu 🍓

avioflags flags (input/output) — Possible values:
- direct — Reduce buffering.
probesize integer (input) — Sets the probing size in bytes, i.e. how much data gets analyzed to get stream information. A higher value lets it detect more info when that info is dispersed into the stream, but it'll increase latency. Must be an integer not lesser than 32. Default is 5000000.
max_probe_packets integer (input) — Sets the max number of buffered packets when probing a codec. Default is 2500 packets.
packetsize integer (output) — Sets the packet size.
fflags flags — Sets format flags. Some of these are only implemented for a limited number of formats, so don't come for me if your format doesn't vibe with them.

Possible values for input files:

discardcorrupt — Discard corrupted packets.
fastseek — Enable fast, but inaccurate seeks for some formats.
genpts — Generate missing PTS if DTS is present.
igndts — Ignore DTS if PTS is also set. In case the PTS is set, the DTS value is set to NOPTS. This is ignored when the nofillin flag is set.
ignidx — Ignore index.
nobuffer — Reduce the latency introduced by buffering during initial input streams analysis.
nofillin — Do not fill in missing values in packet fields that can be exactly calculated.
noparse — Disable AVParsers, this needs +nofillin too.
sortdts — Try to interleave output packets by DTS. At present, available only for AVIs with an index.

Possible values for output files:

autobsf — Automatically apply bitstream filters as required by the output format. Enabled by default.
bitexact — Only write platform-, build- and time-independent data. This ensures that file and data checksums are reproducible and match between platforms. Its primary use is for regression testing.
flush_packets — Write out packets immediately.
shortest — Stop muxing at the end of the shortest stream. It may be needed to increase max_interleave_delta to avoid flushing the longer streams before EOF.

seek2any integer (input) — Allow seeking to non-keyframes on demuxer level when supported if set to 1. Default is 0.
analyzeduration integer (input) — Specify how many microseconds are analyzed to probe the input. Higher value = more accurate info detected, but also more latency. Defaults to 5,000,000 microseconds = 5 seconds.
cryptokey hexadecimal string (input) — Set the decryption key.
indexmem integer (input) — Set max memory used for timestamp index (per stream).
rtbufsize integer (input) — Set max memory used for buffering real-time frames.
fdebug flags (input/output) — Print specific debug info. Possible values:
- ts
max_delay integer (input/output) — Set maximum muxing or demuxing delay in microseconds.
fpsprobesize integer (input) — Set number of frames used to probe fps.
audio_preload integer (output) — Set microseconds by which audio packets should be interleaved earlier.
chunk_duration integer (output) — Set microseconds for each chunk.
chunk_size integer (output) — Set size in bytes for each chunk.
err_detect, f_err_detect flags (input) — Set error detection flags. Heads up: f_err_detect is deprecated and should be used only via the ffmpeg tool. 🙈

Possible values:

crccheck — Verify embedded CRCs.
bitstream — Detect bitstream specification deviations.
buffer — Detect improper bitstream length.
explode — Abort decoding on minor error detection.
careful — Consider things that violate the spec and have not been seen in the wild as errors.
compliant — Consider all spec non compliancies as errors.
aggressive — Consider things that a sane encoder should not do as an error.

max_interleave_delta integer (output) — Set maximum buffering duration for interleaving. Expressed in microseconds, defaults to 10000000 (10 seconds).

Bestie, here's the lore: to make sure all the streams are interleaved correctly, libavformat waits until it has at least one packet for each stream before actually writing any packets to the output file. When some streams are "sparse" (i.e. there are large gaps between successive packets), this can result in excessive buffering. 😭

This field specifies the maximum difference between the timestamps of the first and the last packet in the muxing queue — above that, libavformat will output a packet regardless of whether it has queued a packet for all the streams.

If set to 0, libavformat will continue buffering packets until it has a packet for each stream, no matter the maximum timestamp difference between the buffered packets.

use_wallclock_as_timestamps integer (input) — Use wallclock as timestamps if set to 1. Default is 0.
avoid_negative_ts integer (output) — Possible values:
- make_non_negative — Shift timestamps to make them non-negative. Also note that this affects only leading negative timestamps, and not non-monotonic negative timestamps.
- make_zero — Shift timestamps so that the first timestamp is 0.
- auto (default) — Enables shifting when required by the target format.
- disabled — Disables shifting of timestamp.

When shifting is enabled, all output timestamps get shifted by the same amount. Audio, video, and subtitles desynching and relative timestamp differences are preserved compared to how they would have been without shifting. 🫶

skip_initial_bytes integer (input) — Set number of bytes to skip before reading header and frames if set to 1. Default is 0.
correct_ts_overflow integer (input) — Correct single timestamp overflows if set to 1. Default is 1.
flush_packets integer (output) — Flush the underlying I/O stream after each packet. Default is -1 (auto), which means the underlying protocol decides; 1 enables it and has the effect of reducing latency; 0 disables it and may increase IO throughput in some cases.
output_ts_offset offset (output) — Set the output time offset.

offset must be a time duration specification — see the Time duration section in the ffmpeg-utils(1) manual.

The offset is added by the muxer to the output timestamps.

Specifying a positive offset means the corresponding streams are delayed by the time duration specified in offset. Default value is 0 (meaning no offset is applied).

format_whitelist list (input) — ,-separated list of allowed demuxers. By default all are allowed.
dump_separator string (input) — Separator used to separate the fields printed on the command line about the Stream parameters. For example, to separate the fields with newlines and indentation:

ffprobe -dump_separator "
                          "  -i ~/videos/matrixbench_mpeg2.mpg

max_streams integer (input) — Specifies the maximum number of streams. This can be used to reject files that would require too many resources due to a large number of streams. Iconic gatekeeping behavior honestly 👑
skip_estimate_duration_from_pts bool (input) — Skip estimation of input duration if it requires an additional probing for PTS at end of file. At present, applicable for MPEG-PS and MPEG-TS.
duration_probesize integer (input) — Set probing size, in bytes, for input duration estimation when it actually requires an additional probing for PTS at end of file (at present: MPEG-PS and MPEG-TS). It's aimed at users interested in better durations probing for itself, or indirectly because of using the concat demuxer, for example.

The typical use case is an MPEG-TS CBR with a high bitrate, high video buffering and ending cleaning with similar PTS for video and audio: in such a scenario, the large physical gap between the last video packet and the last audio packet makes it necessary to read many bytes in order to get the video stream duration. Another use case is where the default probing behaviour only reaches a single video frame which is not the last one of the stream due to frame reordering, so the duration is not accurate.

Setting this option has a performance impact even for small files because the probing size is fixed. Default behaviour is a general purpose trade-off, largely adaptive, but the probing size will not be extended to get streams durations at all costs. Must be an integer not lesser than 1, or 0 for default behaviour.

strict, f_strict integer (input/output) — Specify how strictly to follow the standards. f_strict is deprecated and should be used only via the ffmpeg tool. 💀

Possible values:

very — strictly conform to an older more strict version of the spec or reference software
strict — strictly conform to all the things in the spec no matter what consequences
normal
unofficial — allow unofficial extensions
experimental — allow non standardized experimental things, experimental (unfinished/work in progress/not well tested) decoders and encoders.

Note: experimental decoders can pose a security risk — do not use this for decoding untrusted input. Fr fr, be careful out there 🚨

💖 Format stream specifiers 💖

Format stream specifiers let you select one or more streams that match specific properties. So like, choose-your-own-adventure but for streams.

The exact semantics of stream specifiers is defined by the avformat_match_stream_specifier() function declared in the libavformat/avformat.h header and documented in the Stream specifiers section in the ffmpeg(1) manual. Go read that one if you want the full tea ☕

🌈 DEMUXERS 🌈

Demuxers are configured elements in FFmpeg that can read the multimedia streams from a particular type of file. They're the ones doing the reading, the main character energy of the input side, periodt. ✨

Okay so boom — demuxers are on by default when you configure your build. Like, all of them. Every. Single. One. 💅 You can peep the full roster with the configure option --list-demuxers.

Wanna yeet them all and start from scratch? That's --disable-demuxers, then cherry-pick exactly the one you want back in with --enable-demuxer=DEMUXER, or kick an individual one out with --disable-demuxer=DEMUXER. Choose your fighter 🎀

The -demuxers option on the ff* tools shows you the list of enabled demuxers. And if you want the combined demuxers-and-muxers glow-up list, that's -formats.

Below is the tea on some of the demuxers that are currently available. 🍓

aa

Audible Format 2, 3, and 4 demuxer. It's giving audiobook girlie.

This one demuxes Audible Format 2, 3, and 4 (.aa) files.

aac

Raw Audio Data Transport Stream AAC demuxer.

This demuxer is here to demux an ADTS input that's carrying a single AAC stream alongwith any ID3v1/2 or APE tags inside it. Little tags, big vibes. ✨

apng

Animated Portable Network Graphics demuxer.

Used to demux APNG files. All headers — but the PNG signature — up to (but not including) the first fcTL chunk get transmitted as extradata. After that, frames are split as being all the chunks between two fcTL ones, or between the last fcTL and IEND chunks. Weird little system but it ate. 🦋

Options

-ignore_loop bool — Ignores the loop variable in the file if it's set. Default is enabled.
-max_fps int — Maximum framerate in frames per second. A default of 0 imposes no limit at all, go feral.
-default_fps int — Default framerate in frames per second when none is specified in the file (0 meaning as fast as possible). Default is 15.

asf

Advanced Systems Format demuxer.

This bad boy demuxes ASF files and MMS network streams.

-no_resync_search bool — Do not try to resynchronize by looking for a certain optional start code.

concat

Virtual concatenation script demuxer. Lowkey the main character of this section 👑

This demuxer reads a list of files and other directives from a text file, then demuxes them one after the other, as if all their packets had been muxed together. Like a friendship bracelet where every string is a file. 🫶

Timestamps in the files get adjusted so that the first file starts at 0 and each next file starts exactly where the previous one finishes. Heads up though: this is done globally and may cause gaps if all streams do not have exactly the same length.

All files must have the same streams — same codecs, same time base, etc. No mixing and matching, bestie.

The duration of each file is used to adjust the timestamps of the next file. If that duration is incorrect — because it was computed using the bit-rate, or because the file is truncated, for example — it can cause artifacts. The duration directive exists so you can override the duration stored in each file. ✨

Syntax

The script is a text file in extended-ASCII, with one directive per line. Empty lines, leading spaces and lines starting with # are ignored. The following directive is recognized:

file path — Path to a file to read; special characters and spaces must be escaped with backslash or single quotes.

All subsequent file-related directives apply to that file.

ffconcat version 1.0 — Identify the script type and version.

To make FFmpeg recognize the format automatically, this directive must appear exactly as is (no extra space or byte-order-mark) on the very first line of the script.

duration dur — Duration of the file. This information can be specified from the file; specifying it here may be more efficient or help if the information from the file is not available or accurate.

If the duration is set for all files, then it is possible to seek in the whole concatenated video. 👑

inpoint timestamp — In point of the file. When the demuxer opens the file it instantly seeks to the specified timestamp. Seeking is done so that all streams can be presented successfully at In point.

This directive works best with intra frame codecs, because for non-intra frame ones you will usually get extra packets before the actual In point and the decoded content will most likely contain frames before In point too. 😭

For each file, packets before the file In point will have timestamps less than the calculated start timestamp of the file (negative in case of the first file), and the duration of the files (if not specified by the duration directive) will be reduced based on their specified In point.

Because of potential packets before the specified In point, packet timestamps may overlap between two concatenated files.

outpoint timestamp — Out point of the file. When the demuxer reaches the specified decoding timestamp in any of the streams, it handles it as an end of file condition and skips the current and all the remaining packets from all streams.

Out point is exclusive, which means that the demuxer will not output packets with a decoding timestamp greater or equal to Out point.

This directive works best with intra frame codecs and formats where all streams are tightly interleaved. For non-intra frame codecs you will usually get additional packets with presentation timestamp after Out point therefore the decoded content will most likely contain frames after Out point too. If your streams are not tightly interleaved you may not get all the packets from all streams before Out point and you may only will be able to decode the earliest stream until Out point. 💀

The duration of the files (if not specified by the duration directive) will be reduced based on their specified Out point.

file_packet_metadata key=value — Metadata of the packets of the file. The specified metadata will be set for each file packet. You can specify this directive multiple times to add multiple metadata entries. This directive is deprecated, use file_packet_meta instead. 👀
file_packet_meta key value — Metadata of the packets of the file. The specified metadata will be set for each file packet. You can specify this directive multiple times to add multiple metadata entries. (Yes, this is the glow-up version.)
option key value — Option to access, open and probe the file. Can be present multiple times.
stream — Introduce a stream in the virtual file. All subsequent stream-related directives apply to the last introduced stream. Some streams properties must be set in order to allow identifying the matching streams in the subfiles. If no streams are defined in the script, the streams from the first file are copied.
exact_stream_id id — Set the id of the stream. If this directive is given, the string with the corresponding id in the subfiles will be used. This is especially useful for MPEG-PS (VOB) files, where the order of the streams is not reliable. 🔥
stream_meta key value — Metadata for the stream. Can be present multiple times.
stream_codec value — Codec for the stream.
stream_extradata hex_string — Extradata for the string, encoded in hexadecimal.
chapter id start end — Add a chapter. id is an unique identifier, possibly small and consecutive.

Options

This demuxer accepts the following option:

safe — If set to 1, reject unsafe file paths and directives. A file path is considered safe if it does not contain a protocol specification and is relative and all components only contain characters from the portable character set (letters, digits, period, underscore and hyphen) and have no period at the beginning of a component.

If set to 0, any file name is accepted. Live your truth, I guess 🙈

The default is 1.

auto_convert — If set to 1, try to perform automatic conversions on packet data to make the streams concatenable. The default is 1.

Currently, the only conversion is adding the h264_mp4toannexb bitstream filter to H.264 streams in MP4 format. This is necessary in particular if there are resolution changes.

segment_time_metadata — If set to 1, every packet will contain the lavf.concat.start_time and the lavf.concat.duration packet metadata values which are the start_time and the duration of the respective file segments in the concatenated output expressed in microseconds. The duration metadata is only set if it is known based on the concat file. The default is 0.

Examples

Use absolute filenames and include some comments:

# my first filename
file /mnt/share/file-1.wav
# my second filename including whitespace
file '/mnt/share/file 2.wav'
# my third filename including whitespace plus single quote
file '/mnt/share/file 3'\''.wav'

Allow for input format auto-probing, use safe filenames and set the duration of the first file:

ffconcat version 1.0

file file-1.wav
duration 20.0

file subdir/file-2.wav

dash

Dynamic Adaptive Streaming over HTTP demuxer. Streaming queen behavior 🌈

This demuxer presents all AVStreams found in the manifest. By setting the discard flags on AVStreams the caller can decide which streams to actually receive. Each stream mirrors the id and bandwidth properties from the <Representation> as metadata keys named id and variant_bitrate respectively.

Options

This demuxer accepts the following option:

cenc_decryption_key — Default 16-byte key, in hex, to decrypt files encrypted using ISO Common Encryption (CENC/AES-128 CTR; ISO/IEC 23001-7).
cenc_decryption_keys — Dictionary of 16-byte key ID => 16-byte key, both in hex, to decrypt files encrypted using ISO Common Encryption (CENC/AES-128 CTR; ISO/IEC 23001-7).

dvdvideo

DVD-Video demuxer, powered by libdvdnav and libdvdread. Retro tech but make it slay 💿

Can directly ingest DVD titles, specifically sequential PGCs, into a conversion pipeline. Menu assets, such as background video or audio, can also be demuxed given the menu's coordinates (at best effort).

Block devices (DVD drives), ISO files, and directory structures are accepted. Activate with -f dvdvideo in front of one of these inputs.

⚠️ This demuxer does NOT have decryption code of any kind. You are on your own working with encrypted DVDs, and should not expect support on the matter. Fr fr, don't come for me.

Underlying playback is handled by libdvdnav, and structure parsing by libdvdread. FFmpeg must be built with GPL library support available as well as the configure switches --enable-libdvdnav and --enable-libdvdread. 💖

You're gonna need to hand it either the "title number" you want or the exact PGC/PG coordinates — like, pick one, bestie. Lots of open-source DVD players and tools can help you dig that info up. If you don't specify anything, the demuxer just defaults to title 1, which honestly works for a ton of discs. But because this format is so flexible, please check manually — there are lots of discs out there authored weirdly or with straight-up invalid headers. 💀

Oh, and if your input is a real physical DVD drive? Heads up: some drives will silently fail reading bad sectors off the disc and just hand back random bits instead, which is effectively corrupt data. This is especially prominent on aging or rotting discs. You'd need a second pass plus integrity checks to even detect the corruption. And ngl — that is not an FFmpeg issue. 🙈

🎀 Background

DVD-Video is not a directly accessible, linear container format in the traditional sense. Nope! Instead it allows for complex and programmatic playback of carefully muxed MPEG-PS streams that are stored in headerless VOB files. To the end-user these streams are known simply as "titles", but the actual logical playback sequence is defined by one or more "PGCs", or Program Group Chains, within the title. The PGC is in turn comprised of multiple "PGs", or Programs, which are the actual video segments (and for a typical video feature, sequentially ordered). The PGC structure, along with stream layout and metadata, are stored in IFO files that need to be parsed. PGCs can be thought of as playlists, if that's easier — like, that's the vibe. ✨

An actual DVD player relies on user GUI interaction via menus and an internal VM to drive the direction of demuxing. Generally the user would either navigate (via menus) or automatically be redirected to the PGC of their choice. During this process and the subsequent playback, the DVD player's internal VM also maintains a state and executes instructions that can create jumps to different sectors during playback. This is why libdvdnav is involved — because a linear read of the MPEG-PS blobs on the disc (VOBs) is just not enough to produce the right sequence in many cases. 🦋

There are many other DVD structures (a long subject, truly a rabbit hole) that will not be discussed here. NAV packets, in particular, are handled by this demuxer to build accurate timing but are not emitted as a stream. For a good high-level understanding, refer to:

https://code.videolan.org/videolan/libdvdnav/-/blob/master/doc/dvd_structures

💅 Options

This demuxer accepts the following options:

title int — the title number to play. Must be set if pgc and pg are not set. Not applicable to menus. Default is 0 (auto), which currently only selects the first available title (title 1) and notifies the user about the implications.
chapter_start int — the chapter, or PTT (part-of-title), number to start at. Not applicable to menus. Default is 1.
chapter_end int — the chapter, or PTT (part-of-title), number to end at. Not applicable to menus. Default is 0, which is a special value that signals "end at the last possible chapter".
angle int — the video angle number, referring to what is essentially an additional video stream that is composed from alternate frames interleaved in the VOBs. Not applicable to menus. Default is 1.
region int — the region code to use for playback. Some discs may use this to default playback at a particular angle in different regions. This option will not affect the region code of a real DVD drive, if used as an input. Not applicable to menus. Default is 0, "world".
menu bool — demux menu assets instead of navigating a title. Requires exact coordinates of the menu (menu_lu, menu_vts, pgc, pg). Default is false.
menu_lu int — the menu language to demux. In DVD, menus are grouped by language. Default is 1, the first language unit.
menu_vts int — the VTS where the menu lives, or 0 if it is a VMG menu (root-level). Default is 1, menu of the first VTS.
pgc int — the entry PGC to start playback, in conjunction with pg. An alternative to setting title. Chapter markers are not supported at this time. Must be explicitly set for menus. Default is 0, automatically resolve from value of title.
pg int — the entry PG to start playback, in conjunction with pgc. An alternative to setting title. Chapter markers are not supported at this time. Default is 1, the first PG of the PGC.
preindex bool — enable this to have accurate chapter (PTT) markers and duration measurement, which requires a slow second pass read in order to index the chapter marker timestamps from NAV packets. This is non-ideal extra work for real optical drives. It is recommended and faster to use this option with a backup of the DVD structure stored on a hard drive. Not compatible with pgc and pg. Default is 0, false.
trim bool — skip padding cells (i.e. cells shorter than 1 second) from the beginning. There exist many discs with filler segments at the beginning of the PGC, often with junk data intended for controlling a real DVD player's buffering speed and with no other material data value. Not applicable to menus. Default is 1, true.

🌈 Examples

Open title 3 from a given DVD structure:

            ffmpeg -f dvdvideo -title 3 -i <path to DVD> ...

Open chapters 3-6 from title 1 from a given DVD structure:

            ffmpeg -f dvdvideo -chapter_start 3 -chapter_end 6 -title 1 -i <path to DVD> ...

Open only chapter 5 from title 1 from a given DVD structure:

            ffmpeg -f dvdvideo -chapter_start 5 -chapter_end 5 -title 1 -i <path to DVD> ...

Demux menu with language 1 from VTS 1, PGC 1, starting at PG 1:

            ffmpeg -f dvdvideo -menu 1 -menu_lu 1 -menu_vts 1 -pgc 1 -pg 1 -i <path to DVD> ...

🎮 `ea`

Electronic Arts Multimedia format demuxer, bestie.

This format is used by various Electronic Arts games. gamer moment. 🕹️

Options

merge_alpha bool — normally the VP6 alpha channel (if it exists) is returned as a secondary video stream; by setting this option you can make the demuxer return a single video stream which contains the alpha channel in addition to the ordinary video. Two-in-one, iconic behavior tbh.

🎬 `imf`

Interoperable Master Format demuxer.

This demuxer presents audio and video streams found in an IMF Composition, as specified in https://doi.org/10.5594/SMPTE.ST2067-2.2020.

        ffmpeg [-assetmaps <path of ASSETMAP1>,<path of ASSETMAP2>,...] -i <path of CPL> ...

If -assetmaps is not specified, the demuxer looks for a file called ASSETMAP.xml in the same directory as the CPL. It'll find it, don't worry. 💖

📺 `flv`, `live_flv`, `kux`

Adobe Flash Video Format demuxer. Yes, FLV still exists, we're being thorough.

This demuxer is used to demux FLV files and RTMP network streams. In the case of live network streams, if you force the format, you may use live_flv instead of flv to survive timestamp discontinuities. KUX is an flv variant used on the Youku platform.

        ffmpeg -f flv -i myfile.flv ...
        ffmpeg -f live_flv -i rtmp://<any.server>/anything/key ....

-flv_metadata bool — allocate the streams according to the onMetaData array content.
-flv_ignore_prevtag bool — ignore the size of previous tag value.
-flv_full_metadata bool — output all context of the onMetadata.

🧚 `gif`

Animated GIF demuxer. The format we all love to hate, but she's serving. 💅

It accepts the following options:

min_delay — set the minimum valid delay between frames, in hundredths of seconds. Range is 0 to 6000. Default value is 2.
max_gif_delay — set the maximum valid delay between frames, in hundredths of seconds. Range is 0 to 65535. Default value is 65535 (nearly eleven minutes), the maximum value allowed by the specification.
default_delay — set the default delay between frames, in hundredths of seconds. Range is 0 to 6000. Default value is 10.
ignore_loop — GIF files can contain information to loop a certain number of times (or infinitely). If ignore_loop is set to 1, then the loop setting from the input will be ignored and looping will not occur. If set to 0, then looping will occur and will cycle the number of times according to the GIF. Default value is 1.

For example, with the overlay filter, place an infinitely looping GIF over another video:

        ffmpeg -i input.mp4 -ignore_loop 0 -i input.gif -filter_complex overlay=shortest=1 out.mkv

Note that in the above example the shortest option for overlay filter is used to end the output video at the length of the shortest input file, which in this case is input.mp4 as the GIF in this example loops infinitely. Otherwise you'd be there forever, literally. 😭

🍓 `hls`

HLS demuxer. Apple HTTP Live Streaming demuxer, to be precise.

This demuxer presents all AVStreams from all variant streams. The id field is set to the bitrate variant index number. By setting the discard flags on AVStreams (by pressing a or v in ffplay), the caller can decide which variant streams to actually receive. The total bitrate of the variant that the stream belongs to is available in a metadata key named "variant_bitrate". cute, right? ✨

It accepts the following options:

live_start_index — segment index to start live streams at (negative values are from the end).
prefer_x_start — prefer to use #EXT-X-START if it's in playlist instead of live_start_index.
allowed_extensions — ',' separated list of file extensions that hls is allowed to access.
extension_picky — this blocks disallowed extensions from probing. It also requires all available segments to have matching extensions to the format except mpegts, which is always allowed. It is recommended to set the whitelists correctly instead of depending on extensions. Enabled by default.
max_reload — maximum number of times a insufficient list is attempted to be reloaded. Default value is 1000.
m3u8_hold_counters — the maximum number of times to load m3u8 when it refreshes without new segments. Default value is 1000.
http_persistent — use persistent HTTP connections. Applicable only for HTTP streams. Enabled by default.
http_multiple — use multiple HTTP connections for downloading HTTP segments. Enabled by default for HTTP/1.1 servers.
http_seekable — use HTTP partial requests for downloading HTTP segments. 0 = disable, 1 = enable, -1 = auto, Default is auto.
seg_format_options — set options for the demuxer of media segments using a list of key=value pairs separated by ":".
seg_max_retry — maximum number of times to reload a segment on error, useful when segment skip on network error is not desired. Default value is 0.

👑 `image2`

Image file demuxer.

Okay so this demuxer pulls frames from a whole list of image files that you describe with a pattern, and the syntax/meaning of that pattern is decided by the pattern_type option. 💅

The pattern can carry a suffix, and that suffix is what auto-decides the format of the images inside those files.

Heads up bestie: the size, the pixel format, and the format of every image must be identical across all the files in the sequence. No mixing, no cap.

It accepts the following options:

framerate — sets the frame rate for the video stream. Defaults to 25.
loop — if set to 1, it loops over the input. Default value is 0.
pattern_type — picks the pattern type used to interpret the filename you handed over.

pattern_type accepts one of these values:

none — disables pattern matching entirely, so the video will only contain the one image you specified. Use this if you don't wanna build sequences out of multiple images and your filenames might have special pattern characters in them. 🫶
sequence — selects a sequence pattern type, which is what you use to specify a sequence of files indexed by sequential numbers.

A sequence pattern may contain the string %d or %0Nd, which specifies the position of the characters representing a sequential number in each filename matched by the pattern. If the form %d0Nd is used, the string representing the number in each filename is 0-padded and N is the total number of 0-padded digits representing the number. The literal character % can be specified in the pattern with the string %%.

If the sequence pattern contains %d or %0Nd, the first filename of the file list specified by the pattern must contain a number inclusively contained between start_number and start_number+start_number_range-1, and all the following numbers must be sequential. (yes, that's a real requirement, no skipping numbers on me 😭)

For example the pattern img-%03d.bmp will match a sequence of filenames of the form img-001.bmp, img-002.bmp, ..., img-010.bmp, etc.; the pattern i%%m%%g-%d.jpg will match a sequence of filenames of the form i%m%g-1.jpg, i%m%g-2.jpg, ..., i%m%g-10.jpg, etc.

Note that the pattern must not necessarily contain %d or %0Nd, for example to convert a single image file img.jpeg you can employ the command:

ffmpeg -i img.jpeg img.png

glob — selects a glob wildcard pattern type.

The pattern is interpreted like a glob() pattern. This is only selectable if libavformat was compiled with globbing support. (so if it's not working, that's a compile-time thing, not you being dumb 💀)

Default value is sequence.

pixel_format — sets the pixel format of the images to read. If you don't specify it, the pixel format gets guessed from the first image file in the sequence.
start_number — sets the index of the file matched by the image file pattern that reading starts from. Default value is 0.
start_number_range — sets the index interval range to check when hunting for the first image file in the sequence, starting from start_number. Default value is 5.
ts_from_file — if set to 1, it sets the frame timestamp to the modification time of the image file. Note that monotonity of timestamps is not provided: images go in the same order as without this option. Default value is 0. If set to 2, it sets the frame timestamp to the modification time of the image file in nanosecond precision. ⏱️
video_size — sets the video size of the images to read. If not specified, the video size gets guessed from the first image file in the sequence.
export_path_metadata — if set to 1, it adds two extra fields to the metadata found in the input, which also makes them available to other filters (peep the drawtext filter for examples). Default value is 0. The extra fields are:
- lavf.image2dec.source_path — corresponds to the full path to the input file being read.
- lavf.image2dec.source_basename — corresponds to the name of the file being read.

💫 Examples

Use ffmpeg to create a video from the images in the file sequence img-001.jpeg, img-002.jpeg, ..., assuming an input frame rate of 10 frames per second:

ffmpeg -framerate 10 -i 'img-%03d.jpeg' out.mkv

Same as above, but start by reading from a file with index 100 in the sequence:

ffmpeg -framerate 10 -start_number 100 -i 'img-%03d.jpeg' out.mkv

Read images matching the *.png glob pattern, that is all the files terminating with the .png suffix:

ffmpeg -framerate 10 -pattern_type glob -i "*.png" out.mkv

🎮 `libgme`

The Game Music Emu library is a whole collection of video game music file emulators. Iconic honestly.

See <https://bitbucket.org/mpyne/game-music-emu/overview> for more information.

It accepts the following options:

track_index — sets the index of which track to demux. The demuxer can only export one track. Track indexes start at 0. Default is to pick the first track. The number of tracks gets exported as the tracks metadata entry.
sample_rate — sets the sampling rate of the exported track. Range is 1000 to 999999. Default is 44100.
max_size (bytes) — the demuxer buffers the entire file into memory. Adjust this value to set the maximum buffer size, which in turn acts as a ceiling for the size of files that can be read. Default is 50 MiB.

🎛️ `libmodplug`

ModPlug based module demuxer.

See <https://github.com/Konstanty/libmodplug>

It will export one 2-channel 16-bit 44.1 kHz audio stream. Optionally, a pal8 16-color video stream can be exported with or without printed metadata. 🌈

It accepts the following options:

noise_reduction — applies a simple low-pass filter. Can be 1 (on) or 0 (off). Default is 0.
reverb_depth — sets the amount of reverb. Range 0-100. Default is 0.
reverb_delay — sets the delay in ms, clamped to 40-250 ms. Default is 0.
bass_amount — applies bass expansion a.k.a. XBass or megabass. Range is 0 (quiet) to 100 (loud). Default is 0.
bass_range — sets the cutoff, i.e. the upper bound for bass frequencies. Range is 10-100 Hz. Default is 0.
surround_depth — applies a Dolby Pro-Logic surround effect. Range is 0 (quiet) to 100 (heavy). Default is 0.
surround_delay — sets the surround delay in ms, clamped to 5-40 ms. Default is 0.
max_size — the demuxer buffers the entire file into memory. Adjust this value to set the maximum buffer size, which in turn acts as a ceiling for the size of files that can be read. Range is 0 to 100 MiB. 0 removes the buffer size limit (not recommended 🫣). Default is 5 MiB.
video_stream_expr — a string which is evaluated using the eval API to assign colors to the generated video stream. The variables you can use are x, y, w, h, t, speed, tempo, order, pattern and row.
video_stream — generates the video stream. Can be 1 (on) or 0 (off). Default is 0.
video_stream_w — sets the video frame width in 'chars' where one char indicates 8 pixels. Range is 20-512. Default is 30.
video_stream_h — sets the video frame height in 'chars' where one char indicates 8 pixels. Range is 20-512. Default is 30.
video_stream_ptxt — prints metadata on the video stream. Includes speed, tempo, order, pattern, row and ts (time in ms). Can be 1 (on) or 0 (off). Default is 1.

🎼 `libopenmpt`

libopenmpt based module demuxer.

See <https://lib.openmpt.org/libopenmpt/> for more information.

Some files have multiple subsongs (tracks), and that can be set with the subsong option.

It accepts the following options:

subsong — sets the subsong index. This can be either 'all', 'auto', or the index of the subsong. Subsong indexes start at 0. The default is 'auto'.

The default value is to let libopenmpt choose. (letting the library be the main character, we love to see it 👑)

layout — sets the channel layout. Valid values are 1, 2, and 4 channel layouts. The default value is STEREO.
sample_rate — sets the sample rate for libopenmpt to output. Range is from 1000 to INT_MAX. The value default is 48000.

📺 `mcc`

Demuxer for MacCaption MCC files — it supports MCC versions 1.0 and 2.0. MCC files store VANC data, which can include closed captions (EIA-608 and CEA-708), ancillary time code, pan-scan data, etc. By default, for backward compatibility, the MCC demuxer extracts just the EIA-608 and CEA-708 closed captions and returns an EIA_608 stream, ignoring all other VANC data. You can flip it to return all VANC data in a SMPTE_436M_ANC data stream by setting -eia608_extract 0. ✨

💫 Examples

Convert a MCC file to Scenarist (SCC) format:

ffmpeg -i CC.mcc -c:s copy CC.scc

Note that the SCC format only supports EIA-608, so this will discard all other data such as CEA-708 extensions. (rip the extras 😭)

Merge a MCC file into a MXF file:

ffmpeg -i video_and_audio.mxf -eia608_extract 0 -i CC.mcc -c copy -map 0 -map 1 out.mxf

This retains all VANC data and inserts it into the output MXF file as a SMPTE_436M_ANC data stream. Slay.

🎬 `mov`/`mp4`/`3gp`

Demuxer for Quicktime File Format & ISO/IEC Base Media File Format (ISO/IEC 14496-12 or MPEG-4 Part 12, ISO/IEC 15444-12 or JPEG 2000 Part 12).

Registered extensions: mov, mp4, m4a, 3gp, 3g2, mj2, psp, m4b, ism, ismv, isma, f4v

⚙️ Options

This demuxer accepts the following options:

enable_drefs — enables loading of external tracks, disabled by default. Enabling this can theoretically leak information in some use cases. ⚠️ (like, actually think about it first)
use_absolute_path — allows loading of external tracks via absolute paths, disabled by default. Enabling this poses a security risk. It should only be enabled if the source is known to be non-malicious. 🚨
seek_streams_individually — when seeking, identify the closest point in each stream individually and demux packets in that stream from the identified point. This can lead to a different sequence of packets compared to demuxing linearly from the beginning. Default is true.
ignore_editlist — ignore any edit list atoms. The demuxer, by default, modifies the stream index to reflect the timeline described by the edit list. Default is false.
advanced_editlist — modify the stream index to reflect the timeline described by the edit list. ignore_editlist must be set to false for this option to be effective. If both ignore_editlist and this option are set to false, then only the start of the stream index is modified to reflect initial dwell time or starting timestamp described by the edit list. Default is true.

still in the MOV/MP4 demuxer options, bestie — these are the deep-lore file internals 💅

ignore_chapters — don’t parse chapters. This includes GoPro 'HiLight' tags/moments. Note that chapters are only parsed when the input is seekable. Default is false.
use_mfra_for — for seekable fragmented input, set the fragment’s starting timestamp from the media fragment random access box, if present. The options are:
- auto — auto-detect whether to set mfra timestamps as PTS or DTS (default)
- dts — set mfra timestamps as DTS
- pts — set mfra timestamps as PTS
- 0 — don’t use the mfra box to set timestamps
use_tfdt — for fragmented input, set the fragment’s starting timestamp to baseMediaDecodeTime from the tfdt box. Default is enabled, which will prefer to use the tfdt box to set DTS. Disable to use the earliest_presentation_time from the sidx box. In either case, the timestamp from the mfra box will be used if it’s available and use_mfra_for is set to pts or dts.
export_all — export unrecognized boxes within the udta box as metadata entries. The first four characters of the box type are set as the key. Default is false.
export_xmp — export the entire contents of the XMP_ box and uuid box as a string with key xmp. Note that if export_all is set and this option isn’t, the contents of the XMP_ box are still exported but with key XMP_. Default is false.
activation_bytes — the 4-byte key required to decrypt Audible AAX and AAX+ files. See the Audible AAX subsection below.
audible_fixed_key — fixed key used for handling Audible AAX/AAX+ files. It has been pre-set so it should not be necessary to specify it.
decryption_key — default 16-byte key, in hex, to decrypt files encrypted using ISO Common Encryption (CENC/AES-128 CTR; ISO/IEC 23001-7).
decryption_keys — dictionary of 16-byte key ID => 16-byte key, both in hex, to decrypt files encrypted using ISO Common Encryption (CENC/AES-128 CTR; ISO/IEC 23001-7).
max_stts_delta — very high sample deltas written in a trak’s stts box may occasionally be intended, but usually they’re written in error or used to store a negative value for dts correction when treated as signed 32-bit integers. This option lets the user set an upper limit, beyond which the delta is clamped to 1. Values greater than the limit, if negative when cast to int32, are used to adjust onward dts. Unit is the track time scale. Range is 0 to UINT_MAX. Default is UINT_MAX - 48000*10, which allows up to a 10 second dts correction for 48 kHz audio streams while accommodating 99.9% of uint32 range.
interleaved_read — interleave packets from multiple tracks at demuxer level. For badly interleaved files, this prevents playback issues caused by large gaps between packets in different tracks, as MOV/MP4 do not have packet placement requirements. However, this can cause excessive seeking on very badly interleaved files, due to seeking between tracks, so disabling it may prevent I/O issues, at the expense of playback.

🎧 Audible AAX 🎧

Audible AAX files are encrypted M4B files, and they can be decrypted by specifying a 4 byte activation secret. iconic little secret key moment 🔑

ffmpeg -activation_bytes 1CEB00DA -i test.aax -vn -c:a copy output.mp4

📺 `mpegts` 📺

MPEG-2 transport stream demuxer. this one accepts the following options:

resync_size — set size limit for looking up a new synchronization. Default value is 65536.
skip_unknown_pmt — skip PMTs for programs not defined in the PAT. Default value is 0.
fix_teletext_pts — override teletext packet PTS and DTS values with the timestamps calculated from the PCR of the first program which the teletext stream is part of and is not discarded. Default value is 1; set this option to 0 if you want your teletext packet PTS and DTS values untouched.
ts_packetsize — output option carrying the raw packet size in bytes. Show the detected raw packet size, cannot be set by the user.
scan_all_pmts — scan and combine all PMTs. The value is an integer with value from -1 to 1 (-1 means automatic setting, 1 means enabled, 0 means disabled). Default value is -1.
merge_pmt_versions — reuse existing streams when a PMT’s version is updated and elementary streams move to different PIDs. Default value is 0.
max_packet_size — set maximum size, in bytes, of packet emitted by the demuxer. Payloads above this size are split across multiple packets. Range is 184 to INT_MAX/2. Default is 204800 bytes.

🖼️ `mpjpeg` 🖼️

MJPEG encapsulated in multi-part MIME demuxer. this demuxer lets you read MJPEG where each frame is represented as a part of multipart/x-mixed-replace stream.

strict_mime_boundary — the default implementation applies a relaxed standard to multi-part MIME boundary detection, to prevent regression with numerous existing endpoints not generating a proper MIME MJPEG stream. Turning this option on by setting it to 1 will result in a stricter check of the boundary value.

🎥 `rawvideo` 🎥

Raw video demuxer. This demuxer allows one to read raw video data. Since there is no header specifying the assumed video parameters, the user must specify them in order to be able to decode the data correctly. no cap, you gotta tell it everything yourself 😭

This demuxer accepts the following options:

framerate — set input video frame rate. Default value is 25.
pixel_format — set the input video pixel format. Default value is yuv420p.
stride — set frame line size in bytes, only required if there is extra padding. For multiplane formats, stride is a list of line size for each plane.
video_size — set the input video size. This value must be specified explicitly.

For example, to read a rawvideo file input.raw with ffplay, assuming a pixel format of rgb24, a video size of 320x240, and a frame rate of 10 images per second, use the command:

ffplay -f rawvideo -pixel_format rgb24 -video_size 320x240 -framerate 10 input.raw

Read a rawvideo with ffplay, assuming a pixel format of yuv420p, a video size of 1080x1920, 8 bytes of padding in each line of Y plane, and 4 bytes of padding with UV plane,

ffplay -f rawvideo -pixel_format yuv420p -video_size 1080x1920 -stride 1088,544,544 input.raw

💬 `rcwt` 💬

RCWT (Raw Captions With Time) is a format native to ccextractor, a commonly used open source tool for processing 608/708 Closed Captions (CC) sources. For more information on the format, see . the manpage really leaves it at that, bestie 💀

This demuxer implements the specification as of March 2024, which has been stable and unchanged since April 2014.

🌈 Examples 🌈

Render CC to ASS using the built-in decoder:

ffmpeg -i CC.rcwt.bin CC.ass

Note that if your output appears to be empty, you may have to manually set the decoder’s data_field option to pick the desired CC substream.

Convert an RCWT backup to Scenarist (SCC) format:

ffmpeg -i CC.rcwt.bin -c:s copy CC.scc

Note that the SCC format does not support all of the possible CC extensions that can be stored in RCWT (such as EIA-708).

🎵 `sbg` 🎵

SBaGen script demuxer. This demuxer reads the script language used by SBaGen (http://uazu.net/sbagen/) to generate binaural beats sessions. A SBG script looks like that:

-SE
a: 300-2.5/3 440+4.5/0
b: 300-2.5/0 440+4.5/3
off: -
NOW      == a
+0:07:00 == b
+0:14:00 == a
+0:21:00 == b
+0:30:00    off

A SBG script can mix absolute and relative timestamps. If the script uses either only absolute timestamps (including the script start time) or only relative ones, then its layout is fixed, and the conversion is straightforward. On the other hand, if the script mixes both kinds of timestamps, then the NOW reference for relative timestamps will be taken from the current time of day at the time the script is read, and the script layout will be frozen according to that reference. That means that if the script is directly played, the actual times will match the absolute timestamps up to the sound controller’s clock accuracy, but if the user somehow pauses the playback or seeks, all times will be shifted accordingly. clock accuracy tea, periodt ⏰

🗣️ `tedcaptions` 🗣️

JSON captions used for http://www.ted.com/. TED does not provide links to the captions, but they can be guessed from the page. The file tools/bookmarklets.html from the FFmpeg source tree contains a bookmarklet to expose them. bookmarklet magic iykyk ✨

This demuxer accepts the following option:

start_time — set the start time of the TED talk, in milliseconds. The default is 15000 (15s). It is used to sync the captions with the downloadable videos, because they include a 15s intro.

Example: convert the captions to a format most players understand:

ffmpeg -i http://www.ted.com/talks/subtitles/id/1/lang/en talk1-en.srt

🧚 `vapoursynth` 🧚

Vapoursynth wrapper. Due to security concerns, Vapoursynth scripts will not be autodetected, so the input format has to be forced. For ff* CLI tools, add -f vapoursynth before the input -i yourscript.vpy. security said no autodetect, bestie 🔒

This demuxer accepts the following option:

max_script_size — the demuxer buffers the entire script into memory. Adjust this value to set the maximum buffer size, which in turn acts as a ceiling for the size of scripts that can be read. Default is 1 MiB.

🎧 `w64` 🎧

Sony Wave64 Audio demuxer.

This demuxer accepts the following option:

max_size — see the same option for the wav demuxer.

🌊 `wav` 🌊

RIFF Wave Audio demuxer.

This demuxer accepts the following option:

max_size — specify the maximum packet size in bytes for the demuxed packets. By default this is set to 0, which means that a sensible value is chosen based on the input format.

👑 MUXERS 👑

Muxers are configured elements in FFmpeg which allow writing multimedia streams to a particular type of file. the write-it-out girlies 💖

When you configure your FFmpeg build, all the supported muxers are enabled by default. You can list all available muxers using the configure option --list-muxers.

You can disable all the muxers with the configure option --disable-muxers and selectively enable / disable single muxers with the options --enable-muxer=MUXER / --disable-muxer=MUXER.

The option -muxers of the ff* tools will display the list of enabled muxers. Use -formats to view a combined list of enabled demuxers and muxers.

A description of some of the currently available muxers follows. and bestie, there are a lot of them 😭

💿 Raw muxers 💿

okay so this little corner is all about raw muxers. They take a single stream that matches the designated codec, and they do not store timestamps or metadata. The recognized extension is the same as the muxer name unless it says otherwise. ✨

It comprises the following muxers — the media type and any extensions used to auto-select the muxer from the output extensions are shown too. 🫶

ac3 — audio: Dolby Digital, also known as AC-3.
adx — audio: CRI Middleware ADX audio. 🎀 This muxer writes out the total sample count near the start of the first packet when the output is seekable and the count can be stored in 32 bits.
aptx — audio: aptX (Audio Processing Technology for Bluetooth).
aptx_hd — audio (aptxdh): aptX HD (Audio Processing Technology for Bluetooth) audio.
avs2 — video (avs, avs2): AVS2-P2 (Audio Video Standard - Second generation - Part 2) / IEEE 1857.4 video.
avs3 — video (avs3): AVS3-P2 (Audio Video Standard - Third generation - Part 2) / IEEE 1857.10 video.
cavsvideo — video (cavs): Chinese AVS (Audio Video Standard - First generation).
codec2raw — audio: Codec 2 audio. ⚠️ No extension is registered, so the format name has to be supplied, e.g. with the ffmpeg CLI tool -f codec2raw.
data — any: Generic data muxer. This muxer accepts a single stream with any codec of any type. The input stream has to be selected using the -map option with the ffmpeg CLI tool. ⚠️ No extension is registered, so the format name has to be supplied, e.g. with the ffmpeg CLI tool -f data.
dfpwm — audio (dfpwm): Raw DFPWM1a (Dynamic Filter Pulse With Modulation) audio muxer.
dirac — video (drc, vc2): BBC Dirac video. The Dirac Pro codec is a subset and is standardized as SMPTE VC-2.
dnxhd — video (dnxhd, dnxhr): Avid DNxHD video. It is standardized as SMPTE VC-3. Accepts DNxHR streams.
dts — audio: DTS Coherent Acoustics (DCA) audio.
eac3 — audio: Dolby Digital Plus, also known as Enhanced AC-3.
evc — video (evc): MPEG-5 Essential Video Coding (EVC) / EVC / MPEG-5 Part 1 EVC video.
g722 — audio: ITU-T G.722 audio.
g723_1 — audio (tco, rco): ITU-T G.723.1 audio.
g726 — audio: ITU-T G.726 big-endian ("left-justified") audio. ⚠️ No extension is registered, so the format name has to be supplied, e.g. with the ffmpeg CLI tool -f g726.
g726le — audio: ITU-T G.726 little-endian ("right-justified") audio. ⚠️ No extension is registered, so the format name has to be supplied, e.g. with the ffmpeg CLI tool -f g726le.
gsm — audio: Global System for Mobile Communications audio.
h261 — video: ITU-T H.261 video.
h263 — video: ITU-T H.263 / H.263-1996, H.263+ / H.263-1998 / H.263 version 2 video.
h264 — video (h264, 264): ITU-T H.264 / MPEG-4 Part 10 AVC video. Bitstream shall be converted to Annex B syntax if it's in length-prefixed mode.
hevc — video (hevc, h265, 265): ITU-T H.265 / MPEG-H Part 2 HEVC video. Bitstream shall be converted to Annex B syntax if it's in length-prefixed mode.
m4v — video: MPEG-4 Part 2 video.
mjpeg — video (mjpg, mjpeg): Motion JPEG video.
mlp — audio: Meridian Lossless Packing, also known as Packed PCM.
mp2 — audio (mp2, m2a, mpa): MPEG-1 Audio Layer II audio.
mpeg1video — video (mpg, mpeg, m1v): MPEG-1 Part 2 video.
mpeg2video — video (m2v): ITU-T H.262 / MPEG-2 Part 2 video.
obu — video: AV1 low overhead Open Bitstream Units muxer. Temporal delimiter OBUs will be inserted in all temporal units of the stream.
rawvideo — video (yuv, rgb): Raw uncompressed video.
sbc — audio (sbc, msbc): Bluetooth SIG low-complexity subband codec audio.
truehd — audio (thd): Dolby TrueHD audio.
vc1 — video: SMPTE 421M / VC-1 video.

🌟 Examples 🌟

Store raw video frames with the rawvideo muxer using ffmpeg:

ffmpeg -f lavfi -i testsrc -t 10 -s hd1080p testsrc.yuv

Because the rawvideo muxer doesn't store the information related to size and format, that info has to be provided when demuxing the file: 🫶

ffplay -video_size 1920x1080 -pixel_format rgb24 -f rawvideo testsrc.rgb

🎧 Raw PCM muxers 🎧

This section covers raw PCM (Pulse-Code Modulation) audio muxers.

They accept a single stream matching the designated codec. They do not store timestamps or metadata. The recognized extension is the same as the muxer name. ✨

It comprises the following muxers. The optional additional extension used to automatically select the muxer from the output extension is also shown in parentheses.

alaw (al) — PCM A-law.
f32be — PCM 32-bit floating-point big-endian.
f32le — PCM 32-bit floating-point little-endian.
f64be — PCM 64-bit floating-point big-endian.
f64le — PCM 64-bit floating-point little-endian.
mulaw (ul) — PCM mu-law.
s16be — PCM signed 16-bit big-endian.
s16le — PCM signed 16-bit little-endian.
s24be — PCM signed 24-bit big-endian.
s24le — PCM signed 24-bit little-endian.
s32be — PCM signed 32-bit big-endian.
s32le — PCM signed 32-bit little-endian.
s8 (sb) — PCM signed 8-bit.
u16be — PCM unsigned 16-bit big-endian.
u16le — PCM unsigned 16-bit little-endian.
u24be — PCM unsigned 24-bit big-endian.
u24le — PCM unsigned 24-bit little-endian.
u32be — PCM unsigned 32-bit big-endian.
u32le — PCM unsigned 32-bit little-endian.
u8 (ub) — PCM unsigned 8-bit.
vidc — PCM Archimedes VIDC.

📀 MPEG-1/MPEG-2 program stream muxers 📀

This section covers formats belonging to the MPEG-1 and MPEG-2 Systems family.

The MPEG-1 Systems format (also known as ISO/IEEC 11172-1 or MPEG-1 program stream) has been adopted for the format of media track stored in VCD (Video Compact Disc).

The MPEG-2 Systems standard (also known as ISO/IEEC 13818-1) covers two container formats, one known as transport stream and one known as program stream; only the latter is covered here.

The MPEG-2 program stream format (also known as VOB due to the corresponding file extension) is an extension of MPEG-1 program stream: in addition to supporting different codecs for the audio and video streams, it also stores subtitles and navigation metadata. MPEG-2 program stream has been adopted for storing media streams in SVCD and DVD storage devices.

This section comprises the following muxers. 💅

mpeg (mpg,mpeg) — MPEG-1 Systems / MPEG-1 program stream muxer.
vcd — MPEG-1 Systems / MPEG-1 program stream (VCD) muxer. 🎀 This muxer can be used to generate tracks in the format accepted by VCD (Video Compact Disc) storage devices. It is the same as the mpeg muxer with a few differences.
vob — MPEG-2 program stream (VOB) muxer.
dvd — MPEG-2 program stream (DVD VOB) muxer. 🎀 This muxer can be used to generate tracks in the format accepted by DVD (Digital Versatile Disc) storage devices. This is the same as the vob muxer with a few differences.
svcd (vob) — MPEG-2 program stream (SVCD VOB) muxer. 🎀 This muxer can be used to generate tracks in the format accepted by SVCD (Super Video Compact Disc) storage devices. This is the same as the vob muxer with a few differences.

✨ Options ✨

muxrate rate — Set user-defined mux rate expressed as a number of bits/s. If not specified the automatically computed mux rate is employed. Default value is 0.
preload delay — Set initial demux-decode delay in microseconds. Default value is 500000.

🎬 MOV/MPEG-4/ISOMBFF muxers 🎬

This section covers formats belonging to the QuickTime / MOV family, including the MPEG-4 Part 14 format and ISO base media file format (ISOBMFF). These formats share a common structure based on the ISO base media file format (ISOBMFF). ✨

The MOV format was originally developed for use with Apple QuickTime. It was later used as the basis for the MPEG-4 Part 1 (later Part 14) format, also known as ISO/IEC 14496-1. That format was then generalized into ISOBMFF, also named MPEG-4 Part 12 format, ISO/IEC 14496-12, or ISO/IEC 15444-12.

It comprises the following muxers. 🫶

3gp — Third Generation Partnership Project (3GPP) format for 3G UMTS multimedia services.
3g2 — Third Generation Partnership Project 2 (3GP2 or 3GPP2) format for 3G CDMA2000 multimedia services, similar to 3gp with extensions and limitations.
f4v — Adobe Flash Video format.
ipod — MPEG-4 audio file format, as MOV/MP4 but limited to contain only audio streams, typically played with the Apple ipod device.
ismv — Microsoft IIS (Internet Information Services) Smooth Streaming Audio/Video (ISMV or ISMA) format. This is based on MPEG-4 Part 14 format with a few incompatible variants, used to stream media files for the Microsoft IIS server.
mov — QuickTime player format identified by the ".mov" extension.
mp4 — MP4 or MPEG-4 Part 14 format.
psp — PlayStation Portable MP4/MPEG-4 Part 14 format variant. This is based on MPEG-4 Part 14 format with a few incompatible variants, used to play files on PlayStation devices.

🧩 Fragmentation 🧩

The mov, mp4, and ismv muxers support fragmentation. Normally, a MOV/MP4 file has all the metadata about all packets stored in one location.

This data is usually written at the end of the file, but it can be moved to the start for better playback by adding +faststart to the -movflags, or using the qt-faststart tool.

A fragmented file consists of a number of fragments, where packets and metadata about these packets are stored together. Writing a fragmented file has the advantage that the file is decodable even if the writing is interrupted (while a normal MOV/MP4 is undecodable if it is not properly finished), and it requires less memory when writing very long files (since writing normal MOV/MP4 files stores info about every single packet in memory until the file is closed). The downside is that it is less compatible with other applications. 😭

Fragmentation is enabled by setting one of the options that define how to cut the file into fragments:

frag_duration
frag_size
min_frag_duration
movflags +frag_keyframe
movflags +frag_custom

If more than one condition is specified, fragments are cut when one of the specified conditions is fulfilled. The exception to this is the option min_frag_duration, which has to be fulfilled for any of the other conditions to apply.

✨ Options ✨

brand brand_string — overrides the major brand. that's it, that's the whole thing 💅
empty_hdlr_name bool — turn this on and it skips writing the name inside an hdlr box. default is false.
encryption_key key — sets the media encryption key, in hexadecimal format.
encryption_kid kid — sets the media encryption key identifier, also in hexadecimal format.
encryption_scheme scheme — configures the encryption scheme. allowed values are none and cenc-aes-ctr.
frag_duration duration — makes fragments that are duration microseconds long.
frag_interleave number — interleaves samples within fragments (that's the max number of consecutive samples — lower means tighter interleaving, but with more overhead 🫠). it's set to 0 by default.
frag_size size — creates fragments containing up to size bytes of payload data.
iods_audio_profile profile — specifies the iods number for the audio profile atom (from -1 to 255), default is -1.
iods_video_profile profile — same energy but for the video profile atom's iods number (from -1 to 255), default is -1.
ism_lookahead num_entries — specifies the number of lookahead entries for ISM files (from 0 to 255), default is 0.
min_frag_duration duration — will NOT create fragments shorter than duration microseconds long. setting a floor, bestie ✋
moov_size bytes — reserves space for the moov atom at the beginning of the file instead of sticking it at the end. if the space you reserve isn't enough, muxing will fail 😭
mov_gamma gamma — specifies the gamma value for the gama atom (a decimal number from 0 to 10), default is 0.0, and it must be set together with + movflags.
movflags flags — sets the various muxing switches. the following flags can be used:
- cmaf — write CMAF (Common Media Application Format) compatible fragmented MP4 output.
- dash — write DASH (Dynamic Adaptive Streaming over HTTP) compatible fragmented MP4 output.
- default_base_moof — similarly to the omit_tfhd_offset flag, this avoids writing the absolute base_data_offset field in tfhd atoms, but does it by using the new default-base-is-moof flag instead. this flag is new from 14496-12:2012. it may make the fragments easier to parse in certain circumstances (avoiding basing track fragment location calculations on the implicit end of the previous track fragment). lowkey a glow-up for parsing ✨
- delay_moov — delays writing the initial moov until the first fragment is cut, or until the first fragment flush.
- disable_chpl — disables Nero chapter markers (chpl atom). normally both Nero chapters and a QuickTime chapter track get written to the file; with this option set, only the QuickTime chapter track will be written. Nero chapters can cause failures when the file is reprocessed with certain tagging programs, like mp3Tag 2.61a and iTunes 11.3 — most likely other versions are affected as well 💀
- faststart — runs a second pass moving the index (moov atom) to the beginning of the file. this operation can take a while, and will not work in various situations such as fragmented output, so it is not enabled by default.
- frag_custom — lets the caller manually choose when to cut fragments, by calling av_write_frame(ctx, NULL) to write a fragment with the packets written so far. (this is only useful with other applications integrating libavformat, not from ffmpeg.)
- frag_discont — signals that the next fragment is discontinuous from earlier ones.
- frag_every_frame — fragment at every frame.
- frag_keyframe — start a new fragment at each video keyframe.
- global_sidx — write a global sidx index at the start of the file.
- isml — create a live smooth streaming feed (for pushing to a publishing point).
- negative_cts_offsets — enables utilization of version 1 of the CTTS box, in which the CTS offsets can be negative. this lets the initial sample have DTS/CTS of zero, and reduces the need for edit lists for some cases such as video tracks with B-frames. additionally, it eases conformance with the DASH-IF interoperability guidelines. (this option is implicitly set when writing ismv (Smooth Streaming) files 👑)
- omit_tfhd_offset — do not write any absolute base_data_offset in tfhd atoms. this avoids tying fragments to absolute byte positions in the file/streams.
- prefer_icc — if writing a colr atom, prioritise usage of the ICC profile if it exists in the stream packet side data.
- rtphint — add RTP hinting tracks to the output file.
- separate_moof — write a separate moof (movie fragment) atom for each track. normally packets for all tracks are written in one moof atom (which is slightly more efficient), but with this option set, the muxer writes one moof/mdat pair for each track, making it easier to separate tracks 🎀
- skip_sidx — skip writing of the sidx atom. when bitrate overhead due to the sidx atom is high, this option could be used for cases where the sidx atom is not mandatory. when the global_sidx flag is enabled, this option is ignored.
- skip_trailer — skip writing the mfra/tfra/mfro trailer for fragmented files.
- use_metadata_tags — use the mdta atom for metadata.
- write_colr — write the colr atom even if the color info is unspecified. this flag is experimental, may be renamed or changed — do not use it from scripts 🙈
- write_gama — write the deprecated gama atom.
- hybrid_fragmented — for recoverability: write the output file as a fragmented file. this allows the intermediate file to be read while it's being written (in particular, if the writing process is aborted uncleanly). when writing is finished, the file is converted to a regular, non-fragmented file, which is more compatible and allows easier and quicker seeking. if writing is aborted, the intermediate file can manually be remuxed to get a regular, non-fragmented file of what had been written into the unfinished file. literally built-in chaos insurance 🫶
movie_timescale scale — sets the timescale written in the movie header box (mvhd). range is 1 to INT_MAX. default is 1000.
rtpflags flags — adds RTP hinting tracks to the output file. the following flags can be used:
- h264_mode0 — use mode 0 for H.264 in RTP.
- latm — use MP4A-LATM packetization instead of MPEG4-GENERIC for AAC.
- rfc2190 — use RFC 2190 packetization instead of RFC 4629 for H.263.
- send_bye — send RTCP BYE packets when finishing.
- skip_rtcp — do not send RTCP sender reports.
skip_iods bool — skips writing the iods atom (default value is true).
use_editlist bool — uses the edit list (default value is auto).
use_stream_ids_as_track_ids bool — uses stream ids as track ids (default value is false).
video_track_timescale scale — sets the timescale used for video tracks. range is 0 to INT_MAX. if set to 0, the timescale is automatically set based on the native stream time base. default is 0.
write_btrt bool — forces or disables writing the bitrate box inside the stsd box of a track. that box contains the decoding buffer size (in bytes), the maximum bitrate and the average bitrate for the track. the box will be skipped if none of these values can be computed. default is -1 or auto, which will write the box only in MP4 mode.
write_prft option — writes a producer time reference box (PRFT) with a specified time source for the NTP field in the PRFT box. set the value to wallclock to specify the timesource as wallclock time, and pts to specify the timesource as input packets' PTS values.
write_tmcd bool — specify on to force writing a timecode track, off to disable it, and auto to write a timecode track only for mov and mp4 output (default).

setting the value to pts is applicable only for a live encoding use case, where PTS values are set as wallclock time at the source. for example, an encoding use case with a decklink capture source where video_pts and audio_pts are set to abs_wallclock.

✨ Examples ✨

push Smooth Streaming content in real time to a publishing point on IIS with the ismv muxer using ffmpeg:

ffmpeg -re <<normal input/transcoding options>> -movflags isml+frag_keyframe -f ismv http://server/publishingpoint.isml/Streams(Encoder1)

🎮 a64

the A64 Commodore 64 video muxer.

this muxer accepts a single a64_multi or a64_multi5 codec video stream. that's the whole vibe check, retro bestie 🕹️

🔈 ac4

the raw AC-4 audio muxer.

this muxer accepts a single ac4 audio stream.

Options

write_crc bool — when enabled, writes a CRC checksum for each packet to the output. default is false.

📻 adts

the Audio Data Transport Stream muxer.

it accepts a single AAC stream.

Options

write_id3v2 bool — enable to write ID3v2.4 tags at the start of the stream. default is disabled.
write_apetag bool — enable to write APE tags at the end of the stream. default is disabled.
write_mpeg2 bool — enable to set the MPEG version bit in the ADTS frame header to 1, which indicates MPEG-2. default is 0, which indicates MPEG-4.

💾 aea

the MD STUDIO audio muxer.

this muxer accepts a single ATRAC1 audio stream with either one or two channels and a sample rate of 44100Hz.

as AEA supports storing the track title, this muxer will also write the title from the stream's metadata to the container 🎀

🎧 aiff

the Audio Interchange File Format muxer.

Options

write_id3v2 bool — enables ID3v2 tags writing when set to 1. default is 0 (disabled).
id3v2_version bool — selects the ID3v2 version to write. currently only version 3 and 4 (aka. ID3v2.3 and ID3v2.4) are supported. the default is version 4.

🏎️ alp

High Voltage Software's Lego Racers game audio muxer.

it accepts a single ADPCM_IMA_ALP stream with no more than 2 channels and a sample rate not greater than 44100 Hz.

extensions: tun, pcm

Options

type type — sets the file type.

type accepts the following values:

tun — set the file type as music. must have a sample rate of 22050 Hz.
pcm — set the file type as sfx.
auto — set the file type as per the output file extension. .pcm results in type pcm, else type tun is set. (default)

amr 🎧

3GPP AMR (Adaptive Multi-Rate) audio muxer.

It accepts a single audio stream containing an AMR NB stream.

amv 🎀

AMV (Actions Media Video) format muxer.

apm 🍓

Ubisoft Rayman 2 APM audio muxer.

It accepts a single ADPCM IMA APM audio stream.

apng 🦋

Animated Portable Network Graphics muxer.

It accepts a single APNG video stream.

Options

final_delay delay — force a delay expressed in seconds after the last frame of each repetition. Default value is 0.0.
plays repetitions — specify how many times to play the content. 0 causes an infinite loop, and with 1 there is no loop.

Examples

Use ffmpeg to generate an APNG output with 2 repetitions and a delay of half a second after the first repetition:

ffmpeg -i INPUT -final_delay 0.5 -plays 2 out.apng

argo_asf 💜

Argonaut Games ASF audio muxer.

It accepts a single ADPCM audio stream.

Options

version_major version — override the file major version, specified as an integer. Default value is 2.
version_minor version — override the file minor version, specified as an integer. Default value is 1.
name name — embed the file name into the file. If not specified, use the output file name. The name is truncated to 8 characters.

argo_cvg 🌸

Argonaut Games CVG audio muxer.

It accepts a single one-channel ADPCM 22050Hz audio stream.

The loop and reverb options set the corresponding flags in the header, which can be later retrieved to process the audio stream accordingly.

Options

skip_rate_check bool — skip the sample rate check (default is "false").
loop bool — set the loop flag (default is "false").
reverb boolean — set the reverb flag (default is "true").

asf, asf_stream 🎬

Advanced / Active Systems (or Streaming) Format audio muxer.

The asf_stream variant should be selected for streaming.

Note that Windows Media Audio (wma) and Windows Media Video (wmv) use this muxer too. 👀

Options

packet_size size — set the muxer packet size as a number of bytes. By tuning this setting you may reduce data fragmentation or muxer overhead depending on your source. Default value is 3200, minimum is 100, maximum is "64Ki".

ass 💬

ASS/SSA (SubStation Alpha) subtitles muxer.

It accepts a single ASS subtitles stream.

Options

ignore_readorder bool — write dialogue events immediately, even if they are out-of-order. Default is "false"; otherwise they are cached until the expected time event is found.

ast 🌈

AST (Audio Stream) muxer.

This format is used to play audio on some Nintendo Wii games.

It accepts a single audio stream.

The loopstart and loopend options can be used to define a section of the file to loop for players honoring such options.

Options

loopstart start — specify the loop start position expressed in milliseconds, from -1 to INT_MAX. If -1 is set, then no loop is specified (default -1) and the loopend value is ignored.
loopend end — specify the loop end position expressed in milliseconds, from 0 to INT_MAX. Default is 0; if 0 is set, it assumes the total stream duration.

au 🎵

SUN AU audio muxer.

It accepts a single audio stream.

avi 🎞️

Audio Video Interleaved muxer.

AVI is a proprietary format developed by Microsoft, and later formally specified through the Open DML specification.

Because of differences in player implementations, it might be required to set some options to make sure that the generated output can be correctly played by the target player.

Options

flipped_raw_rgb bool — if set to "true", store positive height for raw RGB bitmaps, which indicates the bitmap is stored bottom-up. Note that this option does not flip the bitmap — that has to be done manually beforehand, e.g. by using the vflip filter. Default is "false" and indicates the bitmap is stored top down.
reserve_index_space size — reserve the specified amount of bytes for the OpenDML master index of each stream within the file header. By default, additional master indexes are embedded within the data packets if there is no space left in the first master index and are linked together as a chain of indexes. This index structure can cause problems for some use cases, e.g. third-party software strictly relying on the OpenDML index specification or when file seeking is slow. Reserving enough index space in the file header avoids these problems.

The required index space depends on the output file size and should be about 16 bytes per gigabyte. When this option is omitted or set to zero, the necessary index space is guessed.

Default value is 0.

write_channel_mask bool — write the channel layout mask into the audio stream header.

This option is enabled by default. Disabling the channel mask can be useful in specific scenarios, e.g. when merging multiple audio streams into one for compatibility with software that only supports a single audio stream in AVI (see the "amerge" section in the ffmpeg-filters manual).

avif 🖼️

AV1 (Alliance for Open Media Video codec 1) image format muxer.

This muxer stores images encoded using the AV1 codec.

It accepts one or two video streams. In case two video streams are provided, the second one shall contain a single plane storing the alpha mask.

In case more than one image is provided, the generated output is considered an animated AVIF and the number of loops can be specified with the loop option.

This is based on the specification by Alliance for Open Media at url https://aomediacodec.github.io/av1-avif.

Options

loop count — number of times to loop an animated AVIF. 0 specifies an infinite loop; default is 0.
movie_timescale timescale — set the timescale written in the movie header box ("mvhd"). Range is 1 to INT_MAX. Default is 1000.

avm2 ✨

ShockWave Flash (SWF) / ActionScript Virtual Machine 2 (AVM2) format muxer.

It accepts one audio stream, one video stream, or both.

bit 🧚

G.729 (.bit) file format muxer.

It accepts a single G.729 audio stream.

caf ☕

Apple CAF (Core Audio Format) muxer.

It accepts a single audio stream.

codec2 📻

Codec2 audio audio muxer.

It accepts a single codec2 audio stream.

chromaprint 🎶

Chromaprint fingerprinter muxers.

To enable compilation of this filter you need to configure FFmpeg with --enable-chromaprint.

This muxer feeds audio data to the Chromaprint library, which generates a fingerprint for the provided audio data. See: https://acoustid.org/chromaprint.

It takes a single signed native-endian 16-bit raw audio stream of at most 2 channels.

Options

algorithm version — select the version of the algorithm to fingerprint with. Range is 0 to 4. Version 3 enables silence detection. Default is 1.
fp_format format — format to output the fingerprint as. Accepts the following options:
- base64 — Base64 compressed fingerprint (default)
- compressed — Binary compressed fingerprint
- raw — Binary raw fingerprint
silence_threshold threshold — threshold for detecting silence. Range is from -1 to 32767, where -1 disables silence detection. Silence detection can only be used with version 3 of the algorithm.

Silence detection must be disabled for use with the AcoustID service. Default is -1.

crc 🔍

CRC (Cyclic Redundancy Check) muxer.

This muxer computes and prints the Adler-32 CRC of all the input audio and video frames. By default audio frames are converted to signed 16-bit raw audio and video frames to raw video before computing the CRC.

The output of the muxer consists of a single line of the form: CRC=0xCRC, where CRC is a hexadecimal number 0-padded to 8 digits containing the CRC for all the decoded input frames.

dash 🌐

Dynamic Adaptive Streaming over HTTP (DASH) muxer.

This muxer creates segments and manifest files according to the MPEG-DASH standard ISO/IEC 23009-1:2014 and following standard updates.

For more information see:

ISO DASH Specification: http://standards.iso.org/ittf/PubliclyAvailableStandards/c065274_ISO_IEC_23009-1_2014.zip
WebM DASH Specification: https://sites.google.com/a/webmproject.org/wiki/adaptive-streaming/webm-dash-specification

This muxer creates an MPD (Media Presentation Description) manifest file and segment files for each stream. Segment files are placed in the same directory of the MPD manifest file.

The segment filename might contain pre-defined identifiers used in the manifest "SegmentTemplate" section as defined in section 5.3.9.4.4 of the standard.

Available identifiers are "$RepresentationID$", "$Number$", "$Bandwidth$", and "$Time$". In addition to the standard identifiers, an ffmpeg-specific "$ext$" identifier is also supported. When specified, ffmpeg will replace "$ext$" in the file name with muxing format's extensions such as "mp4", "webm" etc.

Options

adaptation_sets adaptation_sets — assign streams to adaptation sets, specified in the MPD manifest "AdaptationSets" section.

An adaptation set contains a set of one or more streams accessed as a single subset, e.g. corresponding streams encoded at different sizes selectable by the user depending on the available bandwidth, or to different audio streams with a different language.

Each adaptation set is specified with the syntax:

id=<index>,streams=<streams>

where index must be a numerical index, and streams is a sequence of ","-separated stream indices. Multiple adaptation sets can be specified, separated by spaces.

To map all video (or audio) streams to an adaptation set, "v" (or "a") can be used as stream identifier instead of IDs.

When no assignment is defined, this defaults to an adaptation set for each stream.

The following optional fields can also be specified:

Okay so, inside adaptation_sets you can also drop these extra keys — they’re giving fine-grained control:

descriptor — defines the descriptor exactly as ISO/IEC 23009-1:2014/Amd.2:2015 says. And bestie, the descriptor string should be a self-closing XML tag.

For example:

<SupplementalProperty schemeIdUri=\"urn:mpeg:dash:srd:2014\" value=\"0,0,0,1,1,2,2\"/>

frag_duration — overrides the global fragment duration that was set with the frag_duration option.
frag_type — overrides the global fragment type that was set with the frag_type option.
seg_duration — overrides the global segment duration that was set with the seg_duration option.
trick_id — marks an adaptation set as containing streams meant to be used for Trick Mode for the referenced adaptation set.

A few examples of possible values for the adaptation_sets option follow:

id=0,seg_duration=2,frag_duration=1,frag_type=duration,streams=v id=1,seg_duration=2,frag_type=none,streams=a

id=0,seg_duration=2,frag_type=none,streams=0 id=1,seg_duration=10,frag_type=none,trick_id=0,streams=1

✨ DASH muxer options, continued ✨

dash_segment_type type — sets the DASH segment files type. Possible values:
- auto — the DASH segment files format will be selected based on the stream codec. This is the default mode.
- mp4 — the DASH segment files will be in ISOBMFF/MP4 format.
- webm — the DASH segment files will be in WebM format.
extra_window_size size — sets the maximum number of segments kept outside of the manifest before they get removed from disk.
format_options options_list — sets container format (mp4/webm) options using a ":"-separated list of key=value parameters. Values containing ":" special characters must be escaped.
frag_duration duration — sets the length in seconds of fragments within segments. Fractional values can also be set.
frag_type type — sets the type of interval for fragmentation. Possible values:
- auto — set one fragment per segment.
- every_frame — fragment at every frame.
- duration — fragment at specific time intervals.
- pframes — fragment at keyframes and following P-Frame reordering (Video only, experimental).
global_sidx bool — writes global "SIDX" atom. Applicable only for single file, mp4 output, non-streaming mode.
hls_master_name file_name — HLS master playlist name. Default is master.m3u8.
hls_playlist bool — generates HLS playlist files. The master playlist is generated with the filename specified by the hls_master_name option. One media playlist file is generated for each stream with filenames media_0.m3u8, media_1.m3u8, etc.
http_opts http_opts — specifies a list of ":"-separated key=value options to pass to the underlying HTTP protocol. Applicable only for HTTP output.
http_persistent bool — uses persistent HTTP connections. Applicable only for HTTP output.
http_user_agent user_agent — overrides the User-Agent field in the HTTP header. Applicable only for HTTP output.
ignore_io_errors bool — ignores IO errors during open and write. Useful for long-duration runs with network output. This is disabled by default.
index_correction bool — enables or disables segment index correction logic. Applicable only when use_template is enabled and use_timeline is disabled. This is disabled by default.

When enabled, the logic monitors the flow of segment indexes. If a stream’s segment index value is not at the expected real time position, then the logic corrects that index value.

Typically this logic is needed in live streaming use cases. Network bandwidth fluctuations are common during long-run streaming. Each fluctuation can cause the segment indexes to fall behind the expected real time position.

init_seg_name init_name — DASH-templated name to use for the initialization segment. Default is "init-stream$RepresentationID$.$ext$". "$ext$" is replaced with the file name extension specific for the segment format.
ldash bool — enables Low-latency Dash by constraining the presence and values of some elements. This is disabled by default.
lhls bool — enables Low-latency HLS (LHLS). Adds "#EXT-X-PREFETCH" tag with the current segment’s URI. The hls.js player folks are trying to standardize an open LHLS spec. The draft spec is available at https://github.com/video-dev/hlsjs-rfcs/blob/lhls-spec/proposals/0001-lhls.md.

This option tries to comply with the above open spec. It enables streaming and hls_playlist options automatically. This is an experimental feature.

Note: This is not Apple’s version of LHLS. See https://datatracker.ietf.org/doc/html/draft-pantos-hls-rfc8216bis.

master_m3u8_publish_rate segment_intervals_count — publishes the master playlist repeatedly every after the specified number of segment intervals.
max_playback_rate rate — sets the maximum playback rate indicated as appropriate for the purposes of automatically adjusting playback latency and buffer occupancy during normal playback by clients.
media_seg_name segment_name — DASH-templated name to use for the media segments. Default is "chunk-stream$RepresentationID$-$Number%05d$.$ext$". "$ext$" is replaced with the file name extension specific for the segment format.
method method — uses the given HTTP method to create output files. Generally set to "PUT" or "POST".
min_playback_rate rate — sets the minimum playback rate indicated as appropriate for the purposes of automatically adjusting playback latency and buffer occupancy during normal playback by clients.
mpd_profile flags — sets one or more MPD manifest profiles. Possible values:
- dash — MPEG-DASH ISO Base media file format live profile.
- dvb_dash — DVB-DASH profile.

Default value is "dash".

remove_at_exit bool — enables or disables removal of all segments when finished. This is disabled by default.
seg_duration duration — sets the segment length in seconds (fractional value can be set). The value is treated as average segment duration when the use_template option is enabled and the use_timeline option is disabled, and as minimum segment duration for all the other use cases.

Default value is 5.

single_file bool — enables or disables storing all segments in one file, accessed using byte ranges. This is disabled by default.

The name of the single file can be specified with the single_file_name option. If not specified, assume the basename of the manifest file with the output format extension.

single_file_name file_name — DASH-templated name to use for the manifest "baseURL" element. Imply that the single_file option is set to true. In the template, "$ext$" is replaced with the file name extension specific for the segment format.
streaming bool — enables or disables chunk streaming mode of output. In chunk streaming mode, each frame will be a "moof" fragment which forms a chunk. This is disabled by default.
target_latency target_latency — sets an intended target latency in seconds for serving (fractional value can be set). Applicable only when the streaming and write_prft options are enabled. This is an informative field clients can use to measure the latency of the service.
timeout timeout — sets timeout for socket I/O operations expressed in seconds (fractional value can be set). Applicable only for HTTP output.
update_period period — sets the MPD update period, for dynamic content. The unit is second. If set to 0, the period is automatically computed.

Default value is 0.

use_template bool — enables or disables use of "SegmentTemplate" instead of "SegmentList" in the manifest. This is enabled by default.
use_timeline bool — enables or disables use of "SegmentTimeline" within the "SegmentTemplate" manifest section. This is enabled by default.
utc_timing_url url — URL of the page that will return the UTC timestamp in ISO format, for example "https://time.akamai.com/?iso".
window_size size — sets the maximum number of segments kept in the manifest, discarding the oldest one. This is useful for live streaming.

If the value is 0, all segments are kept in the manifest. Default value is 0.

write_prft write_prft — writes Producer Reference Time elements on supported streams. This also enables writing prft boxes in the underlying muxer. Applicable only when the utc_url option is enabled. It is set to auto by default, in which case the muxer will attempt to enable it only in modes that require it.

🌸 Example 🌸

Generate a DASH output reading from an input source in realtime using ffmpeg.

Two multimedia streams are generated from the input file, both containing a video stream encoded through libx264, and an audio stream encoded with libfdk_aac. The first multimedia stream contains video with a bitrate of 800k and audio at the default rate, the second with video scaled to 320x170 pixels at 300k and audio resampled at 22005 Hz.

The window_size option keeps only the latest 5 segments with the default duration of 5 seconds.

ffmpeg -re -i <input> -map 0 -map 0 -c:a libfdk_aac -c:v libx264 \
-b:v:0 800k -profile:v:0 main \
-b:v:1 300k -s:v:1 320x170 -profile:v:1 baseline -ar:a:1 22050 \
-bf 1 -keyint_min 120 -g 120 -sc_threshold 0 -b_strategy 0 \
-use_timeline 1 -use_template 1 -window_size 5 \
-adaptation_sets "id=0,streams=v id=1,streams=a" \
-f dash /path/to/out.mpd

🎬 `daud`

D-Cinema audio muxer.

It accepts a single 6-channels audio stream resampled at 96000 Hz encoded with the pcm_24daud codec.

Example

Use ffmpeg to mux input audio to a 5.1 channel layout resampled at 96000Hz:

ffmpeg -i INPUT -af aresample=96000,pan=5.1 slow.302

For ffmpeg versions before 7.0 you might have to use the asetnsamples filter to limit the muxed packet size, because this format does not support muxing packets larger than 65535 bytes (3640 samples). For newer ffmpeg versions audio is automatically packetized to 36000 byte (2000 sample) packets.

🎬 `dv`

DV (Digital Video) muxer.

It accepts exactly one dvvideo video stream and at most two pcm_s16 audio streams. More constraints are defined by the property of the video, which must correspond to a DV video supported profile, and on the framerate.

Example

Use ffmpeg to convert the input:

ffmpeg -i INPUT -s:v 720x480 -pix_fmt yuv411p -r 29.97 -ac 2 -ar 48000 -y out.dv

🎀 ffmetadata

The FFmpeg metadata muxer, bestie. It writes the streams' metadata out in the ffmetadata format.

Wanna know all about that format? Check the Metadata chapter, it's right there. ✨

Example

Say you wanna pull the metadata out of an input file into a metadata.ffmeta file in ffmetadata format — do this:

ffmpeg -i INPUT -f ffmetadata metadata.ffmeta

🧵 fifo

The FIFO (First-In First-Out) muxer, and honestly it's giving main character energy.

What the fifo pseudo-muxer does is let you separate encoding from muxing — it uses a first-in-first-out queue and runs the actual muxer in a separate thread. 🦋

It's especially iconic when you pair it with the tee muxer, and it lets you send data to several destinations that each have different reliability / writing speed / latency. No cap.

So which target muxer does it use? Either it picks one from the output name, or you tell it directly with the fifo_format option.

Now, what happens when the queue fills up or the output just... fails (like when a packet can't be written to the output)? That's selectable, and you've got options:

Output can be transparently restarted, with a configurable delay between retries, based on real time or on the time of the processed stream.
Encoding can be blocked during a temporary failure, or it can keep going and transparently drop packets when the FIFO queue fills up.

⚠️ API users, listen up: the callback functions ("interrupt_callback", "io_open" and "io_close") used inside its "AVFormatContext" must be thread-safe. Not optional, not a vibe, just facts. 💀

Options

attempt_recovery bool — if a failure happens, try to recover the output. This is especially useful with network output, because it lets you restart streaming transparently. Default is "false".
drop_pkts_on_overflow bool — if it's set to "true" and the fifo queue fills up, packets get dropped instead of blocking the encoder. That means you can keep streaming without delaying the input — at the cost of losing part of the stream. Default is "false", so normally the encoder gets blocked until the muxer processes some packets and nothing is lost. 🫶
fifo_format format_name — set the format name. Handy when it can't be guessed from the output name suffix.
format_opts options — set format options for the underlying muxer. Muxer options go in as a list of key=value pairs separated by ':'.
max_recovery_attempts count — cap the number of successive unsuccessful recovery attempts, after which the output fails permanently. Default is 0, which means unlimited.
queue_size size — the size of the queue, as a number of packets. Default is 60.
recover_any_error bool — if "true", recovery is attempted no matter what type of error caused the failure. Default is "false", and in that case certain (usually permanent) errors won't trigger recovery even when attempt_recovery is "true".
recovery_wait_streamtime bool — if "false", real time is what's used when waiting for the recovery attempt (i.e. recovery gets attempted after the time set by recovery_wait_time). If "true", the time of the processed stream is taken into account instead (i.e. recovery is attempted after discarding the packets corresponding to recovery_wait_time). Default is "false".
recovery_wait_time duration — how long to wait, in seconds, before the next recovery attempt after an unsuccessful one. Default is 5.
restart_with_keyframe bool — whether to wait for a keyframe after recovering from a queue overflow or failure. Default is "false".
timeshift duration — buffer that many packets and delay writing the output. Heads up: your queue_size value has to be big enough to store the packets for the timeshift. At the end of the input, the fifo buffer is flushed at realtime speed. ⏳

Example

Okay so here's the flex: stream to an RTMP server with ffmpeg, keep processing the stream at real-time rate even if there's a temporary failure (network outage moment 😭), and try to recover the stream every second, indefinitely:

ffmpeg -re -i ... -c:v libx264 -c:a aac -f fifo -fifo_format flv \
  -drop_pkts_on_overflow 1 -attempt_recovery 1 -recovery_wait_time 1 \
  -map 0:v -map 0:a rtmp://example.com/live/stream_name

🎮 film_cpk

The Sega film (.cpk) muxer. This format was used as the internal format for several Sega games — retro girlies rise 🕹️✨

Want the lore on the Sega film file format? Go here: http://wiki.multimedia.cx/index.php?title=Sega_FILM.

It accepts at maximum one cinepak or raw video stream, and at maximum one audio stream. That's the limit, don't push it.

🎞️ filmstrip

The Adobe Filmstrip muxer. Several Adobe tools use this format to store a generated filmstrip export. It accepts a single raw video stream.

🔭 fits

The Flexible Image Transport System (FITS) muxer — yes, that's the astrology-girlie of image formats, it's literally used to store astronomical data. 🌌⭐

More info? Here: https://fits.gsfc.nasa.gov.

🍓 flac

The Raw FLAC audio muxer. It accepts exactly one FLAC audio stream. And as a lil bonus, you can add images with the attached_pic disposition. 🖼️

Options

write_header bool — write the file header if set to "true". Default is "true".

Example

Use ffmpeg to store the audio stream from an input file, plus a couple of pictures on attached_pic disposition:

ffmpeg -i INPUT -i pic1.png -i pic2.jpg -map 0:a -map 1 -map 2 -disposition:v attached_pic OUTPUT

🌈 flv

The Adobe Flash Video Format muxer. Flash era representation, honestly.

Options

flvflags flags — the possible values are:
- aac_seq_header_detect — place the AAC sequence header based on the audio stream data.
- no_sequence_end — disable the sequence end tag.
- no_metadata — disable the metadata tag.
- no_duration_filesize — disable duration and filesize in the metadata when they're equal to zero at the end of the stream. (Used for a non-seekable living stream.)
- add_keyframe_index — used to facilitate seeking; especially for HTTP pseudo streaming.

🔍 framecrc

The per-packet CRC (Cyclic Redundancy Check) testing format.

This muxer computes and prints an Adler-32 CRC for every audio and video packet. By default, audio frames get converted to signed 16-bit raw audio and video frames get converted to raw video before the CRC is computed.

The output is one line per audio and video packet, like this:

<stream_index>, <packet_dts>, <packet_pts>, <packet_duration>, <packet_size>, 0x<CRC>

CRC is a hexadecimal number, 0-padded out to 8 digits, holding the CRC of the packet. 💅

Examples

Say you wanna compute the CRC of the audio and video frames in INPUT, converted to raw audio and video packets, and stash it in out.crc:

ffmpeg -i INPUT -f framecrc out.crc

Wanna print it to stdout instead? Run:

ffmpeg -i INPUT -f framecrc -

And with ffmpeg you can pick which output format the audio and video frames get encoded to before the per-packet CRC is computed, just by specifying the audio and video codec. For example, to compute the CRC of each decoded input audio frame converted to PCM unsigned 8-bit, and of each decoded input video frame converted to MPEG-2 video, use:

ffmpeg -i INPUT -c:a pcm_u8 -c:v mpeg2video -f framecrc -

Also go peek at the crc muxer. 👀

#️⃣ framehash

The per-packet hash testing format.

This muxer computes and prints a cryptographic hash for every audio and video packet. Which means you can do packet-by-packet equality checks without having to individually do a binary comparison on each one. Obsessed with the efficiency honestly.

By default, audio frames get converted to signed 16-bit raw audio and video frames to raw video before the hash is computed — but you can also use the output of explicit conversions to other codecs. It uses the SHA-256 cryptographic hash function by default, but supports several other algorithms.

Output is one line for each audio and video packet, like this:

<stream_index>, <packet_dts>, <packet_pts>, <packet_duration>, <packet_size>, <hash>

hash is a hexadecimal number representing the computed hash for the packet.

hash algorithm — use the cryptographic hash function named by the string algorithm. Supported values include "MD5", "murmur3", "RIPEMD128", "RIPEMD160", "RIPEMD256", "RIPEMD320", "SHA160", "SHA224", "SHA256" (default), "SHA512/224", "SHA512/256", "SHA384", "SHA512", "CRC32" and "adler32".

Examples

To compute the SHA-256 hash of the audio and video frames in INPUT, converted to raw audio and video packets, and store it in the file out.sha256:

ffmpeg -i INPUT -f framehash out.sha256

To print the info to stdout using the MD5 hash function:

ffmpeg -i INPUT -f framehash -hash md5 -

Also check out the hash muxer. ✨

💜 framemd5

The per-packet MD5 testing format.

It's a variant of the framehash muxer — except unlike that one, it defaults to the MD5 hash function.

Examples

To compute the MD5 hash of the audio and video frames in INPUT, converted to raw audio and video packets, and store it in the file out.md5:

ffmpeg -i INPUT -f framemd5 out.md5

To print the info to stdout:

ffmpeg -i INPUT -f framemd5 -

Also see the framehash and md5 muxers.

🦋 gif

The Animated GIF muxer. The internet's love language, fr.

⚠️ Heads up: the GIF format has a very large time base — the delay between two frames can't be smaller than one centi second. That's just the format's whole deal.

Options

loop bool — how many times to loop the output. Use -1 for no loop, 0 for looping indefinitely (default). 🔁
final_delay delay — force the delay (expressed in centiseconds) after the last frame. Every frame ends with a delay until the next frame. Default is -1, a special value that tells the muxer to reuse the previous delay. If you're looping, you might wanna customize this value to mark a pause, for instance.

Example

Encode a gif that loops 10 times, with a 5 second delay between the loops:

ffmpeg -i INPUT -loop 10 -final_delay 500 out.gif

Note 1: if you wanna extract the frames into separate GIF files, you need to force the image2 muxer: ffmpeg -i INPUT -c:v gif -f image2 "out%d.gif"

📼 gxf

The General eXchange Format (GXF) muxer.

It accepts at most one video stream with codec mjpeg, or mpeg1video, or mpeg2video, or dvvideo with resolution 512x480 or 608x576, and several audio streams with rate 48000Hz and codec pcm16_le.

hash ✨ the vibe-checker ✨

Hash testing format. No cap, this one is so underrated.

This muxer computes and prints a cryptographic hash of all the input audio and video frames. It's giving "do these two files actually match" energy — used for equality checks without having to do a whole entire binary comparison. Like, why compare every byte when you can just compare the vibe (the hash)? 💅

By default, audio frames get converted to signed 16-bit raw audio and video frames to raw video before computing the hash, but you can also use the output of explicit conversions to other codecs. Timestamps are, respectfully, ignored. 🫶 It uses the SHA-256 cryptographic hash function by default, but supports several other algorithms, iykyk.

The output of the muxer consists of a single line of the form: algo=hash, where algo is a short string representing the hash function used, and hash is a hexadecimal number representing the computed hash.

hash algorithm

Use the cryptographic hash function specified by the string algorithm. Supported values include "MD5", "murmur3", "RIPEMD128", "RIPEMD160", "RIPEMD256", "RIPEMD320", "SHA160", "SHA224", "SHA256" (default), "SHA512/224", "SHA512/256", "SHA384", "SHA512", "CRC32" and "adler32". That's a whole roster, bestie — pick your fighter. 🦋

Examples

To compute the SHA-256 hash of the input converted to raw audio and video, and store it in the file out.sha256:

ffmpeg -i INPUT -f hash out.sha256

To print an MD5 hash to stdout use the command:

ffmpeg -i INPUT -f hash -hash md5 -

hds 🎬 (Adobe's little adaptive-streaming moment)

HTTP Dynamic Streaming (HDS) muxer.

HTTP dynamic streaming, or HDS, is an adaptive bitrate streaming method developed by Adobe. HDS delivers MP4 video content over HTTP connections. HDS can be used for on-demand streaming or live streaming.

This muxer creates an .f4m (Adobe Flash Media Manifest File) manifest, an .abst (Adobe Bootstrap File) for each stream, and segment files in a directory specified as the output.

⚠️ These needs to be accessed by an HDS player through HTTPS for it to be able to perform playback on the generated stream. HTTPS only, bestie — the player will simply not come to the party otherwise.

Options

extra_window_size int — number of fragments kept outside of the manifest before removing from disk
min_frag_duration microseconds — minimum fragment duration (in microseconds), default value is 1 second (10000000)
remove_at_exit bool — remove all fragments when finished when set to "true"
window_size int — number of fragments kept in the manifest, if set to a value different from 0. By default all segments are kept in the output directory.

Example

Use ffmpeg to generate HDS files to the output.hds directory in real-time rate:

ffmpeg -re -i INPUT -f hds -b:v 200k output.hds

hls 👑 the main character muxer 👑

Apple HTTP Live Streaming muxer that segments MPEG-TS according to the HTTP Live Streaming (HLS) specification. She's literally iconic, everyone knows her.

It creates a playlist file, and one or more segment files. The output filename specifies the playlist filename.

By default, the muxer creates a file for each segment produced. These files have the same name as the playlist, followed by a sequential number and a .ts extension.

💀 Make sure to require a closed GOP when encoding and to set the GOP size to fit your segment time constraint. Do not skip this, fr fr — the segments will get feral otherwise.

For example, to convert an input file with ffmpeg:

ffmpeg -i in.mkv -c:v h264 -flags +cgop -g 30 -hls_time 1 out.m3u8

This example will produce the playlist, out.m3u8, and segment files: out0.ts, out1.ts, out2.ts, etc.

See also the segment muxer, which provides a more generic and flexible implementation of a segmenter, and can be used to perform HLS segmentation.

Options

hls_init_time duration Set the initial target segment length. Default value is 0.

duration must be a time duration specification, see the Time duration section in the ffmpeg-utils(1) manual.

Segment will be cut on the next key frame after this time has passed on the first m3u8 list. After the initial playlist is filled, ffmpeg will cut segments at duration equal to hls_time.

hls_time duration Set the target segment length. Default value is 2.

duration must be a time duration specification, see the Time duration section in the ffmpeg-utils(1) manual. Segment will be cut on the next key frame after this time has passed.

hls_list_size size Set the maximum number of playlist entries. If set to 0 the list file will contain all the segments. Default value is 5.

hls_delete_threshold size Set the number of unreferenced segments to keep on disk before "hls_flags delete_segments" deletes them. Increase this to allow continue clients to download segments which were recently referenced in the playlist. Default value is 1, meaning segments older than hls_list_size+1 will be deleted.

hls_start_number_source source Start the playlist sequence number ("#EXT-X-MEDIA-SEQUENCE") according to the specified source. Unless hls_flags single_file is set, it also specifies source of starting sequence numbers of segment and subtitle filenames. In any case, if hls_flags append_list is set and read playlist sequence number is greater than the specified start sequence number, then that value will be used as start value.

It accepts the following values:

generic (default) — Set the start numbers according to the start_number option value.
epoch — Set the start number as the seconds since epoch (1970-01-01 00:00:00).
epoch_us — Set the start number as the microseconds since epoch (1970-01-01 00:00:00).
datetime — Set the start number based on the current date/time as YYYYmmddHHMMSS. e.g. 20161231235759.

start_number number Start the playlist sequence number ("#EXT-X-MEDIA-SEQUENCE") from the specified number when hls_start_number_source value is generic. (This is the default case.) Unless hls_flags single_file is set, it also specifies starting sequence numbers of segment and subtitle filenames. Default value is 0.

hls_allow_cache bool Explicitly set whether the client MAY (1) or MUST NOT (0) cache media segments.

hls_base_url baseurl Append baseurl to every entry in the playlist. Useful to generate playlists with absolute paths.

🎀 Note that the playlist sequence number must be unique for each segment and it is not to be confused with the segment filename sequence number which can be cyclic, for example if the wrap option is specified. Two different numbers, bestie — do not mix them up.

hls_segment_filename filename Set the segment filename. Unless the hls_flags option is set with single_file, filename is used as a string format with the segment number appended.

For example:

ffmpeg -i in.nut -hls_segment_filename 'file%03d.ts' out.m3u8

will produce the playlist, out.m3u8, and segment files: file000.ts, file001.ts, file002.ts, etc.

filename may contain a full path or relative path specification, but only the file name part without any path will be contained in the m3u8 segment list. Should a relative path be specified, the path of the created segment files will be relative to the current working directory. When strftime_mkdir is set, the whole expanded value of filename will be written into the m3u8 segment list.

When var_stream_map is set with two or more variant streams, the filename pattern must contain the string "%v", and this string will be expanded to the position of variant stream index in the generated segment file names. Must contain "%v" — not optional, she'll refuse to work without it.

For example:

ffmpeg -i in.ts -b:v:0 1000k -b:v:1 256k -b:a:0 64k -b:a:1 32k \
  -map 0:v -map 0:a -map 0:v -map 0:a -f hls -var_stream_map "v:0,a:0 v:1,a:1" \
  -hls_segment_filename 'file_%v_%03d.ts' out_%v.m3u8

will produce the playlists segment file sets: file_0_000.ts, file_0_001.ts, file_0_002.ts, etc. and file_1_000.ts, file_1_001.ts, file_1_002.ts, etc.

The string "%v" may be present in the filename or in the last directory name containing the file, but only in one of them. (Additionally, %v may appear multiple times in the last sub-directory or filename.) If the string %v is present in the directory name, then sub-directories are created after expanding the directory name pattern. This enables creation of segments corresponding to different variant streams in subdirectories.

For example:

ffmpeg -i in.ts -b:v:0 1000k -b:v:1 256k -b:a:0 64k -b:a:1 32k \
  -map 0:v -map 0:a -map 0:v -map 0:a -f hls -var_stream_map "v:0,a:0 v:1,a:1" \
  -hls_segment_filename 'vs%v/file_%03d.ts' vs%v/out.m3u8

will produce the playlists segment file sets: vs0/file_000.ts, vs0/file_001.ts, vs0/file_002.ts, etc. and vs1/file_000.ts, vs1/file_001.ts, vs1/file_002.ts, etc.

strftime bool Use strftime() on filename to expand the segment filename with localtime. The segment number is also available in this mode, but to use it, you need to set second_level_segment_index in the hls_flag and %%d will be the specifier.

For example:

ffmpeg -i in.nut -strftime 1 -hls_segment_filename 'file-%Y%m%d-%s.ts' out.m3u8

will produce the playlist, out.m3u8, and segment files: file-20160215-1455569023.ts, file-20160215-1455569024.ts, etc. Note: On some systems/environments, the %s specifier is not available. See strftime() documentation. iykyk, platform stuff happens.

For example:

ffmpeg -i in.nut -strftime 1 -hls_flags second_level_segment_index -hls_segment_filename 'file-%Y%m%d-%%04d.ts' out.m3u8

will produce the playlist, out.m3u8, and segment files: file-20160215-0001.ts, file-20160215-0002.ts, etc.

strftime_mkdir bool Used together with strftime, it will create all subdirectories which are present in the expanded values of option hls_segment_filename.

For example:

ffmpeg -i in.nut -strftime 1 -strftime_mkdir 1 -hls_segment_filename '%Y%m%d/file-%Y%m%d-%s.ts' out.m3u8

will create a directory 201560215 (if it does not exist), and then produce the playlist, out.m3u8, and segment files: 20160215/file-20160215-1455569023.ts, 20160215/file-20160215-1455569024.ts, etc. (yes the manpage really says 201560215 there, we are not fixing it, we are simply reporting it 🙈)

For example:

ffmpeg -i in.nut -strftime 1 -strftime_mkdir 1 -hls_segment_filename '%Y/%m/%d/file-%Y%m%d-%s.ts' out.m3u8

...will create a directory hierarchy 2016/02/15 (if any of them don't already exist, she's not gonna yell at you about it), and then produce the playlist, out.m3u8, and segment files:

2016/02/15/file-20160215-1455569023.ts,
2016/02/15/file-20160215-1455569024.ts, etc.

hls_segment_options options_list — sets output format options using a :-separated list of key=value parameters. Values that contain : special characters have to be escaped 💅
hls_key_info_file key_info_file — uses the info in key_info_file for segment encryption, bestie. The first line of key_info_file specifies the key URI written to the playlist — that key URL is what gets used to access the encryption key during playback. The second line specifies the path to the key file used to obtain the key during the encryption process, and that key file is read as a single packed array of 16 octets in binary format. The optional third line specifies the initialization vector (IV) as a hexadecimal string, used instead of the segment sequence number (the default) for encryption. Changes to key_info_file will result in segment encryption with the new key/IV and an entry in the playlist for the new key URI/IV if hls_flags periodic_rekey is enabled 👑

Key info file format:

<key URI>
<key file path>
<IV> (optional)

Example key URIs:

http://server/file.key
/path/to/file.key
file.key

Example key file paths:

file.key
/path/to/file.key

Example IV:

0123456789ABCDEF0123456789ABCDEF

Key info file example:

http://server/file.key
/path/to/file.key
0123456789ABCDEF0123456789ABCDEF

Example shell script:

#!/bin/sh
BASE_URL=${1:-'.'}
openssl rand 16 > file.key
echo $BASE_URL/file.key > file.keyinfo
echo file.key >> file.keyinfo
echo $(openssl rand -hex 16) >> file.keyinfo
ffmpeg -f lavfi -re -i testsrc -c:v h264 -hls_flags delete_segments \
  -hls_key_info_file file.keyinfo out.m3u8

hls_enc bool — enable (1) or disable (0) the AES128 encryption. When it's enabled, every segment generated gets encrypted and the encryption key is saved as playlist name.key 🔐
hls_enc_key key — specify a 16-octet key to encrypt the segments with. By default it's randomly generated, so if you don't say anything she picks for you ✨
hls_enc_key_url keyurl — if set, keyurl is prepended instead of baseurl to the key filename in the playlist.
hls_enc_iv iv — specify the 16-octet initialization vector for every segment instead of the autogenerated ones.
hls_segment_type flags — possible values:
- mpegts — output segment files in MPEG-2 Transport Stream format. This is compatible with all HLS versions 🫶
- fmp4 — output segment files in fragmented MP4 format, similar to MPEG-DASH. fmp4 files may be used in HLS version 7 and above.
hls_fmp4_init_filename filename — set filename for the fragment files header file; the default filename is init.mp4.

When strftime is enabled, filename is expanded to the segment filename with localtime.

For example:

ffmpeg -i in.nut -hls_segment_type fmp4 -strftime 1 -hls_fmp4_init_filename "%s_init.mp4" out.m3u8

will produce init like this 1602678741_init.mp4 🎀

hls_fmp4_init_resend bool — resend init file after m3u8 file refresh every time, default is 0.

When var_stream_map is set with two or more variant streams, the filename pattern must contain the string "%v" — this string specifies the position of the variant stream index in the generated init file names. The string "%v" may be present in the filename or in the last directory name containing the file. If the string is present in the directory name, then sub-directories are created after expanding the directory name pattern. This enables creation of init files corresponding to different variant streams in subdirectories 🌸

hls_flags flags — possible values:
- single_file — if this flag is set, the muxer will store all segments in a single MPEG-TS file, and will use byte ranges in the playlist. HLS playlists generated this way will have the version number 4.

For example:

ffmpeg -i in.nut -hls_flags single_file out.m3u8

will produce the playlist, out.m3u8, and a single segment file, out.ts.

delete_segments — segment files removed from the playlist are deleted after a period of time equal to the duration of the segment plus the duration of the playlist.
append_list — append new segments into the end of the old segment list, and remove the "#EXT-X-ENDLIST" from the old segment list.
round_durations — round the duration info in the playlist file segment info to integer values, instead of using floating point. If there are no other features requiring higher HLS versions being used, then this will allow ffmpeg to output a HLS version 2 m3u8 ✨
discont_start — add the "#EXT-X-DISCONTINUITY" tag to the playlist, before the first segment's information.
omit_endlist — do not append the "EXT-X-ENDLIST" tag at the end of the playlist.
periodic_rekey — the file specified by "hls_key_info_file" will be checked periodically and updates to the encryption info will be detected. Be sure to replace this file atomically, including the file containing the AES encryption key ⚠️ no cap, do it atomically.
independent_segments — add the "#EXT-X-INDEPENDENT-SEGMENTS" tag to playlists that has video segments and when all the segments of that playlist are guaranteed to start with a key frame.
iframes_only — add the "#EXT-X-I-FRAMES-ONLY" tag to playlists that has video segments and can play only I-frames in the "#EXT-X-BYTERANGE" mode.
split_by_time — allow segments to start on frames other than key frames. This improves behavior on some players when the time between key frames is inconsistent, but may make things worse on others, and can cause some oddities during seeking 😬 This flag should be used with the hls_time option.
program_date_time — generate "EXT-X-PROGRAM-DATE-TIME" tags.
second_level_segment_index — make it possible to use segment indexes as %%d in the hls_segment_filename option expression besides date/time values when strftime option is on. To get fixed width numbers with trailing zeroes, %%0xd format is available where x is the required width.
second_level_segment_size — make it possible to use segment sizes (counted in bytes) as %%s in hls_segment_filename option expression besides date/time values when strftime is on. To get fixed width numbers with trailing zeroes, %%0xs format is available where x is the required width.
second_level_segment_duration — make it possible to use segment duration (calculated in microseconds) as %%t in hls_segment_filename option expression besides date/time values when strftime is on. To get fixed width numbers with trailing zeroes, %%0xt format is available where x is the required width.

For example:

ffmpeg -i sample.mpeg \
   -f hls -hls_time 3 -hls_list_size 5 \
   -hls_flags second_level_segment_index+second_level_segment_size+second_level_segment_duration \
   -strftime 1 -strftime_mkdir 1 -hls_segment_filename "segment_%Y%m%d%H%M%S_%%04d_%%08s_%%013t.ts" stream.m3u8

will produce segments like this: segment_20170102194334_0003_00122200_0000003000000.ts, segment_20170102194334_0004_00120072_0000003000000.ts etc. 🦋 iconic naming honestly.

temp_file — write segment data to filename.tmp and rename to filename only once the segment is complete.

A webserver serving up segments can be configured to reject requests to *.tmp to prevent access to in-progress segments before they have been added to the m3u8 playlist.

This flag also affects how m3u8 playlist files are created. If this flag is set, all playlist files will be written into a temporary file and renamed after they are complete, similarly as segments are handled. But playlists with "file" protocol and with hls_playlist_type type other than vod are always written into a temporary file regardless of this flag.

Master playlist files specified with master_pl_name, if any, with "file" protocol, are always written into temporary file regardless of this flag if master_pl_publish_rate value is other than zero 👑

hls_playlist_type type — if type is event, emit "#EXT-X-PLAYLIST-TYPE:EVENT" in the m3u8 header. This forces hls_list_size to 0; the playlist can only be appended to.

If type is vod, emit "#EXT-X-PLAYLIST-TYPE:VOD" in the m3u8 header. This forces hls_list_size to 0; the playlist must not change. periodt, it's frozen 💀

method method — use the given HTTP method to create the hls files.

For example:

ffmpeg -re -i in.ts -f hls -method PUT http://example.com/live/out.m3u8

will upload all the mpegts segment files to the HTTP server using the HTTP PUT method, and update the m3u8 files every "refresh" times using the same method. Note that the HTTP server must support the given method for uploading files 🙈

http_user_agent agent — override User-Agent field in HTTP header. Applicable only for HTTP output.
var_stream_map stream_map — specify a map string defining how to group the audio, video and subtitle streams into different variant streams. The variant stream groups are separated by space.

Expected string format is like this "a:0,v:0 a:1,v:1 ....". Here a:, v:, s: are the keys to specify audio, video and subtitle streams respectively. Allowed values are 0 to 9 (limited just based on practical usage) 💖

okay so bestie, when there are two or more variant streams, the output filename pattern must contain the string %v — that’s the little slot that tells ffmpeg where to put the variant stream index in the output media playlist filenames. The string %v may be present in the filename or in the last directory name containing the file. If the string is present in the directory name, then sub-directories are created after expanding the directory name pattern. This enables creation of variant streams in subdirectories. 💅✨

a few examples follow 👇

Create two hls variant streams. The first variant stream will contain video stream of bitrate 1000k and audio stream of bitrate 64k, and the second variant stream will contain video stream of bitrate 256k and audio stream of bitrate 32k. Here, two media playlist with file names out_0.m3u8 and out_1.m3u8 will be created.

ffmpeg -re -i in.ts -b:v:0 1000k -b:v:1 256k -b:a:0 64k -b:a:1 32k \
  -map 0:v -map 0:a -map 0:v -map 0:a -f hls -var_stream_map "v:0,a:0 v:1,a:1" \
  http://example.com/live/out_%v.m3u8

If you want something meaningful text instead of indexes in result names, you may specify names for each or some of the variants. The following example will create two hls variant streams as in the previous one. But here, the two media playlist with file names out_my_hd.m3u8 and out_my_sd.m3u8 will be created.

ffmpeg -re -i in.ts -b:v:0 1000k -b:v:1 256k -b:a:0 64k -b:a:1 32k \
  -map 0:v -map 0:a -map 0:v -map 0:a -f hls -var_stream_map "v:0,a:0,name:my_hd v:1,a:1,name:my_sd" \
  http://example.com/live/out_%v.m3u8

Create three hls variant streams. The first variant stream will be a video only stream with video bitrate 1000k, the second variant stream will be an audio only stream with bitrate 64k, and the third variant stream will be a video only stream with bitrate 256k. Here, three media playlist with file names out_0.m3u8, out_1.m3u8 and out_2.m3u8 will be created.

ffmpeg -re -i in.ts -b:v:0 1000k -b:v:1 256k -b:a:0 64k \
  -map 0:v -map 0:a -map 0:v -f hls -var_stream_map "v:0 a:0 v:1" \
  http://example.com/live/out_%v.m3u8

Create the variant streams in subdirectories. Here, the first media playlist is created at http://example.com/live/vs_0/out.m3u8 and the second one at http://example.com/live/vs_1/out.m3u8.

ffmpeg -re -i in.ts -b:v:0 1000k -b:v:1 256k -b:a:0 64k -b:a:1 32k \
  -map 0:v -map 0:a -map 0:v -map 0:a -f hls -var_stream_map "v:0,a:0 v:1,a:1" \
  http://example.com/live/vs_%v/out.m3u8

Create two audio only and two video only variant streams. In addition to the "#EXT-X-STREAM-INF" tag for each variant stream in the master playlist, the "#EXT-X-MEDIA" tag is also added for the two audio only variant streams and they are mapped to the two video only variant streams with audio group names 'aud_low' and 'aud_high'. By default, a single hls variant containing all the encoded streams is created.

ffmpeg -re -i in.ts -b:a:0 32k -b:a:1 64k -b:v:0 1000k -b:v:1 3000k  \
  -map 0:a -map 0:a -map 0:v -map 0:v -f hls \
  -var_stream_map "a:0,agroup:aud_low a:1,agroup:aud_high v:0,agroup:aud_low v:1,agroup:aud_high" \
  -master_pl_name master.m3u8 \
  http://example.com/live/out_%v.m3u8

Create two audio only and one video only variant streams. In addition to the "#EXT-X-STREAM-INF" tag for each variant stream in the master playlist, the "#EXT-X-MEDIA" tag is also added for the two audio only variant streams and they are mapped to the one video only variant streams with audio group name 'aud_low', and the audio group have default stat is NO or YES. By default, a single hls variant containing all the encoded streams is created.

ffmpeg -re -i in.ts -b:a:0 32k -b:a:1 64k -b:v:0 1000k \
  -map 0:a -map 0:a -map 0:v -f hls \
  -var_stream_map "a:0,agroup:aud_low,default:yes a:1,agroup:aud_low v:0,agroup:aud_low" \
  -master_pl_name master.m3u8 \
  http://example.com/live/out_%v.m3u8

Create two audio only and one video only variant streams, but this time with languages. In addition to the "#EXT-X-STREAM-INF" tag for each variant stream in the master playlist, the "#EXT-X-MEDIA" tag is also added for the two audio only variant streams and they are mapped to the one video only variant streams with audio group name 'aud_low', and the audio group have default stat is NO or YES, and one audio have and language is named ENG, the other audio language is named CHN. By default, a single hls variant containing all the encoded streams is created.

ffmpeg -re -i in.ts -b:a:0 32k -b:a:1 64k -b:v:0 1000k \
  -map 0:a -map 0:a -map 0:v -f hls \
  -var_stream_map "a:0,agroup:aud_low,default:yes,language:ENG a:1,agroup:aud_low,language:CHN v:0,agroup:aud_low" \
  -master_pl_name master.m3u8 \
  http://example.com/live/out_%v.m3u8

Create a single variant stream. Add the "#EXT-X-MEDIA" tag with "TYPE=SUBTITLES" in the master playlist with webvtt subtitle group name 'subtitle' and optional subtitle name, e.g. 'English'. Make sure the input file has one text subtitle stream at least.

ffmpeg -y -i input_with_subtitle.mkv \
 -b:v:0 5250k -c:v h264 -pix_fmt yuv420p -profile:v main -level 4.1 \
 -b:a:0 256k \
 -c:s webvtt -c:a mp2 -ar 48000 -ac 2 -map 0:v -map 0:a:0 -map 0:s:0 \
 -f hls -var_stream_map "v:0,a:0,s:0,sgroup:subtitle,sname:English" \
 -master_pl_name master.m3u8 -t 300 -hls_time 10 -hls_init_time 4 -hls_list_size \
 10 -master_pl_publish_rate 10 -hls_flags \
 delete_segments+discont_start+split_by_time ./tmp/video.m3u8

cc_stream_map cc_stream_map 💬

Map string which specifies different closed captions groups and their attributes. The closed captions stream groups are separated by space.

Expected string format is like this "ccgroup:<group name>,instreamid:<INSTREAM-ID>,language:<language code> ....". 'ccgroup' and 'instreamid' are mandatory attributes. 'language' is an optional attribute.

The closed captions groups configured using this option are mapped to different variant streams by providing the same 'ccgroup' name in the var_stream_map string.

For example:

ffmpeg -re -i in.ts -b:v:0 1000k -b:v:1 256k -b:a:0 64k -b:a:1 32k \
  -a53cc:0 1 -a53cc:1 1 \
  -map 0:v -map 0:a -map 0:v -map 0:a -f hls \
  -cc_stream_map "ccgroup:cc,instreamid:CC1,language:en ccgroup:cc,instreamid:CC2,language:sp" \
  -var_stream_map "v:0,a:0,ccgroup:cc v:1,a:1,ccgroup:cc" \
  -master_pl_name master.m3u8 \
  http://example.com/live/out_%v.m3u8

will add two "#EXT-X-MEDIA" tags with "TYPE=CLOSED-CAPTIONS" in the master playlist for the INSTREAM-IDs 'CC1' and 'CC2'. Also, it will add "CLOSED-CAPTIONS" attribute with group name 'cc' for the two output variant streams.

If var_stream_map is not set, then the first available ccgroup in cc_stream_map is mapped to the output variant stream.

For example:

ffmpeg -re -i in.ts -b:v 1000k -b:a 64k -a53cc 1 -f hls \
  -cc_stream_map "ccgroup:cc,instreamid:CC1,language:en" \
  -master_pl_name master.m3u8 \
  http://example.com/live/out.m3u8

this will add "#EXT-X-MEDIA" tag with "TYPE=CLOSED-CAPTIONS" in the master playlist with group name 'cc', language 'en' (english) and INSTREAM-ID 'CC1'. Also, it will add "CLOSED-CAPTIONS" attribute with group name 'cc' for the output variant stream.

master_pl_name name — Create HLS master playlist with the given name. 🎀

For example:

ffmpeg -re -i in.ts -f hls -master_pl_name master.m3u8 http://example.com/live/out.m3u8

creates an HLS master playlist with name master.m3u8 which is published at http://example.com/live/.

master_pl_publish_rate count — Publish master play list repeatedly every after specified number of segment intervals.

For example:

ffmpeg -re -i in.ts -f hls -master_pl_name master.m3u8 \
-hls_time 2 -master_pl_publish_rate 30 http://example.com/live/out.m3u8

creates an HLS master playlist with name master.m3u8 and keeps publishing it repeatedly every after 30 segments i.e. every after 60s.

http_persistent bool — Use persistent HTTP connections. Applicable only for HTTP output.
timeout timeout — Set timeout for socket I/O operations. Applicable only for HTTP output.
ignore_io_errors bool — Ignore IO errors during open, write and delete. Useful for long-duration runs with network output.
headers headers — Set custom HTTP headers, can override built in default headers. Applicable only for HTTP output.

🌈 iamf 💖

Immersive Audio Model and Formats (IAMF) muxer. ✨

IAMF is used to provide immersive audio content for presentation on a wide range of devices in both streaming and offline applications. These applications include internet audio streaming, multicasting/broadcasting services, file download, gaming, communication, virtual and augmented reality, and others. In these applications, audio may be played back on a wide range of devices, e.g., headphones, mobile phones, tablets, TVs, sound bars, home theater systems, and big screens.

This format was promoted and designed by Alliance for Open Media. 🫶

For more information about this format, see https://aomedia.org/iamf/.

🎀 ico 🦋

ICO file muxer.

Microsoft's icon file format (ICO) has some strict limitations that should be noted: 👑

Size cannot exceed 256 pixels in any dimension
Only BMP and PNG images can be stored
If a BMP image is used, it must be one of the following pixel formats:

BMP bit-depth mapping, bestie — BMP Bit Depth → FFmpeg Pixel Format, and the vibe check says match these exactly:

1bit → pal8
4bit → pal8
8bit → pal8
16bit → rgb555le
24bit → bgr24
32bit → bgra

Also, two little rules that are not optional:

If a BMP image is used, it must use the BITMAPINFOHEADER DIB header.
If a PNG image is used, it must use the rgba pixel format.

ilbc 🎀

Internet Low Bitrate Codec (iLBC) raw muxer.

It accepts a single ilbc audio stream. That’s the whole assignment, periodt.

image2, image2pipe 🖼️

Image file muxer.

The image2 muxer writes video frames to image files.

The output filenames are specified by a pattern, which can be used to produce sequentially numbered series of files. The pattern may contain the string %d or %0Nd; that string specifies the position of the characters representing a numbering in the filenames. If the form %0Nd is used, the string representing the number in each filename is 0-padded to N digits. The literal character % can be specified in the pattern with the string %%.

If the pattern contains %d or %0Nd, the first filename of the file list specified will contain the number 1, and all the following numbers will be sequential.

The pattern may contain a suffix which is used to automatically determine the format of the image files to write.

For example, the pattern img-%03d.bmp will specify a sequence of filenames of the form img-001.bmp, img-002.bmp, ..., img-010.bmp, etc. The pattern img%%-%d.jpg will specify a sequence of filenames of the form img%-1.jpg, img%-2.jpg, ..., img%-10.jpg, etc.

The image muxer supports the .Y.U.V image file format. This format is special in that each image frame consists of three files, for each of the YUV420P components. To read or write this image file format, specify the name of the .Y file. The muxer will automatically open the .U and .V files as required.

The image2pipe muxer accepts the same options as the image2 muxer, but ignores the pattern verification and expansion, as it is supposed to write to the command output rather than to an actual stored file.

Options ✨

frame_pts bool — If set to 1, expand the filename with the packet PTS (presentation time stamp). Default value is 0.
start_number count — Start the sequence from the specified number. Default value is 1.
update bool — If set to 1, the filename will always be interpreted as just a filename, not a pattern, and the corresponding file will be continuously overwritten with new images. Default value is 0.
strftime bool — If set to 1, expand the filename with date and time information from strftime(). Default value is 0.
atomic_writing bool — Write output to a temporary file, which is renamed to target filename once writing is completed. Default is disabled.
protocol_opts options_list — Set protocol options as a :-separated list of key=value parameters. Values containing the : special character must be escaped.

Examples 🌟

Use ffmpeg for creating a sequence of files img-001.jpeg, img-002.jpeg, ..., taking one image every second from the input video:

ffmpeg -i in.avi -vsync cfr -r 1 -f image2 'img-%03d.jpeg'

Note that with ffmpeg, if the format is not specified with the -f option and the output filename specifies an image file format, the image2 muxer is automatically selected, so the previous command can be written as:

ffmpeg -i in.avi -vsync cfr -r 1 'img-%03d.jpeg'

Note also that the pattern must not necessarily contain %d or %0Nd, for example to create a single image file img.jpeg from the start of the input video you can employ the command:

ffmpeg -i in.avi -f image2 -frames:v 1 img.jpeg

The strftime option allows you to expand the filename with date and time information. Check the documentation of the strftime() function for the syntax.

To generate image files from the strftime() %Y-%m-%d_%H-%M-%S pattern, the following ffmpeg command can be used:

ffmpeg -f v4l2 -r 1 -i /dev/video0 -f image2 -strftime 1 "%Y-%m-%d_%H-%M-%S.jpg"

Set the file name with current frame’s PTS:

ffmpeg -f v4l2 -r 1 -i /dev/video0 -copyts -f image2 -frame_pts true %d.jpg

Publish contents of your desktop directly to a WebDAV server every second:

ffmpeg -f x11grab -framerate 1 -i :0.0 -q:v 6 -update 1 -protocol_opts method=PUT http://example.com/desktop.jpg

ircam 🎻

Berkeley / IRCAM / CARL Sound Filesystem (BICSF) format muxer.

The Berkeley/IRCAM/CARL Sound Format, developed in the 1980s, is a result of the merging of several different earlier sound file formats and systems including the csound system developed by Dr Gareth Loy at the Computer Audio Research Lab (CARL) at UC San Diego, the IRCAM sound file system developed by Rob Gross and Dan Timis at the Institut de Recherche et Coordination Acoustique / Musique in Paris and the Berkeley Fast Filesystem.

It was developed initially as part of the Berkeley/IRCAM/CARL Sound Filesystem, a suite of programs designed to implement a filesystem for audio applications running under Berkeley UNIX. It was particularly popular in academic music research centres, and was used a number of times in the creation of early computer-generated compositions.

This muxer accepts a single audio stream containing PCM data.

ivf 🦋

On2 IVF muxer.

IVF was developed by On2 Technologies (formerly known as Duck Corporation), to store internally developed codecs.

This muxer accepts a single vp8, vp9, or av1 video stream.

jacosub 💜

JACOsub subtitle format muxer.

This muxer accepts a single jacosub subtitles stream.

For more information about the format, see http://unicorn.us.com/jacosub/jscripts.html.

kvag 🎮

Simon & Schuster Interactive VAG muxer.

This custom VAG container is used by some Simon & Schuster Interactive games such as "Real War", and "Real War: Rogue States".

This muxer accepts a single adpcm_ima_ssi audio stream.

lc3 🎧

Bluetooth SIG Low Complexity Communication Codec audio (LC3), or ETSI TS 103 634 Low Complexity Communication Codec plus (LC3plus).

This muxer accepts a single lc3 audio stream.

lrc 🎤

LRC lyrics file format muxer.

LRC (short for LyRiCs) is a computer file format that synchronizes song lyrics with an audio file, such as MP3, Vorbis, or MIDI.

This muxer accepts a single subrip or text subtitles stream.

Metadata 🏷️

The following metadata tags are converted to the format corresponding metadata:

title
album
artist
author
creator
encoder
encoder_version

If encoder_version is not explicitly set, it is automatically set to the libavformat version.

matroska 👑

Matroska container muxer.

This muxer implements the matroska and webm container specs.

Metadata 🏷️

The recognized metadata settings in this muxer are:

title — Set title name provided to a single track. This gets mapped to the FileDescription element for a stream written as attachment.
language — Specify the language of the track in the Matroska languages form.

The language can be either the 3 letters bibliographic ISO-639-2 (ISO 639-2/B) form (like fre for French), or a language code mixed with a country code for specialities in languages (like fre-ca for Canadian French).

stereo_mode — Set stereo 3D video layout of two views in a single video track.

The following values are recognized:

mono — video is not stereo
left_right — Both views are arranged side by side, Left-eye view is on the left
bottom_top — Both views are arranged in top-bottom orientation, Left-eye view is at bottom
top_bottom — Both views are arranged in top-bottom orientation, Left-eye view is on top
checkerboard_rl — Each view is arranged in a checkerboard interleaved pattern, Left-eye view being first
checkerboard_lr — Each view is arranged in a checkerboard interleaved pattern, Right-eye view being first
row_interleaved_rl — Each view is constituted by a row based interleaving, Right-eye view is first row
row_interleaved_lr — Each view is constituted by a row based interleaving, Left-eye view is first row
col_interleaved_rl — Both views are arranged in a column based interleaving manner, Right-eye view is first column
col_interleaved_lr — Both views are arranged in a column based interleaving manner, Left-eye view is first column
anaglyph_cyan_red — All frames are in anaglyph format viewable through red-cyan filters
right_left — Both views are arranged side by side, Right-eye view is on the left
anaglyph_green_magenta — All frames are in anaglyph format viewable through green-magenta filters
block_lr — Both eyes laced in one Block, Left-eye view is first
block_rl — Both eyes laced in one Block, Right-eye view is first

For example a 3D WebM clip can be created using the following command line:

ffmpeg -i sample_left_right_clip.mpg -an -c:v libvpx -metadata stereo_mode=left_right -y stereo_clip.webm

Options ✨

reserve_index_space size — By default, this muxer writes the index for seeking (called cues in Matroska terms) at the end of the file, because it cannot know in advance how much space to leave for the index at the beginning of the file. However for some use cases — e.g. streaming where seeking is possible but slow — it is useful to put the index at the beginning of the file.

If this option is set to a non-zero value, the muxer will reserve size bytes of space in the file header and then try to write the cues there when the muxing finishes. If the reserved space does not suffice, no Cues will be written, the file will be finalized and writing the trailer will return an error. A safe size for most use cases should be about 50kB per hour of video.

Note that cues are only written if the output is seekable and this option will have no effect if it is not.

cues_to_front bool — If set, the muxer will write the index at the beginning of the file by shifting the main data if necessary. This can be combined with reserve_index_space in which case the data is only shifted if the initially reserved space turns out to be insufficient.

This option is ignored if the output is unseekable.

cluster_size_limit size — caps how many bytes can go into one cluster. If you don't set it, the limit just auto-sets itself to a sensible hardcoded fixed value, no stress 💅
cluster_time_limit duration — caps how many milliseconds can go into one cluster. Same deal, if you don't set it, a sensible hardcoded fixed value swoops in automatically ✨
dash bool — makes a WebM file that conforms to the WebM DASH spec. Default is "false", so it's off unless you say otherwise 🎀
dash_track_number index — sets the track number for the DASH stream. Default is 1, bestie.
live bool — writes files assuming it's a live stream. Default is "false".
allow_raw_vfw bool — lets raw VFW mode through. Default is "false".
flipped_raw_rgb bool — if you set it to "true", it stores a positive height for raw RGB bitmaps, which means the bitmap is stored bottom-up. Careful though: this option does not flip the bitmap for you, you gotta do that yourself beforehand, e.g. with the vflip filter. Default is "false", which means the bitmap is stored top down 🫶
write_crc32 bool — writes a CRC32 element inside every Level 1 element. Default is "true". Heads up: this option gets ignored for WebM 💀
default_mode mode — controls how the FlagDefault of the output tracks gets set. It's the thing that influences which tracks players should play by default. Default mode is passthrough.
- infer — every track with disposition default gets the FlagDefault set. Plus, for each type of track (audio, video or subtitle), if there's no track with disposition default of that type, then the first track of that type gets marked as default (if one exists). This makes sure the default flag is set in a sensible way even if the input came from containers that don't even have the concept of default tracks. Iconic behaviour honestly 👑
- infer_no_subs — same as infer, except if there's no subtitle track with disposition default, then no subtitle track gets marked as default. Simple as that.
- passthrough — in this mode the FlagDefault is set if and only if the AV_DISPOSITION_DEFAULT flag is set in the disposition of the corresponding stream. No guessing, no extras.

🦋 `md5`

The MD5 testing format, bestie. It's a variant of the hash muxer — but unlike that one, it defaults to using the MD5 hash function 💖

See also the hash and framemd5 muxers.

🌸 Examples 🌸

To compute the MD5 hash of the input converted to raw audio and video, and store it in the file out.md5:

ffmpeg -i INPUT -f md5 out.md5

To print the MD5 hash to stdout:

ffmpeg -i INPUT -f md5 -

🎀 `mcc`

A muxer for MacCaption MCC files — it supports MCC versions 1.0 and 2.0. MCC files store VANC data, which can include closed captions (EIA-608 and CEA-708), ancillary time code, pan-scan data, etc. So much packed in there fr fr ✨

💅 Options 💅

The muxer options are:

override_time_code_rate — overrides the "Time Code Rate" value in the output. By default it tries to deduce it from the stream's "time_base", which — ngl — often doesn't work 😭
use_u_alias — uses the "U" alias for the byte sequence E1h 00h 00h 00h. It's disabled by default because some .mcc files disagree on whether it has 2 or 3 zero bytes 🙈
mcc_version — the MCC file format version. Must be either 1 or 2, defaults to 2.
creation_program — the creation program. Defaults to this version of FFmpeg 🫶
creation_time — the creation time. Defaults to the current time.

🌈 Examples 🌈

Extract a MXF SMPTE_436M_ANC stream from a MXF file and write it to a MCC file at 30 fps:

ffmpeg -i input.mxf -c copy -map 0:d -override_time_code_rate 30 out.mcc

Extract EIA-608/CTA-708 closed captions from a .mp4 file and write them to a MCC file at 29.97 fps:

ffmpeg -f lavfi -i "movie=input.mp4[out+subcc]" -c:s copy -map 0:s -override_time_code_rate 30000/1001 out.mcc

🎬 `microdvd`

MicroDVD subtitle format muxer, periodt. This muxer accepts a single microdvd subtitles stream — one, that's it, don't get greedy 💅

🍓 `mmf`

Synthetic music Mobile Application Format (SMAF) format muxer ✨

SMAF is a music data format specified by Yamaha for portable electronic devices, such as mobile phones and personal digital assistants. Very retro of them, we love to see it.

This muxer accepts a single adpcm_yamaha audio stream.

👑 `mp3`

The MP3 muxer writes a raw MP3 stream with the following optional features, and honestly the list is giving main character energy:

An ID3v2 metadata header at the beginning (enabled by default). Versions 2.3 and 2.4 are supported, and the id3v2_version private option controls which one you get (3 or 4). Setting id3v2_version to 0 disables the ID3v2 header completely 🫶

The muxer supports writing attached pictures (APIC frames) to the ID3v2 header. You supply the pictures to the muxer as a video stream with a single packet. There can be any number of those streams, and each one will correspond to a single APIC frame. The stream metadata tags title and comment map to APIC description and picture type respectively. See http://id3.org/id3v2.4.0-frames for the allowed picture types.

Important caveat bestie: the APIC frames must be written at the beginning, so the muxer will buffer the audio frames until it gets all the pictures. So it's advised to provide the pictures as soon as possible to avoid excessive buffering 😭

A Xing/LAME frame right after the ID3v2 header (if present). It's enabled by default, but it will only be written if the output is seekable. The write_xing private option can be used to disable it. The frame contains various information that may be useful to the decoder, like the audio duration or encoder delay ⭐
A legacy ID3v1 tag at the end of the file (disabled by default). You can enable it with the write_id3v1 private option, but since its capabilities are very limited, its usage is not recommended 💀

💖 Examples 💖

Write an mp3 with an ID3v2.3 header and an ID3v1 footer:

ffmpeg -i INPUT -id3v2_version 3 -write_id3v1 1 out.mp3

To attach a picture to an mp3 file, select both the audio and the picture stream with map:

ffmpeg -i input.mp3 -i cover.png -c copy -map 0 -map 1
-metadata:s:v title="Album cover" -metadata:s:v comment="Cover (Front)" out.mp3

Write a "clean" MP3 without any extra features:

ffmpeg -i input.wav -write_xing 0 -id3v2_version 0 out.mp3

📺 `mpegts`

MPEG transport stream muxer, the absolute workhorse of the group chat 🚌✨

This muxer implements ISO 13818-1 and part of ETSI EN 300 468.

The recognized metadata settings in the mpegts muxer are "service_provider" and "service_name". If they're not set, the default for "service_provider" is FFmpeg and the default for "service_name" is Service01.

🎀 Options 🎀

The muxer options are:

mpegts_transport_stream_id integer — sets the transport_stream_id. This identifies a transponder in DVB. Default is 0x0001.
mpegts_original_network_id integer — sets the original_network_id. This is a unique identifier of a network in DVB. Its main use is in the unique identification of a service through the path Original_Network_ID, Transport_Stream_ID. Default is 0x0001.
mpegts_service_id integer — sets the service_id, also known as program in DVB. Default is 0x0001.
mpegts_service_type integer — sets the program service_type. Default is "digital_tv". Accepts the following options:
- hex_value — any hexadecimal value between 0x01 and 0xff as defined in ETSI 300 468.
- digital_tv — Digital TV service.
- digital_radio — Digital Radio service.
- teletext — Teletext service.
- advanced_codec_digital_radio — Advanced Codec Digital Radio service.
- mpeg2_digital_hdtv — MPEG2 Digital HDTV service.
- advanced_codec_digital_sdtv — Advanced Codec Digital SDTV service.
- advanced_codec_digital_hdtv — Advanced Codec Digital HDTV service.
mpegts_pmt_start_pid integer — sets the first PID for PMTs. Default is 0x1000, minimum is 0x0020, maximum is 0x1ffa. Bestie, this option has no effect in m2ts mode, where the PMT PID is fixed at 0x0100 🙈
mpegts_start_pid integer — sets the first PID for elementary streams. Default is 0x0100, minimum is 0x0020, maximum is 0x1ffa. Also no effect in m2ts mode, where the elementary stream PIDs are fixed.
mpegts_m2ts_mode boolean — enables m2ts mode if set to 1. Default value is -1, which disables m2ts mode.
muxrate integer — sets a constant muxrate. Default is VBR.
pes_payload_size integer — sets the minimum PES packet payload in bytes. Default is 2930.
mpegts_flags flags — sets mpegts flags. Accepts the following options:
- resend_headers — reemit PAT/PMT before writing the next packet.
- latm — use LATM packetization for AAC.
- pat_pmt_at_frames — reemit PAT and PMT at each video frame.
- system_b — conform to System B (DVB) instead of System A (ATSC).
- initial_discontinuity — mark the initial packet of each stream as discontinuity.
- nit — emit NIT table.
- omit_rai — disable writing of random access indicator.
mpegts_copyts boolean — preserve original timestamps, if the value is set to 1. Default value is -1, which results in shifting timestamps so that they start from 0 ✨
omit_video_pes_length boolean — omit the PES packet length for video packets. Default is 1 (true).
pcr_period integer — overrides the default PCR retransmission time in milliseconds. Default is -1, which means the PCR interval gets determined automatically: 20 ms is used for CBR streams, and the highest multiple of the frame duration which is less than 100 ms is used for VBR streams. It's giving "we'll handle it, don't worry" 💅
pat_period duration — maximum time in seconds between PAT/PMT tables. Default is 0.1.
sdt_period duration — maximum time in seconds between SDT tables. Default is 0.5.
nit_period duration — maximum time in seconds between NIT tables. Default is 0.5.
tables_version integer — sets PAT, PMT, SDT and NIT version (default 0, valid values are from 0 to 31, inclusively). This option lets you update the stream structure so that a standard consumer may detect the change 🔥 To do so, reopen the output "AVFormatContext" (in case of API usage) or restart the ffmpeg instance, cyclically changing the tables_version value:

               ffmpeg -i source1.ts -codec copy -f mpegts -tables_version 0 udp://1.1.1.1:1111
               ffmpeg -i source2.ts -codec copy -f mpegts -tables_version 1 udp://1.1.1.1:1111
               ...
               ffmpeg -i source3.ts -codec copy -f mpegts -tables_version 31 udp://1.1.1.1:1111
               ffmpeg -i source1.ts -codec copy -f mpegts -tables_version 0 udp://1.1.1.1:1111
               ffmpeg -i source2.ts -codec copy -f mpegts -tables_version 1 udp://1.1.1.1:1111
               ...

Example

Okay so here's the whole metadata glow-up in one go, bestie — watch how every PID and ID gets its own little moment 💅

       ffmpeg -i file.mpg -c copy \
            -mpegts_original_network_id 0x1122 \
            -mpegts_transport_stream_id 0x3344 \
            -mpegts_service_id 0x5566 \
            -mpegts_pmt_start_pid 0x1500 \
            -mpegts_start_pid 0x150 \
            -metadata service_provider="Some provider" \
            -metadata service_name="Some Channel" \
            out.ts

mxf, mxf_d10, mxf_opatom

The MXF muxer, bestie. Broadcast royalty 👑

Options

Here's the muxer option lineup:

store_user_comments bool — tells it whether user comments should be stored when they're available, or literally never. Fun fact: IRT D-10 doesn't allow user comments at all 😭 So the default is to write them for mxf and mxf_opatom, but not for mxf_d10.

null

The null muxer — it's giving "no thoughts, head empty" but make it a container 🌸

This one doesn't generate any output file whatsoever. What's it even for?? Testing and benchmarking, obviously. That's the whole vibe.

Say you wanna benchmark decoding with ffmpeg, you'd run:

       ffmpeg -benchmark -i INPUT -f null out.null

Heads up! That command above does not read or write the out.null file at all — but like, specifying the output file is still required by the ffmpeg syntax. Rules are rules, iykyk.

Alternatively you can write the command like this instead:

       ffmpeg -benchmark -i INPUT -f null -

nut

-syncpoints flags — changes up the syncpoint usage in nut:
- default — uses the normal low-overhead seeking aids. Chill, standard, does the job ✨
- none — doesn't use syncpoints at all, so less overhead, but the stream becomes non-seekable;

Use of this option is not recommended, because the resulting files are very damage sensitive and seeking is not possible 😬 Also, in general the overhead from syncpoints is negligible. Note though, -C<write_index> 0 can be used to disable all growing data tables, which lets you mux endless streams with limited memory and without these disadvantages. Slay.

timestamped — extends the syncpoint with a wallclock field.

Real talk: the none and timestamped flags are experimental. 🙈

-write_index bool — writes the index at the end. Default behavior is to write an index, mwah 💖

           ffmpeg -i INPUT -f_strict experimental -syncpoints none - | processor

ogg

The Ogg container muxer. Classic icon, no notes 🎀

-page_duration duration — your preferred page duration, measured in microseconds. The muxer will try to create pages that are approximately duration microseconds long. This is your lil' dial for compromising between seek granularity and container overhead. Default is 1 second. A value of 0 fills all segments, making pages as large as possible. A value of 1 effectively uses 1 packet-per-page in most situations, which gives you small seek granularity at the cost of extra container overhead. Trade-offs, bestie 💅
-serial_offset value — the serial value that the streams' serial number gets set from. Set it to different and sufficiently large values and you're guaranteed that the produced ogg files can be safely chained together. Obsessed with a safe chain 🔗

rcwt

RCWT (Raw Captions With Time) is a format native to ccextractor — you know, the commonly used open source tool for processing 608/708 Closed Captions (CC) sources. It can be used to archive the original extracted CC bitstream, and to produce a source file for later processing or conversion. The format allows for interoperability between ccextractor and FFmpeg, it's simple to parse, and you can use it to create a backup of the CC presentation. We love a low-effort, high-reward format 🫶

This muxer implements the spec as of March 2024, which has been stable and unchanged since April 2014. Ten years no changes, that's main character energy fr.

So this muxer will have some nuances compared to the way ccextractor muxes RCWT. No compatibility issues have been observed so far when processing the output with ccextractor as a result of this, but mileage may vary and outputs will not be a bit-exact match. Close enough is close enough 💀

A free specification of RCWT can be found here: <https://github.com/CCExtractor/ccextractor/blob/master/docs/BINARY_FILE_FORMAT.TXT>

Examples

Extract Closed Captions to RCWT using lavfi:

                   ffmpeg -f lavfi -i "movie=INPUT.mkv[out+subcc]" -map 0:s:0 -c:s copy -f rcwt CC.rcwt.bin

segment, stream_segment, ssegment

The basic stream segmenter. Chopping things into pieces, but make it cute ✂️✨

This muxer outputs streams into a bunch of separate files of nearly fixed duration. You can set the output filename pattern in a fashion similar to image2, or by using a strftime template if the strftime option is enabled.

stream_segment is a variant of the muxer used to write to streaming output formats — meaning formats that don't require global headers — and it's recommended for outputting to, say, MPEG transport stream segments. ssegment is a shorter alias for stream_segment. Same vibe, less typing 💅

Every segment starts with a keyframe of the selected reference stream, which you set through the reference_stream option.

Real talk: if you want accurate splitting for a video file, you need to make the input key frames correspond to the exact splitting times the segmenter expects — otherwise the segment muxer will start the new segment with the key frame it finds next after the specified start time. No cap, it just does its own thing 😭

The segment muxer works best with a single constant frame rate video. That's its moment.

Optionally it can generate a list of the created segments, by setting the option segment_list. The list type is specified by the segment_list_type option. The entry filenames in the segment list are set by default to the basename of the corresponding segment files.

Also go look at the hls muxer, which provides a more specific implementation for HLS segmentation. Whole different girl 💖

Options

The segment muxer supports the following options:

increment_tc 1|0 — if set to 1, increment the timecode between each segment. If you pick this, the input needs to have a timecode in the first video stream. Default value is 0.
reference_stream specifier — sets the reference stream, as specified by the string specifier. If specifier is set to "auto", the reference is chosen automatically. Otherwise it must be a stream specifier (see the ``Stream specifiers'' chapter in the ffmpeg manual) which specifies the reference stream. Default value is "auto".
segment_format format — override the inner container format. By default it's guessed from the filename extension.
segment_format_options options_list — set output format options using a :-separated list of key=value parameters. Values containing the : special character must be escaped. Escape responsibly 🫶
segment_list name — generate also a listfile named name. If it's not specified, no listfile is generated.
segment_list_flags flags — set flags affecting the segment list generation. It currently supports the following flags:
- cache — allow caching (only affects M3U8 list files).
- live — allow live-friendly file generation.
segment_list_size size — update the list file so it contains at most size segments. If 0, the list file will contain all the segments. Default value is 0.
segment_list_entry_prefix prefix — prepend prefix to each entry. Useful for generating absolute paths. By default no prefix is applied.
segment_list_type type — select the listing format. The following values are recognized:
- flat — generate a flat list for the created segments, one segment per line.
- csv, ext — generate a list for the created segments, one segment per line, each line matching the format (comma-separated values):

                       <segment_filename>,<segment_start_time>,<segment_end_time>

segment_filename is the name of the output file generated by the muxer according to the provided pattern. CSV escaping (according to RFC4180) is applied if required.

segment_start_time and segment_end_time specify the segment start and end time expressed in seconds.

A list file with the suffix ".csv" or ".ext" will auto-select this format.

ext is deprecated in favor of csv. Use csv, bestie, it's the it-girl now 💅

ffconcat — generate an ffconcat file for the created segments. The resulting file can be read using the FFmpeg concat demuxer. A list file with the suffix ".ffcat" or ".ffconcat" will auto-select this format.
m3u8 — generate an extended M3U8 file, version 3, compliant with <http://tools.ietf.org/id/draft-pantos-http-live-streaming>. A list file with the suffix ".m3u8" will auto-select this format.

If not specified, the type is guessed from the list file name suffix. Detective mode 🔍

segment_time time — set segment duration to time; the value must be a duration specification. Default value is "2". See also the segment_times option.

Note that splitting may not be accurate, unless you force the reference stream key-frames at the given time. See the introductory notice and the examples below. It's giving "approximately" 😭

min_seg_duration time — set minimum segment duration to time; the value must be a duration specification. This prevents the muxer from ending segments at a duration below this value. Only effective with segment_time. Default value is "0".
segment_atclocktime 1|0 — if set to "1", split at regular clock time intervals starting from 00:00 o'clock. The time value specified in segment_time is used for setting the length of the splitting interval.

For example, with segment_time set to "900" this makes it possible to create files at 12:00 o'clock, 12:15, 12:30, etc. Clean, scheduled, punctual icon behavior 🕐

Default value is "0".

segment_clocktime_offset duration — delays the segment splitting times by the specified duration when using segment_atclocktime.

For example, with segment_time set to "900" and segment_clocktime_offset set to "300", this makes it possible to create files at 12:05, 12:20, 12:35, etc. Fashionably late ✨

Default value is "0".

segment_clocktime_wrap_duration duration — forces the segmenter to only start a new segment if a packet reaches the muxer within the specified duration after the segmenting clock time. This way you can make the segmenter more resilient to backward local time jumps, like leap seconds or the transition to standard time from daylight savings time. Built different, honestly 🌈

Default is the maximum possible duration, which means it starts a new segment regardless of the elapsed time since the last clock time.

segment_time_delta delta — sets the accuracy time for picking a segment start time, bestie, and it’s a duration spec. Default is "0".

When delta is set, a key-frame starts a new segment if its PTS satisfies this exact relation:

PTS >= start_time - time_delta

This is lowkey iconic when you’re splitting video content that always splits at GOP boundaries, especially if a key frame lands just before your specified split time.

It’s especially useful with the ffmpeg force_key_frames option. The key frame times from force_key_frames might not be set super accurately because of rounding, so a key frame time can end up just before the specified time. For constant frame rate videos, 1/(2*frame_rate) should handle the worst-case mismatch between the specified time and the time set by force_key_frames.

segment_times times — specify a list of split points. times is a comma-separated list of duration specs, in increasing order. Also check segment_time, bestie.
segment_frames frames — specify a list of split video frame numbers. frames is a comma-separated list of integer numbers, in increasing order.

This one says: start a new segment whenever a reference stream key frame is found and the sequential number (starting from 0) of the frame is greater than or equal to the next value in the list.

segment_wrap limit — wrap around the segment index once it hits limit. Cute little loop-de-loop.
segment_start_number number — sets the sequence number of the first segment. Defaults to 0.
strftime 1|0 — use the strftime function to define the name of the new segments to write. If this is selected, the output segment name must contain a strftime function template. Default value is 0.
break_non_keyframes 1|0 — if enabled, segments can start on frames that aren’t keyframes. This improves behavior on some players when the time between keyframes is inconsistent, but it may make things worse on others, and it can cause some oddities during seeking. Defaults to 0.
reset_timestamps 1|0 — reset timestamps at the beginning of each segment so every segment starts with near-zero timestamps. It’s meant to make playback of the generated segments easier. May not work with some combinations of muxers/codecs. Set to 0 by default.
initial_offset offset — timestamp offset to apply to the output packet timestamps. The argument must be a time duration specification, and it defaults to 0.
write_empty_segments 1|0 — if enabled, write an empty segment if there are no packets during the period a segment would usually span. Otherwise, the segment gets filled with the next packet written. Defaults to 0.

Make sure to require a closed GOP when encoding and to set the GOP size to fit your segment time constraint. No cap, this one matters.

💖 Examples 💖

Remux the content of file in.mkv to a list of segments out-000.nut, out-001.nut, etc., and write the list of generated segments to out.list:

ffmpeg -i in.mkv -codec hevc -flags +cgop -g 60 -map 0 -f segment -segment_list out.list out%03d.nut

Segment input and set output format options for the output segments:

ffmpeg -i in.mkv -f segment -segment_time 10 -segment_format_options movflags=+faststart out%03d.mp4

Segment the input file according to the split points specified by the segment_times option:

ffmpeg -i in.mkv -codec copy -map 0 -f segment -segment_list out.csv -segment_times 1,2,3,5,8,13,21 out%03d.nut

Use the ffmpeg force_key_frames option to force key frames in the input at the specified location, together with the segment option segment_time_delta to account for possible roundings operated when setting key frame times:

ffmpeg -i in.mkv -force_key_frames 1,2,3,5,8,13,21 -codec:v mpeg4 -codec:a pcm_s16le -map 0 \
-f segment -segment_list out.csv -segment_times 1,2,3,5,8,13,21 -segment_time_delta 0.05 out%03d.nut

In order to force key frames on the input file, transcoding is required.

Segment the input file by splitting the input file according to the frame numbers sequence specified with the segment_frames option:

ffmpeg -i in.mkv -codec copy -map 0 -f segment -segment_list out.csv -segment_frames 100,200,300,500,800 out%03d.nut

Convert the in.mkv to TS segments using the libx264 and aac encoders:

ffmpeg -i in.mkv -map 0 -codec:v libx264 -codec:a aac -f ssegment -segment_list out.list out%03d.ts

Segment the input file, and create an M3U8 live playlist (can be used as live HLS source):

ffmpeg -re -i in.mkv -codec copy -map 0 -f segment -segment_list playlist.m3u8 \
-segment_list_flags +live -segment_time 10 out%03d.mkv

🌈 smoothstreaming 🌈

The Smooth Streaming muxer generates a set of files (Manifest, chunks) suitable for serving with a conventional web server. Very “it just works” energy.

window_size — specify the number of fragments kept in the manifest. Default 0 (keep all).
extra_window_size — specify the number of fragments kept outside of the manifest before removing them from disk. Default 5.
lookahead_count — specify the number of lookahead fragments. Default 2.
min_frag_duration — specify the minimum fragment duration (in microseconds). Default 5000000.
remove_at_exit — specify whether to remove all fragments when finished. Default 0 (do not remove).

✨ streamhash ✨

Per stream hash testing format. This muxer computes and prints a cryptographic hash of all the input frames, on a per-stream basis. It’s for equality checks without having to do a complete binary comparison. Obsessed.

By default audio frames are converted to signed 16-bit raw audio and video frames to raw video before computing the hash, but the output of explicit conversions to other codecs can also be used. Timestamps are ignored. It uses the SHA-256 cryptographic hash function by default, but supports several other algorithms.

The output of the muxer is one line per stream, like this: streamindex,streamtype,algo=hash, where streamindex is the index of the mapped stream, streamtype is a single character indicating the type of stream, algo is a short string representing the hash function used, and hash is a hexadecimal number representing the computed hash.

hash algorithm — use the cryptographic hash function specified by the string algorithm. Supported values include "MD5", "murmur3", "RIPEMD128", "RIPEMD160", "RIPEMD256", "RIPEMD320", "SHA160", "SHA224", "SHA256" (default), "SHA512/224", "SHA512/256", "SHA384", "SHA512", "CRC32" and "adler32".

🎀 Examples 🎀

To compute the SHA-256 hash of the input converted to raw audio and video, and store it in the file out.sha256:

ffmpeg -i INPUT -f streamhash out.sha256

To print an MD5 hash to stdout use the command:

ffmpeg -i INPUT -f streamhash -hash md5 -

See also the hash and framehash muxers.

👑 tee 👑

The tee muxer can be used to write the same data to several outputs, such as files or streams. It can be used, for example, to stream a video over a network and save it to disk at the same time. Literally the multitasking queen.

It is different from specifying several outputs to the ffmpeg command-line tool. With the tee muxer, the audio and video data will be encoded only once. With conventional multiple outputs, multiple encoding operations in parallel are initiated, which can be a very expensive process. The tee muxer is not useful when using the libavformat API directly because it is then possible to feed the same packets to several muxers directly.

Since the tee muxer does not represent any particular output format, ffmpeg cannot auto-select output streams. So all streams intended for output must be specified using -map. See the examples below.

Some encoders may need different options depending on the output format; the auto-detection of this can not work with the tee muxer, so they need to be explicitly specified. The main example is the global_header flag.

The slave outputs are specified in the file name given to the muxer, separated by |. If any of the slave name contains the | separator, leading or trailing spaces or any special character, those must be escaped (see the "Quoting and escaping" section in the ffmpeg-utils(1) manual).

💅 Options 💅

use_fifo bool — if set to 1, slave outputs will be processed in separate threads using the fifo muxer. This allows you to compensate for different speed/latency/reliability of outputs and setup transparent recovery. By default this feature is turned off.
fifo_options — options to pass to fifo pseudo-muxer instances. See fifo.

Muxer options can be specified for each slave by prepending them as a list of key=value pairs separated by :, between square brackets. If the options values contain a special character or the : separator, they must be escaped; note that this is a second level escaping.

The following special options are also recognized:

f — specify the format name. Required if it cannot be guessed from the output URL.
bsfs[/spec] — specify a list of bitstream filters to apply to the specified output.

It is possible to specify to which streams a given bitstream filter applies, by appending a stream specifier to the option separated by /. spec must be a stream specifier (see Format stream specifiers).

If the stream specifier is not specified, the bitstream filters will be applied to all streams in the output. This will cause that output operation to fail if the output contains streams to which the bitstream filter cannot be applied e.g. h264_mp4toannexb being applied to an output containing an audio stream.

Options for a bitstream filter must be specified in the form of opt=value.

Several bitstream filters can be specified, separated by ,.

use_fifo bool — this allows you to override tee muxer use_fifo option for an individual slave muxer.
fifo_options — this allows you to override tee muxer fifo_options for an individual slave muxer. See fifo.
select — select the streams that should be mapped to the slave output, specified by a stream specifier. If not specified, this defaults to all the mapped streams. This will cause that output operation to fail if the output format does not accept all mapped streams.

You may use multiple stream specifiers separated by commas (,) e.g.: a:0,v.

onfail — tells ffmpeg what to do when an output in this tee group dies 💀. Set it to "abort" (that's the default) or "ignore". "abort" means the whole process fails the moment this slave output fails. "ignore" means the failure gets shrugged off and the other outputs keep vibing along, totally unaffected 🫶.

Example time ✨

💅 Examples

Encode something and both archive it in a WebM file and stream it as MPEG-TS over UDP:

ffmpeg -i ... -c:v libx264 -c:a mp2 -f tee -map 0:v -map 0:a
  "archive-20121107.mkv|[f=mpegts]udp://10.0.1.255:1234/"

Same thing as above, but keep streaming even if the output to the local file fails (say your local drive fills up 😭):

ffmpeg -i ... -c:v libx264 -c:a mp2 -f tee -map 0:v -map 0:a
  "[onfail=ignore]archive-20121107.mkv|[f=mpegts]udp://10.0.1.255:1234/"

Use ffmpeg to encode the input and send the output to three different destinations. The dump_extra bitstream filter is used to add extradata information to all the output video keyframes packets, as requested by the MPEG-TS format. The select option is applied to out.aac in order to make it contain only audio packets 🎧:

ffmpeg -i ... -map 0 -flags +global_header -c:v libx264 -c:a aac
       -f tee "[bsfs/v=dump_extra=freq=keyframe]out.ts|[movflags=+faststart]out.mp4|[select=a]out.aac"

As above, but select only stream "a:1" for the audio output. Note that a second level escaping must be performed, as ":" is a special character used to separate options — so yeah, double-escape or it will not slay 🙈:

ffmpeg -i ... -map 0 -flags +global_header -c:v libx264 -c:a aac
       -f tee "[bsfs/v=dump_extra=freq=keyframe]out.ts|[movflags=+faststart]out.mp4|[select=\'a:1\']out.aac"

🦋 webm_chunk

WebM Live Chunk Muxer, bestie 🎀.

This muxer writes out WebM headers and chunks as separate files, which can be consumed by clients that support WebM Live streams via DASH.

Options

This muxer supports the following options:

chunk_start_index — the index of the first chunk (defaults to 0).
header — filename of the header, i.e. where the initialization data will be written.
audio_chunk_duration — duration of each audio chunk in milliseconds (defaults to 5000).

Example

ffmpeg -f v4l2 -i /dev/video0 \
       -f alsa -i hw:0 \
       -map 0:0 \
       -c:v libvpx-vp9 \
       -s 640x360 -keyint_min 30 -g 30 \
       -f webm_chunk \
       -header webm_live_video_360.hdr \
       -chunk_start_index 1 \
       webm_live_video_360_%d.chk \
       -map 1:0 \
       -c:a libvorbis \
       -b:a 128k \
       -f webm_chunk \
       -header webm_live_audio_128.hdr \
       -chunk_start_index 1 \
       -audio_chunk_duration 1000 \
       webm_live_audio_128_%d.chk

🌈 webm_dash_manifest

WebM DASH Manifest muxer, iconic behaviour only 👑.

This muxer implements the WebM DASH Manifest specification to generate the DASH manifest XML. It also supports manifest generation for DASH live streams 🔥.

For more information see:

WebM DASH Specification: <https://sites.google.com/a/webmproject.org/wiki/adaptive-streaming/webm-dash-specification>
ISO DASH Specification: <http://standards.iso.org/ittf/PubliclyAvailableStandards/c065274_ISO_IEC_23009-1_2014.zip>

Options

This muxer supports the following options:

adaptation_sets — this option has the following syntax: "id=x,streams=a,b,c id=y,streams=d,e" where x and y are the unique identifiers of the adaptation sets and a,b,c,d and e are the indices of the corresponding audio and video streams. Any number of adaptation sets can be added using this option 💖.
live — set this to 1 to create a live stream DASH Manifest. Default: 0.
chunk_start_index — start index of the first chunk. This will go in the startNumber attribute of the SegmentTemplate element in the manifest. Default: 0.
chunk_duration_ms — duration of each chunk in milliseconds. This will go in the duration attribute of the SegmentTemplate element in the manifest. Default: 1000.
utc_timing_url — URL of the page that will return the UTC timestamp in ISO format. This will go in the value attribute of the UTCTiming element in the manifest. Default: None.
time_shift_buffer_depth — smallest time (in seconds) shifting buffer for which any Representation is guaranteed to be available. This will go in the timeShiftBufferDepth attribute of the MPD element. Default: 60.
minimum_update_period — minimum update period (in seconds) of the manifest. This will go in the minimumUpdatePeriod attribute of the MPD element. Default: 0.

Example

ffmpeg -f webm_dash_manifest -i video1.webm \
       -f webm_dash_manifest -i video2.webm \
       -f webm_dash_manifest -i audio1.webm \
       -f webm_dash_manifest -i audio2.webm \
       -map 0 -map 1 -map 2 -map 3 \
       -c copy \
       -f webm_dash_manifest \
       -adaptation_sets "id=0,streams=0,1 id=1,streams=2,3" \
       manifest.xml

⚡ whip

WebRTC (Real-Time Communication) muxer that supports sub-second latency streaming according to the WHIP (WebRTC-HTTP ingestion protocol) specification. Lowkey the coolest one here ngl 🌸.

⚠️ Heads up bestie: this is an experimental feature. Not kidding around.

It uses HTTP as a signaling protocol to exchange SDP capabilities and ICE lite candidates. Then, it uses STUN binding requests and responses to establish a session over UDP. Subsequently, it initiates a DTLS handshake to exchange the SRTP encryption keys. Lastly, it splits video and audio frames into RTP packets and encrypts them using SRTP. Like, the whole pipeline, start to finish 💫.

Ensure that you use H.264 without B frames and Opus for the audio codec. For example, to convert an input file with ffmpeg to WebRTC:

ffmpeg -re -i input.mp4 -acodec libopus -ar 48000 -ac 2 \
  -vcodec libx264 -profile:v baseline -tune zerolatency -threads 1 -bf 0 \
  -f whip "http://localhost:1985/rtc/v1/whip/?app=live&stream=livestream"

For this example, we have employed low latency options, resulting in an end-to-end latency of approximately 150ms. Ate, honestly 🍓.

Options

This muxer supports the following options:

handshake_timeout integer — set the timeout in milliseconds for ICE and DTLS handshake. Default value is 5000.
pkt_size integer — set the maximum size, in bytes, of RTP packets that send out. Default value is 1200.
ts_buffer_size integer — set the buffer size, in bytes, of underlying protocol. Default value is -1(auto). The UDP auto selects a reasonable value.
whip_flags flags — possible values:
- dtls_active — the muxer will try to set dtls active role and send the first client hello.
rtp_history integer — set the number of RTP history items to store. Default value is 512.
authorization string — the optional Bearer token for WHIP Authorization 🔐.
cert_file string — the optional certificate file path for DTLS.
key_file string — the optional private key file path for DTLS.

🎵 METADATA

FFmpeg is able to dump metadata from media files into a simple UTF-8-encoded INI-like text file and then load it back using the metadata muxer/demuxer. It's giving notes-app-but-make-it-a-container-format 📝.

The file format is as follows:

A file consists of a header and a number of metadata tags divided into sections, each on its own line.
The header is a ;FFMETADATA string, followed by a version number (now 1).
Metadata tags are of the form key=value
Immediately after header follows global metadata
After global metadata there may be sections with per-stream/per-chapter metadata.
A section starts with the section name in uppercase (i.e. STREAM or CHAPTER) in brackets ([, ]) and ends with next section or end of file.
At the beginning of a chapter section there may be an optional timebase to be used for start/end values. It must be in form TIMEBASE=num/den, where num and den are integers. If the timebase is missing then start/end times are assumed to be in nanoseconds.

Next a chapter section must contain chapter start and end times in form START=num, END=num, where num is a positive integer.

Empty lines and lines starting with ; or # are ignored.
Metadata keys or values containing special characters (=, ;, #, \ and a newline) must be escaped with a backslash \.
Note that whitespace in metadata (e.g. foo = bar) is considered to be a part of the tag (in the example above key is foo , value is bar). Yeah, the space is part of the vibe, don't @ me 🙈.

A ffmetadata file might look like this:

;FFMETADATA1
title=bike\\shed
;this is a comment
artist=FFmpeg troll team

[CHAPTER]
TIMEBASE=1/1000
START=0
#chapter ends at 0:01:00
END=60000
title=chapter \#1
[STREAM]
title=multi\
line

By using the ffmetadata muxer and demuxer it is possible to extract metadata from an input file to an ffmetadata file, and then transcode the file into an output file with the edited ffmetadata file.

Extracting an ffmetadata file with ffmpeg goes as follows:

ffmpeg -i INPUT -f ffmetadata FFMETADATAFILE

Reinserting edited metadata information from the FFMETADATAFILE file can be done as:

ffmpeg -i INPUT -i FFMETADATAFILE -map_metadata 1 -codec copy OUTPUT

🔌 PROTOCOL OPTIONS

The libavformat library provides some generic global options, which can be set on all the protocols. In addition each protocol may support so-called private options, which are specific for that component 👑.

Options may be set by specifying -option value in the FFmpeg tools, or by setting the value explicitly in the AVFormatContext options or using the libavutil/opt.h API for programmatic use.

The list of supported options follows:

protocol_whitelist list (input) — set a ","-separated list of allowed protocols. "ALL" matches all protocols. Protocols prefixed by - are disabled. All protocols are allowed by default but protocols used by an another protocol (nested protocols) are restricted to a per protocol subset 🫶.

🌐 PROTOCOLS

Protocols are configured elements in FFmpeg that enable access to resources that require specific protocols.

When you configure your FFmpeg build, all the supported protocols are enabled by default. You can list all available ones using the configure option --list-protocols ✨.

okay so bestie, you can disable all the protocols with the configure option --disable-protocols, and then selectively enable one with --enable-protocol=PROTOCOL, or you can disable one particular protocol with --disable-protocol=PROTOCOL. ✨

The -protocols option of the ff* tools will display the list of supported protocols. 💅

All protocols accept the following option:

rw_timeout — maximum time to wait for (network) read/write operations to complete, in microseconds. ⏱️

A description of the currently available protocols follows — let’s go 🫶

🐇 amqp

Advanced Message Queueing Protocol (AMQP) version 0-9-1 is a broker based publish-subscribe communication protocol. iconic little message-broker behavior 💌

FFmpeg must be compiled with --enable-librabbitmq to support AMQP. A separate AMQP broker must also be run. An example open-source AMQP broker is RabbitMQ. 🐰

After starting the broker, an FFmpeg client may stream data to the broker using the command:

ffmpeg -re -i input -f mpegts amqp://[[user]:[password]@]hostname[:port][/vhost]

Where hostname and port (default is 5672) is the address of the broker. The client may also set a user/password for authentication. The default for both fields is "guest". Name of virtual host on broker can be set with vhost. The default value is "/". 🌸

Multiple subscribers may stream from the broker using the command:

ffplay amqp://[[user]:[password]@]hostname[:port][/vhost]

In RabbitMQ all data published to the broker flows through a specific exchange, and each subscribing client has an assigned queue/buffer. When a packet arrives at an exchange, it may be copied to a client’s queue depending on the exchange and routing_key fields. 🎀

The following options are supported:

exchange — sets the exchange to use on the broker. RabbitMQ has several predefined exchanges: "amq.direct" is the default exchange, where the publisher and subscriber must have a matching routing_key; "amq.fanout" is the same as a broadcast operation (i.e. the data is forwarded to all queues on the fanout exchange independent of the routing_key); and "amq.topic" is similar to "amq.direct", but allows for more complex pattern matching (refer to the RabbitMQ documentation). 💫
routing_key — sets the routing key. The default value is "amqp". The routing key is used on the "amq.direct" and "amq.topic" exchanges to decide whether packets are written to the queue of a subscriber. 🔑
pkt_size — maximum size of each packet sent/received to the broker. Default is 131072. Minimum is 4096 and max is any large value (representable by an int). When receiving packets, this sets an internal buffer size in FFmpeg. It should be equal to or greater than the size of the published packets to the broker. Otherwise the received message may be truncated causing decoding errors. 😭
connection_timeout — the timeout in seconds during the initial connection to the broker. The default value is rw_timeout, or 5 seconds if rw_timeout is not set. ⏳
delivery_mode mode — sets the delivery mode of each message sent to broker. The following values are accepted:
- persistent — delivery mode set to "persistent" (2). This is the default value. Messages may be written to the broker’s disk depending on its setup. 💾
- non-persistent — delivery mode set to "non-persistent" (1). Messages will stay in broker’s memory unless the broker is under memory pressure. 🧠

🦋 async

Asynchronous data filling wrapper for input stream. 🌈

Fill data in a background thread, to decouple I/O operation from demux thread. multitasking queen behavior 💜

async:<URL>
async:http://host/resource
async:cache:http://host/resource

💿 bluray

Read BluRay playlist. 👑

The accepted options are:

angle — BluRay angle 🎬
chapter — start chapter (1...N) 📖
playlist — playlist to read (BDMV/PLAYLIST/?????.mpls) 🗂️

Examples:

Read longest playlist from BluRay mounted to /mnt/bluray:

bluray:/mnt/bluray

Read angle 2 of playlist 4 from BluRay mounted to /mnt/bluray, start from chapter 2:

-playlist 4 -angle 2 -chapter 2 bluray:/mnt/bluray

🧺 cache

Caching wrapper for input stream. 🍓

Cache the input stream to temporary file. It brings seeking capability to live streams. lowkey magical ✨

The accepted options are:

read_ahead_limit — amount in bytes that may be read ahead when seeking isn’t supported. Range is -1 to INT_MAX. -1 for unlimited. Default is 65536. 📏

URL Syntax is

cache:<URL>

🔗 concat

Physical concatenation protocol. 🧵

Read and seek from many resources in sequence as if they were a unique resource. one big happy file 💖

A URL accepted by this protocol has the syntax:

concat:<URL1>|<URL2>|...|<URLN>

where URL1, URL2, ..., URLN are the urls of the resource to be concatenated, each one possibly specifying a distinct protocol. 🌟

For example to read a sequence of files split1.mpeg, split2.mpeg, split3.mpeg with ffplay use the command:

ffplay concat:split1.mpeg\|split2.mpeg\|split3.mpeg

Note that you may need to escape the character "|" which is special for many shells. 😭💅

📝 concatf

Physical concatenation protocol using a line break delimited list of resources. 🗒️

Read and seek from many resources in sequence as if they were a unique resource. 🫶

A URL accepted by this protocol has the syntax:

concatf:<URL>

where URL is the url containing a line break delimited list of resources to be concatenated, each one possibly specifying a distinct protocol. Special characters must be escaped with backslash or single quotes. See the "Quoting and escaping" section in the ffmpeg-utils(1) manual. 🧚

For example to read a sequence of files split1.mpeg, split2.mpeg, split3.mpeg listed in separate lines within a file split.txt with ffplay use the command:

ffplay concatf:split.txt

Where split.txt contains the lines:

split1.mpeg
split2.mpeg
split3.mpeg

🔐 crypto

AES-encrypted stream reading protocol. 🕵️‍♀️

The accepted options are:

key — set the AES decryption key binary block from given hexadecimal representation. 🔑
iv — set the AES decryption initialization vector binary block from given hexadecimal representation. 🧬

Accepted URL formats:

crypto:<URL>
crypto+<URL>

🧁 data

Data in-line in the URI. See http://en.wikipedia.org/wiki/Data_URI_scheme. 🌐

For example, to convert a GIF file given inline with ffmpeg:

ffmpeg -i "data:image/gif;base64,R0lGODdhCAAIAMIEAAAAAAAA//8AAP//AP///////////////ywAAAAACAAIAAADF0gEDLojDgdGiJdJqUX02iB4E8Q9jUMkADs=" smiley.png

🪄 fd

File descriptor access protocol. 🎀

The accepted syntax is:

fd: -fd <file_descriptor>

If fd is not specified, by default the stdout file descriptor will be used for writing, stdin for reading. Unlike the pipe protocol, fd protocol has seek support if it corresponding to a regular file. fd protocol doesn’t support pass file descriptor via URL for security. 🔒

This protocol accepts the following options:

blocksize — set I/O operation maximum block size, in bytes. Default value is "INT_MAX", which results in not limiting the requested block size. Setting this value reasonably low improves user termination request reaction time, which is valuable if data transmission is slow. 🧱
fd — set file descriptor. 🪪

📁 file

File access protocol. 💖

Read from or write to a file. 🗂️

A file URL can have the form:

file:<filename>

where filename is the path of the file to read. 🌸

An URL that does not have a protocol prefix will be assumed to be a file URL. Depending on the build, an URL that looks like a Windows path with the drive letter at the beginning will also be assumed to be a file URL (usually not the case in builds for unix-like systems). 🪟

For example to read from a file input.mpeg with ffmpeg use the command:

ffmpeg -i file:input.mpeg output.mpeg

This protocol accepts the following options:

truncate — truncate existing files on write, if set to 1. A value of 0 prevents truncating. Default value is 1. ✂️
blocksize — set I/O operation maximum block size, in bytes. Default value is "INT_MAX", which results in not limiting the requested block size. Setting this value reasonably low improves user termination request reaction time, which is valuable for files on slow medium. 🐢
follow — if set to 1, the protocol will retry reading at the end of the file, allowing reading files that still are being written. In order for this to terminate, you either need to use the rw_timeout option, or use the interrupt callback (for API users). 👀
seekable — controls if seekability is advertised on the file. 0 means non-seekable, -1 means auto (seekable for normal files, non-seekable for named pipes). Many demuxers handle seekable and non-seekable resources differently, overriding this might speed up opening certain files at the cost of losing some features (e.g. accurate seeking). 🧭
pkt_size — set the maximum packet size used for file I/O. A smaller value may reduce memory usage. A higher value may increase throughput especially with networked filesystems. For reading, if explicitly set, it overrides the default internal buffer size (32 KB) and limits the maximum amount of data read per operation. For writing, this sets the size of each write operation. The default is 256 KB for regular files, 32 KB otherwise. 📦

🌐 ftp

FTP (File Transfer Protocol). 📡

Read from or write to remote resources using FTP protocol. 🫶

Following syntax is required.

ftp://[user[:password]@]server[:port]/path/to/remote/resource.mpeg

This protocol accepts the following options. 🎀

timeout — set timeout in microseconds of socket I/O operations used by the underlying low level operation. By default it is set to -1, which means that the timeout is not specified. ⏱️
ftp-user — set a user to be used for authenticating to the FTP server. This is overridden by the user in the FTP URL. 👤
ftp-password — set a password to be used for authenticating to the FTP server. This is overridden by the password in the FTP URL, or by ftp-anonymous-password if no user is set. 🔑
ftp-anonymous-password — password used when login as anonymous user. Typically an e-mail address should be used. 📧
ftp-write-seekable — control seekability of connection during encoding. If set to 1 the resource is supposed to be seekable, if set to 0 it is assumed not to be seekable. Default value is 0. 🎯

🫶 NOTE: Protocol can be used as output, but it is recommended to not do it, unless special care is taken (tests, customized server configuration etc.). Different FTP servers behave in different way during seek operation. ff* tools may produce incomplete content due to server limitations.

gopher 🦋

Gopher protocol. That's it, that's the whole vibe. Old school but still iconic.

gophers ✨

Gophers protocol — literally gopher but with a glow-up.

The Gopher protocol with TLS encapsulation. Secure vibes only 💅

http 👑

HTTP (Hyper Text Transfer Protocol). The main character of protocols, no cap.

This protocol accepts the following options:

seekable — Controls whether the connection can seek. Set it to 1 and the resource is supposed to be seekable, set it to 0 and it's assumed not seekable, and set it to -1 and it'll try to autodetect seekability for you. Default value is -1.
chunked_post — If set to 1 use chunked Transfer-Encoding for posts. Default is 1.
http_proxy — Sets the HTTP proxy to tunnel through, like e.g. http://example.com:1234.
headers — Set custom HTTP headers — you can override the built-in default headers with this. The value must be a string encoding the headers.
content_type — Set a specific content type for the POST messages, or for listen mode.
user_agent — Override the User-Agent header. If you don't specify one, the protocol will use a string describing the libavformat build. ("Lavf/<version>")
referer — Sets the Referer header. Includes the 'Referer: URL' header in the HTTP request.
multiple_requests — Use persistent connections if set to 1. Default is 0.
request_size — Limits the size of requests made. This is useful for some pathological servers that throttle unbounded range requests, and also when you're expecting to seek frequently. Disabled (set to 0) by default.

Note that if you enable this option, it's strongly recommended to also enable the multiple_requests option, as well as setting short_seek_size to the same value or higher. Doing so lets FFmpeg reuse a single HTTP connection wherever possible. Recycle, bestie ♻️

initial_request_size — Limits the size of initial requests. Similar to request_size, but only used during initial format parsing. Useful for formats like MXF or MOV that require frequent seeks during header parsing. It lasts until the demuxer makes a read request larger than this size (without a seek in between), after which the implementation will carry on using requests as usual. Disabled (set to 0) by default.

Note that if you enable this option, it's strongly recommended to also enable the multiple_requests option, as well as setting short_seek_size to the same value or higher.

post_data — Set custom HTTP post data.
mime_type — Exports the MIME type.
http_version — Exports the HTTP response version number. Usually "1.0" or "1.1".
cookies — Sets the cookies to be sent in future requests. The format of each cookie is the same as the value of a Set-Cookie HTTP response field. Multiple cookies can be delimited by a newline character.
icy — If set to 1 request ICY (SHOUTcast) metadata from the server. If the server supports this, the metadata has to be retrieved by the application by reading the icy_metadata_headers and icy_metadata_packet options. The default is 1.
icy_metadata_headers — If the server supports ICY metadata, this contains the ICY-specific HTTP reply headers, separated by newline characters.
icy_metadata_packet — If the server supports ICY metadata, and icy was set to 1, this contains the last non-empty metadata packet sent by the server. Applications interested in mid-stream metadata updates should poll it at regular intervals.
metadata — Set an exported dictionary containing Icecast metadata from the bitstream, if present. Only useful with the C API.
auth_type — Set the HTTP authentication type. There's no option for Digest, because that method requires getting nonce parameters from the server first and can't be used straight away like Basic.
- none — Choose the HTTP authentication type automatically. This is the default.
- basic — Choose HTTP basic authentication.

⚠️ Okay so listen up because this one matters: Basic authentication sends a Base64-encoded string that contains a user name and password for the client. Base64 is not a form of encryption and should be considered the same as sending the user name and password in clear text (Base64 is a reversible encoding). If a resource needs to be protected, strongly consider using an authentication scheme other than basic authentication. HTTPS/TLS should be used with basic authentication. Without these additional security enhancements, basic authentication should not be used to protect sensitive or valuable information. Not me warning you twice 💀

send_expect_100 — Sends an Expect: 100-continue header for POST. If set to 1 it will send, if set to 0 it won't, and if set to -1 it'll try to send if it's applicable. Default value is -1.
location — An exported dictionary containing the content location. Only useful with the C API.
offset — Set initial byte offset.
end_offset — Try to limit the request to bytes preceding this offset.
method — When used as a client option it sets the HTTP method for the request.

When used as a server option it sets the HTTP method that is going to be expected from the client(s). If the expected and the received HTTP method don't match, the client will be given a Bad Request response. When unset, the HTTP method is not checked for now. This will be replaced by autodetection in the future.

reconnect — Reconnect automatically when disconnected before EOF is hit.
reconnect_at_eof — If set then eof is treated like an error and causes reconnection. This is useful for live / endless streams.
reconnect_on_network_error — Reconnect automatically in case of TCP/TLS errors during connect.
reconnect_on_http_error — A comma separated list of HTTP status codes to reconnect on. The list can include specific status codes (e.g. '503') or the strings '4xx' / '5xx'.
reconnect_streamed — If set then even streamed/non seekable streams will be reconnected on errors.
reconnect_delay_max — Set the maximum delay in seconds after which to give up reconnecting.
reconnect_max_retries — Set the maximum number of times to retry a connection. Default unset.
reconnect_delay_total_max — Set the maximum total delay in seconds after which to give up reconnecting.
respect_retry_after — If enabled, and a Retry-After header is encountered, its requested reconnection delay will be honored, rather than using exponential backoff. Useful for 429 and 503 errors. Default enabled.
listen — If set to 1 enables experimental HTTP server. This can be used to send data when used as an output option, or read data from a client with HTTP POST when used as an input option. If set to 2 enables experimental multi-client HTTP server. This is not yet implemented in ffmpeg.c and thus must not be used as a command line option.

# Server side (sending):
ffmpeg -i somefile.ogg -c copy -listen 1 -f ogg http://<server>:<port>

# Client side (receiving):
ffmpeg -i http://<server>:<port> -c copy somefile.ogg

# Client can also be done with wget:
wget http://<server>:<port> -O somefile.ogg

# Server side (receiving):
ffmpeg -listen 1 -i http://<server>:<port> -c copy somefile.ogg

# Client side (sending):
ffmpeg -i somefile.ogg -chunked_post 0 -c copy -f ogg http://<server>:<port>

# Client can also be done with wget:
wget --post-file=somefile.ogg http://<server>:<port>

resource — The resource requested by a client, when the experimental HTTP server is in use.
reply_code — The HTTP code returned to the client, when the experimental HTTP server is in use.
short_seek_size — Set the threshold, in bytes, for when a readahead should be preferred over a seek and new HTTP request. This is useful, for example, to make sure the same connection is used for reading large video packets with small audio packets in between 🎀

HTTP Cookies 🍪

Some HTTP requests will be denied unless cookie values are passed in with the request. The cookies option lets you specify these cookies. At the very least, each cookie must specify a value along with a path and domain. HTTP requests that match both the domain and path will automatically include the cookie value in the HTTP Cookie header field. Multiple cookies can be delimited by a newline.

The required syntax to play a stream specifying a cookie is:

ffplay -cookies "nlqptid=nltid=tsn; path=/; domain=somedomain.com;" http://somedomain.com/somestream.m3u8

Icecast 🌈

Icecast protocol (stream to Icecast servers). Radio-host era, we love to see it ✨

This protocol accepts the following options:

ice_genre — Set the stream genre.
ice_name — Set the stream name.
ice_description — Set the stream description.
ice_url — Set the stream website URL.
ice_public — Set if the stream should be public. The default is 0 (not public).
user_agent — Override the User-Agent header. If not specified a string of the form "Lavf/<version>" will be used.
password — Set the Icecast mountpoint password.
content_type — Set the stream content type. This must be set if it is different from audio/mpeg.
legacy_icecast — This enables support for Icecast versions < 2.4.0, that do not support the HTTP PUT method but the SOURCE method.
tls — Establish a TLS (HTTPS) connection to Icecast.

icecast://[<username>[:<password>]@]<server>:<port>/<mountpoint>

ipfs 🧚

InterPlanetary File System (IPFS) protocol support. You can access files stored on the IPFS network through so-called gateways. These are http(s) endpoints. This protocol wraps the IPFS native protocols (ipfs:// and ipns://) to be sent to such a gateway. Users can (and should) host their own node, which means this protocol will use your local gateway to access files on the IPFS network.

This protocol accepts the following options:

gateway — Defines the gateway to use. When not set, the protocol will first try locating the local gateway by looking at $IPFS_GATEWAY, $IPFS_PATH and "$HOME/.ipfs/", in that order.

You can use this protocol in 2 ways. Using IPFS:

ffplay ipfs://<hash>

Or the IPNS protocol (IPNS is mutable IPFS):

ffplay ipns://<hash>

okay that's the ipns moment wrapped, moving on bestie 💅

`mmst` 🪟

MMS (Microsoft Media Server) protocol, riding on top of TCP. that's it, that's the whole vibe.

`mmsh` 🪟✨

Same MMS energy but over HTTP instead. The required syntax is, and I quote:

mmsh://<server>[:<port>][/<app>][/<playpath>]

`md5` 🧮💖

The MD5 output protocol. It computes the MD5 hash of the data that's being written, and then on close it writes that hash to the designated output — or to stdout if you didn't specify one. It's literally the move for testing muxers without committing to writing an actual file. iconique for debugging, no cap.

Some examples, because bestie loves an example:

# Write the MD5 hash of the encoded AVI file to the file output.avi.md5.
ffmpeg -i input.flv -f avi -y md5:output.avi.md5

# Write the MD5 hash of the encoded AVI file to stdout.
ffmpeg -i input.flv -f avi -y md5:

⚠️ heads up: some formats (typically MOV) require the output protocol to be seekable, so they will fail with the MD5 output protocol. not a vibe, but it is the rules 😭

`pipe` 🚰🫶

The UNIX pipe access protocol. Read and write from UNIX pipes, that's the whole job.

The accepted syntax is:

pipe:[<number>]

If fd isn't specified, number is the number corresponding to the file descriptor of the pipe (e.g. 0 for stdin, 1 for stdout, 2 for stderr). And if number isn't specified at all, then by default the stdout file descriptor gets used for writing, and stdin for reading. slay, sensible defaults 💅

So like, to read from stdin with ffmpeg:

cat test.wav | ffmpeg -i pipe:0
# ...this is the same as...
cat test.wav | ffmpeg -i pipe:

And to write to stdout with ffmpeg:

ffmpeg -i test.wav -f avi pipe:1 | cat > test.avi
# ...this is the same as...
ffmpeg -i test.wav -f avi pipe: | cat > test.avi

This protocol accepts the following options:

blocksize — sets the I/O operation maximum block size, in bytes. Default value is INT_MAX, which results in not limiting the requested block size at all. Setting this value reasonably low improves your user-termination-request reaction time, which is genuinely valuable if the data transmission is slow. lowkey a glow-up for responsiveness ✨
fd — set file descriptor.

⚠️ and again, the classic: some formats (typically MOV) require the output protocol to be seekable, so they will fail with the pipe output protocol. 😭

`prompeg` 📡🔥

The Pro-MPEG Code of Practice #3 Release 2 FEC protocol.

So the Pro-MPEG CoP#3 FEC is a 2D parity-check forward error correction mechanism for MPEG-2 Transport Streams sent over RTP. very "main character of error correction" if you ask me.

This protocol must be used in conjunction with the rtp_mpegts muxer and the rtp protocol. they're a package deal, iykyk.

The required syntax is:

-f rtp_mpegts -fec prompeg=<option>=<val>... rtp://<hostname>:<port>

The destination UDP ports are port + 2 for the column FEC stream and port + 4 for the row FEC stream.

This protocol accepts the following options:

l=n — the number of columns (4-20, LxD <= 100)
d=n — the number of rows (4-20, LxD <= 100)

Example usage:

-f rtp_mpegts -fec prompeg=l=8:d=4 rtp://<hostname>:<port>

`rist` 🌐💜

Reliable Internet Streaming Transport protocol. The accepted options are:

rist_profile — supported values:
- simple
- main — this one is default.
- advanced
buffer_size — set the internal RIST buffer size in milliseconds for retransmission of data. Default value is 0, which means the librist default (1 sec). Maximum value is 30 seconds.
fifo_size — size of the librist receiver output fifo in number of packets. This must be a power of 2. Defaults to 8192 (vs the librist default of 1024).
overrun_nonfatal=1|0 — survive in case of librist fifo buffer overrun. Default value is 0.
pkt_size — set maximum packet size for sending data. 1316 by default.
log_level — set loglevel for RIST logging messages. You only need to set this if you explicitly want to enable debug level messages or packet loss simulation, otherwise the regular loglevel is respected.
secret — set override of encryption secret, by default is unset.
encryption — set encryption type, by default is disabled. Acceptable values are 128 and 256.

`rtmp` 👑📺

Real-Time Messaging Protocol. The Real-Time Messaging Protocol (RTMP) is used for streaming multimedia content across a TCP/IP network. the blueprint, fr fr.

The required syntax is:

rtmp://[<username>:<password>@]<server>[:<port>][/<app>][/<instance>][/<playpath>]

The accepted parameters are:

username — an optional username (mostly for publishing).
password — an optional password (mostly for publishing).
server — the address of the RTMP server.
port — the number of the TCP port to use (by default is 1935).
app — it is the name of the application to access. It usually corresponds to the path where the application is installed on the RTMP server (e.g. /ondemand/, /flash/live/, etc.). You can override the value parsed from the URI through the rtmp_app option, too.
playpath — it is the path or name of the resource to play with reference to the application specified in app, may be prefixed by mp4:. You can override the value parsed from the URI through the rtmp_playpath option, too.
listen — act as a server, listening for an incoming connection.
timeout — maximum time to wait for the incoming connection. Implies listen.

Additionally, the following parameters can be set via command line options (or in code via AVOptions):

rtmp_app — name of application to connect on the RTMP server. This option overrides the parameter specified in the URI.
rtmp_buffer — set the client buffer time in milliseconds. The default is 3000.
rtmp_conn — extra arbitrary AMF connection parameters, parsed from a string, e.g. like B:1 S:authMe O:1 NN:code:1.23 NS:flag:ok O:0. Each value is prefixed by a single character denoting the type: B for Boolean, N for number, S for string, O for object, or Z for null, followed by a colon. For Booleans the data must be either 0 or 1 for FALSE or TRUE, respectively. Likewise for Objects the data must be 0 or 1 to end or begin an object, respectively. Data items in subobjects may be named, by prefixing the type with N and specifying the name before the value (i.e. NB:myFlag:1). This option may be used multiple times to construct arbitrary AMF sequences. very build-your-own-adventure of it 🎀
rtmp_enhanced_codecs — specify the list of codecs the client advertises to support in an enhanced RTMP stream. This option should be set to a comma separated list of fourcc values, like hvc1,av01,vp09 for multiple codecs or hvc1 for only one codec. The specified list will be presented in the fourCcLive property of the Connect Command Message.
rtmp_flashver — version of the Flash plugin used to run the SWF player. The default is LNX 9,0,124,2. (When publishing, the default is FMLE/3.0 (compatible; <libavformat version>).) yeah we're still honoring the Flash lore in this economy 💀
rtmp_flush_interval — number of packets flushed in the same request (RTMPT only). The default is 10.
rtmp_live — specify that the media is a live stream. No resuming or seeking in live streams is possible. The default value is any, which means the subscriber first tries to play the live stream specified in the playpath. If a live stream of that name is not found, it plays the recorded stream. The other possible values are live and recorded.
rtmp_pageurl — URL of the web page in which the media was embedded. By default no value will be sent.
rtmp_playpath — stream identifier to play or to publish. This option overrides the parameter specified in the URI.
rtmp_subscribe — name of live stream to subscribe to. By default no value will be sent. It is only sent if the option is specified or if rtmp_live is set to live.
rtmp_swfhash — SHA256 hash of the decompressed SWF file (32 bytes).
rtmp_swfsize — size of the decompressed SWF file, required for SWFVerification.
rtmp_swfurl — URL of the SWF player for the media. By default no value will be sent.
rtmp_swfverify — URL to player swf file, compute hash/size automatically.
rtmp_tcurl — URL of the target stream. Defaults to proto://host[:port]/app.
tcp_nodelay=1|0 — set TCP_NODELAY to disable Nagle's algorithm. Default value is 0.

remark: writing to the socket is currently not optimized to minimize system calls and reduces the efficiency / effect of TCP_NODELAY. so like, don't expect miracles 🙈

tcp_keepalive=1|0 — enable the TCP keepalive mechanism to detect dead peers and help maintain long-lived idle connections. Default value is 0.

Only the basic keepalive option (SO_KEEPALIVE) can be enabled or disabled. Platform-specific tuning parameters such as TCP_KEEPIDLE, TCP_KEEPINTVL, or TCP_KEEPCNT are not configurable and will use the operating system's default values.

For example to read with ffplay a multimedia resource named "sample" from the application "vod" from an RTMP server "myserver":

ffplay rtmp://myserver/vod/sample

To publish to a password protected server, passing the playpath and app names separately:

ffmpeg -re -i <input> -f flv -rtmp_playpath some/long/path -rtmp_app long/app/name rtmp://username:password@myserver/

`rtmpe` 🔐

Encrypted Real-Time Messaging Protocol. RTMPE is used for streaming multimedia content within standard cryptographic primitives, consisting of Diffie-Hellman key exchange and HMACSHA256, generating a pair of RC4 keys. look at her go with the cryptography 💅✨

`rtmps` 🛡️

Real-Time Messaging Protocol over a secure SSL connection. RTMPS is used for streaming multimedia content across an encrypted connection. the laced-up version, we love her.

`rtmpt` 🚇

Real-Time Messaging Protocol tunneled through HTTP. RTMPT is used for streaming multimedia content within HTTP requests to traverse firewalls. sneaky lil thing, slipping past the firewall 😭

`rtmpte` 🚇🔐

Encrypted Real-Time Messaging Protocol tunneled through HTTP. RTMPTE is used for streaming multimedia content within HTTP requests to traverse firewalls. firewall-avoidance AND encryption? she's doing both.

`rtmpts` 🚇🛡️

Real-Time Messaging Protocol tunneled through HTTPS. RTMPTS is used for streaming multimedia content within HTTPS requests to traverse firewalls. the final boss of the rtmp cinematic universe.

`libsmbclient` 🗂️🌸

libsmbclient permits one to manipulate CIFS/SMB network resources — network shares, bestie. Following syntax is required:

smb://[[domain:]user[:password@]]server[/share[/path[/file]]]

This protocol accepts the following options.

timeout — sets the timeout, in milliseconds, for the socket I/O ops that the underlying low-level stuff does. Default is -1, which literally means "no timeout specified, we're just vibing" ✨
truncate — truncates existing files on write if you set it to 1. Set it to 0 and no truncating happens, no cap. Default is 1.
workgroup — sets the workgroup used for making connections. By default the workgroup is not specified 🫶

For more info, the original points you at http://www.samba.org/ 💖

🧚 libssh

Secure File Transfer Protocol via libssh — yes bestie, that's SFTP 🦋

Read from or write to remote resources using the SFTP protocol.

The following syntax is required, no exceptions:

sftp://[user[:password]@]server[:port]/path/to/remote/resource.mpeg

This protocol accepts the following options:

timeout — sets the timeout of socket I/O operations used by the underlying low-level operation. By default it's -1, meaning the timeout is not specified 💅
truncate — truncates existing files on write if set to 1; a value of 0 prevents truncating. Default value is 1.
private_key — specify the path of the file containing the private key to use during authorization. By default libssh searches for keys in the ~/.ssh/ directory 🔑

Example: play a file stored on a remote server 👑

ffplay sftp://user:password@server_address:22/home/user/resource.mpeg

🎀 librtmp rtmp, rtmpe, rtmps, rtmpt, rtmpte

Real-Time Messaging Protocol and its variants, supported through librtmp. It's giving legacy but iconic 💀

Requires the presence of the librtmp headers and library during configuration. You need to explicitly configure the build with --enable-librtmp. If enabled, this will replace the native RTMP protocol — like, swap it out entirely, heads up.

This protocol provides most client functions and a few server functions needed to support RTMP, RTMP tunneled in HTTP (RTMPT), encrypted RTMP (RTMPE), RTMP over SSL/TLS (RTMPS), and tunneled variants of these encrypted types (RTMPTE, RTMPTS). So many letters, but they all slay in their own way 🌈

The required syntax is:

<rtmp_proto>://<server>[:<port>][/<app>][/<playpath>] <options>

where rtmp_proto is one of the strings rtmp, rtmpt, rtmpe, rtmps, rtmpte, rtmpts, corresponding to each RTMP variant 🦋 and server, port, app and playpath have the same meaning as specified for the RTMP native protocol. options contains a list of space-separated options of the form key=val.

See the librtmp manual page (man 3 librtmp) for more information.

For example, to stream a file in real-time to an RTMP server using ffmpeg:

ffmpeg -re -i myfile -f flv rtmp://myserver/live/mystream

To play the same stream using ffplay:

ffplay "rtmp://myserver/live/mystream live=1"

💅 rtp

Real-time Transport Protocol. Main character of streaming, honestly ✨

The required syntax for an RTP URL is:

rtp://<hostname>[:<port>][?<options>]

port specifies the RTP port to use.

options contains a list of &-separated options of the form key=val. Standard percent-encoding (and using the plus sign for space) can be used to escape keys and values 🫶

Options can also be specified via command line options (or in code via AVOptions).

The list of supported options follows, and yes, we're going through all of them 👑

ttl=n — sets the TTL (Time-To-Live) value (for multicast only).
rtcpport=n — sets the remote RTCP port to n.
localport, local_rtpport, localrtpport=n — sets the local RTP port to n.

Using the localport option name is deprecated and should not be used. Like, don't. 🙈

local_rtcpport, localrtcpport=n' — sets the local RTCP port to n.
pkt_size=n — sets max packet size (in bytes) to n.
buffer_size=size — sets the maximum UDP socket buffer size in bytes.
connect=0|1 — does a connect() on the UDP socket (if set to 1) or not (if set to 0).
sources=ip[,ip] — lists allowed source IP addresses.
block=ip[,ip] — lists disallowed (blocked) source IP addresses. Blocked blocked 💀
write_to_source=0|1 — sends packets to the source address of the latest received packet (if set to 1) or to a default remote address (if set to 0).
localaddr=addr — local IP address of a network interface used for sending packets or joining multicast groups.
timeout=n — sets timeout (in microseconds) of socket I/O operations to n. Microseconds bestie, read it again 😭

Important notes:

If rtcpport is not set, the RTCP port will be set to the RTP port value plus 1.
If localrtpport (the local RTP port) is not set, any available port will be used for the local RTP and RTCP ports.
If localrtcpport (the local RTCP port) is not set, it will be set to the local RTP port value plus 1.

🌸 rtsp

Real-Time Streaming Protocol.

RTSP is not technically a protocol handler in libavformat — plot twist — it's a demuxer and muxer. The demuxer supports both normal RTSP (with data transferred over RTP; this is used by e.g. Apple and Microsoft) and Real-RTSP (with data transferred over RDT) 🍓

The muxer can be used to send a stream using RTSP ANNOUNCE to a server supporting it (currently Darwin Streaming Server and Mischa Spiegelmock's https://github.com/revmischa/rtsp-server).

The required syntax for an RTSP url is:

rtsp://<hostname>[:<port>]/<path>

Options can be set on the ffmpeg/ffplay command line, or set in code via AVOptions or in avformat_open_input.

Muxer

The following options are supported 💖

rtsp_transport — sets RTSP transport protocols.

It accepts the following values:

udp — use UDP as lower transport protocol.
tcp — use TCP (interleaving within the RTSP control channel) as lower transport protocol.

Default value is 0.

rtsp_flags — sets RTSP flags.

The following values are accepted:

latm — use MP4A-LATM packetization instead of MPEG4-GENERIC for AAC.
rfc2190 — use RFC 2190 packetization instead of RFC 4629 for H.263.
skip_rtcp — don't send RTCP sender reports.
h264_mode0 — use mode 0 for H.264 in RTP.
send_bye — send RTCP BYE packets when finishing.

Default value is 0.

min_port — sets minimum local UDP port. Default value is 5000.
max_port — sets maximum local UDP port. Default value is 65000.
buffer_size — sets the maximum socket buffer size in bytes.
pkt_size — sets max send packet size (in bytes). Default value is 1472.

Demuxer

The following options are supported ✨

initial_pause — do not start playing the stream immediately if set to 1. Default value is 0.
rtsp_transport — sets RTSP transport protocols.

It accepts the following values:

udp — use UDP as lower transport protocol.
tcp — use TCP (interleaving within the RTSP control channel) as lower transport protocol.
udp_multicast — use UDP multicast as lower transport protocol.
http — use HTTP tunneling as lower transport protocol, which is useful for passing proxies.
https — use HTTPs tunneling as lower transport protocol, which is useful for passing proxies and widely used for security consideration.

Multiple lower transport protocols may be specified, in that case they are tried one at a time (if the setup of one fails, the next one is tried). For the muxer, only the tcp and udp options are supported.

rtsp_flags — sets RTSP flags.

The following values are accepted:

filter_src — accept packets only from negotiated peer address and port.
listen — act as a server, listening for an incoming connection.
prefer_tcp — try TCP for RTP transport first, if TCP is available as RTSP RTP transport.
satip_raw — export raw MPEG-TS stream instead of demuxing. The flag will simply write out the raw stream, with the original PAT/PMT/PIDs intact.

Default value is none.

allowed_media_types — sets media types to accept from the server.

The following flags are accepted:

video
audio
data
subtitle

By default it accepts all media types.

min_port — sets minimum local UDP port. Default value is 5000.
max_port — sets maximum local UDP port. Default value is 65000.
listen_timeout — sets maximum timeout (in seconds) to establish an initial connection. Setting listen_timeout > 0 sets rtsp_flags to listen. Default is -1, which means an infinite timeout when listen mode is set 🫶
reorder_queue_size — sets number of packets to buffer for handling of reordered packets.
timeout — sets socket TCP I/O timeout in microseconds.
user_agent — overrides the User-Agent header. If not specified, it defaults to the libavformat identifier string.
buffer_size — sets the maximum socket buffer size in bytes.

When receiving data over UDP, the demuxer tries to reorder received packets (since they may arrive out of order, or packets may get lost totally 😭). This can be disabled by setting the maximum demuxing delay to zero (via the max_delay field of AVFormatContext).

When watching multi-bitrate Real-RTSP streams with ffplay, the streams to display can be chosen with -vst n and -ast n for video and audio respectively, and can be switched on the fly by pressing v and a 👑

Examples

The following examples all make use of the ffplay and ffmpeg tools.

Watch a stream over UDP, with a max reordering delay of 0.5 seconds:

ffplay -max_delay 500000 -rtsp_transport udp rtsp://server/video.mp4

Watch a stream tunneled over HTTP:

ffplay -rtsp_transport http rtsp://server/video.mp4

Send a stream in realtime to a RTSP server, for others to watch:

ffmpeg -re -i <input> -f rtsp -muxdelay 0.1 rtsp://server/live.sdp

Receive a stream in realtime:

ffmpeg -rtsp_flags listen -i rtsp://ownaddress/live.sdp <output>

🦋 sap

Session Announcement Protocol (RFC 2974). This is not technically a protocol handler in libavformat — again, it's a muxer and demuxer. It's used for signalling of RTP streams, by announcing the SDP for the streams regularly on a separate port 🌈

Muxer

The syntax for a SAP url given to the muxer is:

sap://<destination>[:<port>][?<options>]

The RTP packets are sent to destination on port port, or to port 5004 if no port is specified. options is a &-separated list. The following options are supported:

announce_addr=address — specify the destination IP address for sending the announcements to. If omitted, the announcements are sent to the commonly used SAP announcement multicast address 224.2.127.254 (sap.mcast.net), or ff0e::2:7ffe if destination is an IPv6 address 🎀
announce_port=port — which port the announcements go out on 🎀 if you don't say anything it defaults to 9875 (no cap, it's just the vibe).
ttl=ttl — the time to live value for the announcements and the RTP packets, defaults to 255 ✨.
same_port=0|1 — set this to 1 and literally all RTP streams get sent on the same port pair. set it to zero (that's the default bestie) and every stream goes out on its own unique port, each one sitting 2 numbers higher than the previous one. 💅 heads up: VLC/Live555 requires this to be 1 to actually receive the stream, and the RTP stack in libavformat for receiving requires all streams to be sent on unique ports. so like, pick your fighter 😭.

Example command lines follow 🫶

to broadcast a stream on the local subnet, for watching in VLC:

ffmpeg -re -i <input> -f sap sap://224.0.0.255?same_port=1

similarly, for watching in ffplay:

ffmpeg -re -i <input> -f sap sap://224.0.0.255

and for watching in ffplay, over IPv6:

ffmpeg -re -i <input> -f sap sap://[ff0e::1:2:3:4]

✨ Demuxer ✨

the syntax for a SAP url handed to the demuxer is:

sap://[<address>][:<port>]

address is the multicast address to listen for announcements on — if you omit it, the default 224.2.127.254 (sap.mcast.net) is used 💖. port is the port that's listened on, 9875 if omitted.

the demuxer listens for announcements on the given address and port. once an announcement shows up, it tries to receive that particular stream 🦋.

Example command lines follow 🌸

to play back the first stream announced on the normal SAP multicast address:

ffplay sap://

to play back the first stream announced on one the default IPv6 SAP multicast address:

ffplay sap://[ff0e::2:7ffe]

🌈 sctp 🌈

Stream Control Transmission Protocol. iconic transport-layer behavior honestly 👑

the accepted URL syntax is:

sctp://<host>:<port>[?<options>]

options contains a list of &-separated options of the form key=val. standard percent-encoding (and using the plus sign for space) can be used to escape keys and values 💜.

options can also can be specified via command line options (or in code via AVOptions).

the list of supported options follows ⭐

listen — if set to any value, listen for an incoming connection. outgoing connection is done by default.
max_streams — set the maximum number of streams. by default no limit is set 🎀.

🧚 srt 🧚

Haivision Secure Reliable Transport Protocol via libsrt 🔥.

the supported syntax for a SRT URL is:

srt://<hostname>:<port>[?<options>]

options contains a list of &-separated options of the form key=val. standard percent-encoding (and using the plus sign for space) can be used to escape keys and values 💖.

options can also can be specified via command line options (or in code via AVOptions).

the list of supported options follows ✨

connect_timeout=milliseconds — connection timeout. fun fact: SRT cannot connect for RTT > 1500 msec (2 handshake exchanges) with the default connect timeout of 3 seconds 😭. this option applies to the caller and rendezvous connection modes. the connect timeout is 10 times the value set for the rendezvous mode (which can be used as a workaround for this connection problem with earlier versions).
ffs=bytes — Flight Flag Size (Window Size), in bytes. FFS is actually an internal parameter and you should set it to not less than recv_buffer_size and mss. the default value is relatively large, so unless you set a very large receiver buffer, you do not need to change this option. default value is 25600.
inputbw=bytes/seconds — sender nominal input rate, in bytes per seconds. used along with oheadbw, when maxbw is set to relative (0), to calculate maximum sending rate when recovery packets are sent along with the main media stream: inputbw * (100 + oheadbw) / 100. if inputbw is not set while maxbw is set to relative (0), the actual input rate is evaluated inside the library. default value is 0.
iptos=tos — IP Type of Service. applies to sender only 🫶. default value is 0xB8.
ipttl=ttl — IP Time To Live. applies to sender only. default value is 64.
latency=microseconds — Timestamp-based Packet Delivery Delay. used to absorb bursts of missed packet retransmissions 💅. this flag sets both rcvlatency and peerlatency to the same value. note that prior to version 1.3.0 this is the only flag to set the latency, however this is effectively equivalent to setting peerlatency when side is sender and rcvlatency when side is receiver, and the bidirectional stream sending is not supported.
listen_timeout=microseconds — set socket listen timeout ⏰.
maxbw=bytes/seconds — maximum sending bandwidth, in bytes per seconds. -1 infinite (CSRTCC limit is 30mbps), 0 relative to input rate (see inputbw), >0 absolute limit value. default value is 0 (relative).
mode=caller|listener|rendezvous — connection mode. caller opens client connection. listener starts server to listen for incoming connections. rendezvous use Rendez-Vous connection mode. default value is caller 👑.
mss=bytes — Maximum Segment Size, in bytes. used for buffer allocation and rate calculation using a packet counter assuming fully filled packets. the smallest MSS between the peers is used. this is 1500 by default in the overall internet 🌐. this is the maximum size of the UDP packet and can be only decreased, unless you have some unusual dedicated network settings. default value is 1500.
nakreport=1|0 — if set to 1, Receiver will send UMSG_LOSSREPORT messages periodically until a lost packet is retransmitted or intentionally dropped. default value is 1.
oheadbw=percents — recovery bandwidth overhead above input rate, in percents. see inputbw. default value is 25%.
passphrase=string — HaiCrypt Encryption/Decryption Passphrase string, length from 10 to 79 characters 🔐. the passphrase is the shared secret between the sender and the receiver. it is used to generate the Key Encrypting Key using PBKDF2 (Password-Based Key Derivation Function). it is used only if pbkeylen is non-zero. it is used on the receiver only if the received data is encrypted. the configured passphrase cannot be recovered (write-only) 😭.
enforced_encryption=1|0 — if true, both connection parties must have the same password set (including empty, that is, with no encryption). if the password doesn't match or only one side is unencrypted, the connection is rejected 💀. default is true.
kmrefreshrate=packets — the number of packets to be transmitted after which the encryption key is switched to a new key. default is -1. -1 means auto (0x1000000 in srt library). the range for this option is integers in the 0 - INT_MAX.
kmpreannounce=packets — the interval between when a new encryption key is sent and when switchover occurs. this value also applies to the subsequent interval between when switchover occurs and when the old encryption key is decommissioned. default is -1. -1 means auto (0x1000 in srt library). the range for this option is integers in the 0 - INT_MAX.
snddropdelay=microseconds — the sender's extra delay before dropping packets. this delay is added to the default drop delay time interval value.

special value -1: do not drop packets on the sender at all 🙈.

payload_size=bytes — sets the maximum declared size of a packet transferred during the single call to the sending function in Live mode. use 0 if this value isn't used (which is default in file mode). default is -1 (automatic), which typically means MPEG-TS; if you are going to use SRT to send any different kind of payload, such as, for example, wrapping a live stream in very small frames, then you can use a bigger maximum frame size, though not greater than 1456 bytes.
pkt_size=bytes — alias for payload_size 🎀.
peerlatency=microseconds — the latency value (as described in rcvlatency) that is set by the sender side as a minimum value for the receiver.
pbkeylen=bytes — sender encryption key length, in bytes. only can be set to 0, 16, 24 and 32. enable sender encryption if not 0. not required on receiver (set to 0), key size obtained from sender in HaiCrypt handshake. default value is 0.
rcvlatency=microseconds — the time that should elapse since the moment when the packet was sent and the moment when it's delivered to the receiver application in the receiving function ⏳. this time should be a buffer time large enough to cover the time spent for sending, unexpectedly extended RTT time, and the time needed to retransmit the lost UDP packet. the effective latency value will be the maximum of this options' value and the value of peerlatency set by the peer side. before version 1.3.0 this option is only available as latency.
recv_buffer_size=bytes — set UDP receive buffer size, expressed in bytes.
send_buffer_size=bytes — set UDP send buffer size, expressed in bytes.
timeout=microseconds — set raise error timeouts for read, write and connect operations. note that the SRT library has internal timeouts which can be controlled separately, the value set here is only a cap on those 💅.
tlpktdrop=1|0 — Too-late Packet Drop. when enabled on receiver, it skips missing packets that have not been delivered in time and delivers the following packets to the application when their time-to-play has come. it also sends a fake ACK to the sender 😳. when enabled on sender and enabled on the receiving peer, the sender drops the older packets that have no chance of being delivered in time. it was automatically enabled in the sender if the receiver supports it.
sndbuf=bytes — set send buffer size, expressed in bytes.
rcvbuf=bytes — set receive buffer size, expressed in bytes.

receive buffer must not be greater than ffs ⚠️.

lossmaxttl=packets — the value up to which the Reorder Tolerance may grow. when Reorder Tolerance is > 0, then packet loss report is delayed until that number of packets come in. Reorder Tolerance increases every time a "belated" packet has come, but it wasn't due to retransmission (that is, when UDP packets tend to come out of order), with the difference between the latest sequence and this packet's sequence, and not more than the value of this option. by default it's 0, which means that this mechanism is turned off, and the loss report is always sent immediately upon experiencing a "gap" in sequences 🦋.
minversion — the minimum SRT version the peer has to have, bestie. If the peer doesn’t satisfy it, the connection gets rejected, no cap. The version format in hex is 0xXXYYZZ for x.y.z in human-readable form.
streamid=string — a string limited to 512 characters that can be set on the socket before connecting. This stream ID will be able to be retrieved by the listener side from the socket that is returned from srt_accept and was connected by a socket with that set stream ID. SRT does not enforce any special interpretation of the contents of this string. This option doesn’t make sense in Rendezvous connection; the result might be that simply one side will override the value from the other side and it’s the matter of luck which one would win 💀
srt_streamid=string — alias for streamid to avoid conflict with ffmpeg command line option. lowkey useful 🫶
smoother=live|file — the type of Smoother used for the transmission for that socket, which is responsible for the transmission and congestion control. The Smoother type must be exactly the same on both connecting parties, otherwise the connection is rejected.
messageapi=1|0 — when set, this socket uses the Message API; otherwise it uses Buffer API. Note that in live mode (see transtype) there’s only message API available. In File mode you can choose one of two modes:
- Stream API (default, when this option is false). In this mode you may send as many data as you wish with one sending instruction, or even use dedicated functions that read directly from a file. The internal facility will take care of any speed and congestion control. When receiving, you can also receive as many data as desired, the data not extracted will be waiting for the next call. There is no boundary between data portions in the Stream mode.
- Message API. In this mode your single sending instruction passes exactly one piece of data that has boundaries (a message). Contrary to Live mode, this message may span across multiple UDP packets and the only size limitation is that it shall fit as a whole in the sending buffer. The receiver shall use as large buffer as necessary to receive the message, otherwise the message will not be given up. When the message is not complete (not all packets received or there was a packet loss) it will not be given up.
transtype=live|file — sets the transmission type for the socket, in particular, setting this option sets multiple other parameters to their default values as required for a particular transmission type.
- live: set options as for live transmission. In this mode, you should send by one sending instruction only so many data that fit in one UDP packet, and limited to the value defined first in payload_size (1316 is default in this mode). There is no speed control in this mode, only the bandwidth control, if configured, in order to not exceed the bandwidth with the overhead transmission (retransmitted and control packets).
- file: set options as for non-live transmission. See messageapi for further explanations.
linger=seconds — the number of seconds that the socket waits for unsent data when closing. Default is -1. -1 means auto (off with 0 seconds in live mode, on with 180 seconds in file mode). The range for this option is integers in the 0 - INT_MAX.
tsbpd=1|0 — when true, use Timestamp-based Packet Delivery mode. The default behavior depends on the transmission type: enabled in live mode, disabled in file mode.

For more information see: https://github.com/Haivision/srt.

✨ srtp ✨

Secure Real-time Transport Protocol. The accepted options are:

srtp_in_suite, srtp_out_suite — select input and output encoding suites. Supported values:

AES_CM_128_HMAC_SHA1_80
SRTP_AES128_CM_HMAC_SHA1_80
AES_CM_128_HMAC_SHA1_32
SRTP_AES128_CM_HMAC_SHA1_32

srtp_in_params, srtp_out_params — set input and output encoding parameters, which are expressed by a base64-encoded representation of a binary block. The first 16 bytes of this binary block are used as master key, the following 14 bytes are used as master salt.

🦋 subfile 🦋

Virtually extract a segment of a file or another stream. The underlying stream must be seekable.

Accepted options:

start — start offset of the extracted segment, in bytes.
end — end offset of the extracted segment, in bytes. If set to 0, extract till end of file.

Examples:

Extract a chapter from a DVD VOB file (start and end sectors obtained externally and multiplied by 2048):

subfile,,start,153391104,end,268142592,,:/media/dvd/VIDEO_TS/VTS_08_1.VOB

Play an AVI file directly from a TAR archive:

subfile,,start,183241728,end,366490624,,:archive.tar

Play a MPEG-TS file from start offset till end:

subfile,,start,32815239,end,0,,:video.ts

🎀 tee 🎀

Writes the output to multiple protocols. The individual outputs are separated by |.

tee:file://path/to/local/this.avi|file://path/to/local/that.avi

💅 tcp 💅

Transmission Control Protocol.

The required syntax for a TCP url is:

tcp://<hostname>:<port>[?<options>]

options contains a list of &-separated options of the form key=val. Standard percent-encoding (and using the plus sign for space) can be used to escape keys and values.

Options can also be specified via command line options (or in code via "AVOption"s).

The list of supported options follows.

listen=2|1|0 — listen for an incoming connection. 0 disables listen, 1 enables listen in single client mode, 2 enables listen in multi-client mode. Default value is 0.
local_addr=addr — local IP address of a network interface used for tcp socket connect.
local_port=port — local port used for tcp socket connect.
timeout=microseconds — set raise error timeout, expressed in microseconds. This option is only relevant in read mode: if no data arrived in more than this time interval, raise error.
listen_timeout=milliseconds — set listen timeout, expressed in milliseconds.
recv_buffer_size=bytes — set receive buffer size, expressed bytes.
send_buffer_size=bytes — set send buffer size, expressed bytes.
tcp_nodelay=1|0 — set TCP_NODELAY to disable Nagle's algorithm. Default value is 0.

Remark: Writing to the socket is currently not optimized to minimize system calls and reduces the efficiency / effect of TCP_NODELAY.

tcp_mss=bytes — set maximum segment size for outgoing TCP packets, expressed in bytes.

The following example shows how to setup a listening TCP connection with ffmpeg, which is then accessed with ffplay:

ffmpeg -i <input> -f <format> tcp://<hostname>:<port>?listen
ffplay tcp://<hostname>:<port>

🌈 tls 🌈

Transport Layer Security (TLS) / Secure Sockets Layer (SSL)

The required syntax for a TLS/SSL url is:

tls://<hostname>:<port>[?<options>]

options contains a list of &-separated options of the form key=val. Standard percent-encoding (and using the plus sign for space) can be used to escape keys and values.

Options can also be specified via command line options (or in code via "AVOption"s).

The list of supported options follows.

ca_file, cafile=filename — a file containing certificate authority (CA) root certificates to treat as trusted. If the linked TLS library contains a default this might not need to be specified for verification to work, but not all libraries and setups have defaults built in. The file must be in OpenSSL PEM format.
tls_verify, verify=1|0 — if enabled, try to verify the peer that we are communicating with. Note, if using OpenSSL, this currently only makes sure that the peer certificate is signed by one of the root certificates in the CA database, but it does not validate that the certificate actually matches the host name we are trying to connect to. (With other backends, the host name is validated as well.) This is disabled by default since it requires a CA database to be provided by the caller in many cases.
cert_file, cert=filename — a file containing a certificate to use in the handshake with the peer. (When operating as server, in listen mode, this is more often required by the peer, while client certificates only are mandated in certain setups.)
key_file, key=filename — a file containing the private key for the certificate.
listen=1|0 — if enabled, listen for connections on the provided port, and assume the server role in the handshake instead of the client role.
http_proxy — the HTTP proxy to tunnel through, e.g. http://example.com:1234. The proxy must support the CONNECT method.

Example command lines:

To create a TLS/SSL server that serves an input stream.

ffmpeg -i <input> -f <format> tls://<hostname>:<port>?listen&cert=<server.crt>&key=<server.key>

To play back a stream from the TLS/SSL server using ffplay:

ffplay tls://<hostname>:<port>

🧚 dtls 🧚

Datagram Transport Layer Security (DTLS)

The required syntax for a DTLS URL is:

dtls://<hostname>:<port>[?<options>]

options contains a list of &-separated options of the form key=val. Standard percent-encoding (and using the plus sign for space) can be used to escape keys and values.

Options can also be specified via command line options (or in code via "AVOption"s).

DTLS shares most options with TLS, but operates over UDP instead of TCP.

The list of supported options follows.

ca_file, cafile=filename — a file containing certificate authority (CA) root certificates to treat as trusted. If the linked TLS library contains a default this might not need to be specified for verification to work, but not all libraries and setups have defaults built in. The file must be in OpenSSL PEM format.
tls_verify, verify=1|0 — if enabled, try to verify the peer that we are communicating with. Note, if using OpenSSL, this currently only makes sure that the peer certificate is signed by one of the root certificates in the CA database, but it does not validate that the certificate actually matches the host name we are trying to connect to. This is disabled by default since it requires a CA database to be provided by the caller in many cases.
cert_file, cert=filename — a file holding the certificate used for the handshake with the peer 🎀 (when you're running as server in listen mode, the peer is usually the one asking for this; client certs are only a must in certain setups, iykyk)
key_file, key=filename — the file with the private key for that certificate 🔑
cert_pem=string — the certificate as a PEM string, right there inline ✨
key_pem=string — the private key as a PEM string too 🫶
listen=1|0 — flip this on and it listens for connections on the port you gave, and takes the server role in the handshake instead of being the client 👑
mtu=size — sets the Maximum Transmission Unit (MTU) for DTLS packets 📦
use_srtp=1|0 — enables the use_srtp DTLS extension. This is the whole WebRTC thing: it establishes SRTP encryption keys through the DTLS handshake. Default is disabled 💅
external_sock=1|0 — uses an external socket instead of making a shiny new one. Lowkey only makes sense when you're talking to the code via API — turning this on from the CLI will cause immediate failure, no cap 💀. Default is disabled.

Example command lines:

To create a DTLS server:

ffmpeg -listen 1 -i dtls://<hostname>:<port> <output>

To create a DTLS client and send data to server:

ffmpeg -i <input> -f <format> dtls://<hostname>:<port>

udp

User Datagram Protocol. The chaotic bestie of transport protocols 🌈

The required syntax for an UDP URL is:

udp://<hostname>:<port>[?<options>]

options is a list of &-separated options of the form key=val. Standard percent-encoding (and the plus sign for space) can be used to escape keys and values 🎀

Options can also be specified via command line options (or in code via AVOptions).

If threading is enabled on your system, a circular buffer is used to store the incoming data, which lets you reduce data loss from UDP socket buffer overruns. The fifo_size and overrun_nonfatal options are the ones tied to this buffer 🦋

The list of supported options follows.

buffer_size=size — sets the UDP maximum socket buffer size in bytes. Used for either the receive or send buffer size, depending on what the socket is doing. Default is 32 KB for output, 384 KB for input. See also fifo_size 🌸
bitrate=bitrate — if set to nonzero, the output gets the specified constant bitrate, as long as the input has enough packets to sustain it 🔥
burst_bits=bits — when you're using bitrate, this sets the maximum number of bits in packet bursts.
localport=port — overrides the local UDP port to bind with.
localaddr=addr — local IP address of a network interface used for sending packets or joining multicast groups.
pkt_size=size — sets the size in bytes of UDP packets.
reuse=1|0 — explicitly allow or disallow reusing UDP sockets.
ttl=ttl — sets the time to live value (multicast only, bestie).
dscp=dscp — sets the 6-bit DSCP field for outgoing packets.
connect=1|0 — initializes the UDP socket with connect(). When you do this, the destination address can't be changed later with ff_udp_set_remote_url. If the destination address isn't known at the start, this option can be specified in ff_udp_set_remote_url too. This lets you find out the source address for the packets with getsockname, and makes writes return with AVERROR(ECONNREFUSED) if "destination unreachable" comes back. For receiving, the serve is that you only receive packets from the specified peer address/port 💅
sources=address[,address] — only receive packets sent from the specified addresses. With multicast, it also subscribes to multicast traffic coming from these addresses only 👑
block=address[,address] — ignores packets sent from the specified addresses. With multicast, it also excludes the source addresses in the multicast subscription. Not me blocking the haters 🚫
fifo_size=units — sets the UDP receiving circular buffer size, expressed as a number of packets with size of 188 bytes. If not specified it defaults to 7*4096 🍓
overrun_nonfatal=1|0 — survive a UDP receiving circular buffer overrun instead of dying. Default value is 0.
timeout=microseconds — sets the raise error timeout, expressed in microseconds. This one is only relevant in read mode: if no data arrived in more than this time interval, raise error 😭
broadcast=1|0 — explicitly allow or disallow UDP broadcasting.

Heads up bestie: broadcasting may not work properly on networks that have broadcast storm protection 💀

Examples

Use ffmpeg to stream over UDP to a remote endpoint:

ffmpeg -i <input> -f <format> udp://<hostname>:<port>

Use ffmpeg to stream in mpegts format over UDP using 188 sized UDP packets, with a large input buffer:

ffmpeg -i <input> -f mpegts udp://<hostname>:<port>?pkt_size=188&buffer_size=65535

Use ffmpeg to receive over UDP from a remote endpoint:

ffmpeg -i udp://[<multicast-address>]:<port> ...

unix

Unix local socket — the introvert of protocols, it never leaves the machine 🧚

The required syntax for a Unix socket URL is:

unix://<filepath>

The following parameters can be set via command line options (or in code via AVOptions):

timeout — timeout in ms.
listen — creates the Unix socket in listening mode.
type — select the socket type ✨
- stream — corresponds to SOCK_STREAM (reliable stream-oriented)
- datagram — corresponds to SOCK_DGRAM (unreliable packet-oriented)
- seqpacket — corresponds to SOCK_SEQPACKET (reliable packet-oriented)
pkt_size — maximum packet size for packet-oriented sockets (SOCK_DGRAM and SOCK_SEQPACKET). If greater than zero, this value is used as max_packet_size. It's ignored for SOCK_STREAM. Default is 0.

zmq

ZeroMQ asynchronous messaging, powered by the libzmq library 🦋

This library supports unicast streaming to multiple clients without ever relying on an external server. Main character energy, honestly.

The required syntax for streaming or connecting to a stream is:

zmq:tcp://ip-address:port

Example: create a localhost stream on port 5555:

ffmpeg -re -i input -f mpegts zmq:tcp://127.0.0.1:5555

Multiple clients may connect to the stream using:

ffplay zmq:tcp://127.0.0.1:5555

Streaming to multiple clients is implemented using a ZeroMQ Pub-Sub pattern. The server side binds to a port and publishes data. Clients connect to the server (via IP address/port) and subscribe to the stream. The order in which the server and client start generally does not matter 🙌

ffmpeg must be compiled with the --enable-libzmq option to support this protocol.

Options can be set on the ffmpeg/ffplay command line. The following options are supported:

pkt_size — forces the maximum packet size for sending/receiving data. The default value is 131,072 bytes. On the server side, this sets the maximum size of sent packets via ZeroMQ. On the clients, it sets an internal buffer size for receiving packets.

Note that pkt_size on the clients should be equal to or greater than pkt_size on the server. Otherwise the received message may get truncated, which causes decoding errors 😭

DEVICE OPTIONS

The libavdevice library gives you the exact same interface as libavformat. Like, an input device is treated like a demuxer, an output device is treated like a muxer, and the interface plus the generic device options are the same ones libavformat provides (see the ffmpeg-formats manual) 🎀

On top of that, each input or output device may support so-called private options, which are specific to that component 💅

Options may be set by specifying -option value in the FFmpeg tools, or by setting the value explicitly in the device AVFormatContext options, or by using the libavutil/opt.h API for programmatic use ✨

INPUT DEVICES

Input devices are configured elements in FFmpeg that let you access the data coming from a multimedia device attached to your system 🎥

When you configure your FFmpeg build, all the supported input devices are enabled by default. You can list all the available ones using the configure option --list-indevs 👑

You can disable all the input devices using the configure option --disable-indevs, and selectively enable an input device using the option --enable-indev=INDEV, or you can disable a particular input device using the option --disable-indev=INDEV 🚫

The option -devices of the ff* tools will display the list of supported input devices.

A description of the currently available input devices follows.

alsa

ALSA (Advanced Linux Sound Architecture) input device 🎧

To enable this input device during configuration you need libasound installed on your system.

This device lets you capture from an ALSA device. The name of the device to capture has to be an ALSA card identifier.

An ALSA identifier has the syntax:

hw:<CARD>[,<DEV>[,<SUBDEV>]]

where the DEV and SUBDEV components are optional.

The three arguments (in order: CARD,DEV,SUBDEV) specify card number or identifier, device number and subdevice number (-1 means any) 🌸

To see the list of cards currently recognized by your system, check the files /proc/asound/cards and /proc/asound/devices 🦋

For example, to capture with ffmpeg from an ALSA device with card id 0, you may run the command:

ffmpeg -f alsa -i hw:0 alsaout.wav

For more information see: http://www.alsa-project.org/alsa-doc/alsa-lib/pcm.html

Options

sample_rate — sets the sample rate in Hz. Default is 48000.
channels — sets the number of channels. Default is 2.

android_camera

Android camera input device 📱

This input device uses the Android Camera2 NDK API, which is available on devices with API level 24+. The availability of android_camera is autodetected during configuration 👀

This device lets you capture from all cameras on an Android device that are integrated into the Camera2 NDK API.

The available cameras are enumerated internally and can be selected with the camera_index parameter. The input file string is discarded — it literally doesn't matter, bestie 🙈

Generally the back facing camera has index 0 while the front facing camera has index 1. Selfie mode, activated 💅

Options

video_size — sets the video size, given as a string such as 640x480 or hd720. Falls back to the first available configuration reported by Android if the requested video size isn't available, or by default.
framerate — sets the video framerate. Falls back to the first available configuration reported by Android if the requested framerate isn't available, or by default (-1).
camera_index — sets the index of the camera to use. Default is 0.
input_queue_size — sets the maximum number of frames to buffer. Default is 5.

avfoundation

The AVFoundation input device — aka Apple's actual fave. 🍎✨ AVFoundation is the currently recommended framework by Apple for streamgrabbing on OSX >= 10.7 as well as on iOS.

The input filename has to be given in the following syntax, no cap:

-i "[[VIDEO]:[AUDIO]]"

The first entry selects the video input while the latter selects the audio input. The stream has to be specified by the device name or the device index as shown by the device list. Alternatively, the video and/or audio input device can be chosen by index using the

-video_device_index <INDEX>

and/or

-audio_device_index <INDEX>

, overriding any device name or index given in the input filename. (Yes, the comma is doing its own thing, she's an icon. 💅)

All available devices can be enumerated by using -list_devices true, listing all device names and corresponding indices.

There are two device name aliases, and they are lowkey everything:

"default" — Select the AVFoundation default device of the corresponding type.
"none" — Do not record the corresponding media type. This is equivalent to specifying an empty device name or index.

Options

AVFoundation supports the following options, bestie:

-list_devices <TRUE|FALSE> — If set to true, a list of all available input devices is given showing all device names and indices.
-video_device_index <INDEX> — Specify the video device by its index. Overrides anything given in the input filename.
-audio_device_index <INDEX> — Specify the audio device by its index. Overrides anything given in the input filename.
-pixel_format <FORMAT> — Request the video device to use a specific pixel format. If the specified format is not supported, a list of available formats is given and the first one in this list is used instead. Available pixel formats are: "monob, rgb555be, rgb555le, rgb565be, rgb565le, rgb24, bgr24, 0rgb, bgr0, 0bgr, rgb0, bgr48be, uyvy422, yuva444p, yuva444p16le, yuv444p, yuv422p16, yuv422p10, yuv444p10, yuv420p, nv12, yuyv422, gray"
-framerate — Set the grabbing frame rate. Default is "ntsc", corresponding to a frame rate of "30000/1001".
-video_size — Set the video frame size.
-capture_cursor — Capture the mouse pointer. Default is 0.
-capture_mouse_clicks — Capture the screen mouse clicks. Default is 0.
-capture_raw_data — Capture the raw device data. Default is 0. Using this option may result in receiving the underlying data delivered to the AVFoundation framework. E.g. for muxed devices that sends raw DV data to the framework (like tape-based camcorders), setting this option to false results in extracted video frames captured in the designated pixel format only. Setting this option to true results in receiving the raw DV stream untouched.

Examples

Print the list of AVFoundation supported devices and exit:

$ ffmpeg -f avfoundation -list_devices true -i ""

Record video from video device 0 and audio from audio device 0 into out.avi:

$ ffmpeg -f avfoundation -i "0:0" out.avi

Record video from video device 2 and audio from audio device 1 into out.avi:

$ ffmpeg -f avfoundation -video_device_index 2 -i ":1" out.avi

Record video from the system default video device using the pixel format bgr0 and do not record any audio into out.avi:

$ ffmpeg -f avfoundation -pixel_format bgr0 -i "default:none" out.avi

Record raw DV data from a suitable input device and write the output into out.dv:

$ ffmpeg -f avfoundation -capture_raw_data true -i "zr100:none" out.dv

decklink

The decklink input device provides capture capabilities for Blackmagic DeckLink devices. 🎥🔥

To enable this input device, you need the Blackmagic DeckLink SDK and you need to configure with the appropriate "--extra-cflags" and "--extra-ldflags". On Windows, you need to run the IDL files through widl.

DeckLink is very picky about the formats it supports — she has standards, respect them. 💅 Pixel format of the input can be set with raw_format. Framerate and video size must be determined for your device with -list_formats 1. Audio sample rate is always 48 kHz and the number of channels can be 2, 8 or 16. Note that all audio channels are bundled in one single audio track.

Options

list_devices — If set to true, print a list of devices and exit. Defaults to false. This option is deprecated, please use the "-sources" option of ffmpeg to list the available input devices.
list_formats — If set to true, print a list of supported formats and exit. Defaults to false.
format_code <FourCC> — This sets the input video format to the format given by the FourCC. To see the supported values of your device(s) use list_formats. Note that there is a FourCC 'pal ' that can also be used as pal (3 letters). Default behavior is autodetection of the input video format, if the hardware supports it.
raw_format — Set the pixel format of the captured video. Available values are:
- auto — This is the default which means 8-bit YUV 422 or 8-bit ARGB if format autodetection is used, 8-bit YUV 422 otherwise.
- uyvy422 — 8-bit YUV 422.
- yuv422p10 — 10-bit YUV 422.
- argb — 8-bit RGB.
- bgra — 8-bit RGB.
- rgb10 — 10-bit RGB.
teletext_lines — If set to nonzero, an additional teletext stream will be captured from the vertical ancillary data. Both SD PAL (576i) and HD (1080i or 1080p) sources are supported. In case of HD sources, OP47 packets are decoded.

This option is a bitmask of the SD PAL VBI lines captured, specifically lines 6 to 22, and lines 318 to 335. Line 6 is the LSB in the mask. Selected lines which do not contain teletext information will be ignored. You can use the special all constant to select all possible lines, or standard to skip lines 6, 318 and 319, which are not compatible with all receivers.

For SD sources, ffmpeg needs to be compiled with "--enable-libzvbi". For HD sources, on older (pre-4K) DeckLink card models you have to capture in 10 bit mode.

channels — Defines number of audio channels to capture. Must be 2, 8 or 16. Defaults to 2.
duplex_mode — Sets the decklink device duplex/profile mode. Must be unset, half, full, one_sub_device_full, one_sub_device_half, two_sub_device_full, four_sub_device_half. Defaults to unset.

Note: DeckLink SDK 11.0 have replaced the duplex property by a profile property. For the DeckLink Duo 2 and DeckLink Quad 2, a profile is shared between any 2 sub-devices that utilize the same connectors. For the DeckLink 8K Pro, a profile is shared between all 4 sub-devices. So DeckLink 8K Pro support four profiles.

Valid profile modes for DeckLink 8K Pro (with DeckLink SDK >= 11.0): one_sub_device_full, one_sub_device_half, two_sub_device_full, four_sub_device_half

Valid profile modes for DeckLink Quad 2 and DeckLink Duo 2: half, full

timecode_format — Timecode type to include in the frame and video stream metadata. Must be none, rp188vitc, rp188vitc2, rp188ltc, rp188hfr, rp188any, vitc, vitc2, or serial. Defaults to none (not included).

In order to properly support 50/60 fps timecodes, the ordering of the queried timecode types for rp188any is HFR, VITC1, VITC2 and LTC for >30 fps content. Note that this is slightly different to the ordering used by the DeckLink API, which is HFR, VITC1, LTC, VITC2. (iykyk 😌)

video_input — Sets the video input source. Must be unset, sdi, hdmi, optical_sdi, component, composite or s_video. Defaults to unset.
audio_input — Sets the audio input source. Must be unset, embedded, aes_ebu, analog, analog_xlr, analog_rca or microphone. Defaults to unset.
video_pts — Sets the video packet timestamp source. Must be video, audio, reference, wallclock or abs_wallclock. Defaults to video.
audio_pts — Sets the audio packet timestamp source. Must be video, audio, reference, wallclock or abs_wallclock. Defaults to audio.
draw_bars — If set to true, color bars are drawn in the event of a signal loss. Defaults to true. This option is deprecated, please use the "signal_loss_action" option.
signal_loss_action — Sets the action to take in the event of a signal loss. Accepts one of the following values:
- 1, none — Do nothing on signal loss. This usually results in black frames.
- 2, bars — Draw color bars on signal loss. Only supported for 8-bit input signals.
- 3, repeat — Repeat the last video frame on signal loss.

Defaults to bars.

queue_size — Sets maximum input buffer size in bytes. If the buffering reaches this value, incoming frames will be dropped. Defaults to 1073741824.
audio_depth — Sets the audio sample bit depth. Must be 16 or 32. Defaults to 16.
decklink_copyts — If set to true, timestamps are forwarded as they are without removing the initial offset. Defaults to false.
timestamp_align — Capture start time alignment in seconds. If set to nonzero, input frames are dropped till the system timestamp aligns with configured value. Alignment difference of up to one frame duration is tolerated. This is useful for maintaining input synchronization across N different hardware devices deployed for 'N-way' redundancy. The system time of different hardware devices should be synchronized with protocols such as NTP or PTP, before using this option. Note that this method is not foolproof. In some border cases input synchronization may not happen due to thread scheduling jitters in the OS. Either sync could go wrong by 1 frame or in a rarer case timestamp_align seconds. Defaults to 0.
wait_for_tc (bool) — Drop frames till a frame with timecode is received. Sometimes serial timecode isn't received with the first input frame. If that happens, the stored stream timecode will be inaccurate. If this option is set to true, input frames are dropped till a frame with timecode is received. Option timecode_format must be specified. Defaults to false.
enable_klv(bool) — If set to true, extracts KLV data from VANC and outputs KLV packets. KLV VANC packets are joined based on MID and PSC fields and aggregated into one KLV packet. Defaults to false.

Examples

List input devices:

ffmpeg -sources decklink

List supported formats:
-a, ... no wait, we're doing examples bestie, not options 😭

ffmpeg -f decklink -list_formats 1 -i 'Intensity Pro'

✨ Capture a video clip at 1080i50:

ffmpeg -format_code Hi50 -f decklink -i 'Intensity Pro' -c:a copy -c:v copy output.avi

💖 Capture a video clip at 1080i50, but make it 10 bit:

ffmpeg -raw_format yuv422p10 -format_code Hi50 -f decklink -i 'UltraStudio Mini Recorder' -c:a copy -c:v copy output.avi

🎀 Capture a video clip at 1080i50 with a full 16 audio channels:

ffmpeg -channels 16 -format_code Hi50 -f decklink -i 'UltraStudio Mini Recorder' -c:a copy -c:v copy output.avi

🌸 dshow 🌸

Okay so this one is the Windows DirectShow input device bestie. 🪟

DirectShow support is only a thing when FFmpeg gets built with the mingw-w64 project. And rn? Only audio and video devices are supported — that's the whole menu, no cap.

You can open multiple devices as separate inputs, but plot twist: you can also open them on the same input, which should improve the synchronism between them. Slay, honestly. 💅

The input name has to be in this format:

<TYPE>=<NAME>[:<TYPE>=<NAME>]

where TYPE can be either audio or video, and NAME is the device's name or alternative name. (Yes, the original says "alternative name.." with two dots — we're keeping the double dots, it's part of the vibe now 💀)

💜 Options 💜

If you specify no options at all, the device's defaults get used. And if the device doesn't support the options you're asking for? It will fail to open. Periodt. No partial credit.

video_size — sets the video size in the captured video.
framerate — sets the frame rate in the captured video.
sample_rate — sets the sample rate (in Hz) of the captured audio.
sample_size — sets the sample size (in bits) of the captured audio.
channels — sets the number of channels in the captured audio.
list_devices — if you set it to true, it prints a list of devices and then exits. Iconic lil recon mission. 🕵️
list_options — if set to true, prints a list of the selected device's options and exits.
video_device_number — sets the video device number for devices that share the same name (starts at 0, defaults to 0).
audio_device_number — sets the audio device number for devices that share the same name (starts at 0, defaults to 0).
pixel_format — picks the pixel format DirectShow uses. Heads up: this may only be set when the video codec is not set, or is set to rawvideo.
audio_buffer_size — sets the audio device buffer size in milliseconds (which can directly impact latency, depending on the device, so like… pick wisely). Defaults to using the audio device's own default buffer size (typically some multiple of 500ms). Setting this value too low can degrade performance — don't do that to yourself 😭. See also <http://msdn.microsoft.com/en-us/library/windows/desktop/dd377582(v=vs.85).aspx>
video_pin_name — selects which video capture pin to use, by name or alternative name.
audio_pin_name — selects which audio capture pin to use, by name or alternative name.
crossbar_video_input_pin_number — selects the video input pin number for a crossbar device. This gets routed to the crossbar device's Video Decoder output pin. ⚠️ Note that changing this value can affect future invocations (it sets a new default) until a system reboot happens. So it's giving "haunted setting," be careful.
crossbar_audio_input_pin_number — selects the audio input pin number for a crossbar device. This gets routed to the crossbar device's Audio Decoder output pin. Same warning bestie: changing it can affect future invocations (new default gets set) until the system reboots.
show_video_device_dialog — if set to true, then before capture starts, it pops up a display dialog for the end user so they can manually change video filter properties and configurations. Note that for crossbar devices, adjusting values in this dialog may sometimes be needed to toggle between PAL (25 fps) and NTSC (29.97) input frame rates, sizes, interlacing, etc. Changing these values can enable different scan rates/frame rates and avoid green bars at the bottom, flickering scan lines, etc. 🟩 And note that with some devices, changing these properties can also affect future invocations (sets new defaults) until a system reboot occurs. Reboot-core fr.
show_audio_device_dialog — if set to true, before capture starts, pops up a display dialog for the end user so they can manually change audio filter properties and configurations.
show_video_crossbar_connection_dialog — if set to true, before capture starts, pops up a display dialog for the end user so they can manually modify crossbar pin routings when it opens a video device.
show_audio_crossbar_connection_dialog — if set to true, before capture starts, pops up a display dialog for the end user so they can manually modify crossbar pin routings when it opens an audio device.
show_analog_tv_tuner_dialog — if set to true, before capture starts, pops up a display dialog for the end user so they can manually modify TV channels and frequencies.
show_analog_tv_tuner_audio_dialog — if set to true, before capture starts, pops up a display dialog for the end user so they can manually modify TV audio (think mono vs. stereo, Language A, B or C).
audio_device_load — loads an audio capture filter device from a file instead of searching for it by name. It may load extra parameters too, if the filter supports serializing its properties. To use this you have to specify an audio capture source, but honestly it can be anything — even a fake one. 🙈
audio_device_save — saves the currently used audio capture filter device and its parameters (if the filter supports it) to a file. If a file with the same name already exists, it gets overwritten, sorry babe.
video_device_load — loads a video capture filter device from a file instead of searching it by name. It may load extra parameters too, if the filter supports serializing its properties. To use this you have to specify a video capture source, but it can be anything — even a fake one.
video_device_save — saves the currently used video capture filter device and its parameters (if the filter supports it) to a file. If a file with the same name exists, it will be overwritten. 💀
use_video_device_timestamps — if set to false, the timestamp for video frames gets derived from the wallclock instead of the timestamp the capture device hands over. This is the workaround for devices that give you unreliable timestamps. Ty for your service. 🫶

🦋 Examples 🦋

Print the list of DirectShow supported devices and exit:

$ ffmpeg -list_devices true -f dshow -i dummy

Open the video device Camera:

$ ffmpeg -f dshow -i video="Camera"

Open the second video device that's also named Camera:

$ ffmpeg -f dshow -video_device_number 1 -i video="Camera"

Open video device Camera and audio device Microphone together:

$ ffmpeg -f dshow -i video="Camera":audio="Microphone"

Print the list of supported options on the selected device and exit:

$ ffmpeg -list_options true -f dshow -i video="Camera"

Specify pin names to capture by name or alternative name, and specify an alternative device name:

$ ffmpeg -f dshow -audio_pin_name "Audio Out" -video_pin_name 2 -i video=video="@device_pnp_\\?\pci#ven_1a0a&dev_6200&subsys_62021461&rev_01#4&e2c7dd6&0&00e1#{65e8773d-8f56-11d0-a3b9-00a0c9223196}\{ca465100-deb0-4d59-818f-8c477184adf6}":audio="Microphone"

Configure a crossbar device, specifying crossbar pins, and let the user adjust video capture properties at startup:

$ ffmpeg -f dshow -show_video_device_dialog true -crossbar_video_input_pin_number 0
     -crossbar_audio_input_pin_number 3 -i video="AVerMedia BDA Analog Capture":audio="AVerMedia BDA Analog Capture"

🧚 fbdev 🧚

This is the Linux framebuffer input device. 🐧

The Linux framebuffer is a graphic hardware-independent abstraction layer for showing graphics on a computer monitor — typically on the console. You access it through a file device node, usually /dev/fb0.

Wanna go deeper? Read the file Documentation/fb/framebuffer.txt that's included in the Linux source tree. 📖

See also <http://linux-fbdev.sourceforge.net/>, and fbset(1).

To record from the framebuffer device /dev/fb0 with ffmpeg:

ffmpeg -f fbdev -framerate 10 -i /dev/fb0 out.avi

And you can grab a single screenshot image with this command:

ffmpeg -f fbdev -framerate 1 -i /dev/fb0 -frames:v 1 screenshot.jpeg

⭐ Options ⭐

framerate — sets the frame rate. Default is 25.

👑 gdigrab 👑

The Win32 GDI-based screen capture device. This one lets you capture a region of the display on Windows. 🖥️✨

Amongst the options for the input filenames are such elements as:

desktop

title=<window_title>

hwnd=<window_hwnd>

So here's the tea ☕: the first option captures the entire desktop, or a fixed region of the desktop. The second and third options instead capture the contents of a single window, regardless of where that window is sitting on the screen. Main character energy, honestly — it doesn't care about its position.

For example, to grab the entire desktop using ffmpeg:

ffmpeg -f gdigrab -framerate 6 -i desktop out.mpg

Grab a 640x480 region at position "10,20":

ffmpeg -f gdigrab -framerate 6 -offset_x 10 -offset_y 20 -video_size vga -i desktop out.mpg

Grab the contents of the window named "Calculator":

ffmpeg -f gdigrab -framerate 6 -i title=Calculator out.mpg

💅 Options 💅

draw_mouse — specifies whether to draw the mouse pointer. Use the value 0 to not draw the pointer. Default value is 1. 🖱️
framerate — sets the grabbing frame rate. Default value is "ntsc", which corresponds to a frame rate of "30000/1001".
show_region — shows the grabbed region on screen.

If show_region is specified with 1, then the grabbing region gets indicated on screen. With this option it's super easy to know what's actually being grabbed if you're only grabbing a portion of the screen. 👀

⚠️ Note that show_region is incompatible with grabbing the contents of a single window. Pick a lane bestie.

For example:

ffmpeg -f gdigrab -show_region 1 -framerate 6 -video_size cif -offset_x 10 -offset_y 20 -i desktop out.mpg

video_size — sets the video frame size. The default is to capture the full screen if desktop is selected, or the full window size if title=window_title is selected.
offset_x — when capturing a region with video_size, sets the distance from the left edge of the screen or desktop.

📍 Note that the offset calculation is from the top left corner of the primary monitor on Windows. So if you've got a monitor positioned to the left of your primary monitor, you'll need to use a negative offset_x value to move the region over to that monitor. Sneaky but true.

offset_y — when capturing a region with video_size, sets the distance from the top edge of the screen or desktop.

Quick note before we dive back in, bestie: the offset math is measured from the top-left corner of the primary monitor on Windows. So if you've got a monitor sitting above your primary one, you're gonna need a negative offset_y value to push the region up onto that monitor. Sneaky, but that's how it works. 💅

iec61883

FireWire DV/HDV input device, powered by libiec61883. 🌈

To switch this input device on, you need libiec61883, libraw1394 and libavc1394 all installed on your system. Then configure with "--enable-libiec61883" to actually compile it in.

The iec61883 capture device supports grabbing from a video device connected over IEEE1394 (FireWire), using libiec61883 plus the new Linux FireWire stack (juju). This is the default DV/HDV input method in Linux Kernel 2.6.37 and later — because the old FireWire stack got yeeted. 💀

Just point it at the FireWire port you wanna use as the input file, or say "auto" and it'll pick the first port that's connected.

Options

dvtype — overrides the DV/HDV autodetection. Only use this when autodetection is being a hater, or when you specifically wanna forbid a different device type. Treating a DV device as HDV (or the other way round) will not work and results in undefined behavior. Supported values: auto, dv, and hdv.
dvbuffer — sets the max buffer size for incoming data, in frames. For DV this is an exact value. For HDV it is not frame exact, because HDV doesn't have a fixed frame size.
dvguid — picks the capture device by specifying its GUID. It'll only capture from that exact device and fails if no device with the given GUID is found. Super useful when you've got multiple devices plugged in at once. Go peek at /sys/bus/firewire/devices to find the GUIDs. 🔍

Examples

Grab and show the input of a FireWire DV/HDV device:

ffplay -f iec61883 -i auto

Grab and record the input of a FireWire DV/HDV device, using a packet buffer of 100000 packets if the source is HDV:

ffmpeg -f iec61883 -i auto -dvbuffer 100000 out.mpg

jack

The JACK input device. 🎧

To enable it at configure time you need libjack installed on your system.

A JACK input device creates one or more JACK writable clients — one for each audio channel — named client_name:input_N, where client_name is the name the application provides and N is a number identifying the channel. Each writable client sends the data it acquires to the FFmpeg input device.

Once you've made one or more JACK readable clients, you gotta connect them to one or more JACK writable clients. 🔗

To connect or disconnect JACK clients you can use the jack_connect and jack_disconnect programs, or do it through a graphical interface — qjackctl, for example.

To list the JACK clients and their properties, just invoke jack_lsp.

Here's an example showing how to capture a JACK readable client with ffmpeg:

# Create a JACK writable client with name "ffmpeg".
$ ffmpeg -f jack -i ffmpeg -y out.wav

# Start the sample jack_metro readable client.
$ jack_metro -b 120 -d 0.2 -f 4000

# List the current JACK clients.
$ jack_lsp -c
system:capture_1
system:capture_2
system:playback_1
system:playback_2
ffmpeg:input_1
metro:120_bpm

# Connect metro to the ffmpeg writable client.
$ jack_connect metro:120_bpm ffmpeg:input_1

For more info, read: <http://jackaudio.org/> 🌐

Options

channels — sets the number of channels. Default is 2.

kmsgrab

The KMS video input device. 🖥️

It captures the KMS scanout framebuffer tied to a specified CRTC or plane as a DRM object, which can then be passed along to other hardware functions.

Requires either DRM master or CAP_SYS_ADMIN to run. 👑

If you don't understand what all of that means, you probably don't want this. Look at x11grab instead. No cap. 🙈

Options

device — the DRM device to capture on. Defaults to /dev/dri/card0.
format — pixel format of the framebuffer. This can be autodetected if you're running Linux 5.7 or later, but it needs to be provided on earlier versions. Defaults to bgr0, which is the most common format used by the Linux console and Xorg X server.
format_modifier — format modifier to signal on output frames. This is necessary to import correctly into some APIs. It can be autodetected on Linux 5.7 or later, but you'll need to provide it explicitly when it's needed on earlier versions. Check the libdrm documentation for the possible values.
crtc_id — KMS CRTC ID that defines the capture source. The first active plane on the given CRTC gets used.
plane_id — KMS plane ID that defines the capture source. Defaults to the first active plane found if neither crtc_id nor plane_id are specified.
framerate — the framerate to capture at. This is not synchronised to any page flipping or framebuffer changes — it literally just defines the interval at which the framebuffer is sampled. Sampling faster than the framebuffer update rate will generate independent frames with the same content. Defaults to 30.

Examples

Capture from the first active plane, download the result to normal frames, and encode. This will only work if the framebuffer is both linear and mappable — if not, the result may be scrambled or fail to download. 😭

ffmpeg -f kmsgrab -i - -vf 'hwdownload,format=bgr0' output.mp4

Capture from CRTC ID 42 at 60fps, map the result to VAAPI, convert to NV12, and encode as H.264:

ffmpeg -crtc_id 42 -framerate 60 -f kmsgrab -i - -vf 'hwmap=derive_device=vaapi,scale_vaapi=w=1920:h=1080:format=nv12' -c:v h264_vaapi output.mp4

Wanna capture only part of a plane? Just crop the output — this can be used to capture a single window, as long as it has a known absolute position and size. For example, to capture and encode the middle quarter of a 1920x1080 plane:

ffmpeg -f kmsgrab -i - -vf 'hwmap=derive_device=vaapi,crop=960:540:480:270,scale_vaapi=960:540:nv12' -c:v h264_vaapi output.mp4

lavfi

The Libavfilter input virtual device. 🦋

This input device reads data from the open output pads of a libavfilter filtergraph.

For each filtergraph open output, the input device creates a corresponding stream mapped to the generated output. The filtergraph itself is specified through the graph option.

Options

graph — specifies the filtergraph to use as input. Each video open output must be labelled with a unique string of the form "outN", where N is a number starting from 0 corresponding to the mapped input stream generated by the device. The first unlabelled output is automatically assigned to the "out0" label, but all the others need to be specified explicitly.

The suffix "+subcc" can be appended to the output label to create an extra stream with the closed captions packets attached to that output (experimental; only for EIA-608 / CEA-708 for now 🧪). The subcc streams are created after all the normal streams, in the order of the corresponding stream. For example, if there is "out19+subcc", "out7+subcc" and up to "out42", the stream #43 is subcc for stream #7 and stream #44 is subcc for stream #19. Wild, right? 😅

If not specified, defaults to the filename specified for the input device.

graph_file — sets the filename of the filtergraph to be read and sent to the other filters. Syntax of the filtergraph is the same as the one specified by the option graph.
dumpgraph — dump graph to stderr.

Examples

Create a color video stream and play it back with ffplay:

ffplay -f lavfi -graph "color=c=pink [out0]" dummy

Same as the previous example, but use the filename for specifying the graph description, and omit the "out0" label:

ffplay -f lavfi color=c=pink

Create three different video test filtered sources and play them:

ffplay -f lavfi -graph "testsrc [out0]; testsrc,hflip [out1]; testsrc,negate [out2]" test3

Read an audio stream from a file using the amovie source and play it back with ffplay:

ffplay -f lavfi "amovie=test.wav"

Read an audio stream and a video stream and play it back with ffplay:

ffplay -f lavfi "movie=test.avi[out0];amovie=test.wav[out1]"

Dump decoded frames to images and Closed Captions to an RCWT backup:

ffmpeg -f lavfi -i "movie=test.ts[out0+subcc]" -map v frame%08d.png -map s -c copy -f rcwt subcc.bin

libcdio

Audio-CD input device based on libcdio. 💿

To enable this input device at configure time you need libcdio installed on your system. It requires the configure option "--enable-libcdio".

This device lets you play and grab from an Audio-CD. Iconic, honestly.

For example, to copy with ffmpeg the entire Audio-CD in /dev/sr0, you may run:

ffmpeg -f libcdio -i /dev/sr0 cd.wav

Options

speed — sets drive reading speed. Default value is 0.

The speed is specified in CD-ROM speed units. The speed is set through the libcdio cdio_cddap_speed_set function. On many CD-ROM drives, specifying a value that's too large will just result in using the fastest speed.

paranoia_mode — sets paranoia recovery mode flags. It accepts one of the following values:
- disable
- verify
- overlap
- neverskip
- full

Default value is disable.

For more info about the available recovery modes, consult the paranoia project documentation. 📚

libdc1394

IIDC1394 input device, based on libdc1394 and libraw1394.

Requires the configure option "--enable-libdc1394".

Options

framerate — sets the frame rate. Default is "ntsc", corresponding to a frame rate of "30000/1001".
pixel_format — selects the pixel format. Default is "uyvy422".
video_size — sets the video size given as a string such as "640x480" or "hd720". Default is "qvga".

openal

The OpenAL input device provides audio capture on all systems with a working OpenAL 1.1 implementation. 🎙️

To enable this input device at configure time, you need the OpenAL headers and libraries installed on your system, and you need to configure FFmpeg with "--enable-openal".

okay so your OpenAL headers and libraries should come bundled with your OpenAL implementation, or you can grab them as an extra download (an SDK, if we're being official about it 🎀). Depending on your install you might need to pass extra flags via --extra-cflags and --extra-ldflags so the build system can actually locate the OpenAL headers and libraries. lowkey a scavenger hunt but we move 💅

An incomplete list of OpenAL implementations follows (yes, incomplete, there are more out there, this is just the starters):

Creative — the official Windows implementation, serving up hardware acceleration with supported devices plus a software fallback. See http://openal.org/.
OpenAL Soft — portable, open source (LGPL) software implementation. Comes with backends for the most common sound APIs on the Windows, Linux, Solaris, and BSD operating systems. See http://kcat.strangesoft.net/openal.html.
Apple — OpenAL is part of Core Audio, the official Mac OS X Audio interface. See http://developer.apple.com/technologies/mac/audio-and-video.html

This device lets you capture from an audio input device that's handled through OpenAL. 🎧

You need to specify the name of the device to capture in the provided filename. If you pass the empty string, the device will automatically select the default device, no thoughts head empty, it just handles it ✨. You can get the list of the supported devices by using the option list_devices.

Options 🌸

channels — sets the number of channels in the captured audio. Only the values 1 (monaural) and 2 (stereo) are currently supported. Defaults to 2.
sample_size — sets the sample size (in bits) of the captured audio. Only the values 8 and 16 are currently supported. Defaults to 16.
sample_rate — sets the sample rate (in Hz) of the captured audio. Defaults to 44.1k.
list_devices — if set to true, prints a list of devices and exits. Defaults to false.

Examples 🦋

Print the list of OpenAL supported devices and exit:

$ ffmpeg -list_devices true -f openal -i dummy out.ogg

Capture from the OpenAL device DR-BT101 via PulseAudio:

$ ffmpeg -f openal -i 'DR-BT101 via PulseAudio' out.ogg

Capture from the default device (note the empty string '' as filename — yes, literally nothing, iconic):

$ ffmpeg -f openal -i '' out.ogg

Capture from two devices simultaneously, writing to two different files, within the same ffmpeg command:

$ ffmpeg -f openal -i 'DR-BT101 via PulseAudio' out1.ogg -f openal -i 'ALSA Default' out2.ogg

Note: not all OpenAL implementations support multiple simultaneous capture — try the latest OpenAL Soft if the above does not work. 💀

oss 🐚

Open Sound System input device.

The filename you hand to the input device is the device node representing the OSS input device, and is usually set to /dev/dsp.

For example to grab from /dev/dsp using ffmpeg use the command:

ffmpeg -f oss -i /dev/dsp /tmp/oss.wav

For more information about OSS see: http://manuals.opensound.com/usersguide/dsp.html

Options 🌈

sample_rate — sets the sample rate in Hz. Default is 48000.
channels — sets the number of channels. Default is 2.

pulse 🫶

PulseAudio input device.

To enable this output device you need to configure FFmpeg with --enable-libpulse. (output device, yes, that's what it says, no I will not be taking questions 🙈)

The filename you provide to the input device is a source device or the string "default".

To list the PulseAudio source devices and their properties you can invoke the command pactl list sources.

More information about PulseAudio can be found on http://www.pulseaudio.org.

Options 💜

server — connects to a specific PulseAudio server, specified by an IP address. Default server is used when not provided.
name — specifies the application name PulseAudio will use when showing active clients, by default it is the "LIBAVFORMAT_IDENT" string.
stream_name — specifies the stream name PulseAudio will use when showing active streams, by default it is "record".
sample_rate — specifies the samplerate in Hz, by default 48kHz is used.
channels — specifies the channels in use, by default 2 (stereo) is set.
frame_size — this option does nothing and is deprecated. it's giving ghost 👻 it's literally not real
fragment_size — specifies the size in bytes of the minimal buffering fragment in PulseAudio, it will affect the audio latency. By default it is set to 50 ms amount of data.
wallclock — sets the initial PTS using the current time. Default is 1.

Examples 🎀

Record a stream from default device:

ffmpeg -f pulse -i default /tmp/pulse.wav

sndio 🌊

sndio input device.

To enable this input device during configuration you need libsndio installed on your system.

The filename you provide to the input device is the device node representing the sndio input device, and is usually set to /dev/audio0.

For example to grab from /dev/audio0 using ffmpeg use the command:

ffmpeg -f sndio -i /dev/audio0 /tmp/oss.wav

Options ⭐

sample_rate — sets the sample rate in Hz. Default is 48000.
channels — sets the number of channels. Default is 2.

video4linux2, v4l2 👑

Video4Linux2 input video device.

v4l2 can be used as alias for video4linux2. same energy, shorter name, slay ✨

If FFmpeg is built with v4l-utils support (by using the --enable-libv4l2 configure option), it is possible to use it with the -use_libv4l2 input device option.

The name of the device to grab is a file device node. Usually Linux systems tend to automatically create such nodes when the device (e.g. an USB webcam) is plugged into the system, and it has a name of the kind /dev/videoN, where N is a number associated to the device.

Video4Linux2 devices usually support a limited set of widthxheight sizes and frame rates. You can check which are supported using -list_formats all for Video4Linux2 devices. Some devices, like TV cards, support one or more standards. It is possible to list all the supported standards using -list_standards all.

The time base for the timestamps is 1 microsecond. Depending on the kernel version and configuration, the timestamps may be derived from the real time clock (origin at the Unix Epoch) or the monotonic clock (origin usually at boot time, unaffected by NTP or manual changes to the clock). The -timestamps abs or -ts abs option can be used to force conversion into the real time clock. ⏰ no cap, clocks are complicated

Some usage examples of the video4linux2 device with ffmpeg and ffplay:

List supported formats for a video4linux2 device:

ffplay -f video4linux2 -list_formats all /dev/video0

Grab and show the input of a video4linux2 device:

ffplay -f video4linux2 -framerate 30 -video_size hd720 /dev/video0

Grab and record the input of a video4linux2 device, leave the frame rate and size as previously set:

ffmpeg -f video4linux2 -input_format mjpeg -i /dev/video0 out.mpeg

For more information about Video4Linux, check http://linuxtv.org/.

Options 💖

standard — sets the standard. Must be the name of a supported standard. To get a list of the supported standards, use the list_standards option.
channel — sets the input channel number. Default to -1, which means using the previously selected channel.
video_size — sets the video frame size. The argument must be a string in the form WIDTHxHEIGHT or a valid size abbreviation.
pixel_format — selects the pixel format (only valid for raw video input).
input_format — sets the preferred pixel format (for raw video) or a codec name. This option allows one to select the input format, when several are available.
framerate — sets the preferred video frame rate.
list_formats — lists available formats (supported pixel formats, codecs, and frame sizes) and exits.

Available values are:

all — show all available (compressed and non-compressed) formats.
raw — show only raw video (non-compressed) formats.
compressed — show only compressed formats.

list_standards — lists supported standards and exits.

Available values are:

all — show all supported standards.

timestamps, ts — sets type of timestamps for grabbed frames.

Available values are:

default — use timestamps from the kernel.
abs — use absolute timestamps (wall clock).
mono2abs — force conversion from monotonic to absolute timestamps.

Default value is "default".

use_libv4l2 — use libv4l2 (v4l-utils) conversion functions. Default is 0.

vfwcap 🎬

VfW (Video for Windows) capture input device.

The filename passed as input is the capture driver number, ranging from 0 to 9. You may use "list" as filename to print a list of drivers. Any other filename will be interpreted as device number 0. (so like, just don't be weird with it and you'll be fine 😭)

Options 🌸

video_size — sets the video frame size.
framerate — sets the grabbing frame rate. Default value is "ntsc", corresponding to a frame rate of "30000/1001".

x11grab 🖥️✨

X11 video input device.

To enable this input device during configuration you need libxcb installed on your system. It will be automatically detected during configuration. it just knows, main character energy fr

This device allows one to capture a region of an X11 display.

The filename passed as input has the syntax:

[<hostname>]:<display_number>.<screen_number>[+<x_offset>,<y_offset>]

hostname:display_number.screen_number specifies the X11 display name of the screen to grab from. hostname can be omitted, and defaults to "localhost". The environment variable DISPLAY contains the default display name.

x_offset and y_offset specify the offsets of the grabbed area with respect to the top-left border of the X11 screen. They default to 0.

Check the X11 documentation (e.g. man X) for more detailed information.

Use the xdpyinfo program for getting basic information about the properties of your X11 display (e.g. grep for "name" or "dimensions").

For example to grab from :0.0 using ffmpeg:

ffmpeg -f x11grab -framerate 25 -video_size cif -i :0.0 out.mpg

Grab at position "10,20":

ffmpeg -f x11grab -framerate 25 -video_size cif -i :0.0+10,20 out.mpg

Options 💅

select_region — specifies whether to select the grabbing area graphically using the pointer. A value of 1 prompts the user to select the grabbing area graphically by clicking and dragging. A single click with no dragging will select the whole screen. A region with zero width or height will also select the whole screen. This option overwrites the video_size, grab_x, and grab_y options. Default value is 0.
draw_mouse — specifies whether to draw the mouse pointer. A value of 0 specifies not to draw the pointer. Default value is 1.
follow_mouse — makes the grabbed area follow the mouse, bestie 💅 The argument can be "centered" or a number of pixels PIXELS.

When it’s specified with "centered", the grabbing region follows the mouse pointer and keeps the pointer at the center of the region; otherwise, the region follows only when the mouse pointer reaches within PIXELS (greater than zero) to the edge of the region.

For example:

ffmpeg -f x11grab -follow_mouse centered -framerate 25 -video_size cif -i :0.0 out.mpg

To follow only when the mouse pointer reaches within 100 pixels to the edge:

ffmpeg -f x11grab -follow_mouse 100 -framerate 25 -video_size cif -i :0.0 out.mpg

framerate — sets the grabbing frame rate. Default value is ntsc, corresponding to a frame rate of 30000/1001 ✨
show_region — shows the grabbed region on screen.

If show_region is specified with 1, then the grabbing region will be indicated on screen. With this option, it’s easy to know what’s being grabbed if only a portion of the screen is grabbed 👀

region_border — sets the region border thickness if -show_region 1 is used. Range is 1 to 128 and default is 3 (XCB-based x11grab only).

For example:

ffmpeg -f x11grab -show_region 1 -framerate 25 -video_size cif -i :0.0+10,20 out.mpg

With follow_mouse:

ffmpeg -f x11grab -follow_mouse centered -show_region 1 -framerate 25 -video_size cif -i :0.0 out.mpg

window_id — grabs this window instead of the whole screen. Default value is 0, which maps to the whole screen (root window).

The id of a window can be found using the xwininfo program, possibly with options -tree and -root.

If the window is later enlarged, the new area is not recorded. Video ends when the window is closed, unmapped (i.e., iconified) or shrunk beyond the video size (which defaults to the initial window size).

This option disables options follow_mouse and select_region 💀

video_size — sets the video frame size. Default is the full desktop or window.
grab_x, grab_y — set the grabbing region coordinates. They are expressed as offset from the top left corner of the X11 window and correspond to the x_offset and y_offset parameters in the device name. The default value for both options is 0.

🌈 OUTPUT DEVICES 🌈

Output devices are configured elements in FFmpeg that can write multimedia data to an output device attached to your system 🫶

When you configure your FFmpeg build, all the supported output devices are enabled by default. You can list all available ones using the configure option --list-outdevs ✨

You can disable all the output devices using the configure option --disable-outdevs, and selectively enable an output device using the option --enable-outdev=OUTDEV, or you can disable a particular input device using the option --disable-outdev=OUTDEV.

The option -devices of the ff* tools will display the list of enabled output devices.

A description of the currently available output devices follows 👑

🎧 alsa

ALSA (Advanced Linux Sound Architecture) output device.

Examples

Play a file on default ALSA device:

ffmpeg -i INPUT -f alsa default

Play a file on soundcard 1, audio device 7:

ffmpeg -i INPUT -f alsa hw:1,7

🎛️ AudioToolbox

AudioToolbox output device.

Allows native output to CoreAudio devices on OSX 🍓

The output filename can be empty (or -) to refer to the default system output device or a number that refers to the device index as shown using: -list_devices true.

Alternatively, the audio input device can be chosen by index using -audio_device_index <INDEX>, overriding any device name or index given in the input filename.

All available devices can be enumerated by using -list_devices true, listing all device names, UIDs and corresponding indices.

Options

AudioToolbox supports the following options:

-audio_device_index <INDEX> — specify the audio device by its index. Overrides anything given in the output filename.

Examples

Print the list of supported devices and output a sine wave to the default device:

$ ffmpeg -f lavfi -i sine=r=44100 -f audiotoolbox -list_devices true -

Output a sine wave to the device with the index 2, overriding any output filename:

$ ffmpeg -f lavfi -i sine=r=44100 -f audiotoolbox -audio_device_index 2 -

🦋 caca

CACA output device.

This output device allows one to show a video stream in a CACA window. Only one CACA window is allowed per application, so you can have only one instance of this output device in an application 😭

To enable this output device you need to configure FFmpeg with --enable-libcaca. libcaca is a graphics library that outputs text instead of pixels.

For more information about libcaca, check: http://caca.zoy.org/wiki/libcaca

Options

window_title — set the CACA window title, if not specified default to the filename specified for the output device.
window_size — set the CACA window size, can be a string of the form widthxheight or a video size abbreviation. If not specified it defaults to the size of the input video.
driver — set display driver.
algorithm — set dithering algorithm. Dithering is necessary because the picture being rendered has usually far more colours than the available palette. The accepted values are listed with -list_dither algorithms.
antialias — set antialias method. Antialiasing smoothens the rendered image and avoids the commonly seen staircase effect. The accepted values are listed with -list_dither antialiases.
charset — set which characters are going to be used when rendering text. The accepted values are listed with -list_dither charsets.
color — set color to be used when rendering text. The accepted values are listed with -list_dither colors.
list_drivers — if set to true, print a list of available drivers and exit.
list_dither — list available dither options related to the argument. The argument must be one of algorithms, antialiases, charsets, colors.

Examples

The following command shows the ffmpeg output is an CACA window, forcing its size to 80x25:

ffmpeg -i INPUT -c:v rawvideo -pix_fmt rgb24 -window_size 80x25 -f caca -

Show the list of available drivers and exit:

ffmpeg -i INPUT -pix_fmt rgb24 -f caca -list_drivers true -

Show the list of available dither colors and exit:

ffmpeg -i INPUT -pix_fmt rgb24 -f caca -list_dither colors -

🎬 decklink

The decklink output device provides playback capabilities for Blackmagic DeckLink devices.

To enable this output device, you need the Blackmagic DeckLink SDK and you need to configure with the appropriate --extra-cflags and --extra-ldflags. On Windows, you need to run the IDL files through widl.

DeckLink is very picky about the formats it supports. Pixel format is always uyvy422, framerate, field order and video size must be determined for your device with -list_formats 1. Audio sample rate is always 48 kHz 😤

Options

list_devices — if set to true, print a list of devices and exit. Defaults to false. This option is deprecated, please use the -sinks option of ffmpeg to list the available output devices.
list_formats — if set to true, print a list of supported formats and exit. Defaults to false.
preroll — amount of time to preroll video in seconds. Defaults to 0.5.
duplex_mode — sets the decklink device duplex/profile mode. Must be unset, half, full, one_sub_device_full, one_sub_device_half, two_sub_device_full, four_sub_device_half. Defaults to unset.

Valid profile modes for DeckLink 8K Pro (with DeckLink SDK >= 11.0): one_sub_device_full, one_sub_device_half, two_sub_device_full, four_sub_device_half

Valid profile modes for DeckLink Quad 2 and DeckLink Duo 2: half, full

timing_offset — sets the genlock timing pixel offset on the used output. Defaults to unset.
link — sets the SDI video link configuration on the used output. Must be unset, single link SDI, dual link SDI or quad link SDI. Defaults to unset.
sqd — enable Square Division Quad Split mode for Quad-link SDI output. Must be unset, true or false. Defaults to unset.
level_a — enable SMPTE Level A mode on the used output. Must be unset, true or false. Defaults to unset.
vanc_queue_size — sets maximum output buffer size in bytes for VANC data. If the buffering reaches this value, outgoing VANC data will be dropped. Defaults to 1048576.

Examples

List output devices:

ffmpeg -sinks decklink

List supported formats:

ffmpeg -i test.avi -f decklink -list_formats 1 'DeckLink Mini Monitor'

Play video clip:

ffmpeg -i test.avi -f decklink -pix_fmt uyvy422 'DeckLink Mini Monitor'

Play video clip with non-standard framerate or video size:

ffmpeg -i test.avi -f decklink -pix_fmt uyvy422 -s 720x486 -r 24000/1001 'DeckLink Mini Monitor'

🖥️ fbdev

Linux framebuffer output device.

The Linux framebuffer is a graphic hardware-independent abstraction layer to show graphics on a computer monitor, typically on the console. It is accessed through a file device node, usually /dev/fb0.

For more detailed information read the file Documentation/fb/framebuffer.txt included in the Linux source tree.

Options

xoffset, yoffset — set x/y coordinate of top left corner. Default is 0.

Examples

Play a file on framebuffer device /dev/fb0. Required pixel format depends on current framebuffer settings.

ffmpeg -re -i INPUT -c:v rawvideo -pix_fmt bgra -f fbdev /dev/fb0

See also http://linux-fbdev.sourceforge.net/, and fbset(1).

🔊 oss

OSS (Open Sound System) output device.

🫧 pulse

PulseAudio output device.

To enable this output device you need to configure FFmpeg with --enable-libpulse.

More information about PulseAudio can be found on http://www.pulseaudio.org

Options

server — points you at a specific PulseAudio server, given by an IP address. Don't pass it? You get the default server, easy peasy 🫶
name — the application name PulseAudio will show off when it's listing active clients. Default is the "LIBAVFORMAT_IDENT" string, iykyk.
stream_name — the name PulseAudio uses when showing active streams. By default it's just set to whatever output name you specified. cute.
device — the device you actually wanna use. Leave it out and the default device gets picked. Wanna see your options? Pull up the list of output devices with pactl list sinks 🔍
buffer_size, buffer_duration — these two control the size and the duration of the PulseAudio buffer. Small buffer = more control but you're gonna be updating way more often, high maintenance queen energy fr.

buffer_size is the size in bytes, while buffer_duration is the duration in milliseconds.

Give it both and the highest value wins (the duration gets recalculated into bytes using the stream parameters, math is mathing ✨). Set them to 0 — which is the default, btw — and the device will just use the default PulseAudio duration value. By default PulseAudio sets buffer duration to around 2 seconds.

prebuf — pre-buffering size in bytes. The server will NOT start playback until at least prebuf bytes are sitting in the buffer, no exceptions bestie. By default this option initializes to the same value as buffer_size or buffer_duration, whichever is bigger.
minreq — minimum request size, in bytes. The server never requests less than minreq bytes from the client — instead it waits until the buffer is free enough to request more bytes all at once. Lowkey recommended that you don't set this one at all; leave it and it initializes to a value the server itself deems sensible 👑

✨ Examples ✨

Play a file on the default device on the default server:

ffmpeg  -i INPUT -f pulse "stream name"

🌈 `sndio` 🌈

sndio audio output device. that's it, that's the whole vibe 🦋

💜 `v4l2` 💜

Video4Linux2 output device. iconic behavior.

🎀 `xv` 🎀

XV (XVideo) output device.

This output device is the one that lets you show a video stream in a X Window System window. very main character of it 💅

Options

display_name — the hardware display name, which decides the display and communications domain being used.

The display name — or the DISPLAY environment variable — can be a string in the format hostname[:number[.screen_number]].

hostname is the name of the host machine the display is physically attached to. number is the number of the display server on that host machine. screen_number is the screen to be used on that server.

Leave it unspecified and it defaults to the value of the DISPLAY environment variable.

For example, "dual-headed:0.1" would specify screen 1 of display 0 on the machine named ``dual-headed''.

Check the X11 specification for the full detailed tea on the display name format 📜

window_id — set it to a non-zero value and the device won't create a new window at all, it'll use the existing one with the window_id you provided. By default this option is 0 and the device creates its own window. independent queen.
window_size — sets the size of the created window. Can be a string of the form widthxheight or a video size abbreviation. Not specified? It defaults to the size of the input video. Ignored when window_id is set 🙈
window_x, window_y — set the X and Y window offsets for the created window. They're both 0 by default. Fair warning: the window manager may just ignore these values, it does what it wants. Also ignored when window_id is set.
window_title — sets the window title. Not specified? Defaults to the filename you specified for the output device. Ignored when window_id is set.

For more info about XVideo, go peek at http://www.x.org/ ⭐

✨ Examples ✨

Decode, display and encode video input with ffmpeg all at the same time, multitasking legend:

ffmpeg -i INPUT OUTPUT -f xv display

Decode and display the input video into multiple X11 windows, because one window is never enough:

ffmpeg -i INPUT -f xv normal -vf negate -f xv negated

🌸 RESAMPLER OPTIONS 🌸

okay so the audio resampler has all these named options and honestly it's giving "customization queen" 💅 here's the whole menu, bestie:

You can set these by passing -option value in the FFmpeg tools, or option=value for the aresample filter, or by setting the value directly in the SwrContext options, or with the libavutil/opt.h API if you're doing that programmatic thing ✨

uchl, used_chlayout — sets the used input channel layout. Default is unset, and ngl this one is only used for special remapping 🎀
isr, in_sample_rate — sets the input sample rate. Default is 0.
osr, out_sample_rate — sets the output sample rate. Default is 0.
isf, in_sample_fmt — specifies the input sample format. Defaults to none 🦋
osf, out_sample_fmt — specifies the output sample format. Also defaults to none.
tsf, internal_sample_fmt — sets the internal sample format. Default is none, and if you don't set it explicitly it just gets chosen automatically for you. main character energy, it picks itself 😌
ichl, in_chlayout and ochl, out_chlayout — set the input/output channel layout fr. Peep the Channel Layout section in the ffmpeg-utils(1) manual for the syntax you need to use.
clev, center_mix_level — center mix level, a value in deciBel that must be in the interval [-32,32].
slev, surround_mix_level — surround mix level, also in deciBel, also gotta be in [-32,32].
lfe_mix_level — sets LFE mix into non-LFE level. It kicks in when you've got an LFE input but no LFE output. deciBel again, and yes, interval [-32,32] 🫶
rmvol, rematrix_volume — sets rematrix volume. Default value is 1.0.
rematrix_maxval — maximum output value for rematrixing. Use this to pick between preventing clipping vs. preventing volume reduction. A value of 1.0 prevents clipping 💀
flags, swr_flags — the flags the converter uses. Default value is 0. Supports this individual flag:
- res — force resampling. This one makes resampling happen even when the input and output sample rates match, no cap.
dither_scale — sets the dither scale. Default value is 1.
dither_method — sets the dither method. Default value is 0. Supported values:
- rectangular — select rectangular dither
- triangular — select triangular dither
- triangular_hp — select triangular dither with high pass
- lipshitz — select Lipshitz noise shaping dither
- shibata — select Shibata noise shaping dither
- low_shibata — select low Shibata noise shaping dither
- high_shibata — select high Shibata noise shaping dither
- f_weighted — select f-weighted noise shaping dither
- modified_e_weighted — select modified-e-weighted noise shaping dither
- improved_e_weighted — select improved-e-weighted noise shaping dither
resampler — sets the resampling engine. Default value is swr. Supported values:
- swr — select the native SW Resampler. heads up: the filter options precision and cheby are not applicable in this case.
- soxr — select the SoX Resampler (where available). Also heads up: compensation, and the filter options filter_size, phase_shift, exact_rational, filter_type & kaiser_beta, are not applicable in this case.
filter_size — for swr only, sets the resampling filter size. Default value is 32.
phase_shift — for swr only, sets the resampling phase shift. Default value is 10, and it must be in the interval [0,30].
linear_interp — use linear interpolation when enabled (that's the default, btw). Disable it if you'd rather preserve speed than quality when exact_rational fails 🏃‍♀️
exact_rational — for swr only. When enabled, it tries to use exact phase_count based on the input and output sample rate. But if that ends up larger than 1 << phase_shift, then phase_count falls back to 1 << phase_shift. Default is enabled.
cutoff — sets the cutoff frequency ratio (swr: 6dB point; soxr: 0dB point). Must be a float between 0 and 1. Default is 0.97 with swr, and 0.91 with soxr (which, at a sample rate of 44100, preserves the entire audio band to 20kHz 👑).
precision — for soxr only. The precision in bits that the resampled signal gets calculated to. The default 20 (which, with suitable dithering, is appropriate for a destination bit-depth of 16) gives SoX's 'High Quality'; a value of 28 gives SoX's 'Very High Quality' ⭐
cheby — for soxr only. Selects passband rolloff none (Chebyshev) & higher-precision approximation for 'irrational' ratios. Default value is 0.
async — for swr only. Simple 1-parameter audio sync to timestamps using stretching, squeezing, filling and trimming. Setting it to 1 enables filling and trimming, and bigger values represent the max amount in samples the data may be stretched or squeezed for each second. Default value is 0, so no compensation is applied to make the samples match the audio timestamps.
first_pts — for swr only. Assume the first pts should be this value. Time unit is 1 / sample rate. This lets you pad/trim at the start of the stream. By default no assumption is made about the first frame's expected pts, so no padding or trimming happens. Like, you could set this to 0 to pad the beginning with silence if an audio stream starts after the video stream, or to trim any samples with a negative pts caused by encoder delay 🎀
min_comp — for swr only. Sets the minimum difference between timestamps and audio data (in seconds) needed to trigger stretching/squeezing/filling or trimming so the data matches the timestamps. The default is that stretching/squeezing/filling and trimming is disabled (min_comp = FLT_MAX).
min_hard_comp — for swr only. Sets the minimum difference between timestamps and audio data (in seconds) needed to trigger adding/dropping samples so it matches the timestamps. This is basically a threshold that picks between hard (trim/fill) and soft (squeeze/stretch) compensation. Note that all compensation is disabled by default through min_comp. The default here is 0.1.
comp_duration — for swr only. Sets the duration (in seconds) over which data gets stretched/squeezed to match the timestamps. Must be a non-negative double float value, default value is 1.0.
max_soft_comp — for swr only. Sets the maximum factor by which data gets stretched/squeezed to match the timestamps. Must be a non-negative double float value, default value is 0.
matrix_encoding — selects matrixed stereo encoding. Accepts these values:
- none — select none
- dolby — select Dolby
- dplii — select Dolby Pro Logic II

Default value is none.

filter_type — for swr only. Selects the resampling filter type. This only affects resampling operations, bestie. Accepts:
- cubic — select cubic
- blackman_nuttall — select Blackman Nuttall windowed sinc
- kaiser — select Kaiser windowed sinc
kaiser_beta — for swr only. Sets the Kaiser window beta value. Must be a double float value in the interval [2,16], default value is 9.
output_sample_bits — for swr only. Sets the number of used output sample bits for dithering. Must be an integer in the interval [0,64], default value is 0, which means it's not used. slay 💜

🌸 SCALER OPTIONS 🌸

Okay bestie, so the video scaler has its own little collection of named options and they're actually kind of a lot. Let's get into it 💅

You can set these by writing -option value in the FFmpeg tools — with a couple of API-only exceptions that I'll flag below, don't worry. And if you're doing the programmatic thing? They can be set explicitly in the "SwsContext" options or through the libavutil/opt.h API. iykyk ✨

sws_flags — sets the scaler flags, and it's also what you use to pick the scaling algorithm. Important: only a single algorithm should be selected, no mixing and matching 🙈 Default value is bicubic.

It accepts the following values:

fast_bilinear — selects the fast bilinear scaling algorithm.
bilinear — selects the bilinear scaling algorithm.
bicubic — selects the bicubic scaling algorithm.
experimental — selects the experimental scaling algorithm.
neighbor — selects the nearest neighbor rescaling algorithm.
area — selects the averaging area rescaling algorithm.
bicublin — selects the bicubic scaling algorithm for the luma component, and bilinear for the chroma components.
gauss — selects the Gaussian rescaling algorithm.
sinc — selects the sinc rescaling algorithm.
lanczos — selects the Lanczos rescaling algorithm. The default width (alpha) is 3 and it can be changed by setting "param0". slay 🔥
spline — selects the natural bicubic spline rescaling algorithm.
print_info — enables printing/debug logging.
accurate_rnd — enables accurate rounding.
full_chroma_int — enables full chroma interpolation.
full_chroma_inp — selects full chroma input.
bitexact — enables bitexact output.
unstable — allows the use of experimental new code. Heads up bestie, this may subtly affect the output or even produce wrong results. For testing only, fr fr 💀

srcw (API only) — set source width.
srch (API only) — set source height.
dstw (API only) — set destination width.
dsth (API only) — set destination height.
src_format (API only) — set source pixel format (must be expressed as an integer).
dst_format (API only) — set destination pixel format (must be expressed as an integer).
src_range (boolean) — if the value is set to 1, it indicates the source is full range. Default value is 0, which indicates the source is limited range.
dst_range (boolean) — if the value is set to 1, enable full range for the destination. Default value is 0, which enables limited range.
gamma (boolean) — if the value is set to 1, enable gamma correct scaling. Default value is 0.
param0, param1 — set scaling algorithm parameters. The values you give are specific to some scaling algorithms and just straight up ignored by others 🙃 The specified values are floating point number values.
sws_dither — set the dithering algorithm. Accepts one of the following values. Default value is auto.
- auto — automatic choice
- none — no dithering
- bayer — bayer dither
- ed — error diffusion dither
- a_dither — arithmetic dither, based using addition
- x_dither — arithmetic dither, based using xor (more random/less apparent patterning than a_dither). lowkey iconic 🦋
alphablend — set the alpha blending to use when the input has alpha but the output does not. Default value is none.
- uniform_color — blend onto a uniform background color
- checkerboard — blend onto a checkerboard
- none — no blending

✨ FILTERING INTRODUCTION ✨

Filtering in FFmpeg is enabled through the libavfilter library. Periodt. 🎀

So in libavfilter, a filter can have multiple inputs and multiple outputs — main character energy honestly. To illustrate the sorts of things that are possible, consider the following filtergraph:

                        [main]
        input --> split ---------------------> overlay --> output
                    |                             ^
                    |[tmp]                  [flip]|
                    +-----> crop --> vflip -------+

This filtergraph splits the input stream into two streams, then sends one stream through the crop filter and the vflip filter, before merging it back with the other stream by overlaying it on top. You can use this command to achieve it:

ffmpeg -i INPUT -vf "split [main][tmp]; [tmp] crop=iw:ih/2:0:0, vflip [flip]; [main][flip] overlay=0:H/2" OUTPUT

The result? The top half of the video gets mirrored onto the bottom half of the output video. like, literally the coolest thing 😭💖

Filters in the same linear chain are separated by commas, and distinct linear chains of filters are separated by semicolons. In our example, crop,vflip are in one linear chain, and split and overlay are separately in another. The points where the linear chains join are labelled by names enclosed in square brackets. In the example, the split filter generates two outputs that are associated to the labels [main] and [tmp].

The stream sent to the second output of split, labelled as [tmp], is processed through the crop filter, which crops away the lower half part of the video, and then vertically flipped. The overlay filter takes in input the first unchanged output of the split filter (which was labelled as [main]), and overlays on its lower half the output generated by the crop,vflip filterchain.

Some filters take in input a list of parameters: they're specified after the filter name and an equal sign, and are separated from each other by a colon.

And just so you know, there exist so-called source filters that do not have an audio/video input, and sink filters that will not have audio/video output. 🫶

🌈 GRAPH 🌈

The graph2dot program included in the FFmpeg tools directory can be used to parse a filtergraph description and issue a corresponding textual representation in the dot language. ngl that's kind of a flex 💅

Invoke the command:

graph2dot -h

to see how to use graph2dot.

You can then pass the dot description to the dot program (from the graphviz suite of programs) and obtain a graphical representation of the filtergraph. obsessed honestly 🦋✨

For example, the sequence of commands:

echo <GRAPH_DESCRIPTION> | \
tools/graph2dot -o graph.tmp && \
dot -Tpng graph.tmp -o graph.png && \
display graph.png

can be used to create and display an image representing the graph described by the GRAPH_DESCRIPTION string. Note that this string must be a complete self-contained graph, with its inputs and outputs explicitly defined. For example, if your command line is of the form:

ffmpeg -i infile -vf scale=640:360 outfile

your GRAPH_DESCRIPTION string will need to be of the form:

nullsrc,scale=640:360,nullsink

you may also need to set the nullsrc parameters and add a format filter in order to simulate a specific input file. sneaky but necessary 🙈💜

💅 FILTERGRAPH DESCRIPTION 💅

okay so bestie, a filtergraph is literally a directed graph of connected filters. and yeah — it can contain cycles, and there can be multiple links between a pair of filters. no cap. 🌈

each link has one input pad on one side connecting it to one filter from which it takes its input, and one output pad on the other side connecting it to one filter accepting its output. that's the whole vibe. 🫶

each filter in a filtergraph is an instance of a filter class registered in the application, and that class is what defines the features and the number of input and output pads of the filter. 👑

and the terminology, because we love vocabulary:

a filter with no input pads is called a "source" ✨
a filter with no output pads is called a "sink" 🕳️

✨ Filtergraph syntax ✨

a filtergraph has a textual representation, and it's recognized by the -filter/-vf/-af and -filter_complex options in ffmpeg and -vf/-af in ffplay, and by the avfilter_graph_parse_ptr() function defined in libavfilter/avfilter.h. iykyk. 💜

a filterchain consists of a sequence of connected filters, each one connected to the previous one in the sequence. a filterchain is represented by a list of ","-separated filter descriptions.

a filtergraph consists of a sequence of filterchains. a sequence of filterchains is represented by a list of ";"-separated filterchain descriptions. 🎀

a filter is represented by a string of the form:

[in_link_1]...[in_link_N]filter_name@id=arguments[out_link_1]...[out_link_M]

filter_name is the name of the filter class of which the described filter is an instance of, and has to be the name of one of the filter classes registered in the program, optionally followed by @id. the name of the filter class is optionally followed by a string =arguments. 🌸

arguments is a string which contains the parameters used to initialize the filter instance. the manpage says it "may have one of two forms" and then lists three bullets — and honestly? we stan a generous queen. here they are:

A ':'-separated list of key=value pairs. 💖
A ':'-separated list of value. in this case, the keys are assumed to be the option names in the order they are declared. E.g. the "fade" filter declares three options in this order -- type, start_frame and nb_frames. then the parameter list in:0:30 means that the value in is assigned to the option type, 0 to start_frame and 30 to nb_frames. 🔥
A ':'-separated list of mixed direct value and long key=value pairs. the direct value must precede the key=value pairs, and follow the same constraints order of the previous point. the following key=value pairs can be set in any preferred order. 💅

if the option value itself is a list of items (e.g. the "format" filter takes a list of pixel formats), the items in the list are usually separated by |. ⭐

the list of arguments can be quoted using the character ' as initial and ending mark, and the character \ for escaping the characters within the quoted text; otherwise the argument string is considered terminated when the next special character (belonging to the set []=;,) is encountered. lowkey important, highkey useful. 🦋

a special syntax implemented in the ffmpeg CLI tool allows loading option values from files. this is done by prepending a slash / to the option name, then the supplied value is interpreted as a path from which the actual value is loaded. E.g.

ffmpeg -i <INPUT> -vf drawtext=/text=/tmp/some_text <OUTPUT>

will load the text to be drawn from /tmp/some_text. API users wishing to implement a similar feature should use the "avfilter_graph_segment_*()" functions together with custom IO code. 🧚

the name and arguments of the filter are optionally preceded and followed by a list of link labels. a link label allows one to name a link and associate it to a filter output or input pad. the preceding labels in_link_1 ... in_link_N are associated to the filter input pads, the following labels out_link_1 ... out_link_M are associated to the output pads. 👑

when two link labels with the same name are found in the filtergraph, a link between the corresponding input and output pad is created. match made in heaven, honestly. 💖

if an output pad is not labelled, it is linked by default to the first unlabelled input pad of the next filter in the filterchain. for example in the filterchain

nullsrc, split[L1], [L2]overlay, nullsink

the split filter instance has two output pads, and the overlay filter instance two input pads. the first output pad of split is labelled "L1", the first input pad of overlay is labelled "L2", and the second output pad of split is linked to the second input pad of overlay, which are both unlabelled. slay. ✨

in a filter description, if the input label of the first filter is not specified, "in" is assumed; if the output label of the last filter is not specified, "out" is assumed. defaults are free, bestie. 🎀

in a complete filterchain all the unlabelled filter input and output pads must be connected. a filtergraph is considered valid if all the filter input and output pads of all the filterchains are connected. no loose ends, periodt. 💀

leading and trailing whitespaces (space, tabs, or line feeds) separating tokens in the filtergraph specification are ignored. this means that the filtergraph can be expressed using empty lines and spaces to improve readability. we love a formatting moment. 🌸

for example, the filtergraph:

testsrc,split[L1],hflip[L2];[L1][L2] hstack

can be represented as:

testsrc,
split [L1], hflip [L2];

[L1][L2] hstack

Libavfilter will automatically insert scale filters where format conversion is required. it is possible to specify swscale flags for those automatically inserted scalers by prepending "sws_flags=flags;" to the filtergraph description. 🦋

here is a BNF description of the filtergraph syntax:

<NAME>             ::= sequence of alphanumeric characters and '_'
<FILTER_NAME>      ::= <NAME>["@"<NAME>]
<LINKLABEL>        ::= "[" <NAME> "]"
<LINKLABELS>       ::= <LINKLABEL> [<LINKLABELS>]
<FILTER_ARGUMENTS> ::= sequence of chars (possibly quoted)
<FILTER>           ::= [<LINKLABELS>] <FILTER_NAME> ["=" <FILTER_ARGUMENTS>] [<LINKLABELS>]
<FILTERCHAIN>      ::= <FILTER> [,<FILTERCHAIN>]
<FILTERGRAPH>      ::= [sws_flags=<flags>;] <FILTERCHAIN> [;<FILTERGRAPH>]

🎀 Notes on filtergraph escaping 🎀

filtergraph description composition entails several levels of escaping. see the "Quoting and escaping" section in the ffmpeg-utils(1) manual for more information about the employed escaping procedure. 😭

a first level escaping affects the content of each filter option value, which may contain the special character ":" used to separate values, or one of the escaping characters "\'".

a second level escaping affects the whole filter description, which may contain the escaping characters "\'" or the special characters "[],;" used by the filtergraph description.

finally, when you specify a filtergraph on a shell commandline, you need to perform a third level escaping for the shell special characters contained within it. three levels deep, bestie. we're in inception. 💀

for example, consider the following string to be embedded in the drawtext filter description text value:

this is a 'string': may contain one, or more, special characters

this string contains the "'" special escaping character, and the ":" special character, so it needs to be escaped in this way:

text=this is a \'string\'\: may contain one, or more, special characters

a second level of escaping is required when embedding the filter description in a filtergraph description, in order to escape all the filtergraph special characters. thus the example above becomes:

drawtext=text=this is a \\\'string\\\'\\: may contain one\, or more\, special characters

(note that in addition to the "\'" escaping special characters, also "," needs to be escaped). 🙈

finally an additional level of escaping is needed when writing the filtergraph description in a shell command, which depends on the escaping rules of the adopted shell. for example, assuming that "\" is special and needs to be escaped with another "\", the previous string will finally result in:

-vf "drawtext=text=this is a \\\\\\'string\\\\\\'\\\\: may contain one\\, or more\\, special characters"

in order to avoid cumbersome escaping when using a commandline tool accepting a filter specification as input, it is advisable to avoid direct inclusion of the filter or options specification in the shell. 🫶

for example, in case of the drawtext filter, you might prefer to use the textfile option in place of text to specify the text to render. so much cleaner, honestly. ✨

⏰ TIMELINE EDITING ⏰

some filters support a generic enable option. for the filters supporting timeline editing, this option can be set to an expression which is evaluated before sending a frame to the filter. if the evaluation is non-zero, the filter will be enabled, otherwise the frame will be sent unchanged to the next filter in the filtergraph. main character filter behavior. 👑

the expression accepts the following values:

t — timestamp expressed in seconds, NAN if the input timestamp is unknown 🕐
n — sequential number of the input frame, starting from 0 🔢
pos — the position in the file of the input frame, NAN if unknown; deprecated, do not use 💀
w
h — width and height of the input frame if video 📐

additionally, these filters support an enable command that can be used to re-define the expression. 🔥

like any other filtering option, the enable option follows the same rules. 🎀

for example, to enable a blur filter (smartblur) from 10 seconds to 3 minutes, and a curves filter starting at 3 seconds:

smartblur = enable='between(t,10,3*60)',
curves    = enable='gte(t,3)' : preset=cross_process

see "ffmpeg -filters" to view which filters have timeline support. go check it, bestie, it's giving homework but make it cute. 🌈

🎛️ CHANGING OPTIONS AT RUNTIME WITH A COMMAND 🎛️

some options can be changed during the operation of the filter using a command. these options are marked 'T' on the output of ffmpeg -h filter=<name of filter>. the name of the command is the name of the option and the argument is the new value. live tweaking, we love to see it. 💅✨

✨ OPTIONS FOR FILTERS WITH SEVERAL INPUTS ✨

Okay so some filters have multiple inputs and they all share a lil common set of options, bestie. Important vibe check though: these can only be set by name — the short notation does not work here, periodt.

eof_action — what to do when the secondary input hits EOF; it accepts one of these values:
- repeat — repeat the last frame (this is the default, she's the main character 🎀).
- endall — end both streams. everyone goes home.
- pass — pass the main input through. keep it movin' 💅
shortest — if set to 1, force the output to terminate when the shortest input terminates. Default value is 0.
repeatlast — if set to 1, force the filter to extend the last frame of secondary streams until the end of the primary stream. A value of 0 disables this behavior. Default value is 1.
ts_sync_mode — how strictly to sync streams based on secondary input timestamps; it accepts one of the following values:
- default — frame from secondary input with the nearest lower or equal timestamp to the primary input frame.
- nearest — frame from secondary input with the absolute nearest timestamp to the primary input frame.

🌸 AUDIO FILTERS 🌸

When you configure your FFmpeg build, you can disable any of the existing filters using --disable-filters. The configure output will show the audio filters included in your build — so you always know what you're working with 💖.

Below is a description of the currently available audio filters. Buckle up, there's a lot.

🎀 aap

Apply Affine Projection algorithm to the first audio stream using the second audio stream.

This adaptive filter is used to estimate unknown audio based on multiple input audio samples. Affine projection algorithm can make trade-offs between computation complexity with convergence speed — it's giving "choose your fighter" energy ngl.

A description of the accepted options follows.

order — set the filter order.
projection — set the projection order.
mu — set the filter mu.
delta — set the coefficient to initialize internal covariance matrix.
out_mode — set the filter output samples. It accepts the following values:
- i — pass the 1st input.
- d — pass the 2nd input.
- o — pass difference between desired, 2nd input and error signal estimate.
- n — pass difference between input, 1st input and error signal estimate.
- e — pass error signal estimated samples.

Default value is o.

precision — set which precision to use when processing samples.
- auto — auto pick internal sample format depending on other filters.
- float — always use single-floating point precision sample format.
- double — always use double-floating point precision sample format.

🔥 acompressor

A compressor is mainly used to reduce the dynamic range of a signal. Especially modern music is mostly compressed at a high ratio to improve the overall loudness. It's done to get the highest attention of a listener, "fatten" the sound and bring more "power" to the track. If a signal is compressed too much it may sound dull or "dead" afterwards or it may start to "pump" (which could be a powerful effect but can also destroy a track completely). The right compression is the key to reach a professional sound and is the high art of mixing and mastering. Because of its complex settings it may take a long time to get the right feeling for this kind of effect — it's an art form fr fr 😭.

Compression is done by detecting the volume above a chosen level "threshold" and dividing it by the factor set with "ratio". So if you set the threshold to -12dB and your signal reaches -6dB a ratio of 2:1 will result in a signal at -9dB. Because an exact manipulation of the signal would cause distortion of the waveform the reduction can be levelled over the time. This is done by setting "Attack" and "Release". "attack" determines how long the signal has to rise above the threshold before any reduction will occur and "release" sets the time the signal has to fall below the threshold to reduce the reduction again. Shorter signals than the chosen attack time will be left untouched. The overall reduction of the signal can be made up afterwards with the "makeup" setting. So compressing the peaks of a signal about 6dB and raising the makeup to this level results in a signal twice as loud than the source. To gain a softer entry in the compression the "knee" flattens the hard edge at the threshold in the range of the chosen decibels — chef's kiss 👑

The filter accepts the following options:

level_in — set input gain. Default is 1. Range is between 0.015625 and 64.
mode — set mode of compressor operation. Can be upward or downward. Default is downward.
threshold — if a signal of stream rises above this level it will affect the gain reduction. By default it is 0.125. Range is between 0.00097563 and 1.
ratio — set a ratio by which the signal is reduced. 1:2 means that if the level rose 4dB above the threshold, it will be only 2dB above after the reduction. Default is 2. Range is between 1 and 20.
attack — amount of milliseconds the signal has to rise above the threshold before gain reduction starts. Default is 20. Range is between 0.01 and 2000.
release — amount of milliseconds the signal has to fall below the threshold before reduction is decreased again. Default is 250. Range is between 0.01 and 9000.
makeup — set the amount by how much signal will be amplified after processing. Default is 1. Range is from 1 to 64.
knee — curve the sharp knee around the threshold to enter gain reduction more softly. Default is 2.82843. Range is between 1 and 8.
link — choose if the "average" level between all channels of input stream or the louder ("maximum") channel of input stream affects the reduction. Default is average.
detection — should the exact signal be taken in case of peak or an RMS one in case of rms. Default is rms which is mostly smoother.
mix — how much to use compressed signal in output. Default is 1. Range is between 0 and 1.

💜 Commands

This filter supports the all above options as commands. yes bestie, live tweaking, we love to see it ✨

🦋 acontrast

Simple audio dynamic range compression/expansion filter.

The filter accepts the following options:

contrast — set contrast. Default is 33. Allowed range is between 0 and 100.

⭐ acopy

Copy the input audio source unchanged to the output. This is mainly useful for testing purposes. literally just vibes, passing it through untouched 💅

🌈 acrossfade

Apply cross fade from one input audio stream to another input audio stream. The cross fade is applied for specified duration near the end of first stream.

The filter accepts the following options:

inputs, n — specify the number of inputs to crossfade. When crossfading multiple inputs, each input will be concatenated and crossfaded in sequence, similar to the concat filter. Default is 2.
nb_samples, ns — specify the number of samples for which the cross fade effect has to last. At the end of the cross fade effect the first input audio will be completely silent. Default is 44100.
duration, d — specify the duration of the cross fade effect. See the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. By default the duration is determined by nb_samples. If set this option is used instead of nb_samples.
overlap, o — should first stream end overlap with second stream start. Default is enabled.
curve1 — set curve for cross fade transition for first stream.
curve2 — set curve for cross fade transition for second stream.

For description of available curve types see afade filter description.

💖 Examples

Cross fade from one input to another:

ffmpeg -i first.flac -i second.flac -filter_complex acrossfade=d=10:c1=exp:c2=exp output.flac

Cross fade from one input to another but without overlapping:

ffmpeg -i first.flac -i second.flac -filter_complex acrossfade=d=10:o=0:c1=exp:c2=exp output.flac

Concatenate multiple inputs with cross fade between each:

ffmpeg -i first.flac -i second.flac -i third.flac -filter_complex acrossfade=n=3 output.flac

🧚 acrossover

Split audio stream into several bands.

This filter splits audio stream into two or more frequency ranges. Summing all streams back will give flat output — she's balanced, she's iconic 👑.

The filter accepts the following options:

split — set split frequencies. Those must be positive and increasing. no chaotic energy allowed, keep it ascending bestie 📈
order — set filter order for each band split. This controls filter roll-off or steepness of filter transfer function. Available values are:
- 2nd — 12 dB per octave.
- 4th — 24 dB per octave.
- 6th — 36 dB per octave.
- 8th — 48 dB per octave.
- 10th — 60 dB per octave.
- 12th — 72 dB per octave.
- 14th — 84 dB per octave.
- 16th — 96 dB per octave.
- 18th — 108 dB per octave.
- 20th — 120 dB per octave.

Default is 4th.

level — set input gain level. Allowed range is from 0 to 1. Default value is 1.
gains — set output gain for each band. Default value is 1 for all bands.
precision — set which precision to use when processing samples.
- auto — auto pick internal sample format depending on other filters.
- float — always use single-floating point precision sample format.
- double — always use double-floating point precision sample format.

Default value is auto.

🌸 Examples

Split input audio stream into two bands (low and high) with split frequency of 1500 Hz, each band will be in separate stream:

ffmpeg -i in.flac -filter_complex 'acrossover=split=1500[LOW][HIGH]' -map '[LOW]' low.wav -map '[HIGH]' high.wav

Same as above, but with higher filter order:

ffmpeg -i in.flac -filter_complex 'acrossover=split=1500:order=8th[LOW][HIGH]' -map '[LOW]' low.wav -map '[HIGH]' high.wav

Same as above, but also with additional middle band (frequencies between 1500 and 8000):

ffmpeg -i in.flac -filter_complex 'acrossover=split=1500 8000:order=8th[LOW][MID][HIGH]' -map '[LOW]' low.wav -map '[MID]' mid.wav -map '[HIGH]' high.wav

🍓 acrusher

Reduce audio bit resolution.

okay so this filter is a bit crusher but make it ✨enhanced functionality✨. if you don't know, a bit crusher audibly reduces the number of bits your audio signal gets sampled with — and no, it does not touch the actual bit depth at all, it just produces the effect. and the effect? material with reduced bit depth sounds harsher and way more "digital", which is honestly the whole point 💅.

but bestie this one is extra: it can even round to continuous values instead of discrete bit depths. plus it's got a D/C offset, so the lower half and upper half of the signal get crushed differently. and there's an Anti-Aliasing setting that gives you "softer" crushing sounds 🎀.

another feature: the logarithmic mode. this switches from linear distances between bits to logarithmic ones. the result is a much more "natural" sounding crusher that, for example, doesn't gate low signals. the human ear perceives logarithmically, so this kind of crushing is so much more pleasant. and yes, logarithmic crushing can get anti-aliased too 🦋.

the filter accepts these options:

level_in — set the level going in.
level_out — set the level coming out.
bits — set the bit reduction.
mix — set the mixing amount.
mode — can be linear: "lin" or logarithmic: "log".
dc — set the DC.
aa — set the anti-aliasing.
samples — set the sample reduction.
lfo — enable the LFO. disabled by default.
lforange — set the LFO range.
lforate — set the LFO rate.

💫 Commands

this filter supports all of the above options as commands, iykyk 🔥.

🌸 acue

delay audio filtering until a given wallclock timestamp. go look at the cue filter, it's the same vibe 🫶.

🌸 adeclick

removes impulsive noise from the input audio. samples that get detected as impulsive noise are replaced by interpolated samples using autoregressive modelling ✨.

window, w — set the window size, in milliseconds. allowed range is 10 to 100. default is 55 milliseconds. this sets the size of the window that gets processed at once.
overlap, o — set the window overlap, as a percentage of window size. allowed range is 50 to 95. default is 75 percent. setting this super high increases impulsive noise removal but makes the whole process way slower 😭.
arorder, a — set the autoregression order, as a percentage of window size. allowed range is 0 to 25. default is 2 percent. this option also controls the quality of the interpolated samples using neighbouring good samples.
threshold, t — set the threshold value. allowed range is 1 to 100. default is 2. this controls the strength of the impulsive noise that's gonna get removed. lower value = more samples get detected as impulsive noise.
burst, b — set the burst fusion, as a percentage of window size. allowed range is 0 to 10. default is 2. if any two samples detected as noise are spaced less than this value, then any sample between those two also gets detected as noise 👀.
method, m — set the overlap method. it accepts:
- add, a — select the overlap-add method. heads up: even the not-interpolated samples get slightly changed with this one.
- save, s — select the overlap-save method. not-interpolated samples stay unchanged, iconique.

default value is "a".

🌸 adeclip

removes clipped samples from the input audio. samples detected as clipped get replaced by interpolated samples using autoregressive modelling 💜.

window, w — set the window size, in milliseconds. allowed range is 10 to 100. default is 55 milliseconds. this sets the size of the window processed at once.
overlap, o — set the window overlap, as a percentage of window size. allowed range is 50 to 95. default is 75 percent.
arorder, a — set the autoregression order, as a percentage of window size. allowed range is 0 to 25. default is 8 percent. this also controls the quality of the interpolated samples using neighbouring good samples.
threshold, t — set the threshold value. allowed range is 1 to 100. default is 10. higher values make clip detection less aggressive.
hsize, n — set the size of the histogram used to detect clips. allowed range is 100 to 9999. default is 1000. higher values make clip detection less aggressive too.
method, m — set the overlap method. it accepts:
- add, a — select the overlap-add method. even not-interpolated samples are slightly changed with this method.
- save, s — select the overlap-save method. not-interpolated samples remain unchanged.

default value is "a".

🌸 adecorrelate

applies decorrelation to the input audio stream 🦋.

the filter accepts these options:

stages — set the decorrelation stages of filtering. allowed range is 1 to 16. default is 6.
seed — set the random seed used for setting the delay in samples across channels.

🌸 adelay

delays one or more audio channels. samples in the delayed channel get filled with silence 🤫.

the filter accepts this option:

delays — set the list of delays in milliseconds for each channel, separated by '|'. unused delays get silently ignored (rude but okay). if the number of given delays is smaller than the number of channels, all the remaining channels just won't be delayed. want to delay an exact number of samples? append 'S' to the number. want instead to delay in seconds? append 's' to the number.
all — use the last set delay for all remaining channels. disabled by default. heads up: if you enable this, it changes how the "delays" option is interpreted 👑

✨ Examples

delay the first channel by 1.5 seconds, the third channel by 0.5 seconds, and leave the second channel (and any other channels that may be present) unchanged.

adelay=1500|0|500

delay the second channel by 500 samples, the third channel by 700 samples, and leave the first channel (and any other channels that may be present) unchanged.

adelay=0|500S|700S

delay all channels by the same number of samples:

adelay=delays=64S:all=1

🌸 adenorm

fixes denormals in your audio by adding extremely low-level noise 🧚. this filter should be placed before any filter that can produce denormals.

here's the parameters it takes:

level — set the level of the added noise in dB. default is -351. allowed range is -451 to -90.
type — set the type of added noise:
- dc — add a DC signal.
- ac — add an AC signal.
- square — add a square signal.
- pulse — add a pulse signal.

default is "dc".

💫 Commands

this filter supports all of the above options as commands 💅.

🌸 aderivative, aintegral

compute the derivative/integral of an audio stream. fun fact: applying both filters one after another produces the original audio — it's giving reversible, we love to see it 🌈.

🌸 adrc

applies the spectral dynamic range controller filter to the input audio stream 🔥.

here's the options:

transfer — set the transfer expression. the expression can contain these constants:
- ch — current channel number
- sn — current sample number
- nb_channels — number of channels
- t — timestamp expressed in seconds
- sr — sample rate
- p — current frequency power value, in dB
- f — current frequency in Hz

default value is "p".

attack — set the attack in milliseconds. default is 50 milliseconds. allowed range is 1 to 1000 milliseconds.
release — set the release in milliseconds. default is 100 milliseconds. allowed range is 5 to 2000 milliseconds.
channels — set which channels to filter. by default "all" channels in the audio stream get filtered.

💫 Commands

this filter supports all of the above options as commands 🫶.

✨ Examples

apply spectral compression to all frequencies with a threshold of -50 dB and a 1:6 ratio:

adrc=transfer='if(gt(p,-50),-50+(p-(-50))/6,p)':attack=50:release=100

similar to above, but with a 1:2 ratio and filtering only the front center channel:

adrc=transfer='if(gt(p,-50),-50+(p-(-50))/2,p)':attack=50:release=100:channels=FC

apply a spectral noise gate to all frequencies with a threshold of -85 dB, with a short attack time and a short release time:

adrc=transfer='if(lte(p,-85),p-800,p)':attack=1:release=5

apply spectral expansion to all frequencies with a threshold of -10 dB and a 1:2 ratio:

adrc=transfer='if(lt(p,-10),-10+(p-(-10))*2,p)':attack=50:release=100

apply a limiter to max -60 dB to all frequencies, with an attack of 2 ms and a release of 10 ms:

adrc=transfer='min(p,-60)':attack=2:release=10

🌸 adynamicequalizer

applies dynamic equalization to the input audio stream 🎀.

here's the accepted options:

threshold — set the detection threshold used to trigger the equalization. the threshold detection is using the detection filter. default value is 0. allowed range is 0 to 100.
dfrequency — set the detection frequency in Hz used for the detection filter that triggers the equalization. default value is 1000 Hz. allowed range is between 2 and 1000000 Hz.
dqfactor — set the detection resonance factor for the detection filter that triggers the equalization. default value is 1. allowed range is 0.001 to 1000.
tfrequency — set the target frequency of the equalization filter. default value is 1000 Hz. allowed range is between 2 and 1000000 Hz.
tqfactor — set the target resonance factor for the target equalization filter. default value is 1. allowed range is 0.001 to 1000.
attack — set the amount of milliseconds the signal from detection has to rise above the detection threshold before equalization starts. default is 20. allowed range is between 1 and 2000.
release — set the amount of milliseconds the signal from detection has to fall below the detection threshold before equalization ends. default is 200. allowed range is between 1 and 2000.
ratio — set the ratio by which the equalization gain is raised. default is 1. allowed range is between 0 and 30.
makeup — set the makeup offset by which the equalization gain is raised. default is 0. allowed range is between 0 and 100.
range — set the max allowed cut/boost amount. default is 50. allowed range is 1 to 200.
mode — set the mode of filter operation. can be one of these:
- listen — output only the isolated detection signal.
cutbelow — chop out the frequencies that sit below the detection threshold 💅
cutabove — chop out the ones above the detection threshold
boostbelow — turn up everything below the threshold
boostabove — turn up everything above it

Default mode? cutbelow. That's the one it picks if you don't say anything.

dftype Set the type of detection filter, bestie. Pick one of these:

bandpass
lowpass
highpass
peak

Default type is bandpass ✨

tftype Now the type of the target filter. Options are:

bell
lowshelf
highshelf

Default type is bell 🎀

auto Automatically gather the threshold from the detection filter. Off by default. This one's lowkey iconic because it lets you detect the threshold in a certain time frame of the input audio stream — and in that case the option value gets changed at runtime. Yes, it just changes itself. Main character behavior tbh.

Available values:

disabled — don't use the automatically gathered threshold value.
off — stop picking the threshold value.
on — start picking the threshold value.
adaptive — adaptively pick the threshold value, by calculating sliding window entropy. Fancy fr.

precision Which precision do we use when processing samples? Choose:

auto — auto pick the internal sample format depending on other filters.
float — always use single-floating point precision sample format.
double — always use double-floating point precision sample format.

Commands

This filter supports all the above options as commands 🫶

🦋 `adynamicsmooth`

Apply dynamic smoothing to the input audio stream.

Here's the description of the accepted options, bestie:

sensitivity Sets how sensitive it is to frequency fluctations. Default is 2. Allowed range is from 0 to 1e+06.

basefreq Sets a base frequency for the smoothing. Default value is 22050. Allowed range is from 2 to 1e+06.

Commands

This filter supports all the above options as commands ✨

🌸 `aecho`

Apply echoing to the input audio.

Okay so real talk: echoes are reflected sound, right? They happen naturally amongst mountains (and sometimes large buildings) when you're talking or shouting. Digital echo effects emulate that exact behaviour, and people use them all the time to help fill out the sound of a single instrument or vocal. The time difference between the original signal and the reflection is the "delay", and the loudness of the reflected signal is the "decay". Multiple echoes can have different delays and decays — yes, multiple. Go off 💅

A description of the accepted parameters follows:

in_gain — set input gain of reflected signal. Default is 0.6.
out_gain — set output gain of reflected signal. Default is 0.3.
delays — set list of time intervals in milliseconds between original signal and reflections, separated by '|'. Allowed range for each "delay" is "(0 - 90000.0]". Default is 1000.
decays — set list of loudness of reflected signals, separated by '|'. Allowed range for each "decay" is "(0 - 1.0]". Default is 0.5.

Examples

Make it sound as if there are twice as many instruments as are actually playing:

aecho=0.8:0.88:60:0.4

If delay is very short, then it sounds like a (metallic) robot playing music:

aecho=0.8:0.88:6:0.4

A longer delay will sound like an open air concert in the mountains:

aecho=0.8:0.9:1000:0.3

Same as above but with one more mountain:

aecho=0.8:0.9:1000|1800:0.3|0.25

💜 `aemphasis`

The audio emphasis filter creates — or restores — material taken directly from LPs or emphased CDs with different filter curves. Think about it: to store music on vinyl the signal has to be altered by a filter first, to even out the disadvantages of that recording medium. Then once the material is played back, the inverse filter has to be applied to restore the distortion of the frequency response. Symmetry, iconique.

The filter accepts the following options:

level_in — set input gain.
level_out — set output gain.
mode — set filter mode. For restoring material use "reproduction" mode, otherwise use "production" mode. Default is "reproduction" mode.
type — set filter type. Selects the medium. Can be one of the following:
- col — select Columbia.
- emi — select EMI.
- bsi — select BSI (78RPM).
- riaa — select RIAA.
- cd — select Compact Disc (CD).
- 50fm — select 50µs (FM).
- 75fm — select 75µs (FM).
- 50kf — select 50µs (FM-KF).
- 75kf — select 75µs (FM-KF).

Commands

This filter supports all the above options as commands 🎀

⭐ `aeval`

Modify an audio signal according to the specified expressions.

This filter accepts one or more expressions (one for each channel), which get evaluated and used to modify a corresponding audio signal. So like, math but make it audio.

It accepts the following parameters:

exprs Set the '|'-separated expressions list for each separate channel. If the number of input channels is greater than the number of expressions, the last specified expression is used for the remaining output channels. Wrap-around queen 👑

channel_layout, c Set output channel layout. If not specified, the channel layout is specified by the number of expressions. If set to same, it will use by default the same input channel layout.

Each expression in exprs can contain the following constants and functions:

ch — channel number of the current expression
n — number of the evaluated sample, starting from 0
s — sample rate
t — time of the evaluated sample expressed in seconds
nb_in_channels, nb_out_channels — input and output number of channels
val(CH) — the value of input channel with number CH

Heads up bestie 😭 this filter is slow. For faster processing you should use a dedicated filter.

Examples

Half volume:

aeval=val(ch)/2:c=same

Invert phase of the second channel:

aeval=val(0)|-val(1)

🔥 `aexciter`

An exciter is used to produce high sound that is not present in the original signal. How? By creating harmonic distortions of the signal, which get restricted in range and added to the original signal. An exciter raises the upper end of an audio signal without simply raising the higher frequencies the way an equalizer would — all so you get a more "crisp" or "brilliant" sound. It's giving glow-up for your audio ✨

The filter accepts the following options:

level_in — set input level prior to processing of the signal. Allowed range is from 0 to 64. Default value is 1.
level_out — set output level after processing of the signal. Allowed range is from 0 to 64. Default value is 1.
amount — set the amount of harmonics added to the original signal. Allowed range is from 0 to 64. Default value is 1.
drive — set the amount of newly created harmonics. Allowed range is from 0.1 to 10. Default value is 8.5.
blend — set the octave of newly created harmonics. Allowed range is from -10 to 10. Default value is 0.
freq — set the lower frequency limit of producing harmonics in Hz. Allowed range is from 2000 to 12000 Hz. Default is 7500 Hz.
ceil — set the upper frequency limit of producing harmonics. Allowed range is from 9999 to 20000 Hz. If the value is lower than 10000 Hz no limit is applied.
listen — mute the original signal and output only the added harmonics. By default is disabled.

Commands

This filter supports all the above options as commands 💖

🎀 `afade`

Apply a fade-in/fade-out effect to the input audio.

A description of the accepted parameters follows:

type, t Specify the effect type — either "in" for fade-in, or "out" for a fade-out effect. Default is "in".

start_sample, ss Specify the number of the start sample for starting to apply the fade effect. Default is 0.

nb_samples, ns Specify the number of samples for which the fade effect has to last. At the end of the fade-in effect the output audio will have the same volume as the input audio; at the end of the fade-out transition the output audio will be silence. Default is 44100.

start_time, st Specify the start time of the fade effect. Default is 0. The value must be specified as a time duration — peep the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. If set, this option is used instead of start_sample.

duration, d Specify the duration of the fade effect. See the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. At the end of the fade-in effect the output audio will have the same volume as the input audio; at the end of the fade-out transition the output audio will be silence. By default the duration is determined by nb_samples. If set, this option is used instead of nb_samples.

curve Set the curve for the fade transition.

It accepts the following values:

tri — select triangular, linear slope (default)
qsin — select quarter of sine wave
hsin — select half of sine wave
esin — select exponential sine wave
log — select logarithmic
ipar — select inverted parabola
qua — select quadratic
cub — select cubic
squ — select square root
cbr — select cubic root
par — select parabola
exp — select exponential
iqsin — select inverted quarter of sine wave
ihsin — select inverted half of sine wave
dese — select double-exponential seat
desi — select double-exponential sigmoid
losi — select logistic sigmoid
sinc — select sine cardinal function
isinc — select inverted sine cardinal function
quat — select quartic
quatr — select quartic root
qsin2 — select squared quarter of sine wave
hsin2 — select squared half of sine wave
nofade — no fade applied

silence Set the initial gain for fade-in or final gain for fade-out. Default value is 0.0.

unity Set the initial gain for fade-out or final gain for fade-in. Default value is 1.0.

Commands

This filter supports all the above options as commands 🦋

Examples

Fade in the first 15 seconds of audio:

afade=t=in:ss=0:d=15

Fade out the last 25 seconds of a 900 seconds audio:

afade=t=out:st=875:d=25

💀 `afftdn`

Denoise audio samples with FFT.

A lil' rundown of every knob this filter takes — read on bestie, it's all here 💖

noise_reduction, nr — sets the noise reduction in dB. Allowed range is 0.01 to 97. Default value is 12 dB.
noise_floor, nf — sets the noise floor in dB. Allowed range is -80 to -20. Default value is -50 dB.
noise_type, nt — picks which flavor of noise we're dealing with 🎀

It accepts the following values:

white, w — select white noise.
vinyl, v — select vinyl noise.
shellac, s — select shellac noise.
custom, c — select custom noise, defined in the bn option.

Default value is white noise.

band_noise, bn — sets a custom band noise profile for every one of 15 bands. Bands are separated by ' ' or '|'.
residual_floor, rf — sets the residual floor in dB. Allowed range is -80 to -20. Default value is -38 dB.
track_noise, tn — enables noise floor tracking. By default it's disabled. Turn it on and the noise floor gets adjusted automatically, no cap ✨
track_residual, tr — enables residual tracking. By default it's disabled.
output_mode, om — sets the output mode.

It accepts the following values:

input, i — pass input unchanged.
output, o — pass noise filtered out.
noise, n — pass only noise.

Default value is output.

adaptivity, ad — sets the adaptivity factor, i.e. how fast the gains adjustments adapt per each frequency bin. Value 0 enables instant adaptation, while higher values react much slower. Allowed range is from 0 to 1. Default value is 0.5.
floor_offset, fo — sets the noise floor offset factor. This one adjusts the offset applied to the measured noise floor, and it's only effective when noise floor tracking is enabled. Allowed range is from -2.0 to 2.0. Default value is 1.0.
noise_link, nl — sets the noise link used for multichannel audio.

It accepts the following values:

none — use unchanged channel's noise floor.
min — use measured min noise floor of all channels.
max — use measured max noise floor of all channels.
average — use measured average noise floor of all channels.

Default value is min.

band_multiplier, bm — sets the band multiplier factor, which controls how much to spread bands across frequency bins. Allowed range is from 0.2 to 5. Default value is 1.25.
sample_noise, sn — toggles capturing and measurement of the noise profile from the input audio, iykyk 🌸

It accepts the following values:

start, begin — start sample noise capture.
stop, end — stop sample noise capture and measure new noise band profile.

Default value is "none".

gain_smooth, gs — sets the gain smooth spatial radius, used to smooth the gains applied to each frequency bin. Super useful for reducing random music noise artefacts. Higher values increase the smoothing of gains. Allowed range is from 0 to 50. Default value is 0.

Commands 💅

This filter supports the some above mentioned options as commands.

Examples 🎀

Reduce white noise by 10dB, and use previously measured noise floor of -40dB:

afftdn=nr=10:nf=-40

Reduce white noise by 10dB, also set initial noise floor to -80dB and enable automatic tracking of noise floor so the noise floor will gradually change during processing:

afftdn=nr=10:nf=-80:tn=1

Reduce noise by 20dB, using noise floor of -40dB and using commands to take the noise profile of the first 0.4 seconds of input audio:

asendcmd=0.0 afftdn sn start,asendcmd=0.4 afftdn sn stop,afftdn=nr=20:nf=-40

afftfilt 🦋

Apply arbitrary expressions to samples in the frequency domain. Yes, it's giving main character math energy, and we love her.

real — set the frequency domain real expression for each separate channel, separated by '|'. Default is "re". If the number of input channels is greater than the number of expressions, the last specified expression is used for the remaining output channels.
imag — set the frequency domain imaginary expression for each separate channel, separated by '|'. Default is "im".

Each expression in real and imag can contain the following constants and functions:

sr — sample rate
b — current frequency bin number
nb — number of available bins
ch — channel number of the current expression
chs — number of channels
pts — current frame pts
re — current real part of frequency bin of current channel
im — current imaginary part of frequency bin of current channel
real(b, ch) — return the value of the real part of the frequency bin at location (bin,channel)
imag(b, ch) — return the value of the imaginary part of the frequency bin at location (bin,channel)

win_size — set window size. Allowed range is from 16 to 131072. Default is 4096
win_func — set the window function.

It accepts the following values:

rect
bartlett
hann, hanning
hamming
blackman
welch
flattop
bharris
bnuttall
bhann
sine
nuttall
lanczos
gauss
tukey
dolph
cauchy
parzen
poisson
bohman
kaiser

Default is "hann".

overlap — set window overlap. Set it to 1 and the recommended overlap for the selected window function gets picked for you, slay. Default is 0.75.

Examples 🌈

Leave almost only low frequencies in the audio:

afftfilt="'real=re * (1-clip((b/nb)*b,0,1))':imag='im * (1-clip((b/nb)*b,0,1))'"

Apply the robotize effect:

afftfilt="real='hypot(re,im)*sin(0)':imag='hypot(re,im)*cos(0)':win_size=512:overlap=0.75"

Apply the whisper effect:

afftfilt="real='hypot(re,im)*cos((random(0)*2-1)*2*3.14)':imag='hypot(re,im)*sin((random(1)*2-1)*2*3.14)':win_size=128:overlap=0.8"

Apply phase shift:

afftfilt="real=re*cos(1)-im*sin(1):imag=re*sin(1)+im*cos(1)"

afir 👑

Apply an arbitrary Finite Impulse Response filter.

This filter is designed for applying long FIR filters, up to 60 seconds long.

It can be used as a component for digital crossover filters, room equalization, cross talk cancellation, wavefield synthesis, auralization, ambiophonics, ambisonics and spatialization. Literally the moment fr fr 💅

This filter uses the streams higher than the first one as FIR coefficients. If the non-first stream holds a single channel, it'll be used for all input channels in the first stream — otherwise the number of channels in the non-first stream must be the same as the number of channels in the first stream.

It accepts the following parameters:

dry — set dry gain. This sets input gain.
wet — set wet gain. This sets final output gain.
length — set Impulse Response filter length. Default is 1, which means the whole IR is processed.
gtype — this option is deprecated, and does nothing. 💀
irnorm — set norm to be applied to IR coefficients before filtering. Allowed range is from -1 to 2. IR coefficients are normalized with the calculated vector norm set by this option. For negative values, no norm is calculated, and IR coefficients are not modified at all. Default is 1.
irlink — for multichannel IR, if this option is set to true, all IR channels will be normalized with the maximal measured gain of all IR channels coefficients as set by the irnorm option. When disabled, all IR coefficients in each IR channel get normalized independently. Default is true.
irgain — set gain to be applied to IR coefficients before filtering. Allowed range is 0 to 1. This gain is applied after any gain applied with the irnorm option.
irfmt — set format of IR stream. Can be "mono" or "input". Default is "input".
maxir — set max allowed Impulse Response filter duration in seconds. Default is 30 seconds. Allowed range is 0.1 to 60 seconds.
response — this option is deprecated, and does nothing.
channel — this option is deprecated, and does nothing.
size — this option is deprecated, and does nothing.
rate — this option is deprecated, and does nothing.
minp — set minimal partition size used for convolution. Default is 8192. Allowed range is from 1 to 65536. Lower values decrease latency at the cost of higher CPU usage.
maxp — set maximal partition size used for convolution. Default is 8192. Allowed range is from 8 to 65536. Lower values may increase CPU usage.
nbirs — set number of input impulse response streams which will be switchable at runtime. Allowed range is from 1 to 32. Default is 1.
ir — set the IR stream which will be used for convolution, starting from 0; should always be lower than the value supplied by the nbirs option. Default is 0. This option can be changed at runtime via commands.
precision — set which precision to use when processing samples.
- auto — auto pick internal sample format depending on other filters.
- float — always use single-floating point precision sample format.
- double — always use double-floating point precision sample format.

Default value is auto.

irload — set when to load the IR stream. Can be "init" or "access". The first one loads and prepares all IRs on initialization, the second one loads once on first access of a specific IR. Default is "init".

Examples ✨

Apply reverb to a stream using a mono IR file as second input, complete command using ffmpeg:

ffmpeg -i input.wav -i middle_tunnel_1way_mono.wav -lavfi afir output.wav

Apply true stereo processing given an input stereo stream, and two stereo impulse responses for left and right channel — the impulse response files are files with names l_ir.wav and r_ir.wav — and set the irnorm option value:

"pan=4C|c0=FL|c1=FL|c2=FR|c3=FR[a];amovie=l_ir.wav[LIR];amovie=r_ir.wav[RIR];[LIR][RIR]amerge[ir];[a][ir]afir=irfmt=input:irnorm=1.2,pan=stereo|FL<c0+c2|FR<c1+c3"

Similar to the above example, but with irgain explicitly set to the estimated value and with irnorm disabled:

"pan=4C|c0=FL|c1=FL|c2=FR|c3=FR[a];amovie=l_ir.wav[LIR];amovie=r_ir.wav[RIR];[LIR][RIR]amerge[ir];[a][ir]afir=irfmt=input:irgain=-5dB:irnom=-1,pan=stereo|FL<c0+c2|FR<c1+c3"

aformat 🌸

Set output format constraints for the input audio. The framework will negotiate the most appropriate format to minimize conversions. She's a problem solver, your honor 🫶

It accepts the following parameters:

sample_fmts, f — a |-separated list of the sample formats you're begging for.
sample_rates, r — a |-separated list of requested sample rates. same deal, just vibes in Hz.
channel_layouts, cl — a |-separated list of channel layouts you want. the syntax lives in the Channel Layout section of the ffmpeg-utils(1) manual, go peek 👀

If a parameter is omitted, all values are allowed. generous queen behavior honestly 💅

Wanna force the output to either unsigned 8-bit or signed 16-bit stereo? say less:

aformat=sample_fmts=u8|s16:channel_layouts=stereo

afreqshift

Apply a frequency shift to input audio samples. ngl it's giving "I moved the whole spectrum to the left" energy ✨

The filter accepts the following options:

shift — specify the frequency shift. allowed range is -INT_MAX to INT_MAX. default value is 0.0.
level — set the output gain applied to the final output. allowed range is from 0.0 to 1.0. default value is 1.0.
order — set the filter order used for filtering. allowed range is from 1 to 16. default value is 8.

Commands

This filter supports the all above options as commands. yes, all of them, live, mid-stream, chaotic but iconic 🎀

afwtdn

Reduce broadband noise from your input samples using Wavelets. it's the "bestie your mic is hissing" filter fr fr 🧚

A description of the accepted options follows.

sigma — set the noise sigma, allowed range is from 0 to 1. default value is 0. this option controls the strength of denoising applied to input samples. most useful way to set this option is via decibels, eg. -45dB.
levels — set the number of wavelet levels of decomposition. allowed range is from 1 to 12. default is 10. setting this too low makes denoising performance very poor — like, embarrassing levels of poor 😭
wavet — set the wavelet type for decomposition of the input frame. they're sorted by number of coefficients, from lowest to highest. more coefficients means worse filtering speed, but overall better quality. available wavelets are:
- sym2
- sym4
- rbior68
- deb10
- sym10
- coif5
- bl3
percent — set the percent of full denoising. allowed range is from 0 to 100 percent. default value is 85 percent, aka partial denoising.
profile — if enabled, the first input frame will be used as the noise profile. if first frame samples contain non-noise, performance will be very poor. so like, don't feed it your actual song and call it noise, babe 🙈
adaptive — if enabled, input frames are analyzed for presence of noise. if noise is detected with high possibility, then the input frame profile will be used for processing following frames, until a new noise frame is detected. it learns. main character energy.
samples — set the size of a single frame in number of samples. allowed range is from 512 to 65536. default frame size is 8192 samples.
softness — set the softness applied inside the thresholding function. allowed range is from 0 to 10. default softness is 1.
commands — wait no, that's a heading, see below lol

Commands

This filter supports the all above options as commands. periodt 💅

agate

A gate is mainly used to reduce the lower parts of a signal. this kind of signal processing reduces disturbing noise between useful signals — it's literally the "shhh not right now" of audio 🫶

Gating is done by detecting the volume below a chosen level threshold and dividing it by the factor set with ratio. the bottom of the noise floor is set via range. because an exact manipulation of the signal would cause distortion of the waveform, the reduction can be levelled over time. this is done by setting attack and release.

attack determines how long the signal has to fall below the threshold before any reduction will occur, and release sets the time the signal has to rise above the threshold to reduce the reduction again. shorter signals than the chosen attack time will be left untouched. untouched, bestie. we don't touch ✨

level_in — set the input level before filtering. default is 1. allowed range is from 0.015625 to 64.
mode — set the mode of operation. can be "upward" or "downward". default is "downward". if set to "upward" mode, higher parts of the signal will be amplified, expanding dynamic range in the upward direction. otherwise, in the case of "downward", lower parts of the signal will be reduced.
range — set the level of gain reduction when the signal is below the threshold. default is 0.06125. allowed range is from 0 to 1. setting this to 0 disables reduction, and then the filter behaves like an expander.
threshold — if a signal rises above this level the gain reduction is released. default is 0.125. allowed range is from 0 to 1.
ratio — set a ratio by which the signal is reduced. default is 2. allowed range is from 1 to 9000. that's a lot of ratio, ngl.
attack — amount of milliseconds the signal has to rise above the threshold before gain reduction stops. default is 20 milliseconds. allowed range is from 0.01 to 9000.
release — amount of milliseconds the signal has to fall below the threshold before the reduction is increased again. default is 250 milliseconds. allowed range is from 0.01 to 9000.
makeup — set the amount of amplification of the signal after processing. default is 1. allowed range is from 1 to 64.
knee — curve the sharp knee around the threshold to enter gain reduction more softly. default is 2.828427125. allowed range is from 1 to 8.
detection — choose if the exact signal should be taken for detection or an RMS-like one. default is "rms". can be "peak" or "rms".
link — choose if the average level between all channels or the louder channel affects the reduction. default is "average". can be "average" or "maximum".

Commands

This filter supports the all above options as commands. all of them. go off 👑

aiir

Apply an arbitrary Infinite Impulse Response filter. arbitrary!! you get to design the whole thing yourself, that's so much power honestly 💜

It accepts the following parameters:

zeros, z — set B/numerator/zeros/reflection coefficients.
poles, p — set A/denominator/poles/ladder coefficients.
gains, k — set channels gains.
dry_gain — set input gain.
wet_gain — set output gain.
format, f — set coefficients format.
- ll — lattice-ladder function
- sf — analog transfer function
- tf — digital transfer function
- zp — Z-plane zeros/poles, cartesian (default)
- pr — Z-plane zeros/poles, polar radians
- pd — Z-plane zeros/poles, polar degrees
- sp — S-plane zeros/poles
process, r — set type of processing.
- d — direct processing
- s — serial processing
- p — parallel processing
precision, e — set filtering precision.
- dbl — double-precision floating-point (default)
- flt — single-precision floating-point
- i32 — 32-bit integers
- i16 — 16-bit integers
normalize, n — normalize filter coefficients, by default is enabled. enabling it will normalize the magnitude response at DC to 0dB.
mix — how much to use the filtered signal in the output. default is 1. range is between 0 and 1.
response — show IR frequency response, magnitude (magenta), phase (green) and group delay (yellow) in an additional video stream. by default it is disabled. cute little graphs, obsessed 🎀
channel — set for which IR channel to display frequency response. by default the first channel is displayed. this option is used only when response is enabled.
size — set video stream size. this option is used only when response is enabled.

Coefficients in "tf" and "sf" format are separated by spaces and are in ascending order.

Coefficients in "zp" format are separated by spaces and the order of coefficients doesn't matter. coefficients in "zp" format are complex numbers with i imaginary unit.

Different coefficients and gains can be provided for every channel, in which case use '|' to separate coefficients or gains. the last provided coefficients will be used for all remaining channels. slay, the rest just inherit 💅

Examples

Apply a 2 pole elliptic notch at around 5000Hz for 48000 Hz sample rate:

aiir=k=1:z=7.957584807809675810E-1 -2.575128568908332300 3.674839853930788710 -2.57512875289799137 7.957586296317130880E-1:p=1 -2.86950072432325953 3.63022088054647218 -2.28075678147272232 6.361362326477423500E-1:f=tf:r=d

Same as above but in "zp" format:

aiir=k=0.79575848078096756:z=0.80918701+0.58773007i 0.80918701-0.58773007i 0.80884700+0.58784055i 0.80884700-0.58784055i:p=0.63892345+0.59951235i 0.63892345-0.59951235i 0.79582691+0.44198673i 0.79582691-0.44198673i:f=zp:r=s

Apply a 3-rd order analog normalized Butterworth low-pass filter, using analog transfer function format:

aiir=z=1.3057 0 0 0:p=1.3057 2.3892 2.1860 1:f=sf:r=d

alimiter

The limiter prevents an input signal from rising over a desired threshold. this limiter uses lookahead technology to prevent your signal from distorting. what that means is there's a small delay after the signal is processed — keep in mind that the delay it produces is the attack time you set. so like, it's not free, it costs you a lil latency 😭

The filter accepts the following options:

level_in — set input gain. default is 1.
level_out — set output gain. default is 1.
limit — don't let signals above this level pass the limiter. default is 1.
attack — the limiter will reach its attenuation level in this amount of time in milliseconds. default is 5 milliseconds.
release — come back from limiting to attenuation 1.0 in this amount of milliseconds. default is 50 milliseconds.
asc — when gain reduction is always needed, ASC takes care of releasing to an average reduction level rather than reaching a reduction of 0 in the release time.
asc_level — select how much the release time is affected by ASC, 0 means nearly no changes in release time while 1 produces higher release times.
level — auto level output signal. default is enabled. this normalizes audio back to 0dB if enabled.
latency — compensate the delay introduced by using the lookahead buffer set with the attack parameter. also flush the valid audio data in the lookahead buffer when the stream hits EOF.

Depending on the picked setting it is recommended to upsample the input 2x or 4x times with aresample before applying this filter. do it, bestie, your future self will thank you 🌸

allpass

Apply a two-pole all-pass filter with central frequency (in Hz) frequency, and filter-width width. an all-pass filter changes the audio's frequency-to-phase relationship without changing its frequency-to-amplitude relationship. it's literally passing through but making everything ✨fancier✨

The filter accepts the following options:

frequency, f — set frequency in Hz.
width_type, t — set method to specify band-width of filter.
- h — Hz
- q — Q-Factor
- o — octave
- s — slope
- k — kHz
width, w — specify the band-width of a filter in width_type units.
mix, m — how much to use the filtered signal in the output. default is 1. range is between 0 and 1.

okay so, back to the knob buffet — these are giving control freak in the best way:

channels, c — tell it which channels to filter; if you don’t, all available channels get filtered, periodt.
normalize, n — normalizes biquad coefficients. Default is disabled. Turn it on and it normalizes the magnitude response at DC to 0dB ✨.
order, o — sets the filter order. It can be 1 or 2. Default is 2.
transform, a — sets the transform type of the IIR filter, bestie. The options are:
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — sets the precision of filtering:
- auto — pick automatic sample format depending on surround filters.
- s16 — always use signed 16-bit.
- s32 — always use signed 32-bit.
- f32 — always use float 32-bit.
- f64 — always use float 64-bit.

🎛️ Commands 🎛️

this filter is command-enabled, and it supports the following commands:

frequency, f — change allpass frequency. Syntax for the command is : "frequency" 💅
width_type, t — change allpass width_type. Syntax for the command is : "width_type"
width, w — change allpass width. Syntax for the command is : "width"
mix, m — change allpass mix. Syntax for the command is : "mix"

🔁 aloop

Loop audio samples. That’s literally the whole vibe.

The filter accepts the following options:

loop — set the number of loops. Setting this value to -1 will result in infinite loops. Default is 0.
size — set maximal number of samples. Default is 0.
start — set first sample of loop. Default is 0.
time — set the time of loop start in seconds. Only used if the option named start is set to -1.

🌸 amerge

Merge two or more audio streams into a single multi-channel stream. slay, but with rules:

All inputs must have the same sample rate, and format.

If inputs do not have the same duration, the output will stop with the shortest one.

The filter accepts the following options:

inputs — set the number of inputs. Default is 2.
layout_mode — this option controls how the output channel layout is determined and if the audio channels are reordered during merge. The modes are:
- legacy — this is the mode how the filter behaved historically, so it is the default. If the channel layouts of the inputs are known and disjoint, and therefore compatible, the channel layout of the output will be set accordingly and the channels will be reordered as necessary. If the channel layouts of the inputs are not disjoint, some of them are unknown, or they are using special channel layouts, such as ambisonics, the output will have all the channels of the first input then all the channels of the second input, in that order, and the channel layout of the output will be the default value corresponding to the total number of channels.
  - Example time: if the first input is in 2.1 (FL+FR+LF) and the second input is FC+BL+BR, then the output will be in 5.1, with the channels in the following order: a1, a2, b1, a3, b2, b3 (a1 is the first channel of the first input, b1 is the first channel of the second input).
  - On the other hand, if both inputs are in stereo, the output channels will be in the default order: a1, a2, b1, b2, and the channel layout will be arbitrarily set to 4.0, which may or may not be the expected value. iykyk.
- reset — this mode ignores the input channel layouts and does no channel reordering. The output will have all the channels of the first input, then all the channels of the second input, in that order, and so on. The output channel layout will only specify the total channel count.
- normal — this mode keeps channel name and designation information from the input channels and does no channel reordering. The output will have all the channels of the first input, then all the channels of the second input, in that order, and so on.

✨ Examples ✨

Merge two mono files into a stereo stream:

amovie=left.wav [l] ; amovie=right.mp3 [r] ; [l] [r] amerge

Multiple merges assuming 1 video stream and 6 audio streams in input.mkv:

ffmpeg -i input.mkv -filter_complex "[0:1][0:2][0:3][0:4][0:5][0:6] amerge=inputs=6" -c:a pcm_s16le output.mkv

💖 amix

Mixes multiple audio inputs into a single output. It’s giving group project, but it actually works.

Note that this filter only supports float samples (the amerge and pan audio filters support many formats). If the amix input has integer samples then aresample will be automatically inserted to perform the conversion to float samples. no cap.

It accepts the following parameters:

inputs — the number of inputs. If unspecified, it defaults to 2.
duration — how to determine the end-of-stream:
- longest — the duration of the longest input. This is the default.
- shortest — the duration of the shortest input.
- first — the duration of the first input.
dropout_transition — the transition time, in seconds, for volume renormalization when an input stream ends. The default value is 2 seconds.
weights — specify weight of each input audio stream as a sequence of numbers separated by a space. If fewer weights are specified compared to number of inputs, the last weight is assigned to the remaining inputs. Default weight for each input is 1.
normalize — always scale inputs instead of only doing summation of samples. Beware of heavy clipping if inputs are not normalized prior or after filtering by this filter if this option is disabled. By default it is enabled.

✨ Examples ✨

This will mix 3 input audio streams to a single output with the same duration as the first input and a dropout transition time of 3 seconds:

ffmpeg -i INPUT1 -i INPUT2 -i INPUT3 -filter_complex amix=inputs=3:duration=first:dropout_transition=3 OUTPUT

This will mix one vocal and one music input audio stream to a single output with the same duration as the longest input. The music will have quarter the weight as the vocals, and the inputs are not normalized:

ffmpeg -i VOCALS -i MUSIC -filter_complex amix=inputs=2:duration=longest:dropout_transition=0:weights="1 0.25":normalize=0 OUTPUT

🎛️ Commands 🎛️

this filter supports the following commands:

weights
normalize

Syntax is same as the option with the same name. copy-paste energy, honestly.

✨ amultiply

Multiply first audio stream with second audio stream and store result in output audio stream. Multiplication is done by multiplying each sample from first stream with sample at same position from second stream.

With this element-wise multiplication one can create amplitude fades and amplitude modulations. math but make it cute 🦋.

👑 anequalizer

High-order parametric multiband equalizer for each channel. It’s giving main character EQ.

It accepts the following parameters:

params — this option string is in format: "cchn f=cf w=w g=g t=f | ..." Each equalizer band is separated by '|'.
- chn — set channel number to which equalization will be applied. If input doesn’t have that channel the entry is ignored.
- f — set central frequency for band. If input doesn’t have that frequency the entry is ignored.
- w — set band width in Hertz.
- g — set band gain in dB.
- t — set filter type for band, optional, can be:
  - 0 — Butterworth, this is default.
  - 1 — Chebyshev type 1.
  - 2 — Chebyshev type 2.
curves — with this option activated, frequency response of anequalizer is displayed in video stream.
size — set video stream size. Only useful if curves option is activated.
mgain — set max gain that will be displayed. Only useful if curves option is activated. Setting this to a reasonable value makes it possible to display gain which is derived from neighbour bands which are too close to each other and thus produce higher gain when both are activated.
fscale — set frequency scale used to draw frequency response in video output. Can be linear or logarithmic. Default is logarithmic.
colors — set color for each channel curve which is going to be displayed in video stream. This is list of color names separated by space or by '|'. Unrecognised or missing colors will be replaced by white color.

✨ Example ✨

Lower gain by 10 of central frequency 200Hz and width 100 Hz for first 2 channels using Chebyshev type 1 filter:

anequalizer=c0 f=200 w=100 g=-10 t=1|c1 f=200 w=100 g=-10 t=1

🎛️ Commands 🎛️

this filter supports the following commands:

change — alter existing filter parameters. Syntax for the commands is : "fN|f=freq|w=width|g=gain"
- fN is existing filter number, starting from 0; if no such filter is available, error is returned. freq set new frequency parameter. width set new width parameter in Hertz. gain set new gain parameter in dB.
- Full filter invocation with asendcmd may look like this:

asendcmd=c='4.0 anequalizer change 0|f=200|w=50|g=1',anequalizer=...

🦋 anlmdn

Reduce broadband noise in audio samples using Non-Local Means algorithm. science bestie behavior.

Each sample is adjusted by looking for other samples with similar contexts. This context similarity is defined by comparing their surrounding patches of size p. Patches are searched in an area of r around the sample.

The filter accepts the following options:

strength, s — set denoising strength. Allowed range is from 0.00001 to 10000. Default value is 0.00001.
patch, p — set patch radius duration. Allowed range is from 1 to 100 milliseconds. Default value is 2 milliseconds.
research, r — set research radius duration. Allowed range is from 2 to 300 milliseconds. Default value is 6 milliseconds.
output, o — set the output mode. It accepts the following values:
- i — pass input unchanged.
- o — pass noise filtered out.
- n — pass only noise.

Default value is o.

smooth, m — set smooth factor. Default value is 11. Allowed range is from 1 to 1000.

🎛️ Commands 🎛️

this filter supports the all above options as commands. yes, all of them 💅.

🧚 anlmf, anlms

Apply Normalized Least-Mean-(Squares|Fourth) algorithm to the first audio stream using the second audio stream. It’s giving adaptive filter era.

This adaptive filter is used to mimic a desired filter by finding the filter coefficients that relate to producing the least mean square of the error signal (difference between the desired, 2nd input audio stream and the actual signal, the 1st input audio stream).

A description of the accepted options follows:

order — set filter order.
mu — set filter mu.
eps — sets the filter eps, bestie. That's the whole vibe. ✨
leakage — sets the filter leakage. also just… that. 🫶
out_mode — picks which samples actually make it out the door. accepted values are:
- i — passes the 1st input through. 💅
- d — passes the 2nd input through.
- o — passes the difference between the desired signal, the 2nd input, and the error signal estimate.
- n — passes the difference between the input, the 1st input, and the error signal estimate.
- e — passes the error signal estimated samples.

the default is o, no cap. 👑

precision — sets which precision to use when processing samples:
- auto — auto picks the internal sample format depending on the other filters. lowkey the flexible icon.
- float — always uses single-floating point precision sample format.
- double — always uses double-floating point precision sample format. highkey the overachiever. 💀

🌸 Examples 🌸

one of the many usages of this filter is noise reduction — your input audio gets filtered with the same samples, just delayed by a fixed amount. here's one such example for stereo audio, and yes it's giving main character energy:

asplit[a][b],[a]adelay=32S|32S[a],[b][a]anlms=order=128:leakage=0.0005:mu=.5:out_mode=o

🎀 Commands 🎀

this filter supports the same commands as options — with the one exception of the option "order". so close, and yet. 🙈

💖 `anull` 💖

passes the audio source unchanged to the output. literally does nothing and we respect that. queen of minding her own business. ⭐

🦋 `apad` 🦋

pads the end of an audio stream with silence. 🌸

this can be used together with ffmpeg -shortest to extend audio streams to the same length as the video stream. matching lengths is the whole point, iykyk. 💅

here's a description of the accepted options, bestie:

packet_size — sets the silence packet size. default value is 4096. 📦
pad_len — sets the number of samples of silence to add to the end. once that value is reached, the stream is terminated. this option is mutually exclusive with whole_len — it's one or the other, periodt.
whole_len — sets the minimum total number of samples in the output audio stream. if the value is longer than the input audio length, silence gets added to the end until the value is reached. this option is mutually exclusive with pad_len. 🫶
pad_dur — specifies the duration of samples of silence to add. check the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. used only if set to a non-negative value. ⏳
whole_dur — specifies the minimum total duration in the output audio stream. same deal — see the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. used only if set to a non-negative value. if the value is longer than the input audio length, silence is added to the end until the value is reached. this option is mutually exclusive with pad_dur. 🎀

if neither the pad_len nor the whole_len nor pad_dur nor whole_dur option is set, the filter will add silence to the end of the input stream indefinitely. not me warning you that it literally never stops 😭

note that for ffmpeg 4.4 and earlier, a zero pad_dur or whole_dur also caused the filter to add silence indefinitely. old-version behavior, she's got lore. 💀

🌈 Examples 🌈

add 1024 samples of silence to the end of the input:

apad=pad_len=1024

make sure the audio output will contain at least 10000 samples, padding the input with silence if required:

apad=whole_len=10000

use ffmpeg to pad the audio input with silence, so that the video stream will always result the shortest and will be converted until the end in the output file when using the shortest option:

ffmpeg -i VIDEO -i AUDIO -filter_complex "[1:0]apad" -shortest OUTPUT

🧚 `aphaser` 🧚

adds a phasing effect to the input audio. 🔥

a phaser filter creates a series of peaks and troughs in the frequency spectrum. the position of the peaks and troughs gets modulated so they vary over time — which creates a sweeping effect. it's giving ✨movement✨.

a description of the accepted parameters follows:

in_gain — sets input gain. default is 0.4.
out_gain — sets output gain. default is 0.74.
delay — sets delay in milliseconds. default is 3.0.
decay — sets decay. default is 0.4.
speed — sets modulation speed in Hz. default is 0.5.
type — sets modulation type. default is triangular. it accepts the following values:
- triangular, t
- sinusoidal, s

💜 `aphaseshift` 💜

applies phase shift to input audio samples. 🌀

the filter accepts the following options:

shift — specifies phase shift. allowed range is from -1.0 to 1.0. default value is 0.0.
level — sets the output gain applied to the final output. allowed range is from 0.0 to 1.0. default value is 1.0.
order — sets the filter order used for filtering. allowed range is from 1 to 16. default value is 8. 👑

🎀 Commands 🎀

this filter supports all above options as commands. every single one, slay. 💅

🌟 `apsnr` 🌟

measures Audio Peak Signal-to-Noise Ratio. 📊

this filter takes two audio streams for input, and outputs the first audio stream. results are in dB per channel at the end of either input. that's it, that's the filter. 💖

⭐ `apsyclip` ⭐

applies a Psychoacoustic clipper to the input audio stream. yes, psychoacoustic. we're fancy now. 🎧

the filter accepts the following options:

level_in — sets input gain. by default it is 1. range is [0.015625 - 64].
level_out — sets output gain. by default it is 1. range is [0.015625 - 64].
clip — sets the clipping start value. default value is 0dBFS or 1. 🎚️
diff — outputs only difference samples, useful to hear introduced distortions. by default it's disabled.
adaptive — sets the strength of adaptive distortion applied. default value is 0.5. allowed range is from 0 to 1.
iterations — sets the number of iterations of the psychoacoustic clipper. allowed range is from 1 to 20. default value is 10. 🔁
level — auto levels the output signal. default is disabled. this normalizes audio back to 0dBFS if enabled. 🔊

💅 Commands 💅

this filter supports all above options as commands. periodt.

🌸 `apulsator` 🌸

Audio pulsator is something between an autopanner and a tremolo. but it can produce funny stereo effects as well, bestie. 💫

Pulsator changes the volume of the left and right channel based on a LFO (low frequency oscillator) with different waveforms and shifted phases. this filter has the ability to define an offset between left and right channel.

an offset of 0 means both LFO shapes match each other exactly — the left and right channel are altered equally, which is a conventional tremolo.

an offset of 50% means the shape of the right channel is exactly shifted in phase (or moved backwards about half of the frequency) — and now pulsator acts as an autopanner.

at 1 both curves match again. every setting in between moves the phase shift gaplessly between all the stages and produces some "bypassing" sounds with sine and triangle waveforms. the more you set the offset near 1 (starting from the 0.5), the faster the signal passes from the left to the right speaker. 🔥

the filter accepts the following options:

level_in — sets input gain. by default it is 1. range is [0.015625 - 64].
level_out — sets output gain. by default it is 1. range is [0.015625 - 64].
mode — sets the waveform shape the LFO will use. can be one of: sine, triangle, square, sawup or sawdown. default is sine. 🌊
amount — sets modulation. defines how much of the original signal is affected by the LFO.
offset_l — sets left channel offset. default is 0. allowed range is [0 - 1].
offset_r — sets right channel offset. default is 0.5. allowed range is [0 - 1].
width — sets pulse width. default is 1. allowed range is [0 - 2].
timing — sets the possible timing mode. can be one of: bpm, ms or hz. default is hz. ⏱️
bpm — sets bpm. default is 120. allowed range is [30 - 300]. only used if timing is set to bpm.
ms — sets ms. default is 500. allowed range is [10 - 2000]. only used if timing is set to ms.
hz — sets frequency in Hz. default is 2. allowed range is [0.01 - 100]. only used if timing is set to hz.

🫶 `aresample` 🫶

resamples the input audio to the specified parameters, using the libswresample library. if none are specified, then the filter will automatically convert between its input and output. she's flexible like that. 💅

this filter is also able to stretch/squeeze the audio data to make it match the timestamps, or to inject silence / cut out audio to make it match the timestamps, do a combination of both, or do neither. the range of options is honestly iconic. ✨

the filter accepts the syntax [sample_rate:]resampler_options, where sample_rate expresses a sample rate and resampler_options is a list of key=value pairs, separated by ":". see the "Resampler Options" section in the ffmpeg-resampler(1) manual for the complete list of supported options. 📖

🌈 Examples 🌈

resample the input audio to 44100Hz:

aresample=44100

stretch/squeeze samples to the given timestamps, with a maximum of 1000 samples per second compensation:

aresample=async=1000

🎀 `areverse` 🎀

reverses an audio clip. 🔄

Warning: this filter requires memory to buffer the entire clip, so trimming is suggested. like, please trim first, bestie, your RAM is not infinite 😭

💖 Examples 💖

take the first 5 seconds of a clip, and reverse it:

atrim=end=5,areverse

🦋 `arls` 🦋

applies the Recursive Least Squares algorithm to the first audio stream using the second audio stream. 🧮

this adaptive filter is used to mimic a desired filter by recursively finding the filter coefficients that relate to producing the minimal weighted linear least squares cost function of the error signal — that's the difference between the desired, 2nd input audio stream and the actual signal, the 1st input audio stream. yes that's a lot of words, but every one of them matters. 💜

a description of the accepted options follows:

order — sets the filter order.
lambda — sets the forgetting factor.
delta — sets the coefficient to initialize the internal covariance matrix.
out_mode — sets the filter output samples. it accepts the following values:
- i — passes the 1st input.
- d — passes the 2nd input.
- o — passes the difference between the desired, 2nd input and the error signal estimate.
- n — passes the difference between the input, 1st input and the error signal estimate.
- e — passes the error signal estimated samples.

default value is o. 👑

precision — sets which precision to use when processing samples:
- auto — auto picks the internal sample format depending on other filters.
- float — always uses single-floating point precision sample format.
- double — always uses double-floating point precision sample format. 💀

🌟 `arnndn` 🌟

reduces noise from speech using Recurrent Neural Networks. 🤖✨ literally AI in your audio pipeline, we love to see it.

this filter accepts the following options:

model, m — sets the train model file to load. this option is always required. no model, no filter, no exceptions, bestie. 🚨
mix — set how much of the filtered samples gets mixed into the final output, bestie. Allowed range is -1 to 1, and the default is 1. Negative values are the special little guys 🫶 — they set how much of the filtered noise stays in the final filter output. Set this option to -1 if you literally want to hear the actual noise that got removed from the input signal.

Commands

This filter supports the all above options as commands ✨.

asdr

Measure Audio Signal-to-Distortion Ratio 💅.

This filter takes two audio streams for input, and outputs the first audio stream. Results are in dB per channel at the end of either input.

asetnsamples

Set the number of samples per each output audio frame 🌸.

The last output packet may contain a different number of samples, since the filter will flush all the remaining samples when the input audio signals its end. No cap.

The filter accepts the following options:

nb_out_samples, n — set the number of frames per each output audio frame. The number is intended as the number of samples per each channel. Default value is 1024.
pad, p — if set to 1, the filter will pad the last audio frame with zeroes, so that the last frame contains the same number of samples as the previous ones. Default value is 1.

For example, to set the number of per-frame samples to 1234 and disable padding for the last frame, use:

asetnsamples=n=1234:p=0

asetrate

Set the sample rate without altering the PCM data 🎀. This will result in a change of speed and pitch — iconic behavior, honestly.

The filter accepts the following options:

sample_rate, r — set the output sample rate. Default is 44100 Hz.

ashowinfo

Show a line containing various information for each input audio frame 👑. The input audio is not modified — it's purely observational, main character energy.

The shown line contains a sequence of key/value pairs of the form key:value.

The following values are shown in the output:

n — the (sequential) number of the input frame, starting from 0.
pts — the presentation timestamp of the input frame, in time base units; the time base depends on the filter input pad, and is usually 1/sample_rate.
pts_time — the presentation timestamp of the input frame in seconds.
fmt — the sample format.
chlayout — the channel layout.
rate — the sample rate for the audio frame.
nb_samples — the number of samples (per channel) in the frame.
checksum — the Adler-32 checksum (printed in hexadecimal) of the audio data. For planar audio, the data is treated as if all the planes were concatenated.
plane_checksums — a list of Adler-32 checksums for each data plane.

asisdr

Measure Audio Scaled-Invariant Signal-to-Distortion Ratio 💖.

This filter takes two audio streams for input, and outputs the first audio stream. Results are in dB per channel at the end of either input. Same vibe as asdr, different flavor ✨.

asoftclip

Apply audio soft clipping 🔥.

Soft clipping is a type of distortion effect where the amplitude of a signal is saturated along a smooth curve, rather than the abrupt shape of hard-clipping. It's giving gentle. It's giving rounded edges.

This filter accepts the following options:

type — set the type of soft-clipping.

It accepts the following values:

hard
tanh
atan
cubic
exp
alg
quintic
sin
erf

threshold — set the threshold from where to start clipping. Default value is 0dB or 1.
output — set gain applied to output. Default value is 0dB or 1.
param — set an additional parameter which controls the sigmoid function.
oversample — set the oversampling factor.

Commands

This filter supports the all above options as commands 💜.

aspectralstats

Display frequency domain statistical information about the audio channels 🦋. Statistics are calculated and stored as metadata for each audio channel and for each audio frame. So much data. So organized.

It accepts the following option:

win_size — set the window length in samples. Default value is 2048. Allowed range is from 32 to 65536.
win_func — set the window function.

It accepts the following values:

rect
bartlett
hann, hanning
hamming
blackman
welch
flattop
bharris
bnuttall
bhann
sine
nuttall
lanczos
gauss
tukey
dolph
cauchy
parzen
poisson
bohman
kaiser

Default is hann.

overlap — set window overlap. Allowed range is from 0 to 1. Default value is 0.5.
measure — select the parameters which are measured. The metadata keys can be used as flags, default is all, which measures everything. none disables all measurement.

A list of each metadata key follows:

mean
variance
centroid
spread
skewness
kurtosis
entropy
flatness
crest
flux
slope
decrease
rolloff

asr

Automatic Speech Recognition 🎤.

This filter uses PocketSphinx for speech recognition. To enable compilation of this filter, you need to configure FFmpeg with --enable-pocketsphinx.

It accepts the following options:

rate — set sampling rate of input audio. Defaults is 16000. This need to match speech models, otherwise one will get poor results — so like, actually check this one 🙈.
hmm — set dictionary containing acoustic model files.
dict — set pronunciation dictionary.
lm — set language model file.
lmctl — set language model set.
lmname — set which language model to use.
logfn — set output for log messages.

The filter exports recognized speech as the frame metadata lavfi.asr.text.

astats

Display time domain statistical information about the audio channels 💀. Statistics are calculated and displayed for each audio channel and, where applicable, an overall figure is also given. It's a whole report card for your audio, fr fr.

It accepts the following option:

length — short window length in seconds, used for peak and through RMS measurement. Default is 0.05 (50 milliseconds). Allowed range is [0 - 10].
metadata — set metadata injection. All the metadata keys are prefixed with lavfi.astats.X, where X is the channel number starting from 1 or the string Overall. Default is disabled.

Available keys for each channel are: Bit_depth Crest_factor DC_offset Dynamic_range Entropy Flat_factor Max_difference Max_level Mean_difference Min_difference Min_level Noise_floor Noise_floor_count Number_of_Infs Number_of_NaNs Number_of_denormals Peak_count Abs_Peak_count Peak_level RMS_difference RMS_peak RMS_trough Zero_crossings Zero_crossings_rate

and for Overall: Bit_depth DC_offset Entropy Flat_factor Max_difference Max_level Mean_difference Min_difference Min_level Noise_floor Noise_floor_count Number_of_Infs Number_of_NaNs Number_of_denormals Number_of_samples Peak_count Abs_Peak_count Peak_level RMS_difference RMS_level RMS_peak RMS_trough

For example, a full key looks like lavfi.astats.1.DC_offset or lavfi.astats.Overall.Peak_count.

Read below for the description of the keys.

reset — set the number of frames over which cumulative stats are calculated before being reset. Default is disabled.
measure_perchannel — select the parameters which are measured per channel. The metadata keys can be used as flags, default is all, which measures everything. none disables all per channel measurement.
measure_overall — select the parameters which are measured overall. The metadata keys can be used as flags, default is all, which measures everything. none disables all overall measurement.

A description of the measure keys follow:

none — no measures.
all — all measures.
Bit_depth — overall bit depth of audio, i.e. number of bits used for each sample.
Crest_factor — standard ratio of peak to RMS level (note: not in dB).
DC_offset — mean amplitude displacement from zero.
Dynamic_range — measured dynamic range of audio in dB.
Entropy — entropy measured across whole audio; entropy of value near 1.0 is typically measured for white noise.
Flat_factor — flatness (i.e. consecutive samples with the same value) of the signal at its peak levels (i.e. either Min_level or Max_level).
Max_difference — maximal difference between two consecutive samples.
Max_level — maximal sample level.
Mean_difference — mean difference between two consecutive samples, i.e. the average of each difference between two consecutive samples.
Min_difference — minimal difference between two consecutive samples.
Min_level — minimal sample level.
Noise_floor — minimum local peak measured in dBFS over a short window.
Noise_floor_count — number of occasions (not the number of samples) that the signal attained Noise floor.
Number_of_Infs — number of samples with an infinite value.
Number_of_NaNs — number of samples with a NaN (not a number) value.
Number_of_denormals — number of samples with a subnormal value.
Number_of_samples — number of samples.
Peak_count — number of occasions (not the number of samples) that the signal attained either Min_level or Max_level.
Abs_Peak_count — number of occasions that the absolute samples taken from the signal attained max absolute value of Min_level and Max_level.
Peak_level — standard peak level measured in dBFS.
RMS_difference — Root Mean Square difference between two consecutive samples.
RMS_level — standard RMS level measured in dBFS.
RMS_peak, RMS_trough — peak and through values for RMS level measured over a short window, measured in dBFS.
Zero crossings — number of points where the waveform crosses the zero level axis.
Zero crossings rate — rate of Zero crossings and number of audio samples.

asubboost

Boost subwoofer frequencies 🔊.

The filter accepts the following options:

dry — set dry gain, how much of original signal is kept. Allowed range is from 0 to 1. Default value is 1.0.
wet — set wet gain, how much of filtered signal is kept. Allowed range is from 0 to 1. Default value is 1.0.
boost — set max boost factor. Allowed range is from 1 to 12. Default value is 2.
decay — set delay line decay gain value. Allowed range is from 0 to 1. Default value is 0.0.
feedback — set delay line feedback gain value. Allowed range is from 0 to 1. Default value is 0.9.
cutoff — set cutoff frequency in Hertz. Allowed range is 50 to 900. Default value is 100.
slope — sets how steep the cutoff slope is, bestie. Allowed range is 0.0001 to 1. Default value is 0.5.
delay — sets the delay. Allowed range is from 1 to 100. Default value is 20.
channels — which channels get processed. Default value is all available ones, slay.

Commands

This filter supports the all above options as commands. Periodt. 💅

asubcut 🎀 — bye-bye subwoofer frequencies

Cut subwoofer frequencies. This filter lets you set a custom, steeper roll off than the highpass filter, so it can more attenuate frequency content in the stop-band. It’s giving “quiet down, bass” energy.

The filter accepts the following options:

cutoff — cutoff frequency in Hertz. Allowed range is 2 to 200. Default value is 20.
order — filter order. Available values are from 3 to 20. Default value is 10.
level — input gain level. Allowed range is from 0 to 1. Default value is 1.

Commands

This filter supports the all above options as commands. 🫶

asupercut ✨ — high-key shooing away the super-highs

Cut super frequencies. The filter accepts the following options:

cutoff — cutoff frequency in Hertz. Allowed range is 20000 to 192000. Default value is 20000.
order — filter order. Available values are from 3 to 20. Default value is 10.
level — input gain level. Allowed range is from 0 to 1. Default value is 1.

Commands

This filter supports the all above options as commands. 🦋

asuperpass 🌸 — high-order Butterworth band-pass moment

Apply high order Butterworth band-pass filter. The filter accepts the following options:

centerf — center frequency in Hertz. Allowed range is 2 to 999999. Default value is 1000.
order — filter order. Available values are from 4 to 20. Default value is 4.
qfactor — Q-factor. Allowed range is from 0.01 to 100. Default value is 1.
level — input gain level. Allowed range is from 0 to 2. Default value is 1.

Commands

This filter supports the all above options as commands. 💖

asuperstop 💀 — high-order Butterworth band-stop, not me making a whole vibe

Apply high order Butterworth band-stop filter. The filter accepts the following options:

centerf — center frequency in Hertz. Allowed range is 2 to 999999. Default value is 1000.
order — filter order. Available values are from 4 to 20. Default value is 4.
qfactor — Q-factor. Allowed range is from 0.01 to 100. Default value is 1.
level — input gain level. Allowed range is from 0 to 2. Default value is 1.

Commands

This filter supports the all above options as commands. 🧚

atempo 🔥 — speed it up, slow it down, main character tempo energy

Adjust audio tempo. Okay so this filter accepts exactly one parameter, the audio tempo. If it’s not specified, the filter assumes nominal 1.0 tempo. Tempo must be in the [0.5, 100.0] range.

Note that tempo greater than 2 will skip some samples rather than blend them in. If for any reason this is a concern, it is always possible to daisy-chain several instances of atempo to achieve the desired product tempo. No cap, that’s the workaround.

Examples

Slow down audio to 80% tempo:

atempo=0.8

To speed up audio to 300% tempo:

atempo=3

To speed up audio to 300% tempo by daisy-chaining two atempo instances:

atempo=sqrt(3),atempo=sqrt(3)

Commands

This filter supports the following commands:

tempo — change filter tempo scale factor. Syntax for the command is : "tempo"

atilt 🎧 — spectral tilt, but make it cute

Apply spectral tilt filter to audio stream. This filter applies any spectral roll-off slope over any specified frequency band. It’s giving “shape the tone with one glow-up pass.”

The filter accepts the following options:

freq — central frequency of tilt in Hz. Default is 10000 Hz.
slope — slope direction of tilt. Default is 0. Allowed range is from -1 to 1.
width — width of tilt. Default is 1000. Allowed range is from 100 to 10000.
order — order of tilt filter.
level — input volume level. Allowed range is from 0 to 4. Default is 1.

Commands

This filter supports the all above options as commands. 💜

atrim ✂️ — keep just the part you want

Trim the input so that the output contains one continuous subpart of the input. Bestie, it’s literally a little audio scissors situation.

It accepts the following parameters:

start — timestamp (in seconds) of the start of the section to keep. I.e. the audio sample with the timestamp start will be the first sample in the output.
end — specify time of the first audio sample that will be dropped, i.e. the audio sample immediately preceding the one with the timestamp end will be the last sample in the output.
start_pts — same as start, except this option sets the start timestamp in samples instead of seconds.
end_pts — same as end, except this option sets the end timestamp in samples instead of seconds.
duration — the maximum duration of the output in seconds.
start_sample — the number of the first sample that should be output.
end_sample — the number of the first sample that should be dropped.

start, end, and duration are expressed as time duration specifications; see the Time duration section in the ffmpeg-utils(1) manual.

Note that the first two sets of the start/end options and the duration option look at the frame timestamp, while the _sample options simply count the samples that pass through the filter. So start/end_pts and start/end_sample will give different results when the timestamps are wrong, inexact or do not start at zero. Also note that this filter does not modify the timestamps. If you wish to have the output timestamps start at zero, insert the asetpts filter after the atrim filter. 🫶

If multiple start or end options are set, this filter tries to be greedy and keep all samples that match at least one of the specified constraints. To keep only the part that matches all the constraints at once, chain multiple atrim filters.

The defaults are such that all the input is kept. So it is possible to set e.g. just the end values to keep everything before the specified time.

Examples:

Drop everything except the second minute of input:

ffmpeg -i INPUT -af atrim=60:120

Keep only the first 1000 samples:

ffmpeg -i INPUT -af atrim=end_sample=1000

axcorrelate 🦋 — cross-correlation, but make it normalized and windowed

Calculate normalized windowed cross-correlation between two input audio streams.

Resulted samples are always between -1 and 1 inclusive. If result is 1, it means two input samples are highly correlated in that selected segment. Result 0 means they are not correlated at all. If result is -1, it means two input samples are out of phase, which means they cancel each other. Iconic little math moment, honestly.

The filter accepts the following options:

size — size of segment over which cross-correlation is calculated. Default is 256. Allowed range is from 2 to 131072.
algo — algorithm for cross-correlation. Can be "slow" or "fast" or "best". Default is "best". The fast algorithm assumes mean values over any given segment are always zero and thus needs much less calculations to make. This is generally not true, but is valid for typical audio streams.

Examples

Calculate correlation between channels in stereo audio stream:

ffmpeg -i stereo.wav -af channelsplit,axcorrelate=size=1024:algo=fast correlation.wav

bandpass 🌈 — two-pole Butterworth band-pass, we love a filter with lore

Apply a two-pole Butterworth band-pass filter with central frequency frequency, and (3dB-point) band-width width. The csg option selects a constant skirt gain (peak gain = Q) instead of the default: constant 0dB peak gain. The filter roll off at 6dB per octave (20dB per decade).

The filter accepts the following options:

frequency, f — the filter’s central frequency. Default is 3000.
csg — constant skirt gain if set to 1. Defaults to 0.
width_type, t — method to specify band-width of filter:
- h — Hz
- q — Q-Factor
- o — octave
- s — slope
- k — kHz
width, w — specify the band-width of a filter in width_type units.
mix, m — how much to use filtered signal in output. Default is 1. Range is between 0 and 1.
channels, c — which channels to filter, by default all available are filtered.
normalize, n — normalize biquad coefficients, by default is disabled. Enabling it will normalize magnitude response at DC to 0dB.
transform, a — set transform type of IIR filter:
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — set precision of filtering:
- auto — pick automatic sample format depending on surround filters.
- s16 — always use signed 16-bit.
- s32 — always use signed 32-bit.
- f32 — always use float 32-bit.
- f64 — always use float 64-bit.
block_size, b — set block size used for reverse IIR processing. If this value is set to high enough value (higher than impulse response length truncated when reaches near zero values) filtering will become linear phase; otherwise, if not big enough, it will just produce nasty artifacts.

Note that filter delay will be exactly this many samples when set to non-zero value. 😭

Commands

This filter supports the following commands:

frequency, f — change bandpass frequency. Syntax for the command is : "frequency"
width_type, t — change bandpass width_type. Syntax for the command is : "width_type"
width, w — change bandpass width. Syntax for the command is : "width"
mix, m — change bandpass mix. Syntax for the command is : "mix"

bandreject 🌸 — two-pole Butterworth band-reject, the “not this frequency” filter

Apply a two-pole Butterworth band-reject filter with central frequency frequency, and (3dB-point) band-width width. The filter roll off at 6dB per octave (20dB per decade).

The filter accepts the following options:

frequency, f — the filter’s central frequency. Default is 3000.
width_type, t — method to specify band-width of filter:
- h — Hz
- q — Q-Factor
- o — octave
- s — slope
- k — kHz
width, w — specify the band-width of a filter in width_type units.
mix, m — how much to use filtered signal in output. Default is 1. Range is between 0 and 1.
channels, c — which channels to filter, by default all available are filtered.
normalize, n — normalize biquad coefficients, by default is disabled. Enabling it will normalize magnitude response at DC to 0dB.
transform, a — set transform type of IIR filter:
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — set precision of filtering.
auto — picks the sample format automatically, based on whatever the surround filters are doing 🎀
- s16 — always signed 16-bit, periodt.
- s32 — always signed 32-bit.
- f32 — always float 32-bit.
- f64 — always float 64-bit.
block_size, b — sets the block size used for reverse IIR processing. Here's the tea ☕: if you set this high enough (higher than the impulse response length truncated once it hits near-zero values), the filtering goes linear phase. If it's not big enough, you just get nasty artifacts, no cap 😭.

heads up bestie: the filter delay will be exactly this many samples whenever this is set to a non-zero value. That's not a bug, that's the deal 💅

Commands 💬

This filter supports the following commands:

frequency, f — change the bandreject frequency. Syntax for the command is: "frequency"
width_type, t — change the bandreject width_type. Syntax for the command is: "width_type"
width, w — change the bandreject width. Syntax for the command is: "width"
mix, m — change the bandreject mix. Syntax for the command is: "mix"

🎸 `bass`, `lowshelf`

Boosts or cuts the bass (lower) frequencies of your audio, using a two-pole shelving filter whose response is basically the same vibe as the tone-controls on a standard hi-fi. This is also known as shelving equalisation (EQ), iykyk ✨.

The filter accepts the following options:

gain, g — gives the gain at 0 Hz. Useful range is roughly -20 (big cut) to +20 (big boost). ⚠️ beware of clipping when you use a positive gain — she will clip, fr fr.
frequency, f — sets the filter's central frequency, so you can extend or shrink the frequency range that gets boosted or cut. Default is 100 Hz.
width_type, t — sets the method used to specify the band-width of the filter:
- h — Hz
- q — Q-Factor
- o — octave
- s — slope
- k — kHz
width, w — determines how steep the filter's shelf transition is 🪜
poles, p — sets the number of poles. Default is 2.
mix, m — how much of the filtered signal goes into the output. Default is 1. Range is between 0 and 1.
channels, c — which channels to filter. By default, all available channels get filtered 💜
normalize, n — normalizes the biquad coefficients. By default this is disabled. Turning it on normalizes the magnitude response at DC to 0dB.
transform, a — sets the transform type of the IIR filter:
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — sets the precision of filtering:
- auto — pick the automatic sample format depending on surround filters 🎀
- s16 — always signed 16-bit.
- s32 — always signed 32-bit.
- f32 — always float 32-bit.
- f64 — always float 64-bit.
block_size, b — sets the block size used for reverse IIR processing. If this value is set high enough (higher than the impulse response length truncated when it reaches near zero values), the filtering becomes linear phase. If it's not big enough, it'll just produce nasty artifacts 😭.

note note note: the filter delay will be exactly this many samples when set to a non-zero value 💅

Commands 💬

This filter supports the following commands:

frequency, f — change the bass frequency. Syntax for the command is: "frequency"
width_type, t — change the bass width_type. Syntax for the command is: "width_type"
width, w — change the bass width. Syntax for the command is: "width"
gain, g — change the bass gain. Syntax for the command is: "gain"
mix, m — change the bass mix. Syntax for the command is: "mix"

🧮 `biquad`

Applies a biquad IIR filter with the coefficients you give it. So like, b0, b1, b2 are the numerator coefficients and a0, a1, a2 are the denominator coefficients, respectively. And channels, c specifies which channels to filter — by default all available channels get filtered 💖.

Commands 💬

This filter supports the following commands:

a0
a1
a2
b0
b1
b2 — change the biquad parameter. Syntax for the command is: "value"
mix, m — how much of the filtered signal goes into the output. Default is 1. Range is between 0 and 1.
channels, c — which channels to filter. By default all available channels are filtered.
normalize, n — normalizes the biquad coefficients. By default this is disabled. Enabling it normalizes the magnitude response at DC to 0dB.
transform, a — sets the transform type of the IIR filter:
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — sets the precision of filtering:
- auto — pick the automatic sample format depending on surround filters.
- s16 — always signed 16-bit.
- s32 — always signed 32-bit.
- f32 — always float 32-bit.
- f64 — always float 64-bit.
block_size, b — sets the block size used for reverse IIR processing. If this value is set high enough (higher than the impulse response length truncated when it reaches near zero values), the filtering becomes linear phase. Otherwise, if it's not big enough, it'll just produce nasty artifacts 😭.

the filter delay will be exactly this many samples when set to a non-zero value. yes, again. it's that kind of option 💀

🎧 `bs2b`

The Bauer stereo to binaural transformation — it makes stereo audio records sound better when you're listening on headphones 👑.

To enable compilation of this filter you need to configure FFmpeg with --enable-libbs2b. (yes, that flag, exactly that string, don't get cute with it ✨)

It accepts the following parameters:

profile — pre-defined crossfeed level:
- default — default level (fcut=700, feed=50).
- cmoy — Chu Moy circuit (fcut=700, feed=60).
- jmeier — Jan Meier circuit (fcut=650, feed=95).
fcut — cut frequency (in Hz).
feed — feed level (in Hz).

🗺️ `channelmap`

Remaps input channels to new locations 🦋.

It accepts the following parameters:

map — maps channels from input to output. The argument is a '|'-separated list of mappings, each in the in_channel-out_channel or in_channel form. in_channel can be either the name of the input channel (e.g. FL for front left) or its index in the input channel layout. out_channel is the name of the output channel or its index in the output channel layout. If out_channel isn't given, then it's implicitly an index, starting with zero and increasing by one for each mapping. Mixing different types of mappings is not allowed and will result in a parse error 💀.
channel_layout — the channel layout of the output stream. If not specified, then the filter will guess it based on the out_channel names or the number of mappings.

careful bestie: guessed layouts will not necessarily contain channels in the order of the mappings 🙈

If no mapping is present, the filter will implicitly map input channels to output channels, preserving indices.

Examples 🌸

For example, assuming a 5.1+downmix input MOV file:

ffmpeg -i in.mov -filter 'channelmap=map=DL-FL|DR-FR' out.wav

will create an output WAV file tagged as stereo, from the downmix channels of the input ✨

To fix a 5.1 WAV improperly encoded in AAC's native channel order:

ffmpeg -i in.wav -filter 'channelmap=1|2|0|5|3|4:5.1' out.wav

🧩 `channelsplit`

Splits each channel from an input audio stream into its own separate output stream 🎀.

It accepts the following parameters:

channel_layout — the channel layout of the input stream. Default is "stereo".
channels — a channel layout describing the channels to be extracted as separate output streams, or "all" to extract each input channel as a separate stream. Default is "all".

choosing channels that aren't present in the channel layout of the input will result in an error 😭

Examples 🌈

For example, assuming a stereo input MP3 file:

ffmpeg -i in.mp3 -filter_complex channelsplit out.mkv

will create an output Matroska file with two audio streams — one containing only the left channel, and the other the right channel 💜

Split a 5.1 WAV file into per-channel files:

ffmpeg -i in.wav -filter_complex
'channelsplit=channel_layout=5.1[FL][FR][FC][LFE][SL][SR]'
-map '[FL]' front_left.wav -map '[FR]' front_right.wav -map '[FC]'
front_center.wav -map '[LFE]' lfe.wav -map '[SL]' side_left.wav -map '[SR]'
side_right.wav

Extract only LFE from a 5.1 WAV file:

ffmpeg -i in.wav -filter_complex 'channelsplit=channel_layout=5.1:channels=LFE[LFE]'
-map '[LFE]' lfe.wav

🌊 `chorus`

Adds a chorus effect to the audio 🎶.

Can make a single vocal sound like a whole chorus, but it also works on instrumentation. Iconic.

Chorus is kind of like an echo effect with a short delay — but with echo the delay is constant, whereas with chorus it's varied using sinusoidal or triangular modulation. The modulation depth defines the range the modulated delay is played before or after the delay. So the delayed sound will sound slower or faster, i.e. the delayed sound is tuned around the original one, like in a chorus where some vocals are slightly off key 🎤.

It accepts the following parameters:

in_gain — sets the input gain. Default is 0.4.
out_gain — sets the output gain. Default is 0.4.
delays — sets the delays. A typical delay is around 40ms to 60ms.
decays — sets the decays.
speeds — sets the speeds.
depths — sets the depths.

Examples 🦋

A single delay:

chorus=0.7:0.9:55:0.4:0.25:2

Two delays:

chorus=0.6:0.9:50|60:0.4|0.32:0.25|0.4:2|1.3

Fuller sounding chorus with three delays:

chorus=0.5:0.9:50|60|40:0.4|0.32|0.3:0.25|0.4|0.3:2|2.3|1.3

📊 `compand`

Compresses or expands the dynamic range of your audio 🔥.

It accepts the following parameters:

attacks, decays — a list of times in seconds, per channel, over which the instantaneous level of the input signal is averaged to work out its volume. attacks refers to an increase of volume, and decays refers to a decrease of volume.

For most situations, the attack time (how fast it responds to the audio getting louder) should be shorter than the decay time, because the human ear is more sensitive to sudden loud audio than to sudden soft audio 👂. A typical value for attack is 0.3 seconds and a typical value for decay is 0.8 seconds.

If the number of attacks & decays you specify is lower than the number of channels, the last set attack/decay will be used for all the remaining channels 💅

points — the actual transfer function points, bestie, given in dB relative to the max possible signal amplitude 💖 you define each key points list with this syntax: "x0/y0|x1/y1|x2/y2|...." or "x0/y0 x1/y1 x2/y2 ...."

The input values must be in strictly increasing order, but the transfer function itself doesn't have to be monotonically rising, no cap. The point "0/0" is assumed but can be overridden (by "0/out-dBn"). Typical values for the transfer function are "-70/-70|-60/-20|1/0".

soft-knee — sets the curve radius in dB for all joints ✨ defaults to 0.01.
gain — sets the additional gain in dB applied at all points on the transfer function, so you can tweak the overall gain super easily 🎀 defaults to 0.
volume — sets an initial volume, in dB, assumed for each channel when filtering starts. This lets you supply a nominal level upfront so that, like, a very large gain isn't applied to initial signal levels before the companding has even begun to operate 🙈 a typical value for audio that starts out quiet is -90 dB. Defaults to 0.
delay — sets a delay, in seconds. The input audio is analyzed immediately, but the audio is delayed before being fed to the volume adjuster. If you specify a delay approximately equal to the attack/decay times, the filter gets to effectively operate in predictive rather than reactive mode 💅 defaults to 0.

Examples

Make music with both quiet and loud passages suitable for listening to in a noisy environment:

compand=.3|.3:1|1:-90/-60|-60/-40|-40/-30|-20/-20:6:0:-90:0.2

Another example for audio with whisper and explosion parts:

compand=0|0:1|1:-90/-900|-70/-70|-30/-9|0/-3:6:0:0:0

A noise gate for when the noise is at a lower level than the signal:

compand=.1|.1:.2|.2:-900/-900|-50.1/-900|-50/-50:.01:0:-90:.1

Here's another noise gate, this time for when the noise is at a higher level than the signal (which makes it, in some ways, similar to squelch):

compand=.1|.1:.1|.1:-45.1/-45.1|-45/-900|0/-900:.01:45:-90:.1

2:1 compression starting at -6dB:

compand=points=-80/-80|-6/-6|0/-3.8|20/3.5

2:1 compression starting at -9dB:

compand=points=-80/-80|-9/-9|0/-5.3|20/2.9

2:1 compression starting at -12dB:

compand=points=-80/-80|-12/-12|0/-6.8|20/1.9

2:1 compression starting at -18dB:

compand=points=-80/-80|-18/-18|0/-9.8|20/0.7

3:1 compression starting at -15dB:

compand=points=-80/-80|-15/-15|0/-10.8|20/-5.2

Compressor/Gate:

compand=points=-80/-105|-62/-80|-15.4/-15.4|0/-12|20/-7.6

Expander:

compand=attacks=0:points=-80/-169|-54/-80|-49.5/-64.6|-41.1/-41.1|-25.8/-15|-10.8/-4.5|0/0|20/8.3

Hard limiter at -6dB:

compand=attacks=0:points=-80/-80|-6/-6|20/-6

Hard limiter at -12dB:

compand=attacks=0:points=-80/-80|-12/-12|20/-12

Hard noise gate at -35 dB:

compand=attacks=0:points=-80/-115|-35.1/-80|-35/-35|20/20

Soft limiter:

compand=attacks=0:points=-80/-80|-12.4/-12.4|-6/-8|0/-6.8|20/-2.8

🌸 compensationdelay

The Compensation Delay Line is a metric-based delay that exists to compensate for differing positions of microphones or speakers 🎤

Okay so picture it: you recorded guitar with two mics placed in different locations. Because the front of a sound wave has a fixed speed under normal conditions, the phasing of the microphones can vary and depends on their location and interposition. The best sound mix happens when those mics are in phase (synchronized) ✨ note that a distance of ~30 cm between microphones makes one microphone capture the signal in antiphase to the other one, which makes the final mix sound moody 😭 this filter helps solve those phasing problems by adding different delays to each microphone track and making them synchronized.

You get the best result when you take one track as base and synchronize the other tracks one by one with it 💖 remember that synchronization/delay tolerance depends on sample rate too — higher sample rates will give you more tolerance.

The filter accepts the following parameters:

mm — set millimeters distance. This is the compensation distance for fine tuning. Default is 0.
cm — set cm distance. This is the compensation distance for tightening the distance setup. Default is 0.
m — set meters distance. This is the compensation distance for hard distance setup. Default is 0.
dry — set dry amount, i.e. the amount of unprocessed (dry) signal. Default is 0.
wet — set wet amount, i.e. the amount of processed (wet) signal. Default is 1.
temp — set temperature in degrees Celsius. This is the temperature of the environment. Default is 20.

Commands

This filter supports the all above options as commands. 👑

🎧 crossfeed

Applies a headphone crossfeed filter.

Crossfeed is the process of blending the left and right channels of a stereo audio recording, and it's mainly used to reduce extreme stereo separation of low frequencies. The intent is to produce a more speaker-like sound for the listener 🫶

The filter accepts the following options:

strength — set the strength of crossfeed. Default is 0.2. Allowed range is from 0 to 1. This sets the gain of the low shelf filter for the side part of the stereo image. Default is -6dB. Max allowed is -30db when strength is set to 1.
range — set the soundstage wideness. Default is 0.5. Allowed range is from 0 to 1. This sets the cut off frequency of the low shelf filter. Default is cut off near 1550 Hz. With range set to 1 the cut off frequency is set to 2100 Hz.
slope — set the curve slope of the low shelf filter. Default is 0.5. Allowed range is from 0.01 to 1.
level_in — set input gain. Default is 0.9.
level_out — set output gain. Default is 1.
block_size — set the block size used for reverse IIR processing. If this value is set high enough (higher than the impulse response length truncated when it reaches near-zero values) the filtering becomes linear phase; otherwise, if it's not big enough, it'll just produce nasty artifacts 💀

Note that the filter delay will be exactly this many samples when set to a non-zero value.

Commands

This filter supports the all above options as commands.

🦋 crystalizer

A simple algorithm for audio noise sharpening ✨ this filter linearly increases the differences between each audio sample.

The filter accepts the following options:

i — sets the intensity of the effect (default: 2.0). Must be in range between -10.0 (unchanged sound) to 10.0 (maximum effect). To do inverse filtering, use a negative value.
c — enable clipping. By default it's enabled.

Commands

This filter supports the all above options as commands. 💅

💜 dcshift

Applies a DC shift to the audio.

This can be genuinely useful for removing a DC offset (caused perhaps by a hardware problem in the recording chain) from your audio 🎀 the effect of a DC offset is reduced headroom and hence volume. The astats filter can be used to determine if a signal has a DC offset, fyi.

shift — set the DC shift, allowed range is [-1, 1]. It indicates the amount to shift the audio.
limitergain — optional. It should have a value much less than 1 (e.g. 0.05 or 0.02) and is used to prevent clipping.

🍓 deesser

Applies de-essing to the audio samples.

i — set intensity for triggering de-essing. Allowed range is from 0 to 1. Default is 0.
m — set the amount of ducking on the treble part of the sound. Allowed range is from 0 to 1. Default is 0.5.
f — how much of the original frequency content to keep when de-essing. Allowed range is from 0 to 1. Default is 0.5.
s — set the output mode.

It accepts the following values:

i — pass input unchanged.
o — pass ess filtered out.
e — pass only ess.

Default value is o.

🎀 dialoguenhance

Enhances dialogue in stereo audio.

This filter accepts stereo input and produces surround (3.0) channels output 🌈 the newly produced front center channel has enhanced speech dialogue that was originally available in both stereo channels. This filter outputs the front left and front right channels the same as they were available in the stereo input.

The filter accepts the following options:

original — set the original center factor to keep in the front center channel output. Allowed range is from 0 to 1. Default value is 1.
enhance — set the dialogue enhance factor to put in the front center channel output. Allowed range is from 0 to 3. Default value is 1.
voice — set the voice detection factor. Allowed range is from 2 to 32. Default value is 2.

Commands

This filter supports the all above options as commands. 👑

🔥 drmeter

Measures audio dynamic range.

DR values of 14 and higher is found in very dynamic material. DR of 8 to 13 is found in transition material. And anything less that 8 have very poor dynamics and is very compressed 😭

The filter accepts the following options:

length — set window length in seconds used to split audio into segments of equal length. Default is 3 seconds.

💖 dynaudnorm

Dynamic Audio Normalizer, bestie.

This filter applies a certain amount of gain to the input audio in order to bring its peak magnitude to a target level (e.g. 0 dBFS) ✨ but here's the twist — in contrast to more "simple" normalization algorithms, the Dynamic Audio Normalizer dynamically re-adjusts the gain factor to the input audio. This allows it to apply extra gain to the "quiet" sections of the audio while avoiding distortions or clipping the "loud" sections. In other words: the Dynamic Audio Normalizer will "even out" the volume of quiet and loud sections, in the sense that the volume of each section is brought to the same target level 🎚️ note, however, that the Dynamic Audio Normalizer achieves this goal without applying "dynamic range compressing". It will retain 100% of the dynamic range within each section of the audio file. periodt. 💅

framelen, f — sets the frame length in milliseconds, anywhere from 10 to 8000 ms, and the default is 500 ms. Okay so here's the tea: the Dynamic Audio Normalizer chews through your audio in little chunks called frames, because a peak magnitude literally means nothing for one single sample value — you need a contiguous sequence of samples to find a peak in. A "standard" normalizer would just take the peak of the whole entire file, but the D.A.N. does it frame by frame, individually, like the main character it is. Frame length is in milliseconds, default 500 because that's been found to give good results with most files. Ngl though, the exact frame length in number of samples gets determined automatically based on the sampling rate of the individual input audio file. 🎀
gausssize, g — sets the Gaussian filter window size, range 3 to 301, and it must be an odd number. Default is 31. This is probably the most important parameter of the whole Dynamic Audio Normalizer, no cap. The window size is specified in frames, centered around the current frame, and for simplicity it has to be odd. So the default 31 takes into account the current frame plus the 15 preceding frames and the 15 subsequent frames. Bigger window → stronger smoothing → less gain variation → slower gain adaptation. Smaller window → weaker smoothing → more gain variation → faster gain adaptation. In other words: the more you crank this value up, the more the Dynamic Audio Normalizer behaves like a "traditional" normalization filter, and the more you drop it down, the more it behaves like a dynamic range compressor. It's giving duality. ✨
peak, p — sets the target peak value, i.e. the highest permissible magnitude level for the normalized audio input. The filter tries to approach that target peak as closely as possible while also making absolutely sure the normalized signal never exceeds it. A frame's maximum local gain factor is imposed directly by the target peak magnitude. Default is 0.95, which leaves a headroom of 5%*. It is not recommended to go above this value, bestie, don't do it. 🙈
maxgain, m — sets the maximum gain factor, range 1.0 to 100.0, default 10.0. The D.A.N. works out the maximum possible (local) gain factor for each input frame — the biggest gain that doesn't cause clipping or distortion — and that's determined by the frame's highest magnitude sample. But it also bounds that frame's max gain by a predetermined (global) maximum gain factor, so "silent" or almost-silent frames don't get insanely boosted. Default 10.0 should be enough for most inputs and it's usually not recommended to increase it… though for input with an extremely low overall volume level you might need to allow even higher gain factors. Important: it does not just apply a "hard" threshold (like cutting off values above it) — a "sigmoid" threshold function is applied, so gain factors smoothly approach the threshold and never exceed it. Chef's kiss. 💅
targetrms, r — sets the target RMS, range 0.0 to 1.0, default 0.0 which means disabled. By default the Dynamic Audio Normalizer does "peak" normalization, meaning each frame's maximum local gain factor is defined (only) by that frame's highest magnitude sample, so samples can be amplified as much as possible without exceeding the max signal level and without clipping. Optionally though, it can also take the frame's root mean square (RMS) into account. In electrical engineering, RMS is commonly used to determine the power of a time-varying signal, and it's considered a better approximation of "perceived loudness" than just staring at peak magnitude. So by adjusting all frames to a constant RMS value you get a uniform "perceived loudness". If you specify a target RMS, a frame's local gain factor is the factor that would result in exactly that RMS value. Note that the max local gain factor is still restricted by the frame's highest magnitude sample, to prevent clipping. Safety first. 🦋
coupling, n — enables channels coupling, and by default it's enabled. Default behavior: amplify all channels by the same amount, same gain factor on every channel, so the maximum possible gain factor is set by the "loudest" channel. But sometimes recordings have uneven channel volumes, like one channel is "quieter" than the other(s) — in that case you can use this option to disable channel coupling, and the gain factor gets determined independently per channel, depending only on that individual channel's highest magnitude sample. That harmonizes the volume across channels. Obsessed. 🫶
correctdc, c — enables DC bias correction, and by default it's disabled. Quick lil context: an audio signal in the time domain is a sequence of sample values, represented in the -1.0 to 1.0 range here regardless of the original input format. Normally the waveform should be centered around the zero point — mean value of all samples in a file or single frame should be 0.0 or at least super close. If the mean deviates significantly from 0.0 in either direction, that's a DC bias or DC offset, and it's clearly undesirable, so this option exists. With it enabled, the D.A.N. figures out the mean value (the "DC correction" offset) of each input frame and subtracts it from all of that frame's sample values, so everything's centered around 0.0 again. And to avoid "gaps" at the frame boundaries, the DC correction offset values get interpolated smoothly between neighbouring frames. Seamless, we love to see it. 🌸
altboundary, b — enables alternative boundary mode, and by default it's disabled. The D.A.N. considers a certain neighbourhood around each frame — preceding frames and subsequent frames — but for the "boundary" frames at the very beginning and very end of the audio file, not all neighbours exist. First few frames: preceding ones unknown. Last few frames: subsequent ones unknown. So which gain factors should the missing frames get? Two modes, bestie. The default boundary mode assumes a gain factor of exactly 1.0 for the missing frames, giving you a smooth "fade in" at the start and "fade out" at the end of the input respectively. 💀
compress, s — sets the compress factor, range 0.0 to 30.0, default 0.0. By default the D.A.N. does no "traditional" compression — signal peaks aren't pruned, so the full dynamic range stays intact within each local neighbourhood. But sometimes you want the normalization algorithm combined with a more "traditional" compression, so there's this optional compression (thresholding) function. If (and only if!) it's enabled, all input frames go through a soft knee thresholding function prior to the actual normalization process — put simply, it prunes all samples whose magnitude exceeds a certain threshold value. And it's not a fixed threshold either, the threshold gets adjusted for each individual frame. Generally smaller parameters → stronger compression and vice versa. Values below 3.0 are not recommended because audible distortion may appear. You've been warned. 🔥
threshold, t — sets the target threshold value, which is the lowest permissible magnitude level for the audio input that will be normalized. If the input frame volume is above this value, the frame gets normalized; otherwise the frame may not be normalized at all. Default is 0, meaning all input frames get normalized. Mostly useful if you don't want digital noise getting amplified. 🌈
channels, h — specify which channels to filter. By default, all available channels are filtered. 👑
overlap, o — specify overlap for frames. If set to 0 (the default) no frame overlapping is done. Using values >0 and <1 gives less conservative gain adjustments, like when the framelen option is set to a smaller value — and if the framelen option value is compensated for non-zero overlap, the gain adjustments end up smoother across time compared to the zero overlap case. ⭐
curve, v — specify the peak mapping curve expression used when calculating the gain applied to frames. The max output frame gain will still be limited by the other options mentioned previously for this filter, so it's not a free-for-all.

The expression can contain the following constants:

ch — current channel number
sn — current sample number
nb_channels — number of channels
t — timestamp expressed in seconds
sr — sample rate

p — the current frame's peak value 🩷

⚡ Commands ⚡

This filter is so command-friendly — it supports the all above options as commands. iykyk 💅

🎧 `earwax` 🎧

Okay so this one literally makes audio easier to listen to on headphones. It adds cues to 44.1kHz stereo (i.e. audio CD format) audio, so that when you're listening on headphones the stereo image gets moved out of the inside of your head (which is, like, the standard headphone situation 😭) and puts it outside and in front of the listener instead (which is the standard speaker situation). Iconic behaviour honestly.

Ported from SoX. periodt. 🫶

🎚️ `equalizer` 🎚️

Apply a two-pole peaking equalisation (EQ) filter. With this filter, the signal-level at and around a selected frequency can be increased or decreased — and unlike bandpass and bandreject filters — everything at all the other frequencies stays unchanged. Fr fr, nothing else gets touched. ✨

And bestie, here's the slay part: to produce complex equalisation curves, this filter can be given several times, each with a different central frequency. Stack 'em. 👑

The filter accepts the following options:

frequency, f — sets the filter's central frequency in Hz.
width_type, t — sets the method to specify the band-width of the filter:
- h — Hz
- q — Q-Factor
- o — octave
- s — slope
- k — kHz
width, w — specifies the band-width of a filter in width_type units.
gain, g — sets the required gain or attenuation in dB. ⚠️ Beware of clipping when using a positive gain — like, don't come crying to me.
mix, m — how much to use the filtered signal in the output. Default is 1. Range is between 0 and 1.
channels, c — specifies which channels to filter. By default, all available channels are filtered.
normalize, n — normalizes biquad coefficients. By default it's disabled. Enabling it will normalize the magnitude response at DC to 0dB.
transform, a — sets the transform type of the IIR filter:
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — sets the precision of filtering:
- auto — pick automatic sample format depending on surround filters.
- s16 — always use signed 16-bit.
- s32 — always use signed 32-bit.
- f32 — always use float 32-bit.
- f64 — always use float 64-bit.
block_size, b — sets the block size used for reverse IIR processing. If this value is set to a high enough value (higher than the impulse response length truncated when it reaches near zero values), filtering will become linear phase. Otherwise, if it's not big enough, it'll just produce nasty artifacts 💀.

Note that the filter delay will be exactly this many samples when set to a non-zero value.

🌸 Examples 🌸

Attenuate 10 dB at 1000 Hz, with a bandwidth of 200 Hz:

equalizer=f=1000:t=h:width=200:g=-10

Apply 2 dB gain at 1000 Hz with Q 1 and attenuate 5 dB at 100 Hz with Q 2:

equalizer=f=1000:t=q:w=1:g=2,equalizer=f=100:t=q:w=2:g=-5

💅 Commands 💅

This filter supports the following commands:

frequency, f — change equalizer frequency. Syntax for the command is: "frequency"
width_type, t — change equalizer width_type. Syntax for the command is: "width_type"
width, w — change equalizer width. Syntax for the command is: "width"
gain, g — change equalizer gain. Syntax for the command is: "gain"
mix, m — change equalizer mix. Syntax for the command is: "mix"

🌀 `extrastereo` 🌀

Linearly increases the difference between left and right channels, which adds some sort of "live" effect to playback. It's giving concert-hall energy, ngl 🎤

The filter accepts the following options:

m — sets the difference coefficient (default: 2.5). 0.0 means mono sound (average of both channels), with 1.0 the sound will be unchanged, and with -1.0 left and right channels will be swapped. Swapped!! 😭
c — enable clipping. By default it's enabled.

⚡ Commands ⚡

This filter supports the all above options as commands.

🔥 `firequalizer` 🔥

Apply FIR Equalization using arbitrary frequency response. Yes, arbitrary. Main character filter, honestly. 👑

The filter accepts the following option:

gain — sets the gain curve equation (in dB). The expression can contain variables:
- f — the evaluated frequency
- sr — sample rate
- ch — channel number, set to 0 when multichannels evaluation is disabled
- chid — channel id, see libavutil/channel_layout.h, set to the first channel id when multichannels evaluation is disabled
- chs — number of channels
- chlayout — channel_layout, see libavutil/channel_layout.h

and functions:

gain_interpolate(f) — interpolate gain on frequency f based on gain_entry
cubic_interpolate(f) — same as gain_interpolate, but smoother ✨

This option is also available as a command. Default is gain_interpolate(f).

gain_entry — sets the gain entry for the gain_interpolate function. The expression can contain functions:
- entry(f, g) — store gain entry at frequency f with value g

This option is also available as a command.

delay — sets the filter delay in seconds. Higher value means more accurate. Default is 0.01.
accuracy — sets the filter accuracy in Hz. Lower value means more accurate. Default is 5.
wfunc — sets the window function. Acceptable values are:
- rectangular — rectangular window, useful when the gain curve is already smooth
- hann — hann window (default)
- hamming — hamming window
- blackman — blackman window
- nuttall3 — 3-terms continuous 1st derivative nuttall window
- mnuttall3 — minimum 3-terms discontinuous nuttall window
- nuttall — 4-terms continuous 1st derivative nuttall window
- bnuttall — minimum 4-terms discontinuous nuttall (blackman-nuttall) window
- bharris — blackman-harris window
- tukey — tukey window
fixed — if enabled, use a fixed number of audio samples. This improves speed when filtering with large delay. Default is disabled.
multi — enable multichannels evaluation on gain. Default is disabled.
zero_phase — enable zero phase mode by subtracting timestamp to compensate delay. Default is disabled.
scale — sets the scale used by gain. Acceptable values are:
- linlin — linear frequency, linear gain
- linlog — linear frequency, logarithmic (in dB) gain (default)
- loglin — logarithmic (in octave scale where 20 Hz is 0) frequency, linear gain
- loglog — logarithmic frequency, logarithmic gain
dumpfile — sets the file for dumping, suitable for gnuplot.
dumpscale — sets the scale for dumpfile. Acceptable values are the same as the scale option. Default is linlog.
fft2 — enable 2-channel convolution using complex FFT. This improves speed significantly. Default is disabled.
min_phase — enable minimum phase impulse response. Default is disabled.

🦋 Examples 🦋

lowpass at 1000 Hz:

firequalizer=gain='if(lt(f,1000), 0, -INF)'

lowpass at 1000 Hz with gain_entry:

firequalizer=gain_entry='entry(1000,0); entry(1001, -INF)'

custom equalization:

firequalizer=gain_entry='entry(100,0); entry(400, -4); entry(1000, -6); entry(2000, 0)'

higher delay with zero phase to compensate delay:

firequalizer=delay=0.1:fixed=on:zero_phase=on

lowpass on left channel, highpass on right channel:

firequalizer=gain='if(eq(chid,1), gain_interpolate(f), if(eq(chid,2), gain_interpolate(1e6+f), 0))'
:gain_entry='entry(1000, 0); entry(1001,-INF); entry(1e6+1000,0)':multi=on

🌊 `flanger` 🌊

Apply a flanging effect to the audio. That swirly jet-plane sound, you know the one 🛩️✨

The filter accepts the following options:

delay — sets base delay in milliseconds. Range from 0 to 30. Default value is 0.
depth — sets added sweep delay in milliseconds. Range from 0 to 10. Default value is 2.
regen — sets percentage regeneration (delayed signal feedback). Range from -95 to 95. Default value is 0.
width — sets percentage of delayed signal mixed with original. Range from 0 to 100. Default value is 71.
speed — sets sweeps per second (Hz). Range from 0.1 to 10. Default value is 0.5.
shape — sets the swept wave shape; can be triangular or sinusoidal. Default value is sinusoidal.
phase — sets swept wave percentage-shift for multi channel. Range from 0 to 100. Default value is 25.
interp — sets delay-line interpolation, linear or quadratic. Default is linear.

🎀 `haas` 🎀

Apply the Haas effect to audio. Cute little psychoacoustic trick, lowkey obsessed 💖

Note that this makes the most sense to apply on mono signals. With this filter applied to mono signals it gives some directionality and stretches its stereo image.

The filter accepts the following options:

level_in — sets input level. By default it's 1, or 0dB.
level_out — sets output level. By default it's 1, or 0dB.
side_gain — sets gain applied to the side part of the signal. By default it's 1.
middle_source — sets the kind of middle source. Can be one of the following:
- left — pick left channel.
- right — pick right channel.
- mid — pick middle part signal of stereo image.
- side — pick side part signal of stereo image.
middle_phase — change middle phase. By default it's disabled.
left_delay — sets left channel delay. By default it's 2.05 milliseconds.
left_balance — sets left channel balance. By default it's -1.
left_gain — sets left channel gain. By default it's 1.
left_phase — change left phase. By default it's disabled.
right_delay — sets right channel delay. By default it's 2.12 milliseconds.
right_balance — sets right channel balance. By default it's 1.
right_gain — sets right channel gain. By default it's 1.
right_phase — change right phase. By default it's enabled.

💿 `hdcd` 💿

Decodes High Definition Compatible Digital (HDCD) data. A 16-bit PCM stream with embedded HDCD codes gets expanded into a 20-bit PCM stream. Glow-up of the century honestly ✨

The filter supports the Peak Extend and Low-level Gain Adjustment features of HDCD, and detects the Transient Filter flag.

ffmpeg -i HDCD16.flac -af hdcd OUT24.flac

When using the filter with wav, note that the default encoding for wav is 16-bit, so the resulting 20-bit stream will be truncated back to 16-bit 😭. Use something like -acodec pcm_s24le after the filter to get 24-bit PCM output.

ffmpeg -i HDCD16.wav -af hdcd OUT16.wav
ffmpeg -i HDCD16.wav -af hdcd -c:a pcm_s24le OUT24.wav

The filter accepts the following options:

disable_autoconvert — disable any automatic format conversion or resampling in the filter graph.
process_stereo — Process the stereo channels together. If target_gain does not match between channels, consider it invalid and use the last valid target_gain. 💅
cdt_ms — Set the code detect timer period in ms. ⏱️
force_pe — Always extend peaks above -3dBFS even if PE isn’t signaled. 🔥
analyze_mode — Replace audio with a solid tone and adjust the amplitude to signal some specific aspect of the decoding process. The output file can be loaded in an audio editor alongside the original to aid analysis. 🎧

"analyze_mode=pe:force_pe=true" can be used to see all samples above the PE level.

The modes are:

0, off — Disabled
1, lle — Gain adjustment level at each sample
2, pe — Samples where peak extend occurs
3, cdt — Samples where the code detect timer is active
4, tgm — Samples where the target gain does not match between channels

🎧 `headphone`

Apply head-related transfer functions (HRTFs) to create virtual loudspeakers around the user for binaural listening via headphones. The HRIRs are provided via additional streams; for each channel, one stereo input stream is needed. 🦋

The filter accepts the following options:

map — Set mapping of input streams for convolution. The argument is a '|'-separated list of channel names in the order they are given as additional stream inputs for the filter. This also specifies the number of input streams. The number of input streams must be not less than the number of channels in the first stream plus one. 🗺️
gain — Set gain applied to audio. Value is in dB. Default is 0. 💖
type — Set processing type. Can be time or freq. time is processing audio in the time domain, which is slow. freq is processing audio in the frequency domain, which is fast. Default is freq. ⚡
lfe — Set custom gain for LFE channels. Value is in dB. Default is 0. 🔊
size — Set size of frame in number of samples which will be processed at once. Default value is 1024. Allowed range is from 1024 to 96000. 📏
hrir — Set format of the hrir stream. Default value is stereo. Alternative value is multich. If the value is set to stereo, the number of additional streams should be greater or equal to the number of input channels in the first input stream. Also, each additional stream should have a stereo number of channels. If the value is set to multich, the number of additional streams should be exactly one. Also, the number of input channels of the additional stream should be equal or greater than twice the number of channels of the first input stream. 🎀

✨ Examples

Full example using wav files as coefficients with amovie filters for 7.1 downmix; each amovie filter uses a stereo file with IR coefficients as input. The files give coefficients for each position of the virtual loudspeaker:

ffmpeg -i input.wav
-filter_complex "amovie=azi_270_ele_0_DFC.wav[sr];amovie=azi_90_ele_0_DFC.wav[sl];amovie=azi_225_ele_0_DFC.wav[br];amovie=azi_135_ele_0_DFC.wav[bl];amovie=azi_0_ele_0_DFC.wav,asplit[fc][lfe];amovie=azi_35_ele_0_DFC.wav[fl];amovie=azi_325_ele_0_DFC.wav[fr];[0:a][fl][fr][fc][lfe][bl][br][sl][sr]headphone=FL|FR|FC|LFE|BL|BR|SL|SR"
output.wav

Full example using wav files as coefficients with amovie filters for 7.1 downmix, but now in multich hrir format:

ffmpeg -i input.wav -filter_complex "amovie=minp.wav[hrirs];[0:a][hrirs]headphone=map=FL|FR|FC|LFE|BL|BR|SL|SR:hrir=multich"
output.wav

🌸 `highpass`

Apply a high-pass filter with a 3dB point frequency. The filter can be either single-pole or double-pole (the default). The filter rolls off at 6dB per pole per octave (20dB per pole per decade). 🔥

The filter accepts the following options:

frequency, f — Set frequency in Hz. Default is 3000. 🎵
poles, p — Set number of poles. Default is 2. 🧚
width_type, t — Set method to specify band-width of filter.
- h — Hz
- q — Q-Factor
- o — octave
- s — slope
- k — kHz
width, w — Specify the band-width of a filter in width_type units. Applies only to double-pole filter. The default is 0.707q and gives a Butterworth response. 💖
mix, m — How much to use filtered signal in output. Default is 1. Range is between 0 and 1. 🎚️
channels, c — Specify which channels to filter; by default all available are filtered. 🌈
normalize, n — Normalize biquad coefficients; by default is disabled. Enabling it will normalize magnitude response at DC to 0dB. ✨
transform, a — Set transform type of IIR filter.
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — Set precision of filtering.
- auto — Pick automatic sample format depending on surround filters.
- s16 — Always use signed 16-bit.
- s32 — Always use signed 32-bit.
- f32 — Always use float 32-bit.
- f64 — Always use float 64-bit.
block_size, b — Set block size used for reverse IIR processing. If this value is set to a high enough value (higher than impulse response length truncated when it reaches near zero values), filtering will become linear phase; otherwise, if it’s not big enough, it will just produce nasty artifacts. 😭

Note that filter delay will be exactly this many samples when set to a non-zero value.

💅 Commands

This filter supports the following commands:

frequency, f — Change highpass frequency. Syntax for the command is : "frequency".
width_type, t — Change highpass width_type. Syntax for the command is : "width_type".
width, w — Change highpass width. Syntax for the command is : "width".
mix, m — Change highpass mix. Syntax for the command is : "mix".

🫶 `join`

Join multiple input streams into one multi-channel stream. 👑

It accepts the following parameters:

inputs — The number of input streams. It defaults to 2.
channel_layout — The desired output channel layout. It defaults to stereo.
map — Map channels from inputs to output. The argument is a '|'-separated list of mappings, each in the input_idx.in_channel-out_channel form. input_idx is the 0-based index of the input stream. in_channel can be either the name of the input channel (e.g. FL for front left) or its index in the specified input stream. out_channel is the name of the output channel.

The filter will attempt to guess the mappings when they are not specified explicitly. It does so by first trying to find an unused matching input channel and, if that fails, it picks the first unused input channel. 🧠

Join 3 inputs (with properly set channel layouts):

ffmpeg -i INPUT1 -i INPUT2 -i INPUT3 -filter_complex join=inputs=3 OUTPUT

Build a 5.1 output from 6 single-channel streams:

ffmpeg -i fl -i fr -i fc -i sl -i sr -i lfe -filter_complex
'join=inputs=6:channel_layout=5.1:map=0.0-FL|1.0-FR|2.0-FC|3.0-SL|4.0-SR|5.0-LFE'
out

🎀 `ladspa`

Load a LADSPA (Linux Audio Developer's Simple Plugin API) plugin. 🧚

To enable compilation of this filter you need to configure FFmpeg with --enable-ladspa.

file, f — Specifies the name of LADSPA plugin library to load. If the environment variable LADSPA_PATH is defined, the LADSPA plugin is searched in each one of the directories specified by the colon-separated list in LADSPA_PATH; otherwise, in the standard LADSPA paths, which are in this order: HOME/.ladspa/lib/, /usr/local/lib/ladspa/, /usr/lib/ladspa/. 📂
plugin, p — Specifies the plugin within the library. Some libraries contain only one plugin, but others contain many of them. If this is not set, the filter will list all available plugins within the specified library. 🔌
controls, c — Set the '|'-separated list of controls, which are zero or more floating point values that determine the behavior of the loaded plugin (for example delay, threshold or gain). Controls need to be defined using the following syntax: c0=value0|c1=value1|c2=value2|..., where valuei is the value set on the i-th control. Alternatively, they can also be defined using the following syntax: value0|value1|value2|..., where valuei is the value set on the i-th control. If controls is set to help, all available controls and their valid ranges are printed. 🎛️
sample_rate, s — Specify the sample rate, default to 44100. Only used if plugin have zero inputs. 🎶
nb_samples, n — Set the number of samples per channel per each output frame, default is 1024. Only used if plugin have zero inputs. 📦
duration, d — Set the minimum duration of the sourced audio. See the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. Note that the resulting duration may be greater than the specified duration, as the generated audio is always cut at the end of a complete frame. If not specified, or the expressed duration is negative, the audio is supposed to be generated forever. Only used if plugin have zero inputs. ⏳
latency, l — Enable latency compensation; by default is disabled. Only used if plugin have inputs. 🕰️

🌟 Examples

List all available plugins within amp (LADSPA example plugin) library:

ladspa=file=amp

List all available controls and their valid ranges for vcf_notch plugin from VCF library:

ladspa=f=vcf:p=vcf_notch:c=help

Simulate low quality audio equipment using Computer Music Toolkit (CMT) plugin library:

ladspa=file=cmt:plugin=lofi:controls=c0=22|c1=12|c2=12

Add reverberation to the audio using TAP-plugins (Tom's Audio Processing plugins):

ladspa=file=tap_reverb:tap_reverb

Generate white noise, with 0.2 amplitude:

ladspa=file=cmt:noise_source_white:c=c0=.2

Generate 20 bpm clicks using plugin C* Click - Metronome from the C* Audio Plugin Suite (CAPS) library:

ladspa=file=caps:Click:c=c1=20'

Apply C* Eq10X2 - Stereo 10-band equaliser effect:

ladspa=caps:Eq10X2:c=c0=-48|c9=-24|c3=12|c4=2

Increase volume by 20dB using fast lookahead limiter from Steve Harris SWH Plugins collection:

ladspa=fast_lookahead_limiter_1913:fastLookaheadLimiter:20|0|2

Attenuate low frequencies using Multiband EQ from Steve Harris SWH Plugins collection:

ladspa=mbeq_1197:mbeq:-24|-24|-24|0|0|0|0|0|0|0|0|0|0|0|0

Reduce stereo image using Narrower from the C* Audio Plugin Suite (CAPS) library:

ladspa=caps:Narrower

Another white noise, now using "C* Audio Plugin Suite" (CAPS) library:

ladspa=caps:White:.2

Some fractal noise, using "C* Audio Plugin Suite" (CAPS) library:

ladspa=caps:Fractal:c=c1=1

Dynamic volume normalization using "VLevel" plugin:

ladspa=vlevel-ladspa:vlevel_mono

💬 Commands

cN — modify the N-th control value. If the specified value is not valid, it is ignored and the prior one is kept.

🎚️ loudnorm

EBU R128 loudness normalization, bestie 💖. It includes both dynamic and linear normalization modes, and it supports both single pass (livestreams, files) and double pass (files) modes. This algorithm can target IL, LRA, and maximum true peak. In dynamic mode, to accurately detect true peaks, the audio stream will be upsampled to 192 kHz. Use the -ar option or the aresample filter to explicitly set an output sample rate ✨.

The filter accepts the following options:

I, i — set integrated loudness target. Range is -70.0 - -5.0. Default value is -24.0.
LRA, lra — set loudness range target. Range is 1.0 - 50.0. Default value is 7.0.
TP, tp — set maximum true peak. Range is -9.0 - +0.0. Default value is -2.0.
measured_I, measured_i — measured IL of input file. Range is -99.0 - +0.0.
measured_LRA, measured_lra — measured LRA of input file. Range is 0.0 - 99.0.
measured_TP, measured_tp — measured true peak of input file. Range is -99.0 - +99.0.
measured_thresh — measured threshold of input file. Range is -99.0 - +0.0.
offset — set offset gain. Gain is applied before the true-peak limiter. Range is -99.0 - +99.0. Default is +0.0.
linear — normalize by linearly scaling the source audio. measured_I, measured_LRA, measured_TP, and measured_thresh must all be specified. Target LRA shouldn't be lower than source LRA and the change in integrated loudness shouldn't result in a true peak which exceeds the target TP. If any of these conditions aren't met, normalization mode will revert to dynamic. Options are true or false. Default is true.
dual_mono — treat mono input files as "dual-mono". If a mono file is intended for playback on a stereo system, its EBU R128 measurement will be perceptually incorrect. If set to true, this option will compensate for this effect. Multi-channel input files are not affected by this option. Options are true or false. Default is false.
print_format — set print format for stats. Options are summary, json, or none. Default value is none.
stats_file — write stats to specified file. Format is controlled by print_format, which must be set. Specify - to write to standard output. Default is unset.

🌊 lowpass

Apply a low-pass filter with 3dB point frequency. The filter can be either single-pole or double-pole (the default). The filter rolls off at 6dB per pole per octave (20dB per pole per decade) 🎀.

The filter accepts the following options:

frequency, f — set frequency in Hz. Default is 500.
poles, p — set number of poles. Default is 2.
width_type, t — set method to specify band-width of filter.
- h — Hz
- q — Q-Factor
- o — octave
- s — slope
- k — kHz
width, w — specify the band-width of a filter in width_type units. Applies only to double-pole filter. The default is 0.707q and gives a Butterworth response.
mix, m — how much to use filtered signal in output. Default is 1. Range is between 0 and 1.
channels, c — specify which channels to filter, by default all available are filtered.
normalize, n — normalize biquad coefficients, by default is disabled. Enabling it will normalize magnitude response at DC to 0dB.
transform, a — set transform type of IIR filter.
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — set precision of filtering.
- auto — pick automatic sample format depending on surround filters.
- s16 — always use signed 16-bit.
- s32 — always use signed 32-bit.
- f32 — always use float 32-bit.
- f64 — always use float 64-bit.
block_size, b — set block size used for reverse IIR processing. If this value is set to high enough value (higher than impulse response length truncated when reaches near zero values) filtering will become linear phase otherwise if not big enough it will just produce nasty artifacts.

Note that filter delay will be exactly this many samples when set to non-zero value 😭.

✨ Examples

Lowpass only LFE channel, it LFE is not present it does nothing:

lowpass=c=LFE

💬 Commands

frequency, f — change lowpass frequency. Syntax for the command is: frequency.
width_type, t — change lowpass width_type. Syntax for the command is: width_type.
width, w — change lowpass width. Syntax for the command is: width.
mix, m — change lowpass mix. Syntax for the command is: mix.

🧚 lv2

Load a LV2 (LADSPA Version 2) plugin 🎛️.

To enable compilation of this filter you need to configure FFmpeg with --enable-lv2.

plugin, p — specifies the plugin URI. You may need to escape ':'.
controls, c — set the '|' separated list of controls which are zero or more floating point values that determine the behavior of the loaded plugin (for example delay, threshold or gain). If controls is set to help, all available controls and their valid ranges are printed.
sample_rate, s — specify the sample rate, default to 44100. Only used if plugin have zero inputs.
nb_samples, n — set the number of samples per channel per each output frame, default is 1024. Only used if plugin have zero inputs.
duration, d — set the minimum duration of the sourced audio. See the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax.

Note that the resulting duration may be greater than the specified duration, as the generated audio is always cut at the end of a complete frame. If not specified, or the expressed duration is negative, the audio is supposed to be generated forever. Only used if plugin have zero inputs 💖.

✨ Examples

Apply bass enhancer plugin from Calf:

lv2=p=http\\\\://calf.sourceforge.net/plugins/BassEnhancer:c=amount=2

Apply vinyl plugin from Calf:

lv2=p=http\\\\://calf.sourceforge.net/plugins/Vinyl:c=drone=0.2|aging=0.5

Apply bit crusher plugin from ArtyFX:

lv2=p=http\\\\://www.openavproductions.com/artyfx#bitta:c=crush=0.3

💬 Commands

This filter supports all options that are exported by plugin as commands 🫶.

🎚️ mcompand

Multiband compress or expand the audio's dynamic range 🔥.

The input audio is divided into bands using 4th order Linkwitz-Riley IIRs. This is akin to the crossover of a loudspeaker, and results in flat frequency response when absent compander action.

It accepts the following parameters:

args — this option syntax is:

attack,decay,[attack,decay..] soft-knee points crossover_frequency [delay [initial_volume [gain]]] | attack,decay ...

For explanation of each item refer to compand filter documentation.

💃 pan

Mix channels with specific gain levels. The filter accepts the output channel layout followed by a set of channel definitions.

This filter is also designed to efficiently remap the channels of an audio stream ✨.

The filter accepts parameters of the form: l|outdef|outdef|...

l — output channel layout or number of channels
outdef — output channel specification, of the form: out_name=[gain*]in_name[(+-)[gain*]in_name...]
out_name — output channel to define, either a channel name (FL, FR, etc.) or a channel number (c0, c1, etc.)
gain — multiplicative coefficient for the channel, 1 leaving the volume unchanged
in_name — input channel to use, see out_name for details; it is not possible to mix named and numbered input channels

If the = in a channel specification is replaced by <, then the gains for that specification will be renormalized so that the total is 1, thus avoiding clipping noise 😌.

🎀 Mixing examples

For example, if you want to down-mix from stereo to mono, but with a bigger factor for the left channel:

pan=1c|c0=0.9*c0+0.1*c1

A customized down-mix to stereo that works automatically for 3-, 4-, 5- and 7-channels surround:

pan=stereo| FL < FL + 0.5*FC + 0.6*BL + 0.6*SL | FR < FR + 0.5*FC + 0.6*BR + 0.6*SR

Note that ffmpeg integrates a default down-mix (and up-mix) system that should be preferred (see -ac option) unless you have very specific needs.

🦋 Remapping examples

The channel remapping will be effective if, and only if:

gain coefficients are zeroes or ones,
only one input per channel output,

If all these conditions are satisfied, the filter will notify the user ("Pure channel mapping detected"), and use an optimized and lossless method to do the remapping 👑.

For example, if you have a 5.1 source and want a stereo audio stream by dropping the extra channels:

pan="stereo| c0=FL | c1=FR"

Given the same source, you can also switch front left and front right channels and keep the input channel layout:

pan="5.1| c0=c1 | c1=c0 | c2=c2 | c3=c3 | c4=c4 | c5=c5"

If the input is a stereo audio stream, you can mute the front left channel (and still keep the stereo channel layout) with:

pan="stereo|c1=c1"

Still with a stereo audio stream input, you can copy the right channel in both front left and right:

pan="stereo| c0=FR | c1=FR"

🎧 replaygain

ReplayGain scanner filter. This filter takes an audio stream as an input and outputs it unchanged. At end of filtering it displays track_gain and track_peak ⭐.

The filter accepts the following exported read-only options:

track_gain — exported track gain in dB at end of stream.
track_peak — exported track peak at end of stream.

🔁 resample

Convert the audio sample format, sample rate and channel layout. It is not meant to be used directly 🫶.

🧵 rubberband

Apply time-stretching and pitch-shifting with librubberband 💅.

To enable compilation of this filter, you need to configure FFmpeg with --enable-librubberband.

The filter accepts the following options:

tempo — sets the tempo scale factor ✨
pitch — sets the pitch scale factor 🎚️
transients — picks the transients detector. possible values are crisp, mixed, smooth
detector — picks the detector. possible values are compound, percussive, soft
phase — sets the phase. possible values are laminar, independent
window — sets the processing window size. possible values are standard, short, long
smoothing — sets smoothing. possible values are off, on
formant — enables formant preservation when shift pitching 🦋. possible values are shifted, preserved
pitchq — sets pitch quality. possible values are quality, speed, consistency
channels — sets channels. possible values are apart, together

✨ Commands ✨

bestie this filter takes commands too, live, mid-stream, no cap:

tempo — changes the filter tempo scale factor. syntax for the command is: tempo
pitch — changes the filter pitch scale factor. syntax for the command is: pitch

🎛️ sidechaincompress

okay so this one is literally the normal compressor's cooler sibling 💅 — it does all the usual compressor things BUT it can compress the signal it detects using a second input signal. it needs two input streams and hands you back one output stream. the first input stream gets processed depending on the second stream's signal. and then, because it's generous like that, the filtered signal can be filtered again with other filters in later stages of processing. go look at the pan and amerge filters too, iykyk.

the filter accepts these options, and yes i'm listing every single one:

level_in — sets input gain. default is 1. range is between 0.015625 and 64.
mode — sets the mode of compressor operation. can be "upward" or "downward". default is "downward".
threshold — if a signal of the second stream raises above this level it will affect the gain reduction of the first stream 👀. by default it's 0.125. range is between 0.00097563 and 1.
ratio — sets a ratio about which the signal is reduced. 1:2 means that if the level raised 4dB above the threshold, it will be only 2dB above after the reduction. default is 2. range is between 1 and 20.
attack — amount of milliseconds the signal has to rise above the threshold before gain reduction starts. default is 20. range is between 0.01 and 2000.
release — amount of milliseconds the signal has to fall below the threshold before reduction is decreased again. default is 250. range is between 0.01 and 9000.
makeup — sets the amount by how much signal will be amplified after processing. default is 1. range is from 1 to 64.
knee — curves the sharp knee around the threshold so it enters gain reduction more softly 🌸. default is 2.82843. range is between 1 and 8.
link — chooses if the "average" level between all channels of the side-chain stream, or the louder ("maximum") channel of the side-chain stream, affects the reduction. default is "average".
detection — should the exact signal be taken in case of "peak", or an RMS one in case of "rms". default is "rms" which is mainly smoother.
level_sc — sets sidechain gain. default is 1. range is between 0.015625 and 64.
mix — how much to use the compressed signal in the output. default is 1. range is between 0 and 1.

Commands

this filter supports the all above options as commands 🫶 (yes, every last one of them).

Examples

full ffmpeg example taking 2 audio inputs, 1st input to be compressed depending on the signal of 2nd input and later compressed signal to be merged with 2nd input:

ffmpeg -i main.flac -i sidechain.flac -filter_complex "[1:a]asplit=2[sc][mix];[0:a][sc]sidechaincompress[compr];[compr][mix]amerge"

🚪 sidechaingate

a sidechain gate acts like a normal (wideband) gate but it has the ability to filter the detected signal before sending it to the gain reduction stage ✨. normally a gate uses the full range signal to detect a level above the threshold. like, for example: if you cut all lower frequencies from your sidechain signal, the gate will decrease the volume of your track only if not enough highs appear. with this technique you're able to reduce the resonation of a natural drum, or remove "rumbling" of muted strokes from a heavily distorted guitar 🎸🔥. it needs two input streams and returns one output stream. first input stream will be processed depending on the second stream's signal.

the filter accepts the following options:

level_in — sets input level before filtering. default is 1. allowed range is from 0.015625 to 64.
mode — sets the mode of operation. can be "upward" or "downward". default is "downward". if set to "upward" mode, higher parts of the signal will be amplified, expanding dynamic range in the upward direction. otherwise, in case of "downward", lower parts of the signal will be reduced 🫡.
range — sets the level of gain reduction when the signal is below the threshold. default is 0.06125. allowed range is from 0 to 1. setting this to 0 disables reduction, and then the filter behaves like an expander.
threshold — if a signal rises above this level the gain reduction is released. default is 0.125. allowed range is from 0 to 1.
ratio — sets a ratio about which the signal is reduced. default is 2. allowed range is from 1 to 9000.
attack — amount of milliseconds the signal has to rise above the threshold before gain reduction stops. default is 20 milliseconds. allowed range is from 0.01 to 9000.
release — amount of milliseconds the signal has to fall below the threshold before the reduction is increased again. default is 250 milliseconds. allowed range is from 0.01 to 9000.
makeup — sets amount of amplification of signal after processing. default is 1. allowed range is from 1 to 64.
knee — curves the sharp knee around the threshold so it enters gain reduction more softly. default is 2.828427125. allowed range is from 1 to 8.
detection — chooses if the exact signal should be taken for detection or an RMS-like one. default is rms. can be peak or rms.
link — chooses if the average level between all channels, or the louder channel, affects the reduction. default is average. can be average or maximum.
level_sc — sets sidechain gain. default is 1. range is from 0.015625 to 64.

Commands

this filter supports the all above options as commands, periodt 💅.

🤫 silencedetect

detects silence in an audio stream. that's it, that's the job.

this filter logs a message when it detects that the input audio volume is less than or equal to a noise tolerance value, for a duration greater than or equal to the minimum detected noise duration. the printed times and duration are expressed in seconds ⏱️.

the lavfi.silence_start or lavfi.silence_start.X metadata key is set on the first frame whose timestamp equals or exceeds the detection duration, and it contains the timestamp of the first frame of the silence.

the lavfi.silence_duration or lavfi.silence_duration.X and lavfi.silence_end or lavfi.silence_end.X metadata keys are set on the first frame after the silence. if mono is enabled, and each channel is evaluated separately, the .X suffixed keys are used, and X corresponds to the channel number 👑

the filter accepts the following options:

noise, n — sets noise tolerance. can be specified in dB (in case "dB" is appended to the specified value) or as an amplitude ratio. default is -60dB, or 0.001.
duration, d — sets silence duration until notification (default is 2 seconds). see the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax 📖
mono, m — processes each channel separately, instead of combined. by default this is disabled.

Examples

detect 5 seconds of silence with -50dB noise tolerance:

silencedetect=n=-50dB:d=5

complete example with ffmpeg to detect silence with 0.0001 noise tolerance in silence.mp3:

ffmpeg -i silence.mp3 -af silencedetect=noise=0.0001 -f null -

✂️ silenceremove

removes silence from the beginning, middle or end of the audio. the glow-up is real 💖

the filter accepts the following options:

start_periods — this value is used to indicate if audio should be trimmed at the beginning of the audio. a value of zero indicates no silence should be trimmed from the beginning. when specifying a non-zero value, it trims audio up until it finds non-silence. normally, when trimming silence from the beginning of audio, the start_periods will be 1, but it can be increased to higher values to trim all audio up to a specific count of non-silence periods. default value is 0.
start_duration — specify the amount of time that non-silence must be detected before it stops trimming audio. by increasing the duration, bursts of noise can be treated as silence and trimmed off 🎀. default value is 0.
start_threshold — this indicates what sample value should be treated as silence. for digital audio, a value of 0 may be fine, but for audio recorded from analog, you may wish to increase the value to account for background noise. can be specified in dB (in case "dB" is appended to the specified value) or as an amplitude ratio. default value is 0.
start_silence — specify max duration of silence at the beginning that will be kept after trimming. default is 0, which is equal to trimming all samples detected as silence.
start_mode — specify mode of detection of silence end at start of multi-channel audio. can be any or all. default is any. with any, any sample from any channel that is detected as non-silence will trigger the end of silence trimming at the start of the audio stream. with all, only if every sample from every channel is detected as non-silence will the end of silence trimming at the start of the audio stream be triggered — limited usage, ngl 🫠
stop_periods — sets the count for trimming silence from the end of audio. when specifying a positive value, it trims audio after it finds the specified silence period. to remove silence from the middle of a file, specify a stop_periods that is negative 👀. this value is then treated as a positive value and is used to indicate the effect should restart processing as specified by stop_periods, making it suitable for removing periods of silence in the middle of the audio. default value is 0.
stop_duration — specify a duration of silence that must exist before audio is not copied any more. by specifying a higher duration, silence that is wanted can be left in the audio. default value is 0.
stop_threshold — the same vibe as start_threshold but for chopping silence off the end of your audio. You can write it in dB (just slap "dB" on the end of the value) or as an amplitude ratio. Default is 0.
stop_silence — how much silence at the end you're allowed to keep after trimming, max duration. Default is 0, which literally means "trim every single sample flagged as silence". No crumbs left. 🍪
stop_mode — sets how silence-start is detected after the start of multi-channel audio. Can be any or all. Default is all. With any, any sample from any channel detected as silence triggers the start of silence trimming after the audio stream starts (limited usage, iykyk). With all, only when every sample from every channel is detected as silence does the trimming kick in.
detection — sets how silence actually gets detected. Bestie, pick your fighter:
- avg — mean of absolute values of samples in the moving window.
- rms — root squared mean of absolute values of samples in the moving window.
- peak — maximum of absolute values of samples in the moving window.
- median — median of absolute values of samples in the moving window.
- ptp — absolute of max peak to min peak difference of samples in the moving window.
- dev — standard deviation of values of samples in the moving window.

Default value is "rms".

window — set the duration, in seconds, used to calculate the window size in number of samples for detecting silence. Using 0 effectively disables any windowing and uses only a single sample per channel for silence detection. In that case you may need to also set start_silence and/or stop_silence to nonzero values, plus start_duration and/or stop_duration to nonzero values too. Default value is 0.02. Allowed range is from 0 to 10.
timestamp — sets the processing mode of every output audio frame's timestamp. Two moods:
- write — full timestamps rewrite, keeping only the start time for the first output frame.
- copy — non-dropped frames keep the same timestamp as the input audio frame.

Default value is "write".

Examples ✨

The example below shows how this filter can start a recording that doesn't have that awkward delay at the beginning — you know, the one between pressing record and the actual performance starting:

silenceremove=start_periods=1:start_duration=5:start_threshold=0.02

Trim all silence from beginning to end where there's more than 1 second of silence in the audio:

silenceremove=stop_periods=-1:stop_duration=1:stop_threshold=-90dB

Trim all digital silence samples, using peak detection, from beginning to end where there are more than 0 samples of digital silence in the audio and digital silence is detected in all channels at the same positions in the stream:

silenceremove=window=0:detection=peak:stop_mode=all:start_mode=all:stop_periods=-1:stop_threshold=0

Trim every 2nd encountered silence period from beginning to end where there's more than 1 second of silence per silence period in the audio:

silenceremove=stop_periods=-2:stop_duration=1:stop_threshold=-90dB

Same as above, but keep a maximum of 0.5 seconds of silence from each trimmed period:

silenceremove=stop_periods=-2:stop_duration=1:stop_threshold=-90dB:stop_silence=0.5

Same as above, but keep a maximum of 1.5 seconds of silence from the start of the audio:

silenceremove=stop_periods=-2:stop_duration=1:stop_threshold=-90dB:stop_silence=0.5:start_periods=1:start_duration=1:start_silence=1.5:stop_threshold=-90dB

Commands ⌨️

This filter supports some of the above options as commands. Slay. 💅

🎧 sofalizer 🎧

SOFAlizer uses head-related transfer functions (HRTFs) to create virtual loudspeakers all around you for binaural listening through headphones (audio formats up to 9 channels supported, no cap). The HRTFs live in SOFA files — peep <http://www.sofacoustics.org/> for a whole database of them. SOFAlizer is developed at the Acoustics Research Institute (ARI) of the Austrian Academy of Sciences. Respect. 👑

To enable compilation of this filter you need to configure FFmpeg with --enable-libmysofa.

The filter accepts the following options:

sofa — set the SOFA file used for rendering.
gain — set gain applied to audio. Value is in dB. Default is 0.
rotation — set rotation of virtual loudspeakers in deg. Default is 0.
elevation — set elevation of virtual speakers in deg. Default is 0.
radius — set distance in meters between the loudspeakers and the listener, using near-field HRTFs. Default is 1.
type — set processing type. Can be time or freq. time processes audio in the time domain, which is slow 🐌. freq processes audio in the frequency domain, which is fast ⚡. Default is freq.
speakers — set custom positions of virtual loudspeakers. Syntax for this option is: <CH> <AZIM> <ELEV>[|<CH> <AZIM> <ELEV>|...]. Each virtual loudspeaker is described with a short channel name followed by azimuth and elevation in degrees. Each virtual loudspeaker description is separated by '|'. For example, to override the front left and front right channel positions use: 'speakers=FL 45 15|FR 345 15'. Descriptions with unrecognised channel names are ignored (they get no attention, periodt).
lfegain — set custom gain for LFE channels. Value is in dB. Default is 0.
framesize — set custom frame size in number of samples. Default is 1024. Allowed range is from 1024 to 96000. Only used if option type is set to freq.
normalize — should all IRs be normalized upon importing the SOFA file? By default is enabled.
interpolate — should nearest IRs be interpolated with neighbor IRs if the exact position doesn't match? By default is disabled.
minphase — minphase all IRs upon loading of the SOFA file. By default is disabled.
anglestep — set neighbor search angle step. Only used if option interpolate is enabled.
radstep — set neighbor search radius step. Only used if option interpolate is enabled.

Examples 🦋

Using the ClubFritz6 sofa file:

sofalizer=sofa=/path/to/ClubFritz6.sofa:type=freq:radius=1

Using the ClubFritz12 sofa file and a bigger radius with a small rotation:

sofalizer=sofa=/path/to/ClubFritz12.sofa:type=freq:radius=2:rotation=5

Similar to above, but with custom speaker positions for front left, front right, back left and back right — plus custom gain:

"sofalizer=sofa=/path/to/ClubFritz6.sofa:type=freq:radius=2:speakers=FL 45|FR 315|BL 135|BR 225:gain=28"

🗣️ speechnorm 🗣️

Speech Normalizer. It expands or compresses each half-cycle of audio samples (a local set of samples all above or all below zero, between two nearest zero crossings) depending on the threshold value, so the audio reaches the target peak value under conditions controlled by the options below. Your voice: louder, but make it make sense. 🎤✨

The filter accepts the following options:

peak, p — set the expansion target peak value. This specifies the highest allowed absolute amplitude level for the normalized audio input. Default value is 0.95. Allowed range is from 0.0 to 1.0.
expansion, e — set the maximum expansion factor. Allowed range is from 1.0 to 50.0. Default value is 2.0. This option controls the maximum local half-cycle-of-samples expansion. The maximum expansion would be such that the local peak value reaches the target peak value but never surpasses it, and that the ratio between new and previous peak value doesn't surpass this option value.
compression, c — set the maximum compression factor. Allowed range is from 1.0 to 50.0. Default value is 2.0. This option controls the maximum local half-cycle-of-samples compression. It's only used if option threshold is set to a value greater than 0.0 — then in those cases where the local peak is lower than or the same as the value set by threshold, all samples belonging to that peak's half-cycle get compressed by the current compression factor.
threshold, t — set the threshold value. Default value is 0.0. Allowed range is from 0.0 to 1.0. This option specifies which half-cycles of samples get compressed and which get expanded. Any half-cycle samples whose local peak value is below or the same as this option value get compressed by the current compression factor — otherwise, if greater than the threshold value, they get expanded with the expansion factor so it could reach the peak target value but never surpass it.
raise, r — set the expansion raising amount per each half-cycle of samples. Default value is 0.001. Allowed range is from 0.0 to 1.0. This controls how fast the expansion factor is raised per each new half-cycle until it reaches the expansion value. Setting this option too high may lead to distortions — highkey be careful. 😭
fall, f — set the compression raising amount per each half-cycle of samples. Default value is 0.001. Allowed range is from 0.0 to 1.0. This controls how fast the compression factor is raised per each new half-cycle until it reaches the compression value.
channels, h — specify which channels to filter. By default all available channels are filtered.
invert, i — enable inverted filtering, by default is disabled. This inverts the interpretation of the threshold option. When enabled, any half-cycle of samples whose local peak value is below or the same as the threshold option gets expanded, otherwise it gets compressed.
link, l — link channels when calculating the gain applied to each filtered channel sample. By default is disabled. When disabled, each filtered channel's gain calculation is independent — otherwise, when this option is enabled, the minimum of all possible gains for each filtered channel is used.
rms, m — set the expansion target RMS value. This specifies the highest allowed RMS level for the normalized audio input. Default value is 0.0, thus disabled. Allowed range is from 0.0 to 1.0.

Commands 🎀

This filter supports all of the above options as commands. All of them. Iconic. 💖

Examples 🌈

Weak and slow amplification:

speechnorm=e=3:r=0.00001:l=1

Moderate and slow amplification:

speechnorm=e=6.25:r=0.00001:l=1

Strong and fast amplification:

speechnorm=e=12.5:r=0.0001:l=1

Very strong and fast amplification:

speechnorm=e=25:r=0.0001:l=1

Extreme and fast amplification:

speechnorm=e=50:r=0.0001:l=1

🎀 `stereotools`

Okay bestie, this filter is literally the Swiss Army knife of stereo. It's got all these handy lil utilities for managing stereo signals — converting M/S stereo recordings to L/R signal while you keep full control over the parameters, or spreading out the stereo image of your master track. Iconic.

The filter accepts the following options:

level_in — sets the input level before filtering, for both channels. Default is 1. Allowed range is from 0.015625 to 64.
level_out — sets the output level after filtering, for both channels. Default is 1. Allowed range is from 0.015625 to 64.
balance_in — sets the input balance between both channels. Default is 0. Allowed range is from -1 to 1.
balance_out — sets the output balance between both channels. Default is 0. Allowed range is from -1 to 1.
softclip — enables softclipping. It gives you analog distortion instead of that harsh digital 0dB clipping. Disabled by default.
mutel — mutes the left channel. Disabled by default.
muter — mutes the right channel. Disabled by default.
phasel — changes the phase of the left channel. Disabled by default.
phaser — changes the phase of the right channel. Disabled by default.
mode — sets the stereo mode. The available values are:
- lr>lr — Left/Right to Left/Right, this is the default one.
- lr>ms — Left/Right to Mid/Side.
- ms>lr — Mid/Side to Left/Right.
- lr>ll — Left/Right to Left/Left.
- lr>rr — Left/Right to Right/Right.
- lr>l+r — Left/Right to Left + Right.
- lr>rl — Left/Right to Right/Left.
- ms>ll — Mid/Side to Left/Left.
- ms>rr — Mid/Side to Right/Right.
- ms>rl — Mid/Side to Right/Left.
- lr>l-r — Left/Right to Left - Right.
slev — sets the level of the side signal. Default is 1. Allowed range is from 0.015625 to 64.
sbal — sets the balance of the side signal. Default is 0. Allowed range is from -1 to 1.
mlev — sets the level of the middle signal. Default is 1. Allowed range is from 0.015625 to 64.
mpan — sets the middle signal pan. Default is 0. Allowed range is from -1 to 1.
base — sets the stereo base between mono and inversed channels. Default is 0. Allowed range is from -1 to 1.
delay — sets the delay in milliseconds, like how much to delay left from right channel and vice versa. Default is 0. Allowed range is from -20 to 20.
sclevel — sets the S/C level. Default is 1. Allowed range is from 1 to 100.
phase — sets the stereo phase in degrees. Default is 0. Allowed range is from 0 to 360.
bmode_in, bmode_out — sets the balance mode for the balance_in/balance_out option. Can be one of the following:
- balance — classic balance mode. Attenuate one channel at time. Gain is raised up to 1.
- amplitude — similar as the classic mode above, but the gain is raised up to 2.
- power — equal power distribution, from -6dB to +6dB range.

💬 Commands

This filter supports all of the above options as commands. Slay.

🌟 Examples

Apply a karaoke like effect:

stereotools=mlev=0.015625

Convert M/S signal to L/R:

"stereotools=mode=ms>lr"

🦋 `stereowiden`

This filter enhances the stereo effect by suppressing the signal that's common to both channels, and by delaying the left signal into the right and vice versa — which widens the stereo effect. It's giving main character energy, fr.

The filter accepts the following options:

delay — time in milliseconds of the delay of left signal into right and vice versa. Default is 20 milliseconds.
feedback — amount of gain in the delayed signal into right and vice versa. It gives a delay effect of the left signal in the right output and vice versa, which gives that widening effect. Default is 0.3.
crossfeed — cross feed of left into right with inverted phase. This helps in suppressing the mono. If the value is 1 it will cancel all the signal common to both channels. Default is 0.3.
drymix — sets the level of the input signal of the original channel. Default is 0.8.

💬 Commands

This filter supports all of the above options except delay as commands. Yeah, that one's not invited to the command party 🙈.

✨ `superequalizer`

Apply an 18 band equalizer. Eighteen. Bands. Obsessed.

The filter accepts the following options:

1b — set the 65Hz band gain.
2b — set the 92Hz band gain.
3b — set the 131Hz band gain.
4b — set the 185Hz band gain.
5b — set the 262Hz band gain.
6b — set the 370Hz band gain.
7b — set the 523Hz band gain.
8b — set the 740Hz band gain.
9b — set the 1047Hz band gain.
10b — set the 1480Hz band gain.
11b — set the 2093Hz band gain.
12b — set the 2960Hz band gain.
13b — set the 4186Hz band gain.
14b — set the 5920Hz band gain.
15b — set the 8372Hz band gain.
16b — set the 11840Hz band gain.
17b — set the 16744Hz band gain.
18b — set the 20000Hz band gain.

🌈 `surround`

Apply the audio surround upmix filter. This filter lets you produce multichannel output from an audio stream. Glow-up for your stereo, honestly.

The filter accepts the following options:

chl_out — sets the output channel layout. By default, this is 5.1. See the Channel Layout section in the ffmpeg-utils(1) manual for the required syntax.
chl_in — sets the input channel layout. By default, this is stereo. See the Channel Layout section in the ffmpeg-utils(1) manual for the required syntax.
level_in — sets the input volume level. By default, this is 1.
level_out — sets the output volume level. By default, this is 1.
lfe — enable LFE channel output if the output channel layout has it. By default, this is enabled.
lfe_low — sets the LFE low cut off frequency. By default, this is 128 Hz.
lfe_high — sets the LFE high cut off frequency. By default, this is 256 Hz.
lfe_mode — sets the LFE mode, can be add or sub. Default is add. In add mode, the LFE channel is created from the input audio and added to the output. In sub mode, the LFE channel is created from the input audio and added to the output, but also all non-LFE output channels are subtracted with the output LFE channel.
smooth — sets the temporal smoothness strength, used to gradually change factors when transforming stereo sound in time. Allowed range is from 0.0 to 1.0. Useful to improve output quality with focus option values greater than 0.0. Default is 0.0. Only values inside this range and without edges are effective.
angle — sets the angle of the stereo surround transform. Allowed range is from 0 to 360. Default is 90.
focus — sets the focus of the stereo surround transform. Allowed range is from -1 to 1. Default is 0.
fc_in — sets the front center input volume. By default, this is 1.
fc_out — sets the front center output volume. By default, this is 1.
fl_in — sets the front left input volume. By default, this is 1.
fl_out — sets the front left output volume. By default, this is 1.
fr_in — sets the front right input volume. By default, this is 1.
fr_out — sets the front right output volume. By default, this is 1.
sl_in — sets the side left input volume. By default, this is 1.
sl_out — sets the side left output volume. By default, this is 1.
sr_in — sets the side right input volume. By default, this is 1.
sr_out — sets the side right output volume. By default, this is 1.
bl_in — sets the back left input volume. By default, this is 1.
bl_out — sets the back left output volume. By default, this is 1.
br_in — sets the back right input volume. By default, this is 1.
br_out — sets the back right output volume. By default, this is 1.
bc_in — sets the back center input volume. By default, this is 1.
bc_out — sets the back center output volume. By default, this is 1.
lfe_in — sets the LFE input volume. By default, this is 1.
lfe_out — sets the LFE output volume. By default, this is 1.
allx — sets spread usage of the stereo image across the X axis, for all channels. Allowed range is from -1 to 15. By default this value is negative -1, and thus unused.
ally — sets spread usage of the stereo image across the Y axis, for all channels. Allowed range is from -1 to 15. By default this value is negative -1, and thus unused.
fcx, flx, frx, blx, brx, slx, srx, bcx — set spread usage of the stereo image across the X axis for each channel. Allowed range is from 0.06 to 15. By default this value is 0.5.
fcy, fly, fry, bly, bry, sly, sry, bcy — set spread usage of the stereo image across the Y axis for each channel. Allowed range is from 0.06 to 15. By default this value is 0.5.
win_size — sets the window size. Allowed range is from 1024 to 65536. Default size is 4096.
win_func — sets the window function. It accepts the following values:

rect
bartlett
hann, hanning
hamming
blackman
welch
flattop
bharris
bnuttall
bhann
sine
nuttall
lanczos
gauss
tukey
dolph
cauchy
parzen
poisson
bohman
kaiser

Default is "hann".

overlap — sets the window overlap. If set to 1, the recommended overlap for the selected window function will be picked. Default is 0.5.

💜 `tiltshelf`

Boost or cut the lower frequencies, and cut or boost the higher frequencies of the audio, using a two-pole shelving filter with a response similar to that of a standard hi-fi's tone-controls. This is also known as shelving equalisation (EQ). Vintage but make it cute 💅.

The filter accepts the following options:

gain, g — give the gain at 0 Hz. Its useful range is about -20 (for a large cut) to +20 (for a large boost). Beware of clipping when using a positive gain — fr fr, watch out for that.
frequency, f — set the filter's central frequency, so it can be used to extend or reduce the frequency range to be boosted or cut. The default value is 3000 Hz.
width_type, t — set the method to specify the band-width of the filter.
- h — Hz
- q — Q-Factor
- o — octave
- s — slope
- k — kHz
width, w — determine how steep the filter's shelf transition is.
poles, p — set the number of poles. Default is 2.
mix, m — how much to use the filtered signal in the output. Default is 1. Range is between 0 and 1.
channels, c — specify which channels to filter; by default all available are filtered.
normalize, n — normalize the biquad coefficients. By default is disabled. Enabling it will normalize the magnitude response at DC to 0dB.
transform, a — set the transform type of the IIR filter.

di
dii
tdi
tdii
latt
svf
zdf

precision, r — sets the precision of the filtering, bestie. pick your sample format:
- auto — picks the sample format automatically, based on the surround filters 💅
- s16 — always signed 16-bit, no cap
- s32 — always signed 32-bit
- f32 — always float 32-bit
- f64 — always float 64-bit
block_size, b — sets the block size used for reverse IIR processing. and listen: if you set this high enough (higher than the impulse response length, truncated when it reaches near-zero values) the filtering goes linear phase ✨ — but if it's not big enough it'll just produce nasty artifacts 😭 fr fr.

heads up: when this is set to a non-zero value, the filter delay will be exactly this many samples. like, exactly. 🫶

Commands

this filter supports some options as commands 👑

🌸 `treble`, `highshelf` 🌸

boosts or cuts the treble (upper) frequencies of the audio, using a two-pole shelving filter with a response similar to a standard hi-fi's tone-controls. this is also known as shelving equalisation (EQ), iykyk 🎀

the filter accepts these options:

gain, g — gives you the gain at whichever is the lower of ~22 kHz and the Nyquist frequency. useful range is like -20 (big cut) to +20 (big boost). ⚠️ beware of clipping when you use a positive gain, okay?
frequency, f — sets the filter's central frequency, so you can extend or reduce the frequency range that gets boosted or cut. default value is 3000 Hz.
width_type, t — sets the method used to specify the band-width of the filter:
- h — Hz
- q — Q-Factor
- o — octave
- s — slope
- k — kHz
width, w — determines how steep the filter's shelf transition is 📐
poles, p — sets the number of poles. default is 2.
mix, m — how much of the filtered signal goes into the output. default is 1. range is between 0 and 1.
channels, c — specifies which channels to filter. by default all available channels get filtered 💖
normalize, n — normalizes the biquad coefficients; by default it's disabled. turning it on normalizes the magnitude response at DC to 0dB. slay.
transform, a — sets the transform type of the IIR filter:
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — sets the precision of filtering:
- auto — picks the automatic sample format depending on the surround filters
- s16 — always use signed 16-bit
- s32 — always use signed 32-bit
- f32 — always use float 32-bit
- f64 — always use float 64-bit
block_size, b — sets the block size used for reverse IIR processing. set it high enough (higher than the impulse response length, truncated when it reaches near-zero values) and the filtering becomes linear phase ✨; if it's not big enough it just produces nasty artifacts 😭

note: the filter delay will be exactly this many samples when set to a non-zero value. same deal as before, bestie.

Commands

this filter supports these commands:

frequency, f — change treble frequency. syntax for the command is: "frequency"
width_type, t — change treble width_type. syntax for the command is: "width_type"
width, w — change treble width. syntax for the command is: "width"
gain, g — change treble gain. syntax for the command is: "gain"
mix, m — change treble mix. syntax for the command is: "mix"

🦋 `tremolo` 🦋

sinusoidal amplitude modulation — literally the wobble filter 🎀

the filter accepts these options:

f — modulation frequency in Hertz. modulation frequencies in the subharmonic range (20 Hz or lower) give you a tremolo effect. you can also use this filter as a ring modulator by specifying a modulation frequency higher than 20 Hz ✨ range is 0.1 - 20000.0. default value is 5.0 Hz.
d — depth of modulation as a percentage. range is 0.0 - 1.0. default value is 0.5.

💜 `vibrato` 💜

sinusoidal phase modulation. the sibling of tremolo, ok?

the filter accepts these options:

f — modulation frequency in Hertz. range is 0.1 - 20000.0. default value is 5.0 Hz.
d — depth of modulation as a percentage. range is 0.0 - 1.0. default value is 0.5.

🔥 `virtualbass` 🔥

applies the audio Virtual Bass filter. it takes stereo input and produces stereo with LFE (2.1) channels on the output 👑

the newly produced LFE channel has enhanced virtual bass, originally obtained from both stereo channels. this filter outputs the front left and front right channels unchanged, exactly as they were in the stereo input — not me preserving your mains, fr fr 💅

the filter accepts these options:

cutoff — sets the virtual bass cutoff frequency. default value is 250 Hz. allowed range is from 100 to 500 Hz.
strength — sets the virtual bass strength. allowed range is from 0.5 to 3. default value is 3.

⭐ `volume` ⭐

adjusts the input audio volume. main character filter, honestly 🎀

it accepts these parameters:

volume — sets the audio volume expression.

output values are clipped to the maximum value.

the output audio volume is given by this relation:

<output_volume> = <volume> * <input_volume>

the default value for volume is "1.0".

precision — this parameter represents the mathematical precision 🧮

it determines which input sample formats will be allowed, which affects the precision of the volume scaling:

fixed — 8-bit fixed-point; this limits input sample format to U8, S16, and S32.
float — 32-bit floating-point; this limits input sample format to FLT. (default)
double — 64-bit floating-point; this limits input sample format to DBL.

replaygain — chooses the behaviour when it encounters ReplayGain side data in input frames:
- drop — removes ReplayGain side data, ignoring its contents (the default).
- ignore — ignores ReplayGain side data, but leaves it in the frame.
- track — prefers the track gain, if present.
- album — prefers the album gain, if present.
replaygain_preamp — pre-amplification gain in dB to apply to the selected replaygain gain. default value for replaygain_preamp is 0.0.
replaygain_noclip — prevents clipping by limiting the gain applied. default value for replaygain_noclip is 1.
eval — sets when the volume expression is evaluated. it accepts these values:
- once — only evaluates the expression once during filter initialization, or when the volume command is sent
- frame — evaluates the expression for each incoming frame

default value is once.

the volume expression can contain these parameters:

n — frame number (starting at zero)
nb_channels — number of channels
nb_consumed_samples — number of samples consumed by the filter
nb_samples — number of samples in the current frame
pos — original frame position in the file; deprecated, do not use 🙈
pts — frame PTS
sample_rate — sample rate
startpts — PTS at start of stream
startt — time at start of stream
t — frame time
tb — timestamp timebase
volume — last set volume value

important, bestie: when eval is set to once, only the sample_rate and tb variables are available — all the other variables will evaluate to NAN. 😭 don't say i didn't warn you.

Commands

this filter supports these commands:

volume — modifies the volume expression. the command accepts the same syntax as the corresponding option.

if the specified expression isn't valid, it's kept at its current value.

Examples

halve the input audio volume:

volume=volume=0.5
volume=volume=1/2
volume=volume=-6.0206dB

in all the above examples the named key for volume can be omitted, like this:

volume=0.5

increase input audio power by 6 decibels using fixed-point precision:

volume=volume=6dB:precision=fixed

fade volume after time 10 with an annihilation period of 5 seconds:

volume='if(lt(t,10),1,max(1-(t-10)/5,0))':eval=frame

🍓 `volumedetect` 🍓

detects the volume of the input video. yes, video — that's what it says 💀

the filter has no parameters. it supports only 16-bit signed integer samples, so the input gets converted when needed. statistics about the volume get printed in the log when the input stream end is reached ✨

in particular it'll show you the mean volume (root mean square), the maximum volume (on a per-sample basis), and the beginning of a histogram of the registered volume values (from the maximum value to a cumulated 1/1000 of the samples).

all volumes are in decibels relative to the maximum PCM value 🎀

Examples

here's an excerpt of the output:

[Parsed_volumedetect_0  0xa23120] mean_volume: -27 dB
[Parsed_volumedetect_0  0xa23120] max_volume: -4 dB
[Parsed_volumedetect_0  0xa23120] histogram_4db: 6
[Parsed_volumedetect_0  0xa23120] histogram_5db: 62
[Parsed_volumedetect_0  0xa23120] histogram_6db: 286
[Parsed_volumedetect_0  0xa23120] histogram_7db: 1042
[Parsed_volumedetect_0  0xa23120] histogram_8db: 2551
[Parsed_volumedetect_0  0xa23120] histogram_9db: 4609
[Parsed_volumedetect_0  0xa23120] histogram_10db: 8409

it means that:

the mean square energy is approximately -27 dB, or 10^-2.7.
the largest sample is at -4 dB, or more precisely between -4 dB and -5 dB.
there are 6 samples at -4 dB, 62 at -5 dB, 286 at -6 dB, etc.

in other words, raising the volume by +4 dB doesn't cause any clipping, raising it by +5 dB causes clipping for 6 samples, etc. 📈

🧚 `whisper` 🧚

it runs automatic speech recognition using the OpenAI's Whisper model 👑

it requires the whisper.cpp library (https://github.com/ggml-org/whisper.cpp) as a prerequisite. after installing the library it can be enabled using: ./configure --enable-whisper. slay.

the filter has these options:

model — the file path of the downloaded whisper.cpp model (mandatory). like, you literally can't skip it 🙈
language — the language to use for transcription ('auto' for auto-detect). default value: "auto"
queue — the maximum size that will be queued into the filter before processing the audio with whisper, bestie. Using a small value means the audio stream gets processed more often, but the transcription quality will be lower and the required processing power will be higher. Using a large value (e.g. 10-20s) produces more accurate results using less CPU (as using the whisper-cli tool), but the transcription latency will be higher, so it’s not useful to process real-time streams. Consider using the vad_model option associated with a large queue value. Default value: "3"
use_gpu — if GPU support should be enabled. Default value: "true"
gpu_device — the GPU device index to use. Default value: "0"
destination — if set, the transcription output will be sent to the specified file or URL (use one of the FFmpeg AVIO protocols); otherwise, the output will be logged as info messages. The output will also be set in the "lavfi.whisper.text" frame metadata. If the destination is a file and it already exists, it will be overwritten. 💅
format — the destination format string; it could be "text" (only the transcribed text will be sent to the destination), "srt" (subtitle format) or "json". Default value: "text"
max_len — maximum segment length in characters. When set to a value greater than 0, transcription segments will be split to not exceed this length. This is useful for generating subtitles with shorter lines. Default value: "0"
vad_model — path to the VAD model file. If set, the filter will load an additional voice activity detection module (https://github.com/snakers4/silero-vad) that will be used to fragment the audio queue; use this option setting a valid path obtained from the whisper.cpp repository (e.g. "../whisper.cpp/models/ggml-silero-v5.1.2.bin") and increase the queue parameter to a higher value (e.g. 20).
vad_threshold — the VAD threshold to use. Default value: "0.5"
vad_min_speech_duration — the minimum VAD speaking duration. Default value: "0.1"
vad_min_silence_duration — the minimum VAD silence duration. Default value: "0.5"

✨ Examples ✨

🌟 Run a transcription with srt file generation:

ffmpeg -i input.mp4 -vn -af "whisper=model=../whisper.cpp/models/ggml-base.en.bin\
:language=en\
:queue=3\
:destination=output.srt\
:format=srt" -f null -

🌟 Run a transcription and send the output in JSON format to an HTTP service:

ffmpeg -i input.mp4 -vn -af "whisper=model=../whisper.cpp/models/ggml-base.en.bin\
:language=en\
:queue=3\
:destination=http\\://localhost\\:3000\
:format=json' -f null -

🌟 Transcribe the microphone input using the VAD option:

ffmpeg -loglevel warning -f pulse -i default \
-af 'highpass=f=200,lowpass=f=3000,whisper=model=../whisper.cpp/models/ggml-medium.bin\
:language=en\
:queue=10\
:destination=-\
:format=json\
:vad_model=../whisper.cpp/models/ggml-silero-v5.1.2.bin' -f null -

🌸 AUDIO SOURCES 🌸

Okay bestie, here’s the description of the currently available audio sources. 🎧✨

🎧 abuffer 🎧

Buffer audio frames, and make them available to the filter chain.

This source is mainly intended for a programmatic use, in particular through the interface defined in libavfilter/buffersrc.h.

It accepts the following parameters:

time_base — the timebase which will be used for timestamps of submitted frames. It must be either a floating-point number or in numerator/denominator form.
sample_rate — the sample rate of the incoming audio buffers.
sample_fmt — the sample format of the incoming audio buffers. Either a sample format name or its corresponding integer representation from the enum AVSampleFormat in libavutil/samplefmt.h.
channel_layout — the channel layout of the incoming audio buffers. Either a channel layout name from channel_layout_map in libavutil/channel_layout.c or its corresponding integer representation from the AV_CH_LAYOUT_* macros in libavutil/channel_layout.h.
channels — the number of channels of the incoming audio buffers. If both channels and channel_layout are specified, then they must be consistent.

🌟 Examples 🌟

abuffer=sample_rate=44100:sample_fmt=s16p:channel_layout=stereo

will instruct the source to accept planar 16bit signed stereo at 44100Hz. Since the sample format with name "s16p" corresponds to the number 6 and the "stereo" channel layout corresponds to the value 0x3, this is equivalent to:

abuffer=sample_rate=44100:sample_fmt=6:channel_layout=0x3

🌈 aevalsrc 🌈

Generate an audio signal specified by an expression.

This source accepts in input one or more expressions (one for each channel), which are evaluated and used to generate a corresponding audio signal.

This source accepts the following options:

exprs — set the '|'-separated expressions list for each separate channel. In case the channel_layout option is not specified, the selected channel layout depends on the number of provided expressions. Otherwise the last specified expression is applied to the remaining output channels.
channel_layout, c — set the channel layout. The number of channels in the specified layout must be equal to the number of specified expressions.
duration, d — set the minimum duration of the sourced audio. See the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. Note that the resulting duration may be greater than the specified duration, as the generated audio is always cut at the end of a complete frame.

If not specified, or the expressed duration is negative, the audio is supposed to be generated forever.

nb_samples, n — set the number of samples per channel per each output frame, default to 1024.
sample_rate, s — specify the sample rate, default to 44100.

Each expression in exprs can contain the following constants:

n — number of the evaluated sample, starting from 0
t — time of the evaluated sample expressed in seconds, starting from 0
s — sample rate

💖 Examples 💖

Generate silence:

aevalsrc=0

Generate a sin signal with frequency of 440 Hz, set sample rate to 8000 Hz:

aevalsrc="sin(440*2*PI*t):s=8000"

Generate a two channels signal, specify the channel layout (Front Center + Back Center) explicitly:

aevalsrc="sin(420*2*PI*t)|cos(430*2*PI*t):c=FC|BC"

Generate white noise:

aevalsrc="-2+random(0)"

Generate an amplitude modulated signal:

aevalsrc="sin(10*2*PI*t)*sin(880*2*PI*t)"

Generate 2.5 Hz binaural beats on a 360 Hz carrier:

aevalsrc="0.1*sin(2*PI*(360-2.5/2)*t) | 0.1*sin(2*PI*(360+2.5/2)*t)"

🦋 afdelaysrc 🦋

Generate a fractional delay FIR coefficients.

The resulting stream can be used with afir filter for filtering the audio signal.

The filter accepts the following options:

delay, d — set the fractional delay. Default is 0.
sample_rate, r — set the sample rate, default is 44100.
nb_samples, n — set the number of samples per each frame. Default is 1024.
taps, t — set the number of filter coefficients in output audio stream. Default value is 0.
channel_layout, c — specifies the channel layout, and can be a string representing a channel layout. The default value of channel_layout is "stereo".

🎀 afireqsrc 🎀

Generate a FIR equalizer coefficients.

The resulting stream can be used with afir filter for filtering the audio signal.

The filter accepts the following options:

preset, p — set equalizer preset. Default preset is "flat".

Available presets are:

custom
flat
acoustic
bass
beats
classic
clear
deep bass
dubstep
electronic
hard-style
hip-hop
jazz
metal
movie
pop
r&b
rock
vocal booster

gains, g — set custom gains for each band. Only used if the preset option is set to "custom". Gains are separated by white spaces and each gain is set in dBFS. Default is "0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0".
bands, b — set the custom bands from where custom equalizer gains are set. This must be in strictly increasing order. Only used if the preset option is set to "custom". Bands are separated by white spaces and each band represent frequency in Hz. Default is "25 40 63 100 160 250 400 630 1000 1600 2500 4000 6300 10000 16000 24000".
taps, t — set number of filter coefficients in output audio stream. Default value is 4096.
sample_rate, r — set sample rate of output audio stream, default is 44100.
nb_samples, n — set number of samples per each frame in output audio stream. Default is 1024.
interp, i — set interpolation method for FIR equalizer coefficients. Can be "linear" or "cubic".
phase, h — set phase type of FIR filter. Can be "linear" or "min": minimum-phase. Default is minimum-phase filter.

🐚 afirsrc 🐚

Generate a FIR coefficients using frequency sampling method.

The resulting stream can be used with afir filter for filtering the audio signal.

The filter accepts the following options:

taps, t — set number of filter coefficients in output audio stream. Default value is 1025.
frequency, f — set frequency points from where magnitude and phase are set. This must be in non decreasing order, and first element must be 0, while last element must be 1. Elements are separated by white spaces.
magnitude, m — set magnitude value for every frequency point set by frequency. Number of values must be same as number of frequency points. Values are separated by white spaces.
phase, p — set phase value for every frequency point set by frequency. Number of values must be same as number of frequency points. Values are separated by white spaces.
sample_rate, r — set sample rate, default is 44100.
nb_samples, n — set number of samples per each frame. Default is 1024.
win_func, w — set window function. Default is blackman.

anullsrc 🫧

okay so this one is the null audio source — it just hands you back unprocessed audio frames, like, it does nothing but it does it flawlessly 💅. it's mainly useful as a template, and for analysis / debugging tools, or as the source for filters that literally ignore the input data (the sox synth filter is the classic example, iykyk).

This source accepts the following options:

channel_layout, cl — tells it the channel layout, and ngl it can be either an integer or a string representing a channel layout. the default value of channel_layout is "stereo". and bestie, check the channel_layout_map definition in libavutil/channel_layout.c for the mapping between strings and channel layout values 👑
sample_rate, r — sets the sample rate, defaults to 44100.
nb_samples, n — sets the number of samples per requested frames.
duration, d — sets the duration of the sourced audio. peep the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. if it's not specified, or the expressed duration is negative, the audio is supposed to be generated forever ♾️

Examples ✨

set the sample rate to 48000 Hz and the channel layout to AV_CH_LAYOUT_MONO:

anullsrc=r=48000:cl=4

do the same operation with a more obvious syntax:

anullsrc=r=48000:cl=mono

All the parameters need to be explicitly defined. no vibes-based defaults here, bestie 🙈

flite 🗣️

synthesize a voice utterance using the libflite library. literally a whole text-to-speech moment 🌸

to enable compilation of this filter you need to configure FFmpeg with --enable-libflite.

heads up: versions of the flite library prior to 2.0 are not thread-safe 😭

The filter accepts the following options:

list_voices — if set to 1, it lists the names of the available voices and exits immediately. default value is 0.
nb_samples, n — sets the maximum number of samples per frame. default value is 512.
textfile — sets the filename containing the text to speak.
text — sets the text to speak.
voice, v — sets the voice to use for the speech synthesis. default value is "kal". see also the list_voices option 👀

Examples 🎀

read from file speech.txt, and synthesize the text using the standard flite voice:

flite=textfile=speech.txt

read the specified text selecting the "slt" voice:

flite=text='So fare thee well, poor devil of a Sub-Sub, whose commentator I am':voice=slt

input text to ffmpeg:

ffmpeg -f lavfi -i flite=text='So fare thee well, poor devil of a Sub-Sub, whose commentator I am':voice=slt

make ffplay speak the specified text, using "flite" and the "lavfi" device:

ffplay -f lavfi flite=text='No more be grieved for which that thou hast done.'

for more information about libflite, check: http://www.festvox.org/flite/ 🦋

anoisesrc 🔊

generate a noise audio signal. yes, noise. on purpose. it's giving static TV at 3am 📺✨

The filter accepts the following options:

sample_rate, r — specify the sample rate. default value is 48000 Hz.
amplitude, a — specify the amplitude (0.0 - 1.0) of the generated audio stream. default value is 1.0.
duration, d — specify the duration of the generated audio stream. not specifying this option results in noise with an infinite length ♾️
color, colour, c — specify the color of noise. available noise colors are white, pink, brown, blue, violet and velvet. default color is white.
seed, s — specify a value used to seed the PRNG.
nb_samples, n — sets the number of samples per each output frame, default is 1024.
density — sets the density (0.0 - 1.0) for the velvet noise generator, default is 0.05.

Examples 💜

generate 60 seconds of pink noise, with a 44.1 kHz sampling rate and an amplitude of 0.5:

anoisesrc=d=60:c=pink:r=44100:a=0.5

hilbert 🌀

generate odd-tap Hilbert transform FIR coefficients.

the resulting stream can be used with the afir filter for phase-shifting the signal by 90 degrees. this is used in many matrix coding schemes and for analytic signal generation. the process is often written as a multiplication by i (or j), the imaginary unit. math girlies rise 🧚

The filter accepts the following options:

sample_rate, s — sets sample rate, default is 44100.
taps, t — sets length of FIR filter, default is 22051.
nb_samples, n — sets number of samples per each frame.
win_func, w — sets window function to be used when generating FIR coefficients.

sinc 🌈

generate a sinc kaiser-windowed low-pass, high-pass, band-pass, or band-reject FIR coefficients. all the flavors, one filter, we love versatility 💅

the resulting stream can be used with the afir filter for filtering the audio signal.

The filter accepts the following options:

sample_rate, r — sets sample rate, default is 44100.
nb_samples, n — sets number of samples per each frame. default is 1024.
hp — sets high-pass frequency. default is 0.
lp — sets low-pass frequency. default is 0. if high-pass frequency is lower than low-pass frequency and low-pass frequency is higher than 0 then the filter will create band-pass filter coefficients, otherwise band-reject filter coefficients. sneaky logic but ok 👀
phase — sets filter phase response. default is 50. allowed range is from 0 to 100.
beta — sets Kaiser window beta.
att — sets stop-band attenuation. default is 120dB, allowed range is from 40 to 180 dB.
round — enable rounding, by default is disabled.
hptaps — sets number of taps for high-pass filter.
lptaps — sets number of taps for low-pass filter.

sine 🎵

generate an audio signal made of a sine wave with amplitude 1/8. and bestie — the audio signal is bit-exact. flawless. no notes. ⭐

The filter accepts the following options:

frequency, f — sets the carrier frequency. default is 440 Hz.
beep_factor, b — enable a periodic beep every second with frequency beep_factor times the carrier frequency. default is 0, meaning the beep is disabled.
sample_rate, r — specify the sample rate, default is 44100.
duration, d — specify the duration of the generated audio stream.
samples_per_frame — sets the number of samples per output frame. the expression can contain the following constants:
- n — the (sequential) number of the output audio frame, starting from 0.
- pts — the PTS (Presentation TimeStamp) of the output audio frame, expressed in TB units.
- t — the PTS of the output audio frame, expressed in seconds.
- TB — the timebase of the output audio frames.

default is 1024.

Examples 🍓

generate a simple 440 Hz sine wave:

sine

generate a 220 Hz sine wave with a 880 Hz beep each second, for 5 seconds:

sine=220:4:d=5
sine=f=220:b=4:d=5
sine=frequency=220:beep_factor=4:duration=5

generate a 1 kHz sine wave following "1602,1601,1602,1601,1602" NTSC pattern:

sine=1000:samples_per_frame='st(0,mod(n,5)); 1602-not(not(eq(ld(0),1)+eq(ld(0),3)))'

🎧 AUDIO SINKS 🎧

below is a description of the currently available audio sinks. they're the end of the line, the final boss of the filter chain 👑

abuffersink 🫶

buffer audio frames, and make them available to the end of the filter chain.

this sink is mainly intended for programmatic use, in particular through the interface defined in libavfilter/buffersink.h or the options system. it accepts a pointer to an AVABufferSinkContext structure, which defines the incoming buffers' formats, to be passed as the opaque parameter to avfilter_init_filter for initialization.

anullsink 😶

null audio sink; does absolutely nothing with the input audio. it is mainly useful as a template and for use in analysis / debugging tools. iconic behaviour honestly — shows up, does nothing, leaves 💀

🎬 VIDEO FILTERS 🎬

when you configure your FFmpeg build, you can disable any of the existing filters using --disable-filters. the configure output will show the video filters included in your build. bestie check your build list, don't come crying when your filter is missing 😭

below is a description of the currently available video filters.

addroi 🎯

mark a region of interest in a video frame.

the frame data is passed through unchanged, but metadata is attached to the frame indicating regions of interest which can affect the behaviour of later encoding. multiple regions can be marked by applying the filter multiple times. lowkey the main character of "i'm not touching your pixels, i'm just leaving notes" 💅

x — region distance in pixels from the left edge of the frame.
y — region distance in pixels from the top edge of the frame.
w — region width in pixels.
h — region height in pixels.

the parameters x, y, w and h are expressions, and may contain the following variables:

iw — width of the input frame.
ih — height of the input frame.

qoffset — quantisation offset to apply within the region. this must be a real value in the range -1 to +1. a value of zero indicates no quality change. a negative value asks for better quality (less quantisation), while a positive value asks for worse quality (greater quantisation).

the range is calibrated so that the extreme values indicate the largest possible offset — if the rest of the frame is encoded with the worst possible quality, an offset of -1 indicates that this region should be encoded with the best possible quality anyway. intermediate values are then interpolated in some codec-dependent way.

for example, in 10-bit H.264 the quantisation parameter varies between -12 and 51. a typical qoffset value of -1/10 therefore indicates that this region should be encoded with a QP around one-tenth of the full range better than the rest of the frame. so, if most of the frame were to be encoded with a QP of around 30, this region would get a QP of around 24 (an offset of approximately -1/10 * (51 - -12) = -6.3). an extreme value of -1 would indicate that this region should be encoded with the best possible quality regardless of the treatment of the rest of the frame — that is, should be encoded at a QP of -12. math slay 🧮✨

clear — if set to true, remove any existing regions of interest marked on the frame before adding the new one.

Examples 🦋

mark the centre quarter of the frame as interesting:

addroi=iw/4:ih/4:iw/2:ih/2:-1/10

mark the 100-pixel-wide region on the left edge of the frame as very uninteresting (to be encoded at much lower quality than the rest of the frame):

addroi=0:0:100:ih:+1/5

alphaextract 🫥

extract the alpha component from the input as a grayscale video. this is especially useful with the alphamerge filter. they're a duo, fr fr 🤝

alphamerge 🧬

add or replace the alpha component of the primary input with the grayscale value of a second input. this is intended for use with alphaextract to allow the transmission or storage of frame sequences that have alpha in a format that doesn't support an alpha channel. big brain workaround energy 👑

so like, say you wanna rebuild full frames from a normal YUV-encoded video plus a separate video you made with alphaextract — adorable, iconic, here's how:

movie=in_alpha.mkv [alpha]; [in][alpha] alphamerge [out]

🌈 amplify

Bestie. This one literally amplifies the differences between the current pixel and the pixels sitting in the same spot in the adjacent frames. Same pixel location, different frames, drama ensues. ✨

It accepts the following options:

radius — sets the frame radius. Default is 2. Allowed range is from 1 to 63. Like, for example, a radius of 3 tells the filter to calculate the average of 7 frames. Math slay.
factor — sets the factor used to amplify the difference. Default is 2. Allowed range is from 0 to 65535.
threshold — sets the threshold for difference amplification. Any difference greater or equal to this value will not alter the source pixel. Default is 10. Allowed range is from 0 to 65535.
tolerance — sets the tolerance for difference amplification. Any difference lower than this value will not alter the source pixel. Default is 0. Allowed range is from 0 to 65535.
low — sets the lower limit for changing the source pixel. Default is 65535. Allowed range is from 0 to 65535. This option controls the maximum possible value that will decrease the source pixel value.
high — sets the high limit for changing the source pixel. Default is 65535. Allowed range is from 0 to 65535. This option controls the maximum possible value that will increase the source pixel value.
planes — sets which planes to filter. Default is all. Allowed range is from 0 to 15.

Commands

This filter supports the following commands, and each one corresponds to the option of the same name:

factor
threshold
tolerance
low
high
planes

🎀 ass

Okay so literally the same as the subtitles filter, except it does not require libavcodec and libavformat to work. Glow-up tradeoff though: it's limited to ASS (Advanced Substation Alpha) subtitles files. No cap, that's the whole deal. 💅

This filter accepts the following option in addition to the common options from the subtitles filter:

shaping — sets the shaping engine. 🧚

Available values are:

auto — the default libass shaping engine, which is the best available. Main character energy.
simple — fast, font-agnostic shaper that can do only substitutions.
complex — slower shaper using OpenType for substitutions and positioning.

The default is "auto".

🦋 atadenoise

Applies an Adaptive Temporal Averaging Denoiser to the video input. Your grainy footage is getting a spa day. ✨

The filter accepts the following options:

0a — sets threshold A for the 1st plane. Default is 0.02. Valid range is 0 to 0.3.
0b — sets threshold B for the 1st plane. Default is 0.04. Valid range is 0 to 5.
1a — sets threshold A for the 2nd plane. Default is 0.02. Valid range is 0 to 0.3.
1b — sets threshold B for the 2nd plane. Default is 0.04. Valid range is 0 to 5.
2a — sets threshold A for the 3rd plane. Default is 0.02. Valid range is 0 to 0.3.
2b — sets threshold B for the 3rd plane. Default is 0.04. Valid range is 0 to 5.

Threshold A is designed to react on abrupt changes in the input signal, and threshold B is designed to react on continuous changes in the input signal. Two different vibes entirely. 💀

s — sets the number of frames the filter will use for averaging. Default is 9. Must be an odd number in range [5, 129]. Even numbers, don't even try. 🙈
p — sets what planes of the frame the filter will use for averaging. Default is all.
a — sets what variant of algorithm the filter will use for averaging. Default is "p" (parallel). Alternatively it can be set to "s" (serial).

Parallel can be faster than serial, while the other way around is never true. Parallel will abort early on the first change being greater than the thresholds, while serial will continue processing the other side of frames if they are equal to or below the thresholds. Serial said "I finish what I start." 🫶

0s, 1s, 2s — set sigma for the 1st plane, 2nd plane, or 3rd plane. Default is 32767. Valid range is from 0 to 32767. This option controls the weight for each pixel in the radius defined by size. The default value means every pixel has the same weight. Setting this option to 0 effectively disables filtering. (So like, careful, bestie — 0 is a full off switch. 😭)

Commands

This filter supports the same commands as options, except the option "s". The command accepts the same syntax of the corresponding option.

🌸 avgblur

Applies the average blur filter. Soft focus, but make it math. ✨

The filter accepts the following options:

sizeX — sets the horizontal radius size.
planes — sets which planes to filter. By default all planes are filtered.
sizeY — sets the vertical radius size; if zero it will be the same as "sizeX". Default is 0.

Commands

This filter supports the same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value. It literally just stays where it is, no drama. 💅

👑 backgroundkey

Turns a static background into transparency. Bye background, you're not the main character anymore. 🫶

The filter accepts the following options:

threshold — threshold for scene change detection.
similarity — similarity percentage with the background.
blend — sets the blend amount for pixels that are not similar.

Commands

This filter supports all the above options as commands. Every single one. Slay. ✨

🍓 bbox

Computes the bounding box for the non-black pixels in the input frame's luma plane. It's giving "where's the content actually at." 👀

This filter computes the bounding box containing all the pixels with a luma value greater than the minimum allowed value. The parameters describing the bounding box are printed on the filter log.

The filter accepts the following option:

min_val — sets the minimal luma value. Default is 16.

Commands

This filter supports all the above options as commands.

💜 bilateral

Applies the bilateral filter — spatial smoothing while preserving edges. So it blurs the boring parts and keeps the outlines crisp. Iconic behaviour, honestly. 🎀

The filter accepts the following options:

sigmaS — sets the sigma of the gaussian function used to calculate spatial weight. Allowed range is 0 to 512. Default is 0.1.
sigmaR — sets the sigma of the gaussian function used to calculate range weight. Allowed range is 0 to 1. Default is 0.1.
planes — sets planes to filter. Default is first only.

Commands

This filter supports all the above options as commands.

⭐ bitplanenoise

Shows and measures bit plane noise. Little detective filter, out here analysing. 🔍

The filter accepts the following options:

bitplane — sets which plane to analyze. Default is 1.
filter — filters out noisy pixels from the "bitplane" set above. Default is disabled.

💀 blackdetect, blackdetect_vulkan

Detects video intervals that are (almost) completely black. Genuinely useful for detecting chapter transitions, commercials, or invalid recordings — like, the filter is literally your editor bestie pointing at the screen going "that part? nothing there." 😭

The filter outputs its detection analysis to both the log and frame metadata. If a black segment of at least the specified minimum duration is found, a line with the start and end timestamps as well as the duration is printed to the log with level "info". In addition, a log line with level "debug" is printed per frame showing the black amount detected for that frame.

The filter also attaches metadata to the first frame of a black segment with key "lavfi.black_start" and to the first frame after the black segment ends with key "lavfi.black_end". The value is the frame's timestamp. This metadata is added regardless of the minimum duration specified. Periodt. ✨

The filter accepts the following options:

black_min_duration, d — sets the minimum detected black duration expressed in seconds. It must be a non-negative floating point number. Default value is 2.0.
picture_black_ratio_th, pic_th — sets the threshold for considering a picture "black". Expresses the minimum value for the ratio:

<nb_black_pixels> / <nb_pixels>

for which a picture is considered black. Default value is 0.98.

pixel_black_th, pix_th — sets the threshold for considering a pixel "black".

The threshold expresses the maximum pixel luma value for which a pixel is considered "black". The provided value is scaled according to the following equation:

<absolute_threshold> = <luma_minimum_value> + <pixel_black_th> * <luma_range_size>

luma_range_size and luma_minimum_value depend on the input video format — the range is [0-255] for YUV full-range formats and [16-235] for YUV non full-range formats.

Default value is 0.10.

alpha — if true, check the alpha channel instead of the luma channel. Detects frames which are (almost) transparent, instead of frames which are almost black. Sneaky little switcheroo. 🙈

Default value is disabled.

The following example sets the maximum pixel threshold to the minimum value, and detects only black intervals of 2 or more seconds:

blackdetect=d=2:pix_th=0.00

🔥 blackframe

Detects frames that are (almost) completely black. Useful for detecting chapter transitions or commercials — same energy as blackdetect but frame-flavoured. Output lines consist of the frame number of the detected frame, the percentage of blackness, the position in the file if known (or -1) and the timestamp in seconds.

In order to display the output lines, you need to set the loglevel at least to the AV_LOG_INFO value. No loglevel, no receipts. 💅

This filter exports frame metadata "lavfi.blackframe.pblack". The value represents the percentage of pixels in the picture that are below the threshold value. Cute.

It accepts the following parameters:

amount — the percentage of the pixels that have to be below the threshold; it defaults to 98.
threshold, thresh — the threshold below which a pixel value is considered black; it defaults to 32.

🧚 blend

Blends two video frames into each other. Two frames, one output, one iconic merge. ✨

The "blend" filter takes two input streams and outputs one stream — the first input is the "top" layer and the second input is the "bottom" layer. By default, the output terminates when the longest input terminates.

The "tblend" (time blend) filter takes two consecutive frames from one single stream, and outputs the result obtained by blending the new frame on top of the old frame. Same stream, different time, still slaying. 🦋

A description of the accepted options follows.

c0_mode, c1_mode, c2_mode, c3_mode, all_mode — sets the blend mode for a specific pixel component, or for all pixel components in the case of all_mode. Default value is "normal".

Available values for component modes are:

addition
and
average
bleach
burn
darken
difference
divide
dodge
exclusion
extremity
freeze
geometric
glow
grainextract
grainmerge
hardlight
hardmix
hardoverlay
harmonic
heat
interpolate
lighten
linearlight
multiply
multiply128
negation
normal
or
overlay
phoenix
pinlight
reflect
screen
softdifference
softlight
stain
subtract
vividlight
xor

okay so now the per-component knobs, and these are the ones where you can really cook 🍳

c0_opacity
c1_opacity
c2_opacity
c3_opacity
all_opacity — sets the blend opacity for one specific pixel component, or literally every pixel component if you use all_opacity. Heads up bestie: this one only does anything when you're pairing it with the pixel-component blend modes 💅
c0_expr
c1_expr
c2_expr
c3_expr
all_expr — sets the blend expression for a single pixel component, or for all of them with all_expr. Important: if you set these, the related mode options just get ignored. Like, they're not even consulted. Bye 👋

The expressions get to use these variables, and yes you will want all of them:

N — the sequential number of the filtered frame, and yeah it starts from 0
X, Y — the coordinates of the current sample
W, H — the width and height of the currently filtered plane
SW, SH — the width and height scale for the plane being filtered. It's the ratio between the dimensions of the current plane and the luma plane. So like, for a yuv420p frame the values are 1,1 for the luma plane and 0.5,0.5 for the chroma planes 🎀
T — time of the current frame, expressed in seconds
TOP, A — the value of the pixel component at the current location for the first video frame (that's the top layer)
BOTTOM, B — the value of the pixel component at the current location for the second video frame (the bottom layer)

Also, the blend filter supports the framesync options too, so go wild ✨

Examples

Apply a transition from the bottom layer to the top layer over the first 10 seconds:

blend=all_expr='A*(if(gte(T,10),1,T/10))+B*(1-(if(gte(T,10),1,T/10)))'

Apply a linear horizontal transition from the top layer to the bottom layer:

blend=all_expr='A*(X/W)+B*(1-X/W)'

Apply a 1x1 checkerboard effect:

blend=all_expr='if(eq(mod(X,2),mod(Y,2)),A,B)'

Apply an uncover-left effect:

blend=all_expr='if(gte(N*SW+X,W),A,B)'

Apply an uncover-down effect:

blend=all_expr='if(gte(Y-N*SH,0),A,B)'

Apply an uncover-up-left effect:

blend=all_expr='if(gte(T*SH*40+Y,H)*gte((T*40*SW+X)*W/H,W),A,B)'

Split the video diagonally and show the top and bottom layer on each side:

blend=all_expr='if(gt(X,Y*(W/H)),A,B)'

Display the differences between the current and the previous frame:

tblend=all_mode=grainextract

Commands

This filter supports the same commands as options. slay 💅

`blockdetect`

Determines the blockiness of frames without altering the input frames. Pure observation, no touching 🫶

It's based on Remco Muijs and Ihor Kirenko: "A no-reference blocking artifact measure for adaptive video processing." 2005 13th European signal processing conference.

The filter accepts the following options:

period_min, period_max — set the minimum and maximum values for determining pixel grids (periods). Default values are [3,24].
planes — set the planes to filter. Default is the first one only.

Examples

Determine the blockiness for the first plane and search for periods within [8,32]:

blockdetect=period_min=8:period_max=32:planes=1

`blurdetect`

Determines the blurriness of frames without altering the input frames 👑 Again: observer only, hands off the pixels.

Based on Marziliano, Pina, et al. "A no-reference perceptual blur metric." And it allows for a block-based abbreviation, iykyk.

The filter accepts the following options:

low, high — set the low and high threshold values used by the Canny thresholding algorithm.

The high threshold selects the "strong" edge pixels, which are then connected through 8-connectivity with the "weak" edge pixels selected by the low threshold.

Note: low and high threshold values must be chosen in the range [0,1], and low should be lesser or equal to high.

The default value for low is 20/255, and the default value for high is 50/255.

radius — define the radius to search around an edge pixel for local maxima.
block_pct — determine blurriness only for the most significant blocks, given in percentage.
block_width — determine blurriness for blocks of width block_width. If set to any value smaller than 1, no blocks are used and the whole image is processed as one, no matter what block_height says.
block_height — determine blurriness for blocks of height block_height. If set to any value smaller than 1, no blocks are used and the whole image is processed as one, no matter what block_width says.
planes — set the planes to filter. Default is the first one only.

Examples

Determine blur for 80% of the most significant 32x32 blocks:

blurdetect=block_width=32:block_height=32:block_pct=80

`bm3d`

Denoise frames using the Block-Matching 3D algorithm. This is the "my footage is noisy and I'm not okay about it" filter 😭

The filter accepts the following options.

sigma — set the denoising strength. Default value is 1. Allowed range is from 0 to 999.9. The denoising algorithm is very sensitive to sigma, so adjust it according to your source. No cap, tune it.
block — set the local patch size. This sets the dimensions in 2D.
bstep — set the sliding step for processing blocks. Default value is 4. Allowed range is from 1 to 64. Smaller values allow processing more reference blocks and are slower (tragic, but that's the trade).
group — set the maximal number of similar blocks for the 3rd dimension. Default value is 1. When set to 1, no block matching is done. Larger values allow more blocks in a single group. Allowed range is from 1 to 256.
range — set the radius for search block matching. Default is 9. Allowed range is from 1 to INT32_MAX.
mstep — set the step between two search locations for block matching. Default is 1. Allowed range is from 1 to 64. Smaller is slower 🐌
thmse — set the threshold of mean square error for block matching. Valid range is 0 to INT32_MAX.
hdthr — set the thresholding parameter for hard thresholding in the 3D transformed domain. Larger values result in stronger hard-thresholding filtering in the frequency domain.
estim — set the filtering estimation mode. Can be basic or final. Default is basic.
ref — if enabled, the filter will use the 2nd stream for block matching. Default is disabled for the basic value of the estim option, and always enabled if the value of estim is final.
planes — set the planes to filter. Default is all available except alpha.

Examples

Basic filtering with bm3d:

bm3d=sigma=3:block=4:bstep=2:group=1:estim=basic

Same as above, but filtering only luma:

bm3d=sigma=3:block=4:bstep=2:group=1:estim=basic:planes=1

Same as above, but with both estimation modes:

split[a][b],[a]bm3d=sigma=3:block=4:bstep=2:group=1:estim=basic[a],[b][a]bm3d=sigma=3:block=4:bstep=2:group=16:estim=final:ref=1

Same as above, but prefilter with the nlmeans filter instead:

split[a][b],[a]nlmeans=s=3:r=7:p=3[a],[b][a]bm3d=sigma=3:block=4:bstep=2:group=16:estim=final:ref=1

`boxblur`

Apply a boxblur algorithm to the input video. Soft, cozy, iconic 🌸

It accepts the following parameters:

luma_radius, lr
luma_power, lp
chroma_radius, cr
chroma_power, cp
alpha_radius, ar
alpha_power, ap

A description of the accepted options follows.

luma_radius, lr / chroma_radius, cr / alpha_radius, ar — set an expression for the box radius in pixels used for blurring the corresponding input plane.

The radius value must be a non-negative number, and it must not be greater than the value of the expression min(w,h)/2 for the luma and alpha planes, and of min(cw,ch)/2 for the chroma planes.

Default value for luma_radius is 2. If it's not specified, chroma_radius and alpha_radius default to the corresponding value set for luma_radius.

The expressions can contain the following constants:

w, h — the input width and height in pixels.
cw, ch — the input chroma image width and height in pixels.
hsub, vsub — the horizontal and vertical chroma subsample values. For example, for the pixel format yuv422p, hsub is 2 and vsub is 1.

luma_power, lp / chroma_power, cp / alpha_power, ap — specify how many times the boxblur filter is applied to the corresponding plane.

Default value for luma_power is 2. If it's not specified, chroma_power and alpha_power default to the corresponding value set for luma_power.

A value of 0 will disable the effect. Zero means off, periodt.

Examples

Apply a boxblur filter with the luma, chroma, and alpha radii set to 2:

boxblur=luma_radius=2:luma_power=1
boxblur=2:1

Set the luma radius to 2, and the alpha and chroma radius to 0:

boxblur=2:1:cr=0:ar=0

Set the luma and chroma radii to a fraction of the video dimension:

boxblur=luma_radius=min(h\,w)/10:luma_power=1:chroma_radius=min(cw\,ch)/10:chroma_power=1

`bwdif`

Deinterlace the input video — and yes, bwdif stands for "Bob Weaver Deinterlacing Filter" 🦋

Motion adaptive deinterlacing based on yadif, with the use of w3fdif and cubic interpolation algorithms. It accepts the following parameters:

mode — the interlacing mode to adopt. It accepts one of the following values:
- 0, send_frame — output one frame for each frame.
- 1, send_field — output one frame for each field.

The default value is send_field.

parity — the picture field parity assumed for the input interlaced video. It accepts one of the following values:
- 0, tff — assume the top field is first.
- 1, bff — assume the bottom field is first.
- -1, auto — enable automatic detection of field parity.

The default value is auto. If the interlacing is unknown, or the decoder doesn't export this information, top field first will be assumed.

deint — specify which frames to deinterlace. Accepts one of the following values:
- 0, all — deinterlace all frames.
- 1, interlaced — only deinterlace frames marked as interlaced.

The default value is "all".

ccrepack 🎀

Repack CEA-708 closed captioning side data.

This filter fixes various issues seen with commercial encoders related to upstream malformed CEA-708 payloads, specifically incorrect number of tuples (wrong cc_count for the target FPS), and incorrect ordering of tuples (i.e., the CEA-608 tuples are not at the first entries in the payload).

cas ✨

Apply Contrast Adaptive Sharpen filter to video stream.

The filter accepts the following options:

strength — Set the sharpening strength. Default value is 0.
planes — Set planes to filter. Default value is to filter all planes except alpha plane.

Commands

This filter supports same commands as options.

chromahold 🌸

Remove all color information for all colors except for certain one.

The filter accepts the following options:

color — The color which will not be replaced with neutral chroma.
similarity — Similarity percentage with the above color. 0.01 matches only the exact key color, while 1.0 matches everything.
blend — Blend percentage. 0.0 makes pixels either fully gray, or not gray at all. Higher values result in more preserved color.
yuv — Signals that the color passed is already in YUV instead of RGB.

Literal colors like "green" or "red" don't make sense with this enabled anymore. This can be used to pass exact YUV values as hexadecimal numbers.

Commands

This filter supports same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value.

chromakey 💅

YUV colorspace color/chroma keying.

The filter accepts the following options:

color — The color which will be replaced with transparency.
similarity — Similarity percentage with the key color.

0.01 matches only the exact key color, while 1.0 matches everything.

blend — Blend percentage.

0.0 makes pixels either fully transparent, or not transparent at all.

Higher values result in semi-transparent pixels, with a higher transparency the more similar the pixels color is to the key color.

yuv — Signals that the color passed is already in YUV instead of RGB.

Literal colors like "green" or "red" don't make sense with this enabled anymore. This can be used to pass exact YUV values as hexadecimal numbers.

Commands

This filter supports same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value.

Examples

Make every green pixel in the input image transparent:

ffmpeg -i input.png -vf chromakey=green out.png

Overlay a greenscreen-video on top of a static black background.

ffmpeg -f lavfi -i color=c=black:s=1280x720 -i video.mp4 -shortest -filter_complex "[1:v]chromakey=0x70de77:0.1:0.2[ckout];[0:v][ckout]overlay[out]" -map "[out]" output.mkv

chromanr 🦋

Reduce chrominance noise.

The filter accepts the following options:

thres — Set threshold for averaging chrominance values. Sum of absolute difference of Y, U and V pixel components of current pixel and neighbour pixels lower than this threshold will be used in averaging. Luma component is left unchanged and is copied to output. Default value is 30. Allowed range is from 1 to 200.
sizew — Set horizontal radius of rectangle used for averaging. Allowed range is from 1 to 100. Default value is 5.
sizeh — Set vertical radius of rectangle used for averaging. Allowed range is from 1 to 100. Default value is 5.
stepw — Set horizontal step when averaging. Default value is 1. Allowed range is from 1 to 50. Mostly useful to speed-up filtering.
steph — Set vertical step when averaging. Default value is 1. Allowed range is from 1 to 50. Mostly useful to speed-up filtering.
threy — Set Y threshold for averaging chrominance values. Set finer control for max allowed difference between Y components of current pixel and neighbour pixels. Default value is 200. Allowed range is from 1 to 200.
threu — Set U threshold for averaging chrominance values. Set finer control for max allowed difference between U components of current pixel and neighbour pixels. Default value is 200. Allowed range is from 1 to 200.
threv — Set V threshold for averaging chrominance values. Set finer control for max allowed difference between V components of current pixel and neighbour pixels. Default value is 200. Allowed range is from 1 to 200.
distance — Set distance type used in calculations.
- manhattan — Absolute difference.
- euclidean — Difference squared.

Default distance type is manhattan.

Commands

This filter supports same commands as options. The command accepts the same syntax of the corresponding option.

chromashift 🧚

Shift chroma pixels horizontally and/or vertically.

The filter accepts the following options:

cbh — Set amount to shift chroma-blue horizontally.
cbv — Set amount to shift chroma-blue vertically.
crh — Set amount to shift chroma-red horizontally.
crv — Set amount to shift chroma-red vertically.
edge — Set edge mode, can be smear, default, or warp.

Commands

This filter supports the all above options as commands.

ciescope 🌈

Display CIE color diagram with pixels overlaid onto it.

The filter accepts the following options:

system — Set color system.
- ntsc, 470m
- ebu, 470bg
- smpte
- 240m
- apple
- widergb
- cie1931
- rec709, hdtv
- uhdtv, rec2020
- dcip3
cie — Set CIE system.
- xyy
- ucs
- luv
gamuts — Set what gamuts to draw.

See system option for available values.

size, s — Set ciescope size, by default set to 512.
intensity, i — Set intensity used to map input pixel values to CIE diagram.
contrast — Set contrast used to draw tongue colors that are out of active color system gamut.
corrgamma — Correct gamma displayed on scope, by default enabled.
showwhite — Show white point on CIE diagram, by default disabled.
gamma — Set input gamma. Used only with XYZ input color space.
fill — Fill with CIE colors. By default is enabled.

codecview 👑

Visualize information exported by some codecs.

Some codecs can export information through frames using side-data or other means. For example, some MPEG based codecs export motion vectors through the export_mvs flag in the codec flags2 option.

The filter accepts the following option:

block — Display block partition structure using the luma plane.
mv — Set motion vectors to visualize.

Available flags for mv are:

pf — forward predicted MVs of P-frames
bf — forward predicted MVs of B-frames
bb — backward predicted MVs of B-frames

qp — Display quantization parameters using the chroma planes.
mv_type, mvt — Set motion vectors type to visualize. Includes MVs from all frames unless specified by frame_type option.

Available flags for mv_type are:

fp — forward predicted MVs
bp — backward predicted MVs

frame_type, ft — Set frame type to visualize motion vectors of.

Available flags for frame_type are:

if — intra-coded frames (I-frames)
pf — predicted frames (P-frames)
bf — bi-directionally predicted frames (B-frames)

Examples

Visualize forward predicted MVs of all frames using ffplay:

ffplay -flags2 +export_mvs input.mp4 -vf codecview=mv_type=fp

Visualize multi-directionals MVs of P and B-Frames using ffplay:

ffplay -flags2 +export_mvs input.mp4 -vf codecview=mv=pf+bf+bb

colorbalance 💖

Modify intensity of primary colors (red, green and blue) of input frames.

The filter allows an input frame to be adjusted in the shadows, midtones or highlights regions for the red-cyan, green-magenta or blue-yellow balance.

A positive adjustment value shifts the balance towards the primary color, a negative value towards the complementary color.

The filter accepts the following options:

rs, gs, bs — Adjust red, green and blue shadows (darkest pixels).
rm, gm, bm — Adjust red, green and blue midtones (medium pixels).
rh, gh, bh — Adjust red, green and blue highlights (brightest pixels).

Allowed ranges for options are [-1.0, 1.0]. Defaults are 0.

pl — Preserve lightness when changing color balance. Default is disabled.

Examples

Add red color cast to shadows:

colorbalance=rs=.3

Commands

This filter supports the all above options as commands.

colorcontrast ⭐

Adjust color contrast between RGB components.

The filter accepts the following options:

rc — Set the red-cyan contrast. Defaults is 0.0. Allowed range is from -1.0 to 1.0.
gm — Set the green-magenta contrast. Defaults is 0.0. Allowed range is from -1.0 to 1.0.
by — Set the blue-yellow contrast. Defaults is 0.0. Allowed range is from -1.0 to 1.0.
rcw, gmw, byw — Set the weight of each rc, gm, by option value. Default value is 0.0. Allowed range is from 0.0 to 1.0. If all weights are 0.0 filtering is disabled.
pl — Set the amount of preserving lightness. Default value is 0.0. Allowed range is from 0.0 to 1.0.

Commands

This filter supports the all above options as commands.

colorcorrect �

Adjust color white balance selectively for blacks and whites. This filter operates in YUV colorspace.

The filter accepts the following options:

rl — Set the red shadow spot. Allowed range is from -1.0 to 1.0. Default value is 0.
bl — Set the blue shadow spot. Allowed range is from -1.0 to 1.0. Default value is 0.
rh — Set the red highlight spot. Allowed range is from -1.0 to 1.0. Default value is 0.
bh — Set the blue highlight spot. Allowed range is from -1.0 to 1.0. Default value is 0.
saturation — Set the amount of saturation. Allowed range is from -3.0 to 3.0. Default value is 1.
analyze — If set to anything other than manual it will analyze every frame and use derived parameters for filtering output frame.

Possible values are:

manual
average
minmax
median

Default value is "manual".

Commands

This filter supports the all above options as commands.

💅 colorchannelmixer

Okay bestie, this one remixes your color channels — like a DJ but for pixels. It adjusts video input frames by re-mixing color channels, and what it does is it modifies a color channel by adding up the values from the other channels of that same pixel. So fr fr, if the value you're modifying is red, your output value is gonna be:

<red>=<red>*<rr> + <blue>*<rb> + <green>*<rg> + <alpha>*<ra>

The filter accepts these options:

rr, rg, rb, ra — adjust how much the input red, green, blue and alpha channels contribute to the output red channel. Default is 1 for rr, and 0 for rg, rb and ra.
gr, gg, gb, ga — same deal but for the output green channel. Default is 1 for gg, and 0 for gr, gb and ga.
br, bg, bb, ba — and for the output blue channel. Default is 1 for bb, and 0 for br, bg and ba.
ar, ag, ab, aa — and finally the output alpha channel. Default is 1 for aa, and 0 for ar, ag and ab.

The allowed ranges for these options are [-2.0, 2.0]. Yes, negative. It's giving chaos but it's legal chaos 💀

pc — set the preserve color mode. Accepted values are:
- none — disable color preserving, this is the default.
- lum — preserve luminance.
- max — preserve max value of RGB triplet.
- avg — preserve average value of RGB triplet.
- sum — preserve sum value of RGB triplet.
- nrm — preserve normalized value of RGB triplet.
- pwr — preserve power value of RGB triplet.
pa — set the preserve color amount when changing colors. Allowed range is [0.0, 1.0]. Default is 0.0, so it's disabled by default, no cap.

✨ Examples

Convert source to grayscale — the classic glow-down:

colorchannelmixer=.3:.4:.3:0:.3:.4:.3:0:.3:.4:.3

Simulate sepia tones — vintage filter but make it math:

colorchannelmixer=.393:.769:.189:0:.349:.686:.168:0:.272:.534:.131

🎀 Commands

This filter supports all the above options as commands. Periodt.

🦋 colordetect

This one's the detective 🔍 — it analyzes the video frames to determine the effective value range and alpha mode. It accepts the following options:

mode — the set of properties to detect. Unavailable properties, like alpha mode on an input image with no alpha channel, get ignored automatically (it's not gonna make stuff up, king behavior).

Accepts a combination of these flags:

color_range — detect if the source contains luma pixels outside the limited (MPEG) range, which indicates that this is a full range YUV source.
alpha_mode — detect if the source contains color values above the alpha channel, which indicates the alpha channel is independent (straight) rather than premultiplied. Also detects if the alpha plane is fully opaque or not.
all — enable detection of all of the above properties. This is the default. 👑

🌸 colorize

Overlay a solid color on the video stream. Simple, iconic, does what it says.

hue — set the color hue. Allowed range is from 0 to 360. Default value is 0.
saturation — set the color saturation. Allowed range is from 0 to 1. Default value is 0.5.
lightness — set the color lightness. Allowed range is from 0 to 1. Default value is 0.5.
mix — set the mix of source lightness. By default it's set to 1.0. Allowed range is from 0.0 to 1.0.

🎀 Commands

This filter supports all the above options as commands. Slay.

💜 colorkey

RGB colorspace color keying, aka greenscreen vibes 🟩. This filter operates on 8-bit RGB format frames by setting the alpha component of each pixel which falls within the similarity radius of the key color to 0. The alpha value for pixels outside the similarity radius depends on the value of the blend option.

color — set the color for which alpha will be set to 0 (full transparency). See the Color section in the ffmpeg-utils manual. Default is black.
similarity — set the radius from the key color within which other colors also have full transparency. The computed distance is related to the unit fractional distance in 3D space between the RGB values of the key color and the pixel's color. Range is 0.01 to 1.0. 0.01 matches within a very small radius around the exact key color, while 1.0 matches everything. Default is 0.01.
blend — set how the alpha value for pixels that fall outside the similarity radius is computed. 0.0 makes pixels either fully transparent or fully opaque. Higher values result in semi-transparent pixels, with greater transparency the more similar the pixel color is to the key color. Range is 0.0 to 1.0. Default is 0.0.

✨ Examples

Make every green pixel in the input image transparent:

ffmpeg -i input.png -vf colorkey=green out.png

Overlay a greenscreen-video on top of a static background image:

ffmpeg -i background.png -i video.mp4 -filter_complex "[1:v]colorkey=0x3BBD1E:0.3:0.2[ckout];[0:v][ckout]overlay[out]" -map "[out]" output.flv

🎀 Commands

This filter supports same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value. No crashes, no drama 🙈.

🌈 colorhold

Removes all color information for all RGB colors except for certain one. Like keeping exactly one friend and ghosting the rest 😭.

color — the color which will not be replaced with neutral gray.
similarity — similarity percentage with the above color. 0.01 matches only the exact key color, while 1.0 matches everything.
blend — blend percentage. 0.0 makes pixels fully gray. Higher values result in more preserved color.

🎀 Commands

This filter supports same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value.

⭐ colorlevels

Adjust video input frames using levels. Very photo-editor-core of it.

rimin, gimin, bimin, aimin — adjust red, green, blue and alpha input black point. Allowed ranges for options are [-1.0, 1.0]. Defaults are 0.
rimax, gimax, bimax, aimax — adjust red, green, blue and alpha input white point. Allowed ranges for options are [-1.0, 1.0]. Defaults are 1.

Input levels are used to lighten highlights (bright tones), darken shadows (dark tones), change the balance of bright and dark tones. It's giving lighting design 💡

romin, gomin, bomin, aomin — adjust red, green, blue and alpha output black point. Allowed ranges for options are [0, 1.0]. Defaults are 0.
romax, gomax, bomax, aomax — adjust red, green, blue and alpha output white point. Allowed ranges for options are [0, 1.0]. Defaults are 1.

Output levels allows manual selection of a constrained output level range. Main character energy for your histogram 🫶

preserve — set preserve color mode. The accepted values are:
- none — disable color preserving, this is the default.
- lum — preserve luminance.
- max — preserve max value of RGB triplet.
- avg — preserve average value of RGB triplet.
- sum — preserve sum value of RGB triplet.
- nrm — preserve normalized value of RGB triplet.
- pwr — preserve power value of RGB triplet.

✨ Examples

Make video output darker:

colorlevels=rimin=0.058:gimin=0.058:bimin=0.058

Increase contrast:

colorlevels=rimin=0.039:gimin=0.039:bimin=0.039:rimax=0.96:gimax=0.96:bimax=0.96

Make video output lighter:

colorlevels=rimax=0.902:gimax=0.902:bimax=0.902

Increase brightness:

colorlevels=romin=0.5:gomin=0.5:bomin=0.5

🎀 Commands

This filter supports all the above options as commands. Obsessed.

🧚 colormap

Apply custom color maps to video stream. This one's a little extra because it needs three input video streams:

First stream is the video stream that is going to be filtered out.
Second and third video stream specify color patches for source color to target color mapping.

patch_size — set the source and target video stream patch size in pixels.
nb_patches — set the max number of used patches from source and target video stream. Default value is number of patches available in additional video streams. Max allowed number of patches is 64.
type — set the adjustments used for target colors. Can be relative or absolute. Defaults is absolute.
kernel — set the kernel used to measure color differences between mapped colors. The accepted values are:
- euclidean
- weuclidean

Default is euclidean.

💀 colormatrix

Convert color matrix. Short king of a filter.

src, dst — specify the source and destination color matrix. Both values must be specified.

The accepted values are:

bt709 — BT.709
fcc — FCC
bt601 — BT.601
bt470 — BT.470
bt470bg — BT.470BG
smpte170m — SMPTE-170M
smpte240m — SMPTE-240M
bt2020 — BT.2020

For example to convert from BT.601 to SMPTE-240M, use the command:

colormatrix=bt601:smpte240m

🐛 colorspace

Convert colorspace, transfer characteristics or color primaries. Heads up bestie: input video needs to have an even size. Not optional, that's the vibe check.

all — specify all color properties at once. The accepted values are:
- bt470m — BT.470M
- bt470bg — BT.470BG
- bt601-6-525 — BT.601-6 525
- bt601-6-625 — BT.601-6 625
- bt709 — BT.709
- smpte170m — SMPTE-170M
- smpte240m — SMPTE-240M
- bt2020 — BT.2020
space — specify output colorspace. The accepted values are:
- bt709 — BT.709
- fcc — FCC
- bt470bg — BT.470BG or BT.601-6 625
- smpte170m — SMPTE-170M or BT.601-6 525
- smpte240m — SMPTE-240M
- ycgco — YCgCo
- bt2020ncl — BT.2020 with non-constant luminance
trc — specify output transfer characteristics. The accepted values are:
- bt709 — BT.709
- bt470m — BT.470M
- bt470bg — BT.470BG
- gamma22 — Constant gamma of 2.2
- gamma28 — Constant gamma of 2.8
- smpte170m — SMPTE-170M, BT.601-6 625 or BT.601-6 525
- smpte240m — SMPTE-240M
- srgb — SRGB
iec61966-2-1 — iec61966-2-1 💅
iec61966-2-4 — iec61966-2-4
xvycc — xvycc
bt2020-10 — BT.2020 for 10-bits content
bt2020-12 — BT.2020 for 12-bits content
primaries — sets the output color primaries, bestie 🎨

These are the values you can pick from:

bt709 — BT.709
bt470m — BT.470M
bt470bg — BT.470BG or BT.601-6 625
smpte170m — SMPTE-170M or BT.601-6 525
smpte240m — SMPTE-240M
film — film
smpte431 — SMPTE-431
smpte432 — SMPTE-432
bt2020 — BT.2020
jedec-p22 — JEDEC P22 phosphors

range — sets the output color range 🌈

Accepted values, go off:

tv — TV (restricted) range
mpeg — MPEG (restricted) range
pc — PC (full) range
jpeg — JPEG (full) range

format — sets the output color format, no cap 💖

Accepted values:

yuv420p — YUV 4:2:0 planar 8-bits
yuv420p10 — YUV 4:2:0 planar 10-bits
yuv420p12 — YUV 4:2:0 planar 12-bits
yuv422p — YUV 4:2:2 planar 8-bits
yuv422p10 — YUV 4:2:2 planar 10-bits
yuv422p12 — YUV 4:2:2 planar 12-bits
yuv444p — YUV 4:4:4 planar 8-bits
yuv444p10 — YUV 4:4:4 planar 10-bits
yuv444p12 — YUV 4:4:4 planar 12-bits

fast — does a fast conversion, which skips gamma/primary correction ⚡ This one takes significantly less CPU, but lowkey it's gonna be mathematically incorrect. If you want output compatible with whatever the colormatrix filter produces, use fast=1. iykyk.
dither — picks your dithering mode ✨

Accepted values:

none — no dithering
fsb — Floyd-Steinberg dithering

wpadapt — whitepoint adaptation mode 🎀

Accepted values:

bradford — Bradford whitepoint adaptation
vonkries — von Kries whitepoint adaptation
identity — identity whitepoint adaptation (i.e. no whitepoint adaptation)

clipgamut — controls how to clip out-of-gamut colors that pop up as a result of the colorspace conversion 🦋

Accepted values:

none — no clipping of out of gamut colors
rgb — clips the RGB values to the [0, 1] range when building the gamma transfer LUTs

iall — overrides all input properties at once. Same accepted values as all. 💀
ispace — overrides the input colorspace. Same accepted values as space.
iprimaries — overrides the input color primaries. Same accepted values as primaries.
itrc — overrides the input transfer characteristics. Same accepted values as trc.
irange — overrides the input color range. Same accepted values as range.

Okay so here's the tea on what this filter actually does 🫖 it converts the transfer characteristics, color space and color primaries to the values you specified. The output value, if you didn't specify one, gets set to a default based on the all property. And if that property isn't specified either, the filter will log an error 😭. The output color range and format default to the same value as the input color range and format. Your input transfer characteristics, color space, color primaries and color range should be set on the input data — if any of these are missing, the filter logs an error and no conversion will take place. Like, none. Zero. 🙈

So for example, to convert the input to SMPTE-240M, you'd run:

colorspace=smpte240m

colortemperature 🌡️

Adjusts the color temperature in video to simulate variations in ambient color temperature. It's giving cozy lamp lighting, honestly. 🕯️

This filter accepts the following options:

temperature — sets the temperature in Kelvin. Allowed range is from 1000 to 40000. Default value is 6500 K. 👑
mix — sets mixing with filtered output. Allowed range is from 0 to 1. Default value is 1.
pl — sets the amount of preserving lightness. Allowed range is from 0 to 1. Default value is 0.

Commands

This filter supports the same commands as options. periodt. 💅

convolution 🌀

Applies convolution of 3x3, 5x5, 7x7 or horizontal/vertical up to 49 elements. Math but make it fashion.

This filter accepts the following options:

0m, 1m, 2m, 3m — sets the matrix for each plane. The matrix is a sequence of 9, 25 or 49 signed integers in square mode, and from 1 to 49 odd number of signed integers in row mode. 🔢
0rdiv, 1rdiv, 2rdiv, 3rdiv — sets the multiplier for the calculated value for each plane. If unset or 0, it will be 1/sum of all matrix elements.
0bias, 1bias, 2bias, 3bias — sets the bias for each plane. This value is added to the result of the multiplication. Super useful for making the overall image brighter or darker. Default is 0.0. ✨
0mode, 1mode, 2mode, 3mode — sets the matrix mode for each plane. Can be square, row or column. Default is square.

Commands

This filter supports all of the above options as commands 🫶

Examples

Apply sharpen:

convolution="0 -1 0 -1 5 -1 0 -1 0:0 -1 0 -1 5 -1 0 -1 0:0 -1 0 -1 5 -1 0 -1 0:0 -1 0 -1 5 -1 0 -1 0"

Apply blur:

convolution="1 1 1 1 1 1 1 1 1:1 1 1 1 1 1 1 1 1:1 1 1 1 1 1 1 1 1:1 1 1 1 1 1 1 1 1:1/9:1/9:1/9:1/9"

Apply edge enhance:

convolution="0 0 0 -1 1 0 0 0 0:0 0 0 -1 1 0 0 0 0:0 0 0 -1 1 0 0 0 0:0 0 0 -1 1 0 0 0 0:5:1:1:1:0:128:128:128"

Apply edge detect:

convolution="0 1 0 1 -4 1 0 1 0:0 1 0 1 -4 1 0 1 0:0 1 0 1 -4 1 0 1 0:0 1 0 1 -4 1 0 1 0:5:5:5:1:0:128:128:128"

Apply laplacian edge detector which includes diagonals:

convolution="1 1 1 1 -8 1 1 1 1:1 1 1 1 -8 1 1 1 1:1 1 1 1 -8 1 1 1 1:1 1 1 1 -8 1 1 1 1:5:5:5:1:0:128:128:0"

Apply emboss:

convolution="-2 -1 0 -1 1 1 0 1 2:-2 -1 0 -1 1 1 0 1 2:-2 -1 0 -1 1 1 0 1 2:-2 -1 0 -1 1 1 0 1 2"

convolve 🦋

Applies a 2D convolution of the video stream in the frequency domain, using a second stream as the impulse. Fancy fr fr.

This filter accepts the following options:

planes — sets which planes to process.
impulse — sets which impulse video frames will be processed. Can be first or all. Default is all.

The convolve filter also supports the framesync options 💜

copy 📋

Copies the input video source unchanged to the output. Mainly useful for testing purposes — it's literally the "no thoughts, head empty" filter and I respect it.

coreimage 🍎

Video filtering on GPU using Apple's CoreImage API on OSX. It's giving main character energy, but only on a Mac.

Hardware acceleration is based on an OpenGL context. Usually this means it's processed by video hardware. However, software-based OpenGL implementations exist too, which means there's no guarantee of hardware processing — it depends on the respective OSX. 😭

Apple ships a ton of filters and image generators with a huge variety of options. You reference the filter by its name along with its options.

The coreimage filter accepts the following options:

list_filters — lists all available filters and generators along with all their respective options, plus possible minimum and maximum values along with the default values. 👑

list_filters=true

filter — specifies all filters by their respective name and options. Use list_filters to figure out all valid filter names and options. Numerical options are specified by a float value and are automatically clamped to their respective value range. Vector and color options have to be specified by a list of space separated float values. Character escaping has to be done. A special option name default is available to use default options for a filter. ✨

It is required to specify either default or at least one of the filter options. All omitted options are used with their default values. The syntax of the filter string is as follows:

filter=<NAME>@<OPTION>=<VALUE>[@<OPTION>=<VALUE>][@...][#<NAME>@<OPTION>=<VALUE>[@<OPTION>=<VALUE>][@...]][#...]

output_rect — specifies a rectangle where the output of the filter chain is copied into the input image. It's given by a list of space separated float values:

output_rect=x\ y\ width\ height

If not given, the output rectangle equals the dimensions of the input image. The output rectangle is automatically cropped at the borders of the input image. Negative values are valid for each component. 🙈

output_rect=25\ 25\ 100\ 100

Several filters can be chained for successive processing without GPU-HOST transfers, which is what allows fast processing of complex filter chains. Currently, only filters with zero (generators) or exactly one (filters) input image and one output image are supported. Also, transition filters are not yet usable as intended. 💀

Some filters generate output images with additional padding depending on the respective filter kernel. That padding is automatically removed to make sure the filter output has the same size as the input image. 🎀

For image generators, the size of the output image is determined by the previous output image of the filter chain or the input image of the whole filterchain, respectively. The generators don't use the pixel information of this image to generate their output. However, the generated output is blended onto this image, resulting in partial or complete coverage of the output image. 🦋

The coreimagesrc video source can be used for generating input images which are fed directly into the filter chain. By using it, you don't need to provide input images by another video source or an input video at all. Slay.

Examples

List all filters available:

coreimage=list_filters=true

Use the CIBoxBlur filter with default options to blur an image:

coreimage=filter=CIBoxBlur@default

Use a filter chain with CISepiaTone at default values and CIVignetteEffect with its center at 100x100 and a radius of 50 pixels:

coreimage=filter=CIBoxBlur@default#CIVignetteEffect@inputCenter=100\ 100@inputRadius=50

Use nullsrc and CIQRCodeGenerator to create a QR code for the FFmpeg homepage, given as complete and escaped command-line for Apple's standard bash shell:

ffmpeg -f lavfi -i nullsrc=s=100x100,coreimage=filter=CIQRCodeGenerator@inputMessage=https\\\\\://FFmpeg.org/@inputCorrectionLevel=H -frames:v 1 QRCode.png

corr 📊

Obtains the correlation between two input videos. Bestie, it literally just compares them.

This filter takes two input videos. 🫶

Both input videos have to match in resolution and pixel format or this filter just won't do its thing correctly, bestie. It also assumes both inputs carry the same number of frames, and it compares them one by one. 💅

The per-component, average, min and max correlation it comes up with gets printed through the logging system — yes, through the logs, keep an eye on stderr.

It also stashes the calculated correlation of each frame into frame metadata ✨

Oh and this filter supports the framesync options too 🫶

Down here the input file main.mpg gets compared against the reference file ref.mpg:

ffmpeg -i main.mpg -i ref.mpg -lavfi corr -f null -

cover_rect

Covers up a rectangular object 💅

It takes these options:

cover — filepath of the optional cover image. Heads up: it needs to be in yuv420.
mode — sets the covering mode. The values it accepts are:
- cover — cover it by the supplied image
- blur — cover it by interpolating the surrounding pixels

Default value is blur.

Examples

Cover a rectangular object using the supplied image of a given video, with ffmpeg:

ffmpeg -i file.ts -vf find_rect=newref.pgm,cover_rect=cover.jpg:mode=cover new.mkv

crop

Crops the input video down to whatever dimensions you want ✂️

It takes these parameters:

w, out_w — the width of the output video. Defaults to iw. This expression gets evaluated only once, during the filter configuration, or when the w or out_w command is sent.
h, out_h — the height of the output video. Defaults to ih. Same deal: evaluated only once during filter config, or when the h or out_h command is sent.
x — the horizontal position, in the input video, of the left edge of the output video. Defaults to (in_w-out_w)/2. This one is evaluated per-frame.
y — the vertical position, in the input video, of the top edge of the output video. Defaults to (in_h-out_h)/2. Also evaluated per-frame.
keep_aspect — if you set it to 1 it forces the output display aspect ratio to be the same as the input's, by changing the output sample aspect ratio. Defaults to 0.
exact — enables exact cropping. When it's on, subsampled videos get cropped at the exact width/height/x/y you specified and will not be rounded down to the nearest smaller value. Defaults to 0.

The out_w, out_h, x, y parameters are expressions that can contain these constants:

x, y — the computed values for x and y. They're evaluated for each new frame.
in_w, in_h — the input width and height.
iw, ih — these are the same as in_w and in_h.
out_w, out_h — the output (cropped) width and height.
ow, oh — these are the same as out_w and out_h.
a — same as iw / ih
sar — input sample aspect ratio
dar — input display aspect ratio, it's the same as (iw / ih) * sar
hsub, vsub — horizontal and vertical chroma subsample values. For example, for the pixel format yuv422p, hsub is 2 and vsub is 1.
n — the number of the input frame, starting from 0.
pos — the position in the file of the input frame, NAN if unknown; deprecated, do not use 🙅‍♀️
t — the timestamp expressed in seconds. It's NAN if the input timestamp is unknown.

The expression for out_w may depend on the value of out_h, and the expression for out_h may depend on out_w — but they cannot depend on x and y, because x and y get evaluated after out_w and out_h. Order matters, bestie 💅

The x and y parameters specify the expressions for the position of the top-left corner of the output (non-cropped) area. They're evaluated for each frame. If the evaluated value isn't valid, it gets approximated to the nearest valid value.

The expression for x may depend on y, and the expression for y may depend on x. Mutual dependency, iconic.

Examples

Crop an area of size 100x100 at position (12,34):

crop=100:100:12:34

Using named options, the example above becomes:

crop=w=100:h=100:x=12:y=34

Crop the central input area with size 100x100:

crop=100:100

Crop the central input area with size 2/3 of the input video:

crop=2/3*in_w:2/3*in_h

Crop the input video central square:

crop=out_w=in_h
crop=in_h

Delimit the rectangle with the top-left corner placed at position 100:100 and the right-bottom corner corresponding to the right-bottom corner of the input image:

crop=in_w-100:in_h-100:100:100

Crop 10 pixels off the left and right borders, and 20 pixels off the top and bottom borders:

crop=in_w-2*10:in_h-2*20

Keep only the bottom right quarter of the input image:

crop=in_w/2:in_h/2:in_w/2:in_h/2

Crop height for getting Greek harmony 🇬🇷:

crop=in_w:1/PHI*in_w

Apply a trembling effect:

crop=in_w/2:in_h/2:(in_w-out_w)/2+((in_w-out_w)/2)*sin(n/10):(in_h-out_h)/2 +((in_h-out_h)/2)*sin(n/7)

Apply an erratic camera effect depending on timestamp:

crop=in_w/2:in_h/2:(in_w-out_w)/2+((in_w-out_w)/2)*sin(t*10):(in_h-out_h)/2 +((in_h-out_h)/2)*sin(t*13)

Set x depending on the value of y:

crop=in_w/2:in_h/2:y:10+10*sin(n/10)

Commands

This filter supports the following commands:

w, out_w, h, out_h, x, y — set the width/height of the output video and the horizontal/vertical position in the input video. The command accepts the same syntax as the corresponding option.

If the specified expression isn't valid, it stays at its current value. No crashing, no drama 💅

cropdetect

Auto-detects the crop size 🔍✨

It calculates the cropping parameters you need and prints the recommended ones through the logging system. The detected dimensions correspond to the non-black / video area of the input video, according to mode.

It takes these parameters:

mode — depending on the mode, crop detection is based on either the mere black value of surrounding pixels, or a combination of motion vectors and edge pixels.
- black — detect black pixels surrounding the playing video. For fine control use the limit option.
- mvedges — detect the playing video by the motion vectors inside the video, and scan for edge pixels typically forming the border of a playing video.
limit — sets a higher black value threshold, which you can optionally specify anywhere from nothing (0) to everything (255 for 8-bit based formats). Any intensity value greater than the set value is considered non-black. Defaults to 24. You can also give a value between 0.0 and 1.0, which gets scaled depending on the bitdepth of the pixel format.
round — the value the width/height should be divisible by. Defaults to 16. The offset is automatically adjusted to center the video. Use 2 to get only even dimensions (needed for 4:2:2 video). 16 is best when encoding to most video codecs 👑
skip — sets the number of initial frames for which evaluation is skipped. Default is 2. Range is 0 to INT_MAX.
reset_count, reset — sets the counter that determines after how many frames cropdetect will reset the previously detected largest video area and start over to detect the current optimal crop area. Default value is 0. This can be handy when channel logos distort the video area. 0 means 'never reset', and returns the largest area encountered during playback.
mv_threshold — sets motion in pixel units as the threshold for motion detection. Defaults to 8.
low, high — set the low and high threshold values used by the Canny thresholding algorithm. The high threshold selects the "strong" edge pixels, which then get connected through 8-connectivity with the "weak" edge pixels selected by the low threshold. low and high threshold values must be chosen in the range [0,1], and low should be lesser or equal to high. Default value for low is 5/255, and default value for high is 15/255.

Examples

Find the video area surrounded by black borders:

ffmpeg -i file.mp4 -vf cropdetect,metadata=mode=print -f null -

Find an embedded video area, generating motion vectors beforehand:

ffmpeg -i file.mp4 -vf mestimate,cropdetect=mode=mvedges,metadata=mode=print -f null -

Find an embedded video area, using motion vectors from the decoder:

ffmpeg -flags2 +export_mvs -i file.mp4 -vf cropdetect=mode=mvedges,metadata=mode=print -f null -

Commands

This filter supports the following commands:

limit — the command accepts the same syntax as the corresponding option.

If the specified expression isn't valid, it's kept at its current value. Same energy as before 💅

cue

Delays video filtering until a given wallclock timestamp ⏰

Here's the vibe: the filter first passes on preroll amount of frames, then it buffers at most buffer amount of frames and waits for the cue. After reaching the cue, it forwards the buffered frames and also any subsequent frames coming in on its input.

You'd use this to synchronize the output of multiple ffmpeg processes for realtime output devices like decklink. By putting the delay in the filtering chain and pre-buffering frames, the process can pass data on to the output almost immediately after the target wallclock timestamp is reached. Main character scheduling energy 👑

Perfect frame accuracy cannot be guaranteed — but the result is good enough for some use cases. Fr fr, no cap.

cue — the cue timestamp, expressed in a UNIX timestamp in microseconds. Default is 0.
preroll — the duration of content to pass on as preroll, expressed in seconds. Default is 0.
buffer — the maximum duration of content to buffer before waiting for the cue, expressed in seconds. Default is 0.

curves

Applies colour adjustments using curves 🎨🦋

This filter is similar to the Adobe Photoshop and GIMP curves tools. Each component (red, green and blue) has its values defined by N key points, tied to each other using a smooth curve. The x-axis represents the pixel values from the input frame, and the y-axis represents the new pixel values to be set for the output frame.

By default, a component curve is defined by the two points (0;0) and (1;1). That makes a straight line where each original pixel value is "adjusted" to its own value, which means... no change to the image at all. Literally nothing happens, it's giving "I'm just here for the vibes" 💀

Okay so this filter lets you redefine those two points and add some more ✨. A brand-new curve gets defined to pass all smooth and slay through every one of these new coordinates. The new points have to be strictly increasing along the x-axis, and their x and y values must sit in the [0;1] interval — no exceptions, bestie. The curve gets built using either a natural or a monotonic cubic spline interpolation, depending on the interp option (default: "natural"). The "natural" spline gives you a smoother curve in general, while the monotonic ("pchip") spline guarantees the transitions between the points you specified are monotonic. And if the computed curves happen to wander outside the vector spaces? The values get clipped accordingly. It's giving responsible adult behavior 💅.

The filter accepts the following options:

preset — pick one of the available color presets 🌈. You can use this in addition to the r, g, b parameters — and when you do, the later options take priority over the preset values, so the preset is more of a vibe suggestion. The available presets are:
- none
- color_negative
- cross_process
- darker
- increase_contrast
- lighter
- linear_contrast
- medium_contrast
- negative
- strong_contrast
- vintage

Default is "none".

master, m — set the master key points 👑. These bad boys define a second pass mapping. People sometimes call it a "luminance" or "value" mapping. It works great with r, g, b or all, since it acts like a post-processing LUT.
red, r — set the key points for the red component 🍓.
green, g — set the key points for the green component 🧚.
blue, b — set the key points for the blue component 🦋.
all — set the key points for all components (master is not included, fyi). Can be used in addition to the other per-component key points options — and in that case, the unset component(s) will fallback on this all setting. Iconic fallback behavior, honestly.
psfile — specify a Photoshop curves file (.acv) to import the settings from 🎀.
plot — save a Gnuplot script of the curves in the specified file.
interp — specify the kind of interpolation. Available algorithms are:
- natural — natural cubic spline using a piece-wise cubic polynomial that is twice continuously differentiable.
- pchip — monotonic cubic spline using a piecewise cubic Hermite interpolating polynomial (PCHIP).

To dodge some filtergraph syntax conflicts 🙈, each key points list needs to be defined using this exact syntax: "x0/y0 x1/y1 x2/y2 ...".

Commands

This filter supports same commands as options. Short and sweet, we love that for us 🫶.

Examples

Increase slightly the middle level of blue 💙:

curves=blue='0/0 0.5/0.58 1/1'

Vintage effect ✨:

curves=r='0/0.11 .42/.51 1/0.95':g='0/0 0.50/0.48 1/1':b='0/0.22 .49/.44 1/0.8'

Here we obtain the following coordinates for each component:

red "(0;0.11) (0.42;0.51) (1;0.95)"

green "(0;0) (0.50;0.48) (1;1)"

blue "(0;0.22) (0.49;0.44) (1;0.80)"

The previous example can also be achieved with the associated built-in preset 🎀:

curves=preset=vintage

Or simply, because we're efficient like that:

curves=vintage

Use a Photoshop preset and redefine the points of the green component 💚:

curves=psfile='MyCurvesPresets/purple.acv':green='0/0 0.45/0.53 1/1'

Check out the curves of the "cross_process" profile using ffmpeg and gnuplot 🔥:

ffmpeg -f lavfi -i color -vf curves=cross_process:plot=/tmp/curves.plt -frames:v 1 -f null -
gnuplot -p /tmp/curves.plt

datascope

Video data analysis filter, aka the let me see the raw truth filter. This filter shows hexadecimal pixel values of part of video 👀.

The filter accepts the following options:

size, s — set output video size.
x — set x offset from where to pick pixels.
y — set y offset from where to pick pixels.
mode — set scope mode, can be one of the following:
- mono — draw hexadecimal pixel values with white color on black background.
- color — draw hexadecimal pixel values with input video pixel color on black background.
- color2 — draw hexadecimal pixel values on color background picked from input video, and the text color is picked in such a way so it's always visible. Accessibility king 👑.
axis — draw rows and columns numbers on left and top of video.
opacity — set background opacity.
format — set display number format. Can be "hex", or "dec". Default is "hex".
components — set pixel components to display. By default all pixel components are displayed.

Commands

This filter supports same commands as options excluding the size option ⚠️.

dblur

Apply Directional blur filter 🌫️.

The filter accepts the following options:

angle — set angle of directional blur. Default is 45.
radius — set radius of directional blur. Default is 5.
planes — set which planes to filter. By default all planes are filtered.

Commands

This filter supports same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value — no drama, no crash 💅.

dctdnoiz

Denoise frames using 2D DCT (frequency domain filtering) 🎛️.

This filter is not designed for real time. Like, respectfully, do not.

The filter accepts the following options:

sigma, s — set the noise sigma constant. This sigma defines a hard threshold of 3 * sigma; every DCT coefficient (absolute value) below this threshold will be dropped 🗑️. If you need a more advanced filtering, see expr. Default is 0.
overlap — set number overlapping pixels for each block. Since the filter can be slow, you may want to reduce this value — at the cost of a less effective filter and the risk of various artefacts 😬. If the overlapping value doesn't permit processing the whole input width or height, a warning will be displayed and the according borders won't be denoised. Default value is blocksize-1, which is the best possible setting ✨.
expr, e — set the coefficient factor expression. For each coefficient of a DCT block, this expression will be evaluated as a multiplier value for the coefficient. If this option is set, the sigma option will be ignored. The absolute value of the coefficient can be accessed through the c variable.
n — set the blocksize using the number of bits. 1<<n defines the blocksize, which is the width and height of the processed blocks. The default value is 3 (8x8) and can be raised to 4 for a blocksize of 16x16. Note that changing this setting has huge consequences on the speed processing 👀. Also, a larger block size does not necessarily mean a better de-noising.

Examples

Apply a denoise with a sigma of 4.5:

dctdnoiz=4.5

The same operation can be achieved using the expression system:

dctdnoiz=e='gte(c, 4.5*3)'

Violent denoise using a block size of "16x16" 💀:

dctdnoiz=15:n=4

deband

Remove banding artifacts from input video. It works by replacing banded pixels with average value of referenced pixels 🩹.

The filter accepts the following options:

1thr
2thr
3thr
4thr

Set banding detection threshold for each plane. Default is 0.02. Valid range is 0.00003 to 0.5. If the difference between current pixel and reference pixel is less than threshold, it will be considered as banded.

range, r — banding detection range in pixels. Default is 16. If positive, a random number in range 0 to the set value will be used. If negative, the exact absolute value will be used. The range defines square of four pixels around current pixel.
direction, d — set direction in radians from which four pixels will be compared. If positive, a random direction from 0 to set direction will be picked. If negative, exact of absolute value will be picked. For example direction 0, -PI or -2*PI radians will pick only pixels on same row and -PI/2 will pick only pixels on same column 📐.
blur, b — if enabled, current pixel is compared with average value of all four surrounding pixels. The default is enabled. If disabled, current pixel is compared with all four surrounding pixels. The pixel is considered banded if only all four differences with surrounding pixels are less than threshold.
coupling, c — if enabled, current pixel is changed if and only if all pixel components are banded, e.g. banding detection threshold is triggered for all color components. The default is disabled.

Commands

This filter supports the all above options as commands.

deblock

Remove blocking artifacts from input video 🧱➡️✨.

The filter accepts the following options:

filter — set filter type, can be weak or strong. Default is strong. This controls what kind of deblocking is applied.
block — set size of block, allowed range is from 4 to 512. Default is 8.
alpha
beta
gamma
delta

Set blocking detection thresholds. Allowed range is 0 to 1. Defaults are: 0.098 for alpha and 0.05 for the rest. Using a higher threshold gives more deblocking strength. Setting alpha controls threshold detection at exact edge of block. The remaining options control threshold detection near the edge — each one for below/above or left/right. Setting any of those to 0 disables deblocking.

planes — set planes to filter. Default is to filter all available planes.

Examples

Deblock using weak filter and block size of 4 pixels:

deblock=filter=weak:block=4

Deblock using strong filter, block size of 4 pixels and custom thresholds for deblocking more edges:

deblock=filter=strong:block=4:alpha=0.12:beta=0.07:gamma=0.06:delta=0.05

Similar as above, but filter only first plane:

deblock=filter=strong:block=4:alpha=0.12:beta=0.07:gamma=0.06:delta=0.05:planes=1

Similar as above, but filter only second and third plane:

deblock=filter=strong:block=4:alpha=0.12:beta=0.07:gamma=0.06:delta=0.05:planes=6

Commands

This filter supports the all above options as commands.

decimate

Drop duplicated frames at regular intervals 🗑️🎞️.

The filter accepts the following options:

cycle — set the number of frames from which one will be dropped. Setting this to N means one frame in every batch of N frames will be dropped. Default is 5.
dupthresh — sets the duplicate-detection threshold, bestie. if the difference metric for a frame comes out less than or equal to this value, that frame gets declared a duplicate. default is 1.1.
scthresh — sets the scene change threshold. default is 15.
blockx, blocky — the size of the x-axis and y-axis blocks used while the metrics are being calculated. bigger blocks = better noise suppression, but also = worse detection of small movements, so pick your fighter 💅. must be a power of two. default is 32.
ppsrc — marks the main input as a pre-processed input and activates the clean source input stream. this is the whole slay: it lets you pre-process the input with whatever filters you want to help the metrics calculation, while keeping the frame selection lossless. when it's set to 1, the first stream is the pre-processed input and the second stream is the clean source that the kept frames are actually chosen from. default is 0.
chroma — sets whether chroma is considered in the metric calculations or not. default is 1.
mixed — sets whether the input only partially contains content to be decimated. default is "false". if you enable it, the video output stream will be in variable frame rate. lowkey important, don't skip it 🫶.

✨ `deconvolve` ✨

applies a 2D deconvolution of the video stream in the frequency domain, using the second stream as the impulse. iconic math behaviour fr fr 🧚.

the filter accepts the following options:

planes — set which planes to process.
impulse — set which impulse video frames will be processed. can be first or all. default is all.
noise — set the noise when doing divisions. default is 0.0000001. this one's useful when width and height are not the same and not a power of 2, or if the stream prior to convolving had noise in it.

the deconvolve filter also supports the framesync options 💖.

✨ `dedot` ✨

reduces cross-luminance (that's dot-crawl) and cross-color (that's rainbows) from video. yes, both. we're eating good 🌈.

it accepts the following options:

m — set the mode of operation. can be a combination of dotcrawl for cross-luminance reduction and/or rainbows for cross-color reduction.
lt — set the spatial luma threshold. lower values increase the reduction of cross-luminance.
tl — set the tolerance for temporal luma. higher values increase the reduction of cross-luminance.
tc — set the tolerance for chroma temporal variation. higher values increase the reduction of cross-color.
ct — set the temporal chroma threshold. lower values increase the reduction of cross-color.

✨ `deflate` ✨

applies a deflate effect to the video.

this filter replaces the pixel with the local (3x3) average, taking into account only values lower than the pixel. subtle. tasteful. obsessed.

it accepts the following options:

threshold0, threshold1, threshold2, threshold3 — limit the maximum change for each plane. default is 65535. if 0, that plane will remain unchanged.

Commands

this filter supports all the above options as commands 👑

✨ `deflicker` ✨

removes temporal frame luminance variations. the flickering? gone. not on my watch 💅.

it accepts the following options:

size, s — set the moving-average filter size in frames. default is 5. allowed range is 2 - 129.
mode, m — set the averaging mode, for smoothing out those temporal luminance variations.

available values are:

am — arithmetic mean
gm — geometric mean
hm — harmonic mean
qm — quadratic mean
cm — cubic mean
pm — power mean
median — median

bypass — do not actually modify the frame. useful when you only want the metadata and not the glow-up.

✨ `deinterlace_d3d12` ✨

deinterlaces the input video using D3D12 hardware-accelerated video processing. literally GPU brain 🧠🔥.

this filter uses the DirectX 12 Video Processor to perform deinterlacing on the GPU, supporting both simple bob and driver-defined advanced (motion-adaptive) algorithms. it requires the input to be in d3d12 hardware pixel format. no format, no entry, sorry.

the filter automatically queries the hardware for the required number of temporal reference frames (if needed) and manages a frame queue to supply them. it handles all that for you, bestie, you don't have to.

it accepts the following options:

method — the deinterlacing method. it accepts one of the following values:
- default — select the best deinterlacing method available on the hardware. if the driver supports custom, that method is used; otherwise it falls back to bob. this is the default.
- bob — bob deinterlacing. each field is independently scaled to full frame height. simple and fast, but it may produce visible bob artifacts on moving content 😬.
- custom — driver-defined advanced deinterlacing. the exact algorithm is hardware-specific and typically employs motion-adaptive techniques with temporal reference frames for higher quality.

the default value is default.

mode — specify the interlacing mode. it accepts one of the following values:
- frame — send one frame for each frame. the output frame rate equals the input frame rate. this is the default.
- field — send one frame for each field. the output frame rate is double the input frame rate. double the frames, double the fun.

the default value is frame.

deint — specify which frames to deinterlace. it accepts one of the following values:
- all — deinterlace all frames. this is the default.
- interlaced — only deinterlace frames marked as interlaced; progressive frames are passed through unchanged.

the default value is all.

Examples 🎀

deinterlace an interlaced video using software decoding and the best available hardware deinterlacing method:

ffmpeg -init_hw_device d3d12va=d3d12 -filter_hw_device d3d12 -i input.ts \
       -vf "format=nv12,hwupload,deinterlace_d3d12=method=default,hwdownload,format=nv12" \
       -c:v libx264 -crf 18 output.mp4

deinterlace at field rate (double the frame rate) using the bob method with hardware decoding (full d3d12 pipeline):

ffmpeg -hwaccel d3d12va -hwaccel_output_format d3d12 -i input.ts \
       -vf "deinterlace_d3d12=method=bob:mode=field" \
       -c:v h264_d3d12va output.mp4

deinterlace only interlaced frames, passing progressive frames through:

ffmpeg -init_hw_device d3d12va=d3d12 -filter_hw_device d3d12 -i input.ts \
       -vf "format=nv12,hwupload,deinterlace_d3d12=deint=interlaced,hwdownload,format=nv12" \
       -c:v libx264 -crf 18 output.mp4

✨ `dejudder` ✨

removes judder produced by partially interlaced telecined content. the shake? evicted 🦋.

judder can be introduced, for instance, by the pullup filter. if the original source was partially telecined content, then the output of pullup,dejudder will have a variable frame rate. it may change the recorded frame rate of the container. aside from that change, this filter will not affect constant frame rate video. no cap, that's the whole deal.

the option available in this filter is:

cycle — specify the length of the window over which the judder repeats.

accepts any integer greater than 1. useful values are:

4 — if the original was telecined from 24 to 30 fps (Film to NTSC).
5 — if the original was telecined from 25 to 30 fps (PAL to NTSC).
20 — if it's a mixture of the two.

the default is 4.

✨ `delogo` ✨

suppresses a TV station logo by a simple interpolation of the surrounding pixels. you literally just set a rectangle covering the logo and watch it disappear (and sometimes something even uglier appears instead — your mileage may vary, and honestly 😭). we love a chaotic queen.

it accepts the following parameters:

x, y — specify the top left corner coordinates of the logo. they must be specified.
w, h — specify the width and height of the logo to clear. they must be specified.
show — when set to 1, a green rectangle is drawn on the screen so you can actually find the right x, y, w and h parameters instead of guessing. the default value is 0.

the rectangle is drawn on the outermost pixels which will be (partly) replaced with interpolated values. the values of the next pixels immediately outside this rectangle in each direction will be used to compute the interpolated pixel values inside the rectangle. iykyk 👑

Examples 🎀

set a rectangle covering the area with top left corner coordinates 0,0 and size 100x77:

delogo=x=0:y=0:w=100:h=77

✨ `derain` ✨

removes the rain in the input image/video by applying the derain methods based on convolutional neural networks. neural nets vs. weather. the nets are winning 🌧️➡️🌸.

supported models:

Recurrent Squeeze-and-Excitation Context Aggregation Net (RESCAN). see http://openaccess.thecvf.com/content_ECCV_2018/papers/Xia_Li_Recurrent_Squeeze-and-Excitation_Context_ECCV_2018_paper.pdf.

training as well as model generation scripts are provided in the repository at https://github.com/XueweiMeng/derain_filter.git.

the filter accepts the following options:

filter_type — specify which filter to use. this option accepts the following values:
- derain — derain filter. to conduct a derain filter, you need to use a derain model.
- dehaze — dehaze filter. to conduct a dehaze filter, you need to use a dehaze model.

default value is derain.

dnn_backend — specify which DNN backend to use for model loading and execution. this option accepts the following values:
- tensorflow — TensorFlow backend. to enable this backend you need to install the TensorFlow for C library (see https://www.tensorflow.org/install/lang_c) and configure FFmpeg with --enable-libtensorflow.
model — set the path to the model file specifying the network architecture and its parameters. note that different backends use different file formats. TensorFlow can load files for only its format. no cross-format sneaky business 💀.

to get full functionality (such as async execution), please use the dnn_processing filter instead.

✨ `deshake` ✨

attempts to fix small changes in horizontal and/or vertical shift. this filter helps remove camera shake from hand-holding a camera, bumping a tripod, moving on a vehicle, etc. goodbye shaky hands energy 🙈🫶.

the filter accepts the following options:

x, y, w, h — specify a rectangular area where to limit the search for motion vectors. if you want, the search for motion vectors can be limited to a rectangular area of the frame defined by its top left corner, width and height. these parameters have the same meaning as the drawbox filter, which can be used to visualise the position of the bounding box.

this is useful when simultaneous movement of subjects within the frame might get confused for camera motion by the motion vector search. main character energy from the subjects, honestly — they're stealing the scene 🌟.

If any or all of x, y, w and h are set to -1 then the full frame is used. This is lowkey a genius move because it lets you set the later options without having to specify the bounding box for the motion vector search first. 🎀

Default: search the whole frame.

rx, ry — set the max extent of movement in the x and y directions, in the range 0-64 pixels. Default is 16. 🦋
edge — tells it how to generate pixels to fill the blanks at the edge of the frame. The available values are:
- blank, 0 — fill zeroes at blank locations
- original, 1 — original image at blank locations
- clamp, 2 — extruded edge value at blank locations
- mirror, 3 — mirrored edge at blank locations
- Default value is mirror.
blocksize — the blocksize to use for motion search. Range 4-128 pixels, default 8.
contrast — the contrast threshold for blocks. Only blocks with more than the specified contrast (that's the difference between the darkest and lightest pixels) get considered. Range 1-255, default 125.
search — the search strategy. Available values are:
- exhaustive, 0 — set exhaustive search
- less, 1 — set less exhaustive search
- Default value is exhaustive.
filename — if it's set, a detailed log of the motion search gets written to the specified file. 💖

🌸 despill

Removes that unwanted contamination of foreground colors that happens when the greenscreen or bluescreen reflects its color back onto your subject. Not the vibe, we fix it. ✨

This filter accepts the following options:

type — set what type of despill to use.
mix — set how spillmap will be generated.
expand — set how much to get rid of still remaining spill.
red — controls amount of red in spill area.
green — controls amount of green in spill area. Should be -1 for greenscreen. 💚
blue — controls amount of blue in spill area. Should be -1 for bluescreen. 💙
brightness — controls brightness of spill area, preserving colors.
alpha — modify alpha from generated spillmap.

Commands

This filter supports the all above options as commands. 🫶

🎀 detelecine

Applies an exact inverse of the telecine operation. It needs a predefined pattern specified with the pattern option, and that pattern has to be the same one you passed to the telecine filter. No cap. 🦋

This filter accepts the following options:

first_field — which field comes first:
- top, t — top field first
- bottom, b — bottom field first
- The default value is "top".
pattern — a string of numbers representing the pulldown pattern you wish to apply. The default value is 23.
start_frame — a number representing position of the first frame with respect to the telecine pattern. Use this if the stream is cut. The default value is 0.

💅 dilation

Applies a dilation effect to the video. What it actually does: it replaces the pixel by the local (3x3) maximum. Iconic, honestly.

It accepts the following options:

threshold0, threshold1, threshold2, threshold3 — limit the maximum change for each plane, default is 65535. If it's 0, that plane will remain unchanged.
coordinates — a flag specifying the pixel to refer to. Default is 255, i.e. all eight pixels are used. 🌈

Flags to local 3x3 coordinates map like this:

1 2 3
4   5
6 7 8

Commands

This filter supports the all above options as commands.

🦋 displace

Displaces pixels as indicated by the second and third input streams. It takes three input streams and outputs one stream: the first input is the source, and the second and third inputs are displacement maps. 🗺️

The second input specifies how much to displace pixels along the x-axis, while the third input specifies how much to displace pixels along the y-axis. If one of the displacement map streams terminates, the last frame from that displacement map will be used.

Note that once generated, displacement maps can be reused over and over again. Free real estate, bestie. 💖

A description of the accepted options follows.

edge — set the displace behavior for pixels that are out of range. Available values are:
- blank — missing pixels are replaced by black pixels.
- smear — adjacent pixels will spread out to replace missing pixels.
- wrap — out of range pixels are wrapped so they point to pixels of the other side.
- mirror — out of range pixels will be replaced with mirrored pixels.
- Default is smear.

✨ Examples

Add ripple effect to rgb input of video size hd720:

ffmpeg -i INPUT -f lavfi -i nullsrc=s=hd720,lutrgb=128:128:128 -f lavfi -i nullsrc=s=hd720,geq='r=128+30*sin(2*PI*X/400+T):g=128+30*sin(2*PI*X/400+T):b=128+30*sin(2*PI*X/400+T)' -lavfi '[0][1][2]displace' OUTPUT

Add wave effect to rgb input of video size hd720:

ffmpeg -i INPUT -f lavfi -i nullsrc=hd720,geq='r=128+80*(sin(sqrt((X-W/2)*(X-W/2)+(Y-H/2)*(Y-H/2))/220*2*PI+T)):g=128+80*(sin(sqrt((X-W/2)*(X-W/2)+(Y-H/2)*(Y-H/2))/220*2*PI+T)):b=128+80*(sin(sqrt((X-W/2)*(X-W/2)+(Y-H/2)*(Y-H/2))/220*2*PI+T))' -lavfi '[1]split[x][y],[0][x][y]displace' OUTPUT

🧠 dnn_classify

Does classification with deep neural networks, based on bounding boxes. Big brain filter fr fr. 💜

The filter accepts the following options:

dnn_backend — specify which DNN backend to use for model loading and execution. This option accepts only openvino right now; tensorflow backends will be added later.
model — set path to the model file specifying network architecture and its parameters. Note that different backends use different file formats.
input — set the input name of the dnn network.
output — set the output name of the dnn network.
confidence — set the confidence threshold (default: 0.5).
labels — set path to a label file specifying the mapping between label id and name. Each label name is written in one line; tailing spaces and empty lines are skipped. The first line is the name of label id 0, and the second line is the name of label id 1, etc. The label id is considered as the name if the label file is not provided. 🏷️
backend_configs — set the configs to be passed into backend. For tensorflow backend, you can set its configs with sess_config options — please use tools/python/tf_sess_config.py to get the configs for your system.

🔥 dnn_detect

Does object detection with deep neural networks. Yes bestie, it literally finds the things. 👑

The filter accepts the following options:

dnn_backend — specify which DNN backend to use for model loading and execution. This option accepts only openvino right now; tensorflow backends will be added later.
model — set path to the model file specifying network architecture and its parameters. Note that different backends use different file formats.
input — set the input name of the dnn network.
output — set the output name of the dnn network.
confidence — set the confidence threshold (default: 0.5).
labels — set path to a label file specifying the mapping between label id and name. Each label name is written in one line; tailing spaces and empty lines are skipped. The first line is the name of label id 0 (usually it's 'background'), and the second line is the name of label id 1, etc. The label id is considered as the name if the label file is not provided.
backend_configs — set the configs to be passed into backend. To use async execution, set async (default: set). Roll back to sync execution if the backend does not support async.

🎀 dnn_processing

Does image processing with deep neural networks. It works together with another filter which converts the pixel format of the Frame to what the dnn network requires. Teamwork makes the dream work. 🫶

The filter accepts the following options:

dnn_backend — specify which DNN backend to use for model loading and execution. This option accepts the following values:
- tensorflow — TensorFlow backend. To enable this backend you need to install the TensorFlow for C library (see <https://www.tensorflow.org/install/lang_c>) and configure FFmpeg with --enable-libtensorflow
- openvino — OpenVINO backend. To enable this backend you need to build and install the OpenVINO for C library (see <https://github.com/openvinotoolkit/openvino/blob/master/build-instruction.md>) and configure FFmpeg with --enable-libopenvino (--extra-cflags=-I... --extra-ldflags=-L... might be needed if the header files and libraries are not installed into system path)
- torch — Libtorch backend. To enable this backend you need to build and install Libtroch for C++ library. Please download cxx11 ABI version (see <https://pytorch.org/get-started/locally>) and configure FFmpeg with --enable-libtorch --extra-cflags=-I/libtorch_root/libtorch/include --extra-cflags=-I/libtorch_root/libtorch/include/torch/csrc/api/include --extra-ldflags=-L/libtorch_root/libtorch/lib/
model — set path to the model file specifying network architecture and its parameters. Note that different backends use different file formats. TensorFlow, OpenVINO and Libtorch backend can load files for only its format. 🙈
input — set the input name of the dnn network.
output — set the output name of the dnn network.
backend_configs — set the configs to be passed into backend. To use async execution, set async (default: set). Roll back to sync execution if the backend does not support async. For tensorflow backend, you can set its configs with sess_config options — please use tools/python/tf_sess_config.py to get the configs of TensorFlow backend for your system.

✨ Examples

Remove rain in rgb24 frame with can.pb (see derain filter):

./ffmpeg -i rain.jpg -vf format=rgb24,dnn_processing=dnn_backend=tensorflow:model=can.pb:input=x:output=y derain.jpg

Handle the Y channel with srcnn.pb (see sr filter) for frame with yuv420p (planar YUV formats supported):

./ffmpeg -i 480p.jpg -vf format=yuv420p,scale=w=iw*2:h=ih*2,dnn_processing=dnn_backend=tensorflow:model=srcnn.pb:input=x:output=y -y srcnn.jpg

Handle the Y channel with espcn.pb (see sr filter), which changes frame size, for format yuv420p (planar YUV formats supported). Please use tools/python/tf_sess_config.py to get the configs of TensorFlow backend for your system.

./ffmpeg -i 480p.jpg -vf format=yuv420p,dnn_processing=dnn_backend=tensorflow:model=espcn.pb:input=x:output=y:backend_configs=sess_config=0x10022805320e09cdccccccccccec3f20012a01303801 -y tmp.espcn.jpg

drawbox 🎨

okay so this one literally just draws a colored box on top of your input image. iconic behavior honestly.

It accepts the following parameters:

x
y — the expressions that set the top-left corner coords of the box. Defaults to 0.
width, w
height, h — the expressions that set the width and height of the box. If they're 0, they get interpreted as the input width and height. Defaults to 0.
color, c — pick the color of the box you're writing. For the general syntax of this option, check the "Color" section in the ffmpeg-utils manual. And if you use the special value invert, the box edge color is the same as the video with inverted luma. 💅
thickness, t — expression that sets the thickness of the box edge. A value of fill will create a filled box. Default value is 3. See below for the list of accepted constants.
replace — only applies if your input has alpha. With value 1, the pixels of the painted box will overwrite the video's color and alpha pixels. Default is 0, which composites the box onto the input and leaves the video's alpha intact.

The parameters for x, y, w and h and t are expressions containing the following constants:

dar — the input display aspect ratio, same as (w / h) * sar.
hsub
vsub — horizontal and vertical chroma subsample values. Like for the pixel format yuv422p, hsub is 2 and vsub is 1.
in_h, ih
in_w, iw — the input width and height.
sar — the input sample aspect ratio.
x
y — the x and y offset coordinates where the box is drawn.
w
h — the width and height of the drawn box.
box_source — box source can be set as side_data_detection_bboxes if you want to use box data in detection bboxes of side data. If box_source is set, the x, y, width and height will be ignored and it'll still use box data in detection bboxes of side data. So bestie, please do not use this parameter if you were not sure about the box source. 🙈
t — the thickness of the drawn box.

These constants allow the x, y, w, h and t expressions to refer to each other, so you may for example specify y=x/dar or h=w/dar. 🫶

Examples

Draw a black box around the edge of the input image:

drawbox

Draw a box with color red and an opacity of 50%:

drawbox=10:20:200:60:red@0.5

The previous example can be specified as:

drawbox=x=10:y=20:w=200:h=60:color=red@0.5

Fill the box with pink color:

drawbox=x=10:y=10:w=100:h=100:color=pink@0.5:t=fill

Draw a 2-pixel red 2.40:1 mask:

drawbox=x=-t:y=0.5*(ih-iw/2.4)-t:w=iw+t*2:h=iw/2.4+t*2:t=2:c=red

Commands

This filter supports same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value. no cap.

drawgraph 📈

Draws a graph using input video metadata. it's giving data visualization era.

It accepts the following parameters:

m1 — set 1st frame metadata key from which metadata values will be used to draw a graph.
fg1 — set 1st foreground color expression.
m2 — set 2nd frame metadata key from which metadata values will be used to draw a graph.
fg2 — set 2nd foreground color expression.
m3 — set 3rd frame metadata key from which metadata values will be used to draw a graph.
fg3 — set 3rd foreground color expression.
m4 — set 4th frame metadata key from which metadata values will be used to draw a graph.
fg4 — set 4th foreground color expression.
min — set minimal value of metadata value.
max — set maximal value of metadata value.
bg — set graph background color. Default is white.
mode — set graph mode.

Available values for mode is:

bar
dot
line

Default is line.

slide — set slide mode.

Available values for slide is:

frame — draw new frame when right border is reached.
replace — replace old columns with new ones.
scroll — scroll from right to left.
rscroll — scroll from left to right.
picture — draw single picture.

Default is frame.

size — set size of graph video. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. The default value is 900x256.
rate, r — set the output frame rate. Default value is 25.

The foreground color expressions can use the following variables:

MIN — minimal value of metadata value.
MAX — maximal value of metadata value.
VAL — current metadata key value.

The color is defined as 0xAABBGGRR. 🎀

Example using metadata from `signalstats` filter

signalstats,drawgraph=lavfi.signalstats.YAVG:min=0:max=255

Example using metadata from `ebur128` filter

ebur128=metadata=1,adrawgraph=lavfi.r128.M:min=-120:max=5

drawgrid 🔲

Draws a grid on the input image. grid supremacy honestly.

It accepts the following parameters:

x
y — the expressions that set the coordinates of some point of grid intersection (meant to configure offset). Both default to 0.
width, w
height, h — the expressions that set the width and height of the grid cell. If 0, they're interpreted as the input width and height, respectively, minus thickness, so the image gets framed. Default to 0.
color, c — specify the color of the grid. For the general syntax of this option, check the "Color" section in the ffmpeg-utils manual. If the special value invert is used, the grid color is the same as the video with inverted luma.
thickness, t — the expression that sets the thickness of the grid line. Default value is 1. See below for the list of accepted constants.
replace — applies if the input has alpha. With 1 the pixels of the painted grid will overwrite the video's color and alpha pixels. Default is 0, which composites the grid onto the input and leaves the video's alpha intact.

The parameters for x, y, w and h and t are expressions containing the following constants:

dar — the input display aspect ratio, same as (w / h) * sar.
hsub
vsub — horizontal and vertical chroma subsample values. Like for the pixel format yuv422p, hsub is 2 and vsub is 1.
in_h, ih
in_w, iw — the input grid cell width and height.
sar — the input sample aspect ratio.
x
y — the x and y coordinates of some point of grid intersection (meant to configure offset).
w
h — the width and height of the drawn cell.
t — the thickness of the drawn cell.

These constants allow the x, y, w, h and t expressions to refer to each other, so you may for example specify y=x/dar or h=w/dar. ✨

Examples

Draw a grid with cell 100x100 pixels, thickness 2 pixels, with color red and an opacity of 50%:

drawgrid=width=100:height=100:thickness=2:color=red@0.5

Draw a white 3x3 grid with an opacity of 50%:

drawgrid=w=iw/3:h=ih/3:t=2:c=white@0.5

Commands

This filter supports same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value. 💯

drawtext ✍️👑

Draws a text string, or text from a specified file, on top of a video — using the libfreetype library. this is the main character filter, i fear.

To enable compilation of this filter, you need to configure FFmpeg with --enable-libfreetype and --enable-libharfbuzz. To enable default font fallback and the font option you need to configure FFmpeg with --enable-libfontconfig. To enable the text_shaping option, you need to configure FFmpeg with --enable-libfribidi. 🛠️

Syntax

It accepts the following parameters:

box — used to draw a box around text using the background color. The value must be either 1 (enable) or 0 (disable). The default value of box is 0.
boxborderw — set the width of the border to be drawn around the box using boxcolor. The value must be specified using one of the following formats:
- boxborderw=10 sets the width of all the borders to 10
- boxborderw=10|20 sets the width of the top and bottom borders to 10 and the width of the left and right borders to 20
- boxborderw=10|20|30 sets the width of the top border to 10, the width of the bottom border to 30 and the width of the left and right borders to 20
- boxborderw=10|20|30|40 sets the borders width to 10 (top), 20 (right), 30 (bottom), 40 (left)

The default value of boxborderw is 0.

boxcolor — the color to be used for drawing box around text. For the syntax of this option, check the "Color" section in the ffmpeg-utils manual. The default value of boxcolor is white.
line_spacing — set the line spacing in pixels. The default value of line_spacing is 0.
text_align — set the vertical and horizontal alignment of the text with respect to the box boundaries. The value is combination of flags, one for the vertical alignment (T=top, M=middle, B=bottom) and one for the horizontal alignment (L=left, C=center, R=right). Please note that tab characters are only supported with the left horizontal alignment. 💅
y_align — specify what the y value is referred to. Possible values are:
- text — the top of the highest glyph of the first text line is placed at y
- baseline — the baseline of the first text line is placed at y
- font — the baseline of the first text line is placed at y plus the ascent (in pixels) defined in the font metrics

The default value of y_align is text for backward compatibility.

borderw — set the width of the border to be drawn around the text using bordercolor. The default value of borderw is 0.
bordercolor — set the color to be used for drawing border around text. For the syntax of this option, check the "Color" section in the ffmpeg-utils manual. The default value of bordercolor is black.
expansion — select how the text is expanded. Can be either none, strftime (deprecated) or normal (default). See the drawtext_expansion, Text expansion section below for details. 👀
basetime — set a start time for the count. Value is in microseconds. Only applied in the deprecated strftime expansion mode. To emulate in normal expansion mode use the pts function, supplying the start time (in seconds) as the second argument.
fix_bounds — if true, it checks and fixes the text coords so nothing gets clipped. very considerate of it honestly 💅
fontcolor — the color used to draw the font. for the syntax of this option, check the "Color" section in the ffmpeg-utils manual. the default value of fontcolor is "black".
fontcolor_expr — a string that gets expanded the same way as text to obtain a dynamic fontcolor value. by default this option has an empty value and is not processed. when this option is set, it overrides the fontcolor option. (yes bestie, it's giving "i'm the main character now") 👑
font — the font family used for drawing text. by default Sans.
fontfile — the font file used for drawing text. the path must be included. this parameter is mandatory if the fontconfig support is disabled.
alpha — draw the text applying alpha blending. the value can be a number between 0.0 and 1.0. the expression accepts the same variables x, y as well. the default value is 1. please see fontcolor_expr. 🫶
fontsize — the font size used for drawing text. the default value of fontsize is 16.
text_shaping — if set to 1, it attempts to shape the text (for example, reverse the order of right-to-left text and join Arabic characters) before drawing it. otherwise, it just draws the text exactly as given. by default 1 (if supported).
ft_load_flags — the flags used for loading the fonts. these flags map the corresponding flags supported by libfreetype, and are a combination of the following values:
- default
- no_scale
- no_hinting
- render
- no_bitmap
- vertical_layout
- force_autohint
- crop_bitmap
- pedantic
- ignore_global_advance_width
- no_recurse
- ignore_transform
- monochrome
- linear_design
- no_autohint

default value is "default". for more information consult the documentation for the FT_LOAD_* libfreetype flags.

shadowcolor — the color used for drawing a shadow behind the drawn text. for the syntax of this option, check the "Color" section in the ffmpeg-utils manual. the default value of shadowcolor is "black".
boxw — set the width of the box to be drawn around text. the default value of boxw is computed automatically to match the text width.
boxh — set the height of the box to be drawn around text. the default value of boxh is computed automatically to match the text height.
shadowx, shadowy — the x and y offsets for the text shadow position with respect to the position of the text. they can be either positive or negative values. the default value for both is "0".
start_number — the starting frame number for the n/frame_num variable. the default value is "0".
tabsize — the size in number of spaces used for rendering the tab. default value is 4.
timecode — set the initial timecode representation in "hh:mm:ss[:;.]ff" format. it can be used with or without text parameter. timecode_rate option must be specified.
timecode_rate, rate, r — set the timecode frame rate (timecode only). the value will be rounded to nearest integer. minimum value is "1". drop-frame timecode is supported for frame rates 30 & 60.
tc24hmax — if set to 1, the output of the timecode option will wrap around at 24 hours. default is 0 (disabled).
text — the text string to be drawn. the text must be a sequence of UTF-8 encoded characters. this parameter is mandatory if no file is specified with the parameter textfile.
textfile — a text file containing text to be drawn. the text must be a sequence of UTF-8 encoded characters. this parameter is mandatory if no text string is specified with the parameter text. if both text and textfile are specified, an error is thrown. 😭 (pick one bestie, the filter is not doing polyamory today)
text_source — text source should be set as side_data_detection_bboxes if you want to use text data in detection bboxes of side data. if text source is set, text and textfile will be ignored and it'll still use text data in detection bboxes of side data. so please do not use this parameter if you are not sure about the text source. > fr fr, don't just press buttons on this one 🙈
reload — the textfile will be reloaded at specified frame interval. be sure to update textfile atomically, or it may be read partially, or even fail. range is 0 to INT_MAX. default is 0. 💀
x, y — the expressions which specify the offsets where text will be drawn within the video frame. they are relative to the top/left border of the output image. the default value of x and y is "0". see below for the list of accepted constants and functions. 👇

the parameters for x and y are expressions containing the following constants and functions:

dar — input display aspect ratio, it is the same as (w / h) * sar.
hsub, vsub — horizontal and vertical chroma subsample values. for example for the pixel format "yuv422p", hsub is 2 and vsub is 1.
line_h, lh — the height of each text line.
main_h, h, H — the input height.
main_w, w, W — the input width.
max_glyph_a, ascent — the maximum distance from the baseline to the highest/upper grid coordinate used to place a glyph outline point, for all the rendered glyphs. it is a positive value, due to the grid's orientation with the Y axis upwards.
max_glyph_d, descent — the maximum distance from the baseline to the lowest grid coordinate used to place a glyph outline point, for all the rendered glyphs. this is a negative value, due to the grid's orientation, with the Y axis upwards.
max_glyph_h — maximum glyph height, that is the maximum height for all the glyphs contained in the rendered text, it is equivalent to ascent - descent.
max_glyph_w — maximum glyph width, that is the maximum width for all the glyphs contained in the rendered text.
font_a — the ascent size defined in the font metrics.
font_d — the descent size defined in the font metrics.
top_a — the maximum ascender of the glyphs of the first text line.
bottom_d — the maximum descender of the glyphs of the last text line.
n — the number of input frame, starting from 0.
rand(min, max) — return a random number included between min and max.
sar — the input sample aspect ratio.
t — timestamp expressed in seconds, NAN if the input timestamp is unknown.
text_h, th — the height of the rendered text.
text_w, tw — the width of the rendered text.
x, y — the x and y offset coordinates where the text is drawn. these parameters allow the x and y expressions to refer to each other, so you can for example specify y=x/dar. (mutual reference?? it's giving collaborative queen energy ✨)
pict_type — a one character description of the current frame's picture type.
pkt_pos — the current packet's position in the input file or stream (in bytes, from the start of the input). a value of -1 indicates this info is not available.
duration — the current packet's duration, in seconds.
pkt_size — the current packet's size (in bytes).

✨ Text expansion ✨

if expansion is set to "strftime", the filter recognizes sequences accepted by the strftime C function in the provided text and expands them accordingly. check the documentation of strftime. this feature is deprecated in favor of "normal" expansion with the gmtime or localtime expansion functions. 💀 (we don't do that anymore bestie)

if expansion is set to "none", the text is printed verbatim.

if expansion is set to "normal" (which is the default), the following expansion mechanism is used.

the backslash character \, followed by any character, always expands to the second character.

sequences of the form %{...} are expanded. the text between the braces is a function name, possibly followed by arguments separated by ':'. if the arguments contain special characters or delimiters (':' or '}'), they should be escaped.

note that they probably must also be escaped as the value for the text option in the filter argument string and as the filter argument in the filtergraph description, and possibly also for the shell, that makes up to four levels of escaping; using a text file with the textfile option avoids these problems. 😭 four levels, we're literally in inception

the following functions are available:

expr, e — the expression evaluation result. it must take one argument specifying the expression to be evaluated, which accepts the same constants and functions as the x and y values. note that not all constants should be used, for example the text size is not known when evaluating the expression, so the constants text_w and text_h will have an undefined value.
expr_int_format, eif — evaluate the expression's value and output as formatted integer. the first argument is the expression to be evaluated, just as for the expr function. the second argument specifies the output format. allowed values are x, X, d and u. they are treated exactly as in the printf function. the third parameter is optional and sets the number of positions taken by the output. it can be used to add padding with zeros from the left.
gmtime — the time at which the filter is running, expressed in UTC. it can accept an argument: a strftime C function format string. the format string is extended to support the variable %[1-6]N which prints fractions of the second with optionally specified number of digits.
localtime — the time at which the filter is running, expressed in the local time zone. it can accept an argument: a strftime C function format string. the format string is extended to support the variable %[1-6]N which prints fractions of the second with optionally specified number of digits. 🌸 (gmtime's local-time girlie twin)
metadata — frame metadata. takes one or two arguments. the first argument is mandatory and specifies the metadata key. the second argument is optional and specifies a default value, used when the metadata key is not found or empty. available metadata can be identified by inspecting entries starting with TAG included within each frame section printed by running ffprobe -show_frames. string metadata generated in filters leading to the drawtext filter are also available. 🎀
n, frame_num — the frame number, starting from 0.
pict_type — a one character description of the current picture type.
pts — the timestamp of the current frame. it can take up to three arguments.

The first argument is the format of the timestamp — it defaults to "flt" (seconds as a decimal number, microsecond accuracy, cute and simple); "hms" gives you a formatted [-]HH:MM:SS.mmm timestamp with millisecond accuracy. "gmtime" is the timestamp of the frame formatted as UTC time, and "localtime" is the timestamp of the frame formatted as local time zone time. 🌍

The second argument is an offset added to the timestamp.

If the format is set to "hms", a third argument "24HH" may be supplied to present the hour part of the formatted timestamp in 24h format (00-23). ⏰

If the format is set to "localtime" or "gmtime", a third argument may be supplied: a strftime C function format string. By default, YYYY-MM-DD HH:MM:SS format will be used.

Commands

Not only does this filter do the most, it also supports altering parameters via commands:

reinit — Alter existing filter parameters. 💅

The syntax for the argument is the same as for filter invocation, e.g.

fontsize=56:fontcolor=green:text='Hello World'

Full filter invocation with sendcmd would look like this:

sendcmd=c='56.0 drawtext reinit fontsize=56\:fontcolor=green\:text=Hello\\ World'

If the entire argument can't be parsed or applied as valid values then the filter will continue with its existing parameters. (No crash, no drama — it just keeps doing what it was doing. 🫶)

The following options are also supported as commands:

*<x>
*<y>
*<alpha>
*<fontsize>
*<fontcolor>
*<boxcolor>
*<bordercolor>
*<shadowcolor>
*<box>
*<boxw>
*<boxh>
*<boxborderw>
*<line_spacing>
*<text_align>
*<shadowx>
*<shadowy>
*<borderw>

Examples

Okay bestie, the example section ate and left no crumbs — here's every single one:

Draw "Test Text" with font FreeSerif, using the default values for the optional parameters.

drawtext="fontfile=/usr/share/fonts/truetype/freefont/FreeSerif.ttf: text='Test Text'"

Draw 'Test Text' with font FreeSerif of size 24 at position x=100 and y=50 (counting from the top-left corner of the screen), text is yellow with a red box around it. Both the text and the box have an opacity of 20%. ✨

drawtext="fontfile=/usr/share/fonts/truetype/freefont/FreeSerif.ttf: text='Test Text':\
          x=100: y=50: fontsize=24: fontcolor=yellow@0.2: box=1: boxcolor=red@0.2"

Note that the double quotes are not necessary if spaces are not used within the parameter list. Save yourself the keystrokes, bestie.

Show the text at the center of the video frame:

drawtext="fontsize=30:fontfile=FreeSerif.ttf:text='hello world':x=(w-text_w)/2:y=(h-text_h)/2"

Show the text at a random position, switching to a new position every 30 seconds:

drawtext="fontsize=30:fontfile=FreeSerif.ttf:text='hello world':x=if(eq(mod(t\,30)\,0)\,rand(0\,(w-text_w))\,x):y=if(eq(mod(t\,30)\,0)\,rand(0\,(h-text_h))\,y)"

Show a text line sliding from right to left in the last row of the video frame. The file LONG_LINE is assumed to contain a single line with no newlines. 🛝

drawtext="fontsize=15:fontfile=FreeSerif.ttf:text=LONG_LINE:y=h-line_h:x=-50*t"

Show the content of file CREDITS off the bottom of the frame and scroll up.

drawtext="fontsize=20:fontfile=FreeSerif.ttf:textfile=CREDITS:y=h-20*t"

Draw a single green letter "g", at the center of the input video. The glyph baseline is placed at half screen height.

drawtext="fontsize=60:fontfile=FreeSerif.ttf:fontcolor=green:text=g:x=(w-max_glyph_w)/2:y=h/2-ascent"

Show text for 1 second every 3 seconds: 👀

drawtext="fontfile=FreeSerif.ttf:fontcolor=white:x=100:y=x/dar:enable=lt(mod(t\,3)\,1):text='blink'"

Use fontconfig to set the font. Note that the colons need to be escaped.

drawtext='fontfile=Linux Libertine O-40\\:style=Semibold:text=FFmpeg'

Draw "Test Text" with font size dependent on height of the video.

drawtext="text='Test Text': fontsize=h/30: x=(w-text_w)/2: y=(h-text_h*2)"

Print the date of a real-time encoding (see documentation for the strftime C function):

drawtext='fontfile=FreeSans.ttf:text=%{localtime\:%a %b %d %Y}'

Show text fading in and out (appearing/disappearing):

#!/bin/sh
DS=1.0 # display start
DE=10.0 # display end
FID=1.5 # fade in duration
FOD=5 # fade out duration
ffplay -f lavfi "color,drawtext=text=TEST:fontsize=50:fontfile=FreeSerif.ttf:fontcolor_expr=ff0000%{eif\\\\: clip(255*(1*between(t\\, $DS + $FID\\, $DE - $FOD) + ((t - $DS)/$FID)*between(t\\, $DS\\, $DS + $FID) + (-(t - $DE)/$FOD)*between(t\\, $DE - $FOD\\, $DE) )\\, 0\\, 255) \\\\: x\\\\: 2 }"

Horizontally align multiple separate texts. Note that max_glyph_a and the fontsize value are included in the y offset. 📐

drawtext=fontfile=FreeSans.ttf:text=DOG:fontsize=24:x=10:y=20+24-max_glyph_a,
drawtext=fontfile=FreeSans.ttf:text=cow:fontsize=24:x=80:y=20+24-max_glyph_a

Plot special lavf.image2dec.source_basename metadata onto each frame if such metadata exists. Otherwise, plot the string "NA". Note that image2 demuxer must have option -export_path_metadata 1 for the special metadata fields to be available for filters. 🏷️

drawtext="fontsize=20:fontcolor=white:fontfile=FreeSans.ttf:text='%{metadata\:lavf.image2dec.source_basename\:NA}':x=10:y=10"

For more information about libfreetype, check: http://www.freetype.org/.

For more information about fontconfig, check: http://freedesktop.org/software/fontconfig/fontconfig-user.html.

For more information about libfribidi, check: http://fribidi.org/.

For more information about libharfbuzz, check: https://github.com/harfbuzz/harfbuzz.

drawvg

Draw vector graphics on top of video frames, by executing a script written in a custom language called VGS (Vector Graphics Script). 🎨 Iconic, honestly.

The documentation for the language can be found in drawvg - Language Reference. A version of this reference with rendered examples can be found at https://ayosec.github.io/ffmpeg-drawvg/playground/docs/langref.html.

Graphics are rendered using https://cairographics.org/.

To enable compilation of this filter, you need to configure FFmpeg with --enable-cairo. (No cairo, no vibes. 💀)

Parameters

Either script or file must be set. Like, pick one, bestie.

s, script — Script source to draw the graphics.
file — Path of the file to load the script source.

Pixel Formats

Since Cairo only supports RGB images, if the input video is something else (like YUV 4:2:0), before executing the script the video is converted to a format compatible with Cairo. Then, you have to use either the format filter, or the -pix_fmt option, to convert it to the expected format in the output. 🖼️

Examples

Draw the outline of an ellipse.

ffmpeg -i input.webm \
    -vf 'drawvg=ellipse (w/2) (h/2) (w/3) (h/3) stroke' \
    -pix_fmt yuv420p \
    output.webm

Draw a square rotating in the middle of the frame.

The script for drawvg is in a file draw.vgs:

translate (w/2) (h/2)
rotate t
rect -100 -100 200 200
setcolor red@0.5
fill

Then:

ffmpeg -i input.webm -vf 'drawvg=file=draw.vgs,format=yuv420p' output.webm

edgedetect

Detect and draw edges. The filter uses the Canny Edge Detection algorithm. 🔍

The filter accepts the following options:

low, high — Set low and high threshold values used by the Canny thresholding algorithm.

The high threshold selects the "strong" edge pixels, which are then connected through 8-connectivity with the "weak" edge pixels selected by the low threshold.

low and high threshold values must be chosen in the range [0,1], and low should be lesser or equal to high.

Default value for low is "20/255", and default value for high is "50/255".

mode — Define the drawing mode.
- wires — Draw white/gray wires on black background.
- colormix — Mix the colors to create a paint/cartoon effect.
- canny — Apply Canny edge detector on all selected planes.

Default value is wires.

planes — Select planes for filtering. By default all available planes are filtered.

Examples

Standard edge detection with custom values for the hysteresis thresholding:

edgedetect=low=0.1:high=0.4

Painting effect without thresholding:

edgedetect=mode=colormix:high=0

elbg

Apply a posterize effect using the ELBG (Enhanced LBG) algorithm. 🌈

For each input image, the filter will compute the optimal mapping from the input to the output given the codebook length, that is the number of distinct output colors.

This filter accepts the following options.

codebook_length, l — Set codebook length. The value must be a positive integer, and represents the number of distinct output colors. Default value is 256.
nb_steps, n — Set the maximum number of iterations to apply for computing the optimal mapping. The higher the value the better the result and the higher the computation time. Default value is 1.
seed, s — Set a random seed, must be an integer included between 0 and UINT32_MAX. If not specified, or if explicitly set to -1, the filter will try to use a good random seed on a best effort basis.
pal8 — Set pal8 output pixel format. This option does not work with codebook length greater than 256. Default is disabled.
use_alpha — Include alpha values in the quantization calculation. Allows creating palettized output images (e.g. PNG8) with multiple alpha smooth blending. 🫧

entropy

Measure graylevel entropy in histogram of color channels of video frames. Nerd mode: activated. 📊

It accepts the following parameters:

mode — Can be either normal or diff. Default is normal.

diff mode measures entropy of histogram delta values, absolute differences between neighbour histogram values.

epx

Apply the EPX magnification filter which is designed for pixel art. 🕹️

It accepts the following option:

n — Set the scaling dimension: 2 for "2xEPX", 3 for "3xEPX". Default is 3.

eq

Set brightness, contrast, saturation and approximate gamma adjustment.

The filter accepts the following options:

contrast — Set the contrast expression. The value must be a float value in range -1000.0 to 1000.0. The default value is "1".
brightness — Set the brightness expression. The value must be a float value in range -1.0 to 1.0. The default value is "0".
saturation — the saturation expression, bestie. Gotta be a float in range 0.0 to 3.0, and the default is "1". 💅
gamma — the gamma expression. Float in range 0.1 to 10.0, default "1".
gamma_r — the gamma expression for red. Float in range 0.1 to 10.0, default "1".
gamma_g — the gamma expression for green. Float in range 0.1 to 10.0, default "1".
gamma_b — the gamma expression for blue. Float in range 0.1 to 10.0, default "1".
gamma_weight — the gamma weight expression, and this one's lowkey the main character 💫 It can be used to reduce the effect of a high gamma value on bright image areas — like, keep them from getting overamplified and just plain white. The value must be a float in range 0.0 to 1.0. A value of 0.0 turns the gamma correction all the way down while 1.0 leaves it at its full strength. Default is "1".
eval — sets when the expressions for brightness, contrast, saturation and gamma expressions are evaluated. It accepts these values:
- init — only evaluate expressions once during the filter initialization or when a command is processed
- frame — evaluate expressions for each incoming frame

Default value is init.

The expressions accept the following parameters:

n — frame count of the input frame starting from 0
pos — byte position of the corresponding packet in the input file, NAN if unspecified; deprecated, do not use 🙈
r — frame rate of the input video, NAN if the input frame rate is unknown
t — timestamp expressed in seconds, NAN if the input timestamp is unknown

🎮 Commands

The filter supports the following commands:

contrast — set the contrast expression.
brightness — set the brightness expression.
saturation — set the saturation expression.
gamma — set the gamma expression.
gamma_r — set the gamma_r expression.
gamma_g — set gamma_g expression.
gamma_b — set gamma_b expression.
gamma_weight — set gamma_weight expression.

The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value. No take-backs, it just stays where it is 🫶

🌸 erosion 🌸

Apply erosion effect to the video. Obsessed with this one ngl.

This filter replaces the pixel by the local(3x3) minimum. Fr fr, that's the whole vibe.

It accepts the following options:

threshold0, threshold1, threshold2, threshold3 — limit the maximum change for each plane, default is 65535. If 0, plane will remain unchanged. Respect the boundaries bestie 💀
coordinates — flag which specifies the pixel to refer to. Default is 255 i.e. all eight pixels are used.

Flags to local 3x3 coordinates maps like this:

1 2 3
4   5
6 7 8

🎮 Commands

This filter supports the all above options as commands.

🦋 estdif 🦋

Deinterlace the input video ("estdif" stands for "Edge Slope Tracing Deinterlacing Filter").

Spatial only filter that uses edge slope tracing algorithm to interpolate missing lines. It accepts the following parameters:

mode — the interlacing mode to adopt. It accepts one of the following values:
- frame — output one frame for each frame.
- field — output one frame for each field.

The default value is "field".

parity — the picture field parity assumed for the input interlaced video. It accepts one of the following values:
- tff — assume the top field is first.
- bff — assume the bottom field is first.
- auto — enable automatic detection of field parity.

The default value is "auto". If the interlacing is unknown or the decoder does not export this information, top field first will be assumed. iykyk ✨

deint — specify which frames to deinterlace. Accepts one of the following values:
- all — deinterlace all frames.
- interlaced — only deinterlace frames marked as interlaced.

The default value is "all".

rslope — specify the search radius for edge slope tracing. Default value is 1. Allowed range is from 1 to 15.
redge — specify the search radius for best edge matching. Default value is 2. Allowed range is from 0 to 15.
ecost — specify the edge cost for edge matching. Default value is 2. Allowed range is from 0 to 50.
mcost — specify the middle cost for edge matching. Default value is 1. Allowed range is from 0 to 50.
dcost — specify the distance cost for edge matching. Default value is 1. Allowed range is from 0 to 50.
interp — specify the interpolation used. Default is 4-point interpolation. It accepts one of the following values:
- 2p — two-point interpolation.
- 4p — four-point interpolation.
- 6p — six-point interpolation.

🎮 Commands

This filter supports same commands as options.

✨ exposure ✨

Adjust exposure of the video stream. Little glow-up filter, literally.

The filter accepts the following options:

exposure — set the exposure correction in EV. Allowed range is from -3.0 to 3.0 EV. Default value is 0 EV.
black — set the black level correction. Allowed range is from -1.0 to 1.0. Default value is 0.

🎮 Commands

This filter supports same commands as options.

🎀 extractplanes 🎀

Extract color channel components from input video stream into separate grayscale video streams. It's giving assembly line energy honestly.

The filter accepts the following option:

planes — set plane(s) to extract. Available values for planes are:
- y
- u
- v
- a
- r
- g
- b

Choosing planes not available in the input will result in an error. That means you cannot select "r", "g", "b" planes with "y", "u", "v" planes at same time. Pick a lane bestie 🫶

Examples

Extract luma, u and v color channel component from input video frame into 3 grayscale outputs:

ffmpeg -i video.avi -filter_complex 'extractplanes=y+u+v[y][u][v]' -map '[y]' y.avi -map '[u]' u.avi -map '[v]' v.avi

🌈 fade 🌈

Apply a fade-in/out effect to the input video. The drama of it all 💅

It accepts the following parameters:

type, t — the effect type can be either "in" for a fade-in, or "out" for a fade-out effect. Default is "in".
start_frame, s — specify the number of the frame to start applying the fade effect at. Default is 0.
nb_frames, n — the number of frames that the fade effect lasts. At the end of the fade-in effect, the output video will have the same intensity as the input video. At the end of the fade-out transition, the output video will be filled with the selected color. Default is 25.
alpha — if set to 1, fade only alpha channel, if one exists on the input. Default value is 0.
start_time, st — specify the timestamp (in seconds) of the frame to start to apply the fade effect. If both start_frame and start_time are specified, the fade will start at whichever comes last. Default is 0.
duration, d — the number of seconds for which the fade effect has to last. At the end of the fade-in effect the output video will have the same intensity as the input video, at the end of the fade-out transition the output video will be filled with the selected color. If both duration and nb_frames are specified, duration is used. Default is 0 (nb_frames is used by default).
color, c — specify the color of the fade. Default is "black".

Examples

Fade in the first 30 frames of video:

fade=in:0:30

The command above is equivalent to:

fade=t=in:s=0:n=30

Fade out the last 45 frames of a 200-frame video:

fade=out:155:45
fade=type=out:start_frame=155:nb_frames=45

Fade in the first 25 frames and fade out the last 25 frames of a 1000-frame video:

fade=in:0:25, fade=out:975:25

Make the first 5 frames yellow, then fade in from frame 5-24:

fade=in:5:20:color=yellow

Fade in alpha over first 25 frames of video:

fade=in:0:25:alpha=1

Make the first 5.5 seconds black, then fade in for 0.5 seconds:

fade=t=in:st=5.5:d=0.5

💜 feedback 💜

Apply feedback video filter. This one's kinda mind-bending, stay with me 🌀

This filter pass cropped input frames to 2nd output. From there it can be filtered with other video filters. After filter receives frame from 2nd input, that frame is combined on top of original frame from 1st input and passed to 1st output.

The typical usage is filter only part of frame. Selective glow-up, we love to see it ✨

The filter accepts the following options:

x, y — set the top left crop position.
w, h — set the crop size.

Examples

Blur only top left rectangular part of video frame size 100x100 with gblur filter.

[in][blurin]feedback=x=0:y=0:w=100:h=100[out][blurout];[blurout]gblur=8[blurin]

Draw black box on top left part of video frame of size 100x100 with drawbox filter.

[in][blurin]feedback=x=0:y=0:w=100:h=100[out][blurout];[blurout]drawbox=x=0:y=0:w=100:h=100:t=100[blurin]

Pixelize rectangular part of video frame of size 100x100 with pixelize filter.

[in][blurin]feedback=x=320:y=240:w=100:h=100[out][blurout];[blurout]pixelize[blurin]

🧚 fftdnoiz 🧚

Denoise frames using 3D FFT (frequency domain filtering). Science girlie hours 🔬

The filter accepts the following options:

sigma — set the noise sigma constant. This sets denoising strength. Default value is 1. Allowed range is from 0 to 30. Using very high sigma with low overlap may give blocking artifacts. Not the cute kind of blocky, be warned 😭
amount — set amount of denoising. By default all detected noise is reduced. Default value is 1. Allowed range is from 0 to 1.
block — set size of block in pixels, default is 32, can be 8 to 256.
overlap — set block overlap. Default is 0.5. Allowed range is from 0.2 to 0.8.
method — set denoising method. Default is "wiener", can also be "hard".
prev — set number of previous frames to use for denoising. By default is set to 0.
next — set number of next frames to to use for denoising. By default is set to 0.
planes — set planes which will be filtered, by default are all available filtered except alpha.

👑 fftfilt 👑

Apply arbitrary expressions to samples in frequency domain

dc_Y — tweaks the dc value (that's the gain, bestie) of the image's luma plane 💅 it takes an integer from 0 to 1000, and the default is 0.
dc_U — same energy but for the 1st chroma plane ✨ integer, 0 to 1000, default 0.
dc_V — and this one's for the 2nd chroma plane 🎀 integer, 0 to 1000, default 0.
weight_Y — sets the frequency domain weight expression for the luma plane, iykyk.
weight_U — the frequency domain weight expression for the 1st chroma plane 🦋
weight_V — frequency domain weight expression for the 2nd chroma plane, no cap.
eval — controls when the expressions get evaluated ⏰ it accepts:
- init — evaluate the expressions exactly once, during filter initialization.
- frame — evaluate the expressions for every single incoming frame.

Default is init.

The filter is also like, "here are the variables you can use in those expressions" 🫶

X, Y — the coordinates of the current sample.
W, H — the width and height of the image.
N — the number of the input frame, counting from 0.
WS, HS — the size of the FFT array for horizontal and vertical processing.

✨ Examples ✨

High-pass — literally the classic:

fftfilt=dc_Y=128:weight_Y='squish(1-(Y+X)/100)'

Low-pass — the soft girl of filters:

fftfilt=dc_Y=0:weight_Y='squish((Y+X)/100-1)'

Sharpen — crisp, defined, iconic:

fftfilt=dc_Y=0:weight_Y='1+squish(1-(Y+X)/100)'

Blur — put it in a dream sequence, it's giving haze:

fftfilt=dc_Y=0:weight_Y='exp(-4 * ((Y+X)/(W+H)))'

💫 field

Extracts a single field from an interlaced image using stride arithmetic, so you're not out here wasting CPU time, bestie 💅 the output frames get marked as non-interlaced.

It takes this option:

type — pick whether to extract the top field (value 0 or "top") or the bottom field (value 1 or "bottom").

🎀 fieldhint

Creates new frames by copying the top and bottom fields from the surrounding frames — and those surrounding frames are supplied as numbers by a hint file. Obsessed with how specific that is.

hint — sets the file containing the hints: absolute/relative frame numbers.

There has to be one line for each frame in the clip 🌸 each line must contain two numbers separated by the comma, optionally followed by "-" or "+". The numbers on each line of the file can not be out of [N-1,N+1], where N is the current frame number for "absolute" mode, or out of the [-1, 1] range for "relative" mode. The first number tells it which frame to pick up the top field from, and the second number tells it which frame to pick up the bottom field from.

If optionally followed by "+" the output frame will be marked as interlaced; else if followed by "-" the output frame will be marked as progressive; else it'll be marked the same as the input frame. And if optionally followed by "t" the output frame will use only the top field, or in the case of "b" it will use only the bottom field 👑 if a line starts with "#" or ";", that line is skipped.

mode — can be "absolute" or "relative" or "pattern". Default is "absolute" ✨ the "pattern" mode is the same as "relative" mode, except at the last entry of the file, if there are more frames to process than the "hint" file, it seeks back to the start.

Example of the first several lines of a "hint" file for "relative" mode 🧚

0,0 - # first frame
1,0 - # second frame, use third's frame top field and second's frame bottom field
1,0 - # third frame, use fourth's frame top field and third's frame bottom field
1,0 -
0,0 -
0,0 -
1,0 -
1,0 -
1,0 -
0,0 -
0,0 -
1,0 -
1,0 -
1,0 -
0,0 -

🔥 fieldmatch

Field matching filter for inverse telecine 🎬 it's meant to reconstruct the progressive frames from a telecined stream. The filter does not drop duplicated frames, so to achieve a complete inverse telecine, fieldmatch needs to be followed by a decimation filter such as decimate in the filtergraph. Periodt.

The separation of the field matching and the decimation is notably motivated by the possibility of inserting a de-interlacing filter fallback between the two 👀 if the source has mixed telecined and real interlaced content, fieldmatch will not be able to match fields for the interlaced parts. But those leftover combed frames will be marked as interlaced, and so they can be de-interlaced by a later filter such as yadif before decimation. Slay.

In addition to the various configuration options, fieldmatch can take an optional second stream, activated through the ppsrc option 💖 if enabled, the frame reconstruction will be based on the fields and frames from this second stream. This lets the first input be pre-processed in order to help the filter's various algorithms, while keeping the output lossless (assuming the fields are matched properly). Typically a field-aware denoiser, or brightness/contrast adjustments, can help.

Heads up bestie: note that this filter uses the same algorithms as TIVTC/TFM (AviSynth project) and VIVTC/VFM (VapourSynth project). The latter is a light clone of TFM, which fieldmatch is based on. The semantics and usage are very close, but some behaviour and option names can differ. Don't @ me.

The decimate filter currently only works for constant frame rate input 😭 if your input has mixed telecined (30fps) and progressive content with a lower framerate like 24fps, use the following filterchain to produce the necessary cfr stream: dejudder,fps=30000/1001,fieldmatch,decimate.

The filter accepts these options:

order — specifies the assumed field order of the input stream 🦋 available values:
- auto — auto detect parity (uses FFmpeg's internal parity value).
- bff — assume bottom field first.
- tff — assume top field first.

Ngl it's sometimes recommended not to trust the parity announced by the stream.

Default value is auto.

mode — sets the matching mode or strategy to use 👑 pc mode is the safest in the sense that it won't risk creating jerkiness due to duplicate frames when possible, but if there are bad edits or blended fields it'll end up outputting combed frames when a good match might actually exist. On the other hand, pcn_ub mode is the most risky in terms of creating jerkiness, but it will almost always find a good frame if there is one. All the other values are somewhere in between pc and pcn_ub in terms of risking jerkiness and creating duplicate frames versus finding good matches in sections with bad edits, orphaned fields, blended fields, etc.

More details about p/c/n/u/b live in the p/c/n/u/b meaning section 💫

Available values:

pc — 2-way matching (p/c)
pc_n — 2-way matching, and trying 3rd match if still combed (p/c + n)
pc_u — 2-way matching, and trying 3rd match (same order) if still combed (p/c + u)
pc_n_ub — 2-way matching, trying 3rd match if still combed, and trying 4th/5th matches if still combed (p/c + n + u/b)
pcn — 3-way matching (p/c/n)
pcn_ub — 3-way matching, and trying 4th/5th matches if all 3 of the original matches are detected as combed (p/c/n + u/b)

The parenthesis at the end indicates the matches that would be used for that mode, assuming order=tff (and field on auto or top).

In terms of speed, pc mode is by far the fastest and pcn_ub is the slowest 🐌

Default value is pc_n.

ppsrc — marks the main input stream as a pre-processed input, and enables the secondary input stream as the clean source to pick the fields from ✨ see the filter introduction for more details. It's similar to the clip2 feature from VFM/TFM.

Default value is 0 (disabled).

field — sets the field to match from 🎀 it's recommended to set this to the same value as order unless you experience matching failures with that setting. In certain circumstances, changing the field that's used to match from can have a large impact on matching performance. Available values:
- auto — automatic (same value as order).
- bottom — match from the bottom field.
- top — match from the top field.

Default value is auto.

mchroma — sets whether or not chroma is included during the match comparisons 💅 in most cases it's recommended to leave this enabled. You should set this to 0 only if your clip has bad chroma problems such as heavy rainbowing or other artifacts. Setting this to 0 could also be used to speed things up, at the cost of some accuracy.

Default value is 1.

y0, y1 — these define an exclusion band which excludes the lines between y0 and y1 from being included in the field matching decision 🌸 an exclusion band can be used to ignore subtitles, a logo, or other things that may interfere with the matching. y0 sets the starting scan line and y1 sets the ending line; all lines in between y0 and y1 (including y0 and y1) will be ignored. Setting y0 and y1 to the same value disables the feature. y0 and y1 default to 0.
scthresh — sets the scene change detection threshold as a percentage of maximum change on the luma plane 🔥 good values are in the [8.0, 14.0] range. Scene change detection is only relevant in case combmatch=sc. The range for scthresh is [0.0, 100.0].

Default value is 12.0.

combmatch — when combmatch is not none, fieldmatch will take into account the combed scores of matches when deciding which match to use as the final match 👑 available values:
- none — no final matching based on combed scores.
- sc — combed scores are only used when a scene change is detected.
- full — use combed scores all the time.

Default is sc.

combdbg — forces fieldmatch to calculate the combed metrics for certain matches and print them 💀 this setting is known as micout in TFM/VFM vocabulary. Available values:
- none — no forced calculation.
- pcn — force p/c/n calculations.
- pcnub — force p/c/n/u/b calculations.

Default value is none.

cthresh — okay so this is the area combing threshold used for combed-frame detection. It basically controls how “strong” or “visible” combing has to be before it gets detected. Larger values mean the combing must be more visible; smaller values mean it can be less visible or strong and still get caught. Valid settings are from -1 (every pixel will be detected as combed) to 255 (no pixel will be detected as combed). It’s basically a pixel difference value. A good range is "[8, 12]".

Default value is 9. 💅

chroma — sets whether or not chroma is considered in the combed-frame decision. Only disable this if your source has chroma problems (rainbowing, etc.) that are causing problems for the combed-frame detection with chroma enabled. Actually, using chroma=0 is usually more reliable, except for the case where there is chroma-only combing in the source.

Default value is 0. ✨

blockx, blocky — respectively set the x-axis and y-axis size of the window used during combed-frame detection. This has to do with the size of the area in which combpel pixels are required to be detected as combed for a frame to be declared combed. See the combpel parameter description for more info. Possible values are any number that is a power of 2 starting at 4 and going up to 512.

Default value is 16. 🎀

combpel — the number of combed pixels inside any of the blocky by blockx size blocks on the frame for the frame to be detected as combed. While cthresh controls how “visible” the combing must be, this setting controls “how much” combing there must be in any localized area (a window defined by the blockx and blocky settings) on the frame. Minimum value is 0 and maximum is blocky x blockx (at which point no frames will ever be detected as combed). This setting is known as MI in TFM/VFM vocabulary.

Default value is 80. 🔥

p/c/n/u/b meaning 🧚

p/c/n

We assume the following telecined stream:

Top fields:     1 2 2 3 4
Bottom fields:  1 2 3 4 4

The numbers correspond to the progressive frame the fields relate to. Here, the first two frames are progressive, the 3rd and 4th are combed, and so on.

When fieldmatch is configured to run a matching from bottom (field=bottom) this is how this input stream gets transformed:

Input stream:
                T     1 2 2 3 4
                B     1 2 3 4 4   <-- matching reference

Matches:              c c n n c

Output stream:
                T     1 2 3 4 4
                B     1 2 3 4 4

As a result of the field matching, we can see that some frames get duplicated. To perform a complete inverse telecine, you need to rely on a decimation filter after this operation. See for instance the decimate filter.

The same operation now matching from top fields (field=top) looks like this:

Input stream:
                T     1 2 2 3 4   <-- matching reference
                B     1 2 3 4 4

Matches:              c c p p c

Output stream:
                T     1 2 2 3 4
                B     1 2 2 3 4

In these examples, we can see what p, c and n mean; basically, they refer to the frame and field of the opposite parity:

p — matches the field of the opposite parity in the previous frame
c — matches the field of the opposite parity in the current frame
n — matches the field of the opposite parity in the next frame

u/b

The u and b matching are a bit special in the sense that they match from the opposite parity flag. In the following examples, we assume that we are currently matching the 2nd frame (Top:2, bottom:2). According to the match, an 'x' is placed above and below each matched fields.

With bottom matching (field=bottom):

Match:           c         p           n          b          u

                 x       x               x        x          x
 Top          1 2 2     1 2 2       1 2 2      1 2 2      1 2 2
 Bottom       1 2 3     1 2 3       1 2 3      1 2 3      1 2 3
                 x         x           x        x              x

Output frames:
                 2          1          2          2          2
                 2          2          2          1          3

With top matching (field=top):

Match:           c         p           n          b          u

                 x         x           x        x              x
 Top          1 2 2     1 2 2       1 2 2      1 2 2      1 2 2
 Bottom       1 2 3     1 2 3       1 2 3      1 2 3      1 2 3
                 x       x               x        x          x

Output frames:
                 2          2          2          1          2
                 2          1          3          2          2

Examples

Simple IVTC of a top field first telecined stream:

fieldmatch=order=tff:combmatch=none, decimate

Advanced IVTC, with fallback on yadif for still combed frames:

fieldmatch=order=tff:combmatch=full, yadif=deint=interlaced, decimate

fieldorder 🌸

Transform the field order of the input video.

It accepts the following parameters:

order — the output field order. Valid values are tff for top field first or bff for bottom field first.

The default value is tff.

The transformation is done by shifting the picture content up or down by one line, and filling the remaining line with appropriate picture content. This method is consistent with most broadcast field order converters.

If the input video is not flagged as being interlaced, or it is already flagged as being of the required output field order, then this filter does not alter the incoming video.

It is very useful when converting to or from PAL DV material, which is bottom field first.

For example:

ffmpeg -i in.vob -vf "fieldorder=bff" out.dv

fillborders 🦋

Fill borders of the input video, without changing video stream dimensions. Sometimes video can have garbage at the four edges and you may not want to crop video input to keep size multiple of some number.

This filter accepts the following options:

left — number of pixels to fill from left border.
right — number of pixels to fill from right border.
top — number of pixels to fill from top border.
bottom — number of pixels to fill from bottom border.
mode — set fill mode. It accepts the following values:
- smear — fill pixels using outermost pixels
- mirror — fill pixels using mirroring (half sample symmetric)
- fixed — fill pixels with constant value
- reflect — fill pixels using reflecting (whole sample symmetric)
- wrap — fill pixels using wrapping
- fade — fade pixels to constant value
- margins — fill pixels at top and bottom with weighted averages pixels near borders

Default is smear.

color — set color for pixels in fixed or fade mode. Default is black.

Commands 💬

This filter supports same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value.

find_rect 🔎

Find a rectangular object in the input video.

The object to search for must be specified as a gray8 image specified with the object option.

For each possible match, a score is computed. If the score reaches the specified threshold, the object is considered found.

If the input video contains multiple instances of the object, the filter will find only one of them.

When an object is found, the following metadata entries are set in the matching frame:

lavfi.rect.w — width of object
lavfi.rect.h — height of object
lavfi.rect.x — x position of object
lavfi.rect.y — y position of object
lavfi.rect.score — match score of the found object

It accepts the following options:

object — filepath of the object image; needs to be in gray8.
threshold — detection threshold, expressed as a decimal number in the range 0-1. A threshold value of 0.01 means only exact matches, a threshold of 0.99 means almost everything matches. Default value is 0.5.
mipmaps — number of mipmaps, default is 3.
xmin, ymin, xmax, ymax — specifies the rectangle in which to search.
discard — discard frames where object is not detected. Default is disabled.

Examples

Cover a rectangular object by the supplied image of a given video using ffmpeg:

ffmpeg -i file.ts -vf find_rect=newref.pgm,cover_rect=cover.jpg:mode=cover new.mkv

Find the position of an object in each frame using ffprobe and write it to a log file:

ffprobe -f lavfi movie=test.mp4,find_rect=object=object.pgm:threshold=0.3 \
  -show_entries frame=pkt_pts_time:frame_tags=lavfi.rect.x,lavfi.rect.y \
  -of csv -o find_rect.csv

floodfill 🌊

Flood area with values of same pixel components with another values.

It accepts the following options:

x — set pixel x coordinate.
y — set pixel y coordinate.
s0 — set source #0 component value.
s1 — set source #1 component value.
s2 — set source #2 component value.
s3 — set source #3 component value.
d0 — set destination #0 component value.
d1 — set destination #1 component value.
d2 — set destination #2 component value.
d3 — set destination #3 component value.

format 🎨

Convert the input video to one of the specified pixel formats. Libavfilter will try to pick one that is suitable as input to the next filter.

It accepts the following parameters:

pix_fmts — a '|'-separated list of pixel format names, such as pix_fmts=yuv420p|monow|rgb24.
color_spaces — a '|'-separated list of color space names, such as color_spaces=bt709|bt470bg|bt2020nc.
color_ranges — a '|'-separated list of color range names, such as color_ranges=tv|pc.
alpha_modes — a '|'-separated list of color range names, such as alpha_modes=straight|premultiplied.

Examples

Convert the input video to the yuv420p format:

format=pix_fmts=yuv420p

Convert the input video to any of the formats in the list:

format=pix_fmts=yuv420p|yuv444p|yuv410p

🌸 fps 🌸

Okay so bestie, this one turns your video into a constant frame rate by duplicating or dropping frames whenever it needs to. It's giving "I will make your timeline behave." ✨

It accepts these parameters:

fps — the output frame rate you want. It takes expressions containing these constants, and yes you can use them:
- source_fps — the input's frame rate
- ntsc — NTSC frame rate of 30000/1001
- pal — PAL frame rate of 25.0
- film — film frame rate of 24.0
- ntsc_film — NTSC-film frame rate of 24000/1001

The default is 25.

start_time — assume the first PTS should be this value, in seconds. This lets you pad/trim at the start of the stream. By default no assumption is made about the first frame's expected PTS, so no padding or trimming happens. For example, you could set this to 0 to pad the beginning with duplicates of the first frame if a video stream starts after the audio stream, or to trim any frames with a negative PTS. 💅
round — the timestamp (PTS) rounding method. Possible values:
- zero — round towards 0
- inf — round away from 0
- down — round towards -infinity
- up — round towards +infinity
- near — round to nearest

The default is near.

eof_action — what gets performed when reading the last frame. Possible values:
- round — use the same timestamp rounding method as used for other frames.
- pass — pass through the last frame if input duration has not been reached yet.

The default is round.

Alternatively you can specify the options as a flat string, like fps[:start_time[:round]].

Also go peek at the setpts filter, they're neighbors 🫶

Examples

A typical usage in order to set the fps to 25:

fps=fps=25

Sets the fps to 24, using abbreviation and rounding method to round to nearest:

fps=fps=film:round=near

🦋 framepack 🦋

Pack two different video streams into a stereoscopic video, setting proper metadata on supported codecs. The two views should have the same size and framerate, and processing stops when the shorter video ends — so like, the shorter one is the vibe-setter, no cap. Pls note you can conveniently adjust view properties with the scale and fps filters. 💖

It accepts the following parameters:

format — the desired packing format. Supported values are:
- sbs — the views are next to each other (default).
- tab — the views are on top of each other.
- lines — the views are packed by line.
- columns — the views are packed by column.
- frameseq — the views are temporally interleaved.

Some examples:

# Convert left and right views into a frame-sequential video
ffmpeg -i LEFT -i RIGHT -filter_complex framepack=frameseq OUTPUT

# Convert views into a side-by-side video with the same output resolution as the input
ffmpeg -i LEFT -i RIGHT -filter_complex [0:v]scale=w=iw/2[left],[1:v]scale=w=iw/2[right],[left][right]framepack=sbs OUTPUT

⭐ framerate ⭐

Change the frame rate by interpolating new video output frames from the source frames. It's literally inventing frames, main character energy. 😭

Okay this is the important part: this filter is not designed to function correctly with interlaced media. If you want to change the frame rate of interlaced media then you are required to deinterlace before this filter and re-interlace after this filter. No skipping steps, bestie.

A description of the accepted options follows.

fps — specify the output frames per second. This option can also be specified as a value alone. The default is 50.
interp_start — specify the start of a range where the output frame will be created as a linear interpolation of two frames. The range is [0-255], the default is 15.
interp_end — specify the end of a range where the output frame will be created as a linear interpolation of two frames. The range is [0-255], the default is 240.
scene — specify the level at which a scene change is detected as a value between 0 and 100 to indicate a new scene; a low value reflects a low probability for the current frame to introduce a new scene, while a higher value means the current frame is more likely to be one. The default is 8.2.
flags — specify flags influencing the filter process. Available value for flags is:
- scene_change_detect, scd — enable scene change detection using the value of the option scene. This flag is enabled by default.

🎀 framestep 🎀

Select one frame every N-th frame. Snip snip ✂️

This filter accepts the following option:

step — select frame after every step frames. Allowed values are positive integers higher than 0. Default value is 1.

💀 freezedetect 💀

Detect frozen video. Yes, it's watching for when your video just... stops living. 🙈

This filter logs a message and sets frame metadata when it detects that the input video has no significant change in content during a specified duration. Video freeze detection calculates the mean average absolute difference of all the components of video frames and compares it to a noise floor.

The printed times and duration are expressed in seconds. The lavfi.freezedetect.freeze_start metadata key is set on the first frame whose timestamp equals or exceeds the detection duration, and it contains the timestamp of the first frame of the freeze. The lavfi.freezedetect.freeze_duration and lavfi.freezedetect.freeze_end metadata keys are set on the first frame after the freeze.

The filter accepts the following options:

noise, n — set noise tolerance. Can be specified in dB (in case dB is appended to the specified value) or as a difference ratio between 0 and 1. Default is -60dB, or 0.001.
duration, d — set freeze duration until notification (default is 2 seconds).

🧚 freezeframes 🧚

Freeze video frames. Ice ice baby 🧊

This filter freezes video frames using frame from 2nd input.

The filter accepts the following options:

first — set number of first frame from which to start freeze.
last — set number of last frame from which to end freeze.
replace — set number of frame from 2nd input which will be used instead of replaced frames.

🌈 frei0r 🌈

Apply a frei0r effect to the input video. This is the "plug in your favorite effects" filter and it ate, honestly.

To enable the compilation of this filter, you need to install the frei0r header and configure FFmpeg with --enable-frei0r.

It accepts the following parameters:

filter_name — the name of the frei0r effect to load. If the environment variable FREI0R_PATH is defined, the frei0r effect is searched for in each of the directories specified by the colon-separated list in FREI0R_PATH. Otherwise, the standard frei0r paths are searched, in this order: HOME/.frei0r-1/lib/, /usr/local/lib/frei0r-1/, /usr/lib/frei0r-1/.
filter_params — a '|'-separated list of parameters to pass to the frei0r effect.

A frei0r effect parameter can be a boolean (its value is either y or n), a double, a color (specified as R/G/B, where R, G, and B are floating point numbers between 0.0 and 1.0, inclusive) or a color description as specified in the Color section in the ffmpeg-utils manual, a position (specified as X/Y, where X and Y are floating point numbers) and/or a string.

The number and types of parameters depend on the loaded effect. If an effect parameter is not specified, the default value is set.

Examples

Apply the distort0r effect, setting the first two double parameters:

frei0r=filter_name=distort0r:filter_params=0.5|0.01

Apply the colordistance effect, taking a color as the first parameter:

frei0r=colordistance:0.2/0.3/0.4
frei0r=colordistance:violet
frei0r=colordistance:0x112233

Apply the perspective effect, specifying the top left and top right image positions:

frei0r=perspective:0.2/0.2|0.8/0.2

For more information, see http://frei0r.dyne.org

Commands

This filter supports the filter_params option as commands. 👑

💅 fspp 💅

Apply fast and simple postprocessing. It's a faster version of spp — same energy, less waiting.

It splits (I)DCT into horizontal/vertical passes. Unlike the simple post-processing filter, one of them is performed once per block, not per pixel. This allows for much higher speed. Slay. 🔥

The filter accepts the following options:

quality — set quality. This option defines the number of levels for averaging. It accepts an integer in the range 4-5. Default value is 4.
qp — force a constant quantization parameter. It accepts an integer in range 0-63. If not set, the filter will use the QP from the video stream (if available).
strength — set filter strength. It accepts an integer in range -15 to 32. Lower values mean more details but also more artifacts, while higher values make the image smoother but also blurrier. Default value is 0 − PSNR optimal.
use_bframe_qp — enable the use of the QP from the B-Frames if set to 1. Using this option may cause flicker since the B-Frames have often larger QP. Default is 0 (not enabled).

That flicker warning is real, bestie — don't come for me if you turn it on and things get jumpy.

🍓 fsync 🍓

Synchronize video frames with an external mapping from a file. It's giving "match my frames to the map, exactly." 🗺️

For each input PTS given in the map file it either drops or creates as many frames as necessary to recreate the sequence of output frames given in the map file.

This filter is useful to recreate the output frames of a framerate conversion by the fps filter, recorded into a map file using the ffmpeg option -stats_mux_pre, and do further processing to the corresponding frames e.g. quality comparison.

Each line of the map file must contain three items per input frame, the input PTS (decimal), the output PTS (decimal) and the output TIMEBASE (decimal/decimal), separated by a space. This file format corresponds to the output of -stats_mux_pre_fmt="{ptsi} {pts} {tb}".

The filter assumes the map file is sorted by increasing input PTS.

The filter accepts the following options:

file, f — the filename of the map file to be used.

Example

# Convert a video to 25 fps and record a MAP_FILE file with the default format of this filter
ffmpeg -i INPUT -vf fps=fps=25 -stats_mux_pre MAP_FILE -stats_mux_pre_fmt "{ptsi} {pts} {tb}" OUTPUT

# Sort MAP_FILE by increasing input PTS
sort -n MAP_FILE

# Use INPUT, OUTPUT and the MAP_FILE from above to compare the corresponding frames in INPUT and OUTPUT via SSIM
ffmpeg -i INPUT -i OUTPUT -filter_complex '[0:v]fsync=file=MAP_FILE[ref];[1:v][ref]ssim' -f null -

✨ gblur ✨

Apply Gaussian blur filter. Soft focus, dreamy, iconic. 🌸

The filter accepts the following options:

sigma — set horizontal sigma, standard deviation of Gaussian blur. Default is 0.5.
steps — how many steps you want for the Gaussian approximation. Default is 1.
planes — which planes get filtered. By default, all planes are filtered.
sigmaV — vertical sigma. If it's negative it just uses the same value as "sigma". Default is -1.

🎀 Commands

This filter supports the same commands as options, bestie. The command takes the exact same syntax as the corresponding option.

If the expression you fed it isn't valid, it just stays at its current value. No drama. 💅

🌈 geq

Apply a generic equation to each pixel. That's it, that's the filter. Math homework but make it ✨aesthetic✨.

The filter accepts the following options:

lum_expr, lum — set the luma expression.
cb_expr, cb — set the chrominance blue expression.
cr_expr, cr — set the chrominance red expression.
alpha_expr, a — set the alpha expression.
red_expr, r — set the red expression.
green_expr, g — set the green expression.
blue_expr, b — set the blue expression.

The colorspace gets picked based on whatever options you specified, like it's reading the room. If you specify one of lum_expr, cb_expr, or cr_expr, the filter automatically selects a YCbCr colorspace. If you specify one of red_expr, green_expr, or blue_expr, it selects an RGB colorspace instead.

If one of the chrominance expressions isn't defined, it falls back on the other one. If no alpha expression is specified, it evaluates to opaque. And if none of the chrominance expressions are specified at all, they all just evaluate to the luma expression. Iconic teamwork honestly. 🫶

The expressions can use the following variables and functions:

N — the sequential number of the filtered frame, starting from 0.
X, Y — the coordinates of the current sample.
W, H — the width and height of the image.
SW, SH — width and height scale depending on the currently filtered plane. It's the ratio between the corresponding luma plane's number of pixels and the current plane's ones. E.g. for YUV4:2:0 the values are "1,1" for the luma plane, and "0.5,0.5" for chroma planes.
T — time of the current frame, expressed in seconds.
p(x, y) — return the value of the pixel at location (x,y) of the current plane.
lum(x, y) — return the value of the pixel at location (x,y) of the luma plane.
cb(x, y) — return the value of the pixel at location (x,y) of the blue-difference chroma plane. Returns 0 if there's no such plane.
cr(x, y) — return the value of the pixel at location (x,y) of the red-difference chroma plane. Returns 0 if there's no such plane.
r(x, y), g(x, y), b(x, y) — return the value of the pixel at location (x,y) of the red/green/blue component. Returns 0 if there's no such component.
alpha(x, y) — return the value of the pixel at location (x,y) of the alpha plane. Returns 0 if there's no such plane.
psum(x,y), lumsum(x, y), cbsum(x,y), crsum(x,y), rsum(x,y), gsum(x,y), bsum(x,y), alphasum(x,y) — sum of sample values in the rectangle from (0,0) to (x,y). This lets you get sums of samples within a rectangle. See the functions without the sum postfix for the version you're prob used to.
interpolation — set one of the interpolation methods:
- nearest, n
- bilinear, b

Default is bilinear.

For functions, if x and y are outside the area, the value gets automatically clipped to the closer edge. No rage-quitting on out-of-bounds, it just snaps to the nearest edge like a good little function.

Please note that this filter can use multiple threads, in which case each slice has its own expression state. If you want only a single expression state because your expressions depend on previous state, then you should limit the number of filter threads to 1. fr fr, don't skip this one.

✨ Examples ✨

Flip the image horizontally:

geq=p(W-X\,Y)

Generate a bidimensional sine wave, with angle "PI/3" and a wavelength of 100 pixels:

geq=128 + 100*sin(2*(PI/100)*(cos(PI/3)*(X-50*T) + sin(PI/3)*Y)):128:128

Generate a fancy enigmatic moving light:

nullsrc=s=256x256,geq=random(1)/hypot(X-cos(N*0.07)*W/2-W/2\,Y-sin(N*0.09)*H/2-H/2)^2*1000000*sin(N*0.02):128:128

Generate a quick emboss effect:

format=gray,geq=lum_expr='(p(X,Y)+(256-p(X-4,Y-4)))/2'

Modify RGB components depending on pixel position:

geq=r='X/W*r(X,Y)':g='(1-X/W)*g(X,Y)':b='(H-Y)/H*b(X,Y)'

Create a radial gradient that is the same size as the input (also see the vignette filter):

geq=lum=255*gauss((X/W-0.5)*3)*gauss((Y/H-0.5)*3)/gauss(0)/gauss(0),format=gray

🦋 gradfun

Fixes those banding artifacts that sometimes show up in nearly flat regions when everything gets truncated to 8-bit color depth. It interpolates the gradients that should go where the bands are, and dithers them. Basically it does damage control for bit-depth trauma 😭✨.

It's designed for playback only. Do not use it prior to lossy compression, because compression tends to lose the dither and bring back the bands. That's the whole point, bestie, don't undo it.

It accepts the following parameters:

strength — the maximum amount by which the filter will change any one pixel. This is also the threshold for detecting nearly flat regions. Acceptable values range from .51 to 64; the default value is 1.2. Out-of-range values get clipped to the valid range.
radius — the neighborhood to fit the gradient to. A larger radius makes for smoother gradients, but also prevents the filter from modifying the pixels near detailed regions. Acceptable values are 8-32; the default value is 16. Out-of-range values get clipped to the valid range.

Alternatively, the options can be specified as a flat string: strength[:radius]

✨ Examples ✨

Apply the filter with a 3.5 strength and radius of 8:

gradfun=3.5:8

Specify radius, omitting the strength (which falls back to the default value):

gradfun=radius=8

👑 graphmonitor

Shows you various filtergraph stats. It's literally the main character of debugging.

With this filter you can debug a complete filtergraph. Especially issues with links filling up with queued frames. So when things are getting sus and backing up, this is your girl. 💅

The filter accepts the following options:

size, s — set video output size. Default is hd720.
opacity, o — set video opacity. Default is 0.9. Allowed range is from 0 to 1.
mode, m — set output mode flags.

Available values for flags are:

full — no any filtering. Default.
compact — show only filters with queued frames.
nozero — show only filters with non-zero stats.
noeof — show only filters with non-eof stat.
nodisabled — show only filters that are enabled in timeline.

flags, f — set flags which enable which stats are shown in video.

Available values for flags are:

none — all flags turned off.
all — all flags turned on.
queue — display number of queued frames in each link.
frame_count_in — display number of frames taken from filter.
frame_count_out — display number of frames given out from filter.
frame_count_delta — display delta number of frames between above two values.
pts — display current filtered frame pts.
pts_delta — display pts delta between current and previous frame.
time — display current filtered frame time.
time_delta — display time delta between current and previous frame.
timebase — display time base for filter link.
format — display used format for filter link.
size — display video size or number of audio channels in case of audio used by filter link.
rate — display video frame rate or sample rate in case of audio used by filter link.
eof — display link output status.
sample_count_in — display number of samples taken from filter.
sample_count_out — display number of samples given out from filter.
sample_count_delta — display delta number of samples between above two values.
disabled — show the timeline filter status.

rate, r — set upper limit for video rate of output stream. Default value is 25. This guarantees that output video frame rate will not be higher than this value. No cap, it's capped. 💀

🌸 grayworld

A color constancy filter that applies color correction based on the grayworld assumption. It's giving "the average of the world is gray" and honestly? It works.

See: https://www.researchgate.net/publication/275213614_A_New_Color_Correction_Method_for_Underwater_Imaging

The algorithm uses linear light, so input data should be linearized beforehand (and possibly correctly tagged). Don't skip this step, bestie, it matters.

ffmpeg -i INPUT -vf zscale=transfer=linear,grayworld,zscale=transfer=bt709,format=yuv420p OUTPUT

🧚 greyedge

A color constancy variation filter which estimates scene illumination via the grey edge algorithm and corrects the scene colors accordingly. Same energy as grayworld but it went to a different school.

See: https://staff.science.uva.nl/th.gevers/pub/GeversTIP07.pdf

The filter accepts the following options:

difford — the order of differentiation to be applied on the scene. Must be chosen in the range [0,2] and default value is 1.
minknorm — the Minkowski parameter to be used for calculating the Minkowski distance. Must be chosen in the range [0,20] and default value is 1. Set to 0 for getting max value instead of calculating Minkowski distance.
sigma — the standard deviation of Gaussian blur to be applied on the scene. Must be chosen in the range [0,1024.0] and default value = 1. floor( sigma * break_off_sigma(3) ) can't be equal to 0 if difford is greater than 0.

✨ Examples ✨

Grey Edge:

greyedge=difford=1:minknorm=5:sigma=2

Max Edge:

greyedge=difford=1:minknorm=0:sigma=2

🍓 guided

Apply a guided filter for edge-preserving smoothing, dehazing and so on. A whole little multitool, honestly obsessed.

The filter accepts the following options:

radius — set the box radius in pixels. Allowed range is 1 to 20. Default is 3.
eps — set regularization parameter (with square). Allowed range is 0 to 1. Default is 0.01.
mode — set filter mode. Can be "basic" or "fast". Default is "basic".
sub — set subsampling ratio for "fast" mode. Range is 2 to 64. Default is 4. No subsampling occurs in "basic" mode.
guidance — set guidance mode. Can be "off" or "on". Default is "off". If "off", single input is required. If "on", two inputs of the same resolution and pixel format are required. The second input serves as the guidance. Like a little buddy holding your hand 🫶
planes — set planes to filter. Default is first only.

🎀 Commands

This filter supports the all above options as commands. Yup, all of them.

✨ Examples ✨

Edge-preserving smoothing with guided filter:

ffmpeg -i in.png -vf guided out.png

Dehazing, structure-transferring filtering, detail enhancement with guided filter. For the generation of guidance image, refer to paper "Guided Image Filtering". See: http://kaiminghe.com/publications/pami12guidedfilter.pdf.

ffmpeg -i in.png -i guidance.png -filter_complex guided=guidance=on out.png

haldclut 💅

Apply a Hald CLUT to a video stream. <sparkle>✨</sparkle>

First input is the video stream to process, and second one is the Hald CLUT. The Hald CLUT input can be a simple picture or a complete video stream.

The filter accepts the following options:

clut — Set which CLUT video frames will be processed from second input stream, can be first or all. Default is all.
shortest — Force termination when the shortest input terminates. Default is 0.
repeatlast — Continue applying the last CLUT after the end of the stream. A value of 0 disable the filter after the last frame of the CLUT is reached. Default is 1.

haldclut also has the same interpolation options as lut3d (both filters share the same internals).

This filter also supports the framesync options.

More information about the Hald CLUT can be found on Eskil Steenberg's website (Hald CLUT author) at http://www.quelsolaar.com/technology/clut.html.

Commands

This filter supports the interp option as commands. 🫶

Workflow examples

Hald CLUT video stream

Generate an identity Hald CLUT stream altered with various effects:

ffmpeg -f lavfi -i B<haldclutsrc>=8 -vf "hue=H=2*PI*t:s=sin(2*PI*t)+1, curves=cross_process" -t 10 -c:v ffv1 clut.nut

Note: make sure you use a lossless codec. 💖

Then use it with haldclut to apply it on some random stream:

ffmpeg -f lavfi -i mandelbrot -i clut.nut -filter_complex '[0][1] haldclut' -t 20 mandelclut.mkv

The Hald CLUT will be applied to the 10 first seconds (duration of clut.nut), then the latest picture of that CLUT stream will be applied to the remaining frames of the mandelbrot stream.

Hald CLUT with preview

A Hald CLUT is supposed to be a squared image of Level*Level*Level by Level*Level*Level pixels. For a given Hald CLUT, FFmpeg will select the biggest possible square starting at the top left of the picture. The remaining padding pixels (bottom or right) will be ignored. This area can be used to add a preview of the Hald CLUT.

Typically, the following generated Hald CLUT will be supported by the haldclut filter:

ffmpeg -f lavfi -i B<haldclutsrc>=8 -vf "
   pad=iw+320 [padded_clut];
   smptebars=s=320x256, split [a][b];
   [padded_clut][a] overlay=W-320:h, curves=color_negative [main];
   [main][b] overlay=W-320" -frames:v 1 clut.png

It contains the original and a preview of the effect of the CLUT: SMPTE color bars are displayed on the right-top, and below the same color bars processed by the color changes.

Then, the effect of this Hald CLUT can be visualized with:

ffplay input.mkv -vf "movie=clut.png, [in] haldclut"

hflip 🦋

Flip the input video horizontally.

For example, to horizontally flip the input video with ffmpeg:

ffmpeg -i in.avi -vf "hflip" out.avi

histeq 🌈

This filter applies a global color histogram equalization on a per-frame basis.

It can be used to correct video that has a compressed range of pixel intensities. The filter redistributes the pixel intensities to equalize their distribution across the intensity range. It may be viewed as an "automatically adjusting contrast filter". This filter is useful only for correcting degraded or poorly captured source video. fr fr

The filter accepts the following options:

strength — Determine the amount of equalization to be applied. As the strength is reduced, the distribution of pixel intensities more-and-more approaches that of the input frame. The value must be a float number in the range [0,1] and defaults to 0.200.
intensity — Set the maximum intensity that can generated and scale the output values appropriately. The strength should be set as desired and then the intensity can be limited if needed to avoid washing-out. The value must be a float number in the range [0,1] and defaults to 0.210.
antibanding — Set the antibanding level. If enabled the filter will randomly vary the luminance of output pixels by a small amount to avoid banding of the histogram. Possible values are none, weak or strong. It defaults to none.

histogram ✨

Compute and draw a color distribution histogram for the input video.

The computed histogram is a representation of the color component distribution in an image.

Standard histogram displays the color components distribution in an image. Displays color graph for each color component. Shows distribution of the Y, U, V, A or R, G, B components, depending on input format, in the current frame. Below each graph a color component scale meter is shown.

The filter accepts the following options:

level_height — Set height of level. Default value is 200. Allowed range is [50, 2048].
scale_height — Set height of color scale. Default value is 12. Allowed range is [0, 40].
display_mode — Set display mode. It accepts the following values:
- stack — Per color component graphs are placed below each other.
- parade — Per color component graphs are placed side by side.
- overlay — Presents information identical to that in the parade, except that the graphs representing color components are superimposed directly over one another.

Default is stack.

levels_mode — Set mode. Can be either linear, or logarithmic. Default is linear.
components — Set what color components to display. Default is 7.
fgopacity — Set foreground opacity. Default is 0.7.
bgopacity — Set background opacity. Default is 0.5.
colors_mode — Set colors mode. It accepts the following values:
- whiteonblack
- blackonwhite
- whiteongray
- blackongray
- coloronblack
- coloronwhite
- colorongray
- blackoncolor
- whiteoncolor
- grayoncolor

Default is whiteonblack.

Examples

Calculate and draw histogram:

ffplay -i input -vf histogram

hqdn3d 💜

This is a high precision/quality 3d denoise filter. It aims to reduce image noise, producing smooth images and making still images really still. It should enhance compressibility. <sparkle>✨</sparkle>

It accepts the following optional parameters:

luma_spatial — A non-negative floating point number which specifies spatial luma strength. It defaults to 4.0.
chroma_spatial — A non-negative floating point number which specifies spatial chroma strength. It defaults to 3.0*luma_spatial/4.0.
luma_tmp — A floating point number which specifies luma temporal strength. It defaults to 6.0*luma_spatial/4.0.
chroma_tmp — A floating point number which specifies chroma temporal strength. It defaults to luma_tmp*chroma_spatial/luma_spatial.

Commands

This filter supports same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value.

hwdownload 🫶

Download hardware frames to system memory.

The input must be in hardware frames, and the output a non-hardware format. Not all formats will be supported on the output - it may be necessary to insert an additional format filter immediately following in the graph to get the output in a supported format.

hwmap 🧚

Map hardware frames to system memory or to another device.

This filter has several different modes of operation; which one is used depends on the input and output formats:

Hardware frame input, normal frame output — Map the input frames to system memory and pass them to the output. If the original hardware frame is later required (for example, after overlaying something else on part of it), the hwmap filter can be used again in the next mode to retrieve it.
Normal frame input, hardware frame output — If the input is actually a software-mapped hardware frame, then unmap it - that is, return the original hardware frame.

Otherwise, a device must be provided. Create new hardware surfaces on that device for the output, then map them back to the software format at the input and give those frames to the preceding filter. This will then act like the hwupload filter, but may be able to avoid an additional copy when the input is already in a compatible format.

Hardware frame input and output — A device must be supplied for the output, either directly or with the derive_device option. The input and output devices must be of different types and compatible - the exact meaning of this is system-dependent, but typically it means that they must refer to the same underlying hardware context (for example, refer to the same graphics card).

If the input frames were originally created on the output device, then unmap to retrieve the original frames.

Otherwise, map the frames to the output device - create new hardware frames on the output corresponding to the frames on the input.

The following additional parameters are accepted:

mode — Set the frame mapping mode. Some combination of:
- read — The mapped frame should be readable.
- write — The mapped frame should be writeable.
- overwrite — The mapping will always overwrite the entire frame.

This may improve performance in some cases, as the original contents of the frame need not be loaded.

direct — The mapping must not involve any copying.

Indirect mappings to copies of frames are created in some cases where either direct mapping is not possible or it would have unexpected properties. Setting this flag ensures that the mapping is direct and will fail if that is not possible.

Defaults to read+write if not specified.

derive_device type — Rather than using the device supplied at initialisation, instead derive a new device of type type from the device the input frames exist on.
reverse — In a hardware to hardware mapping, map in reverse - create frames in the sink and map them back to the source. This may be necessary in some cases where a mapping in one direction is required but only the opposite direction is supported by the devices being used.

⚠️ Heads up bestie — this option is dangerous. It may break the preceding filter in undefined ways if there are any additional constraints on that filter's output. Do not use it without fully understanding the implications of its use. Like, fr fr. 🙈

hwupload 💅

Uploads system memory frames onto hardware surfaces. Glow-up but for your frames, no cap.

The device you're uploading to must be supplied when the filter is initialised. If you're using ffmpeg, pick the right device with the -filter_hw_device option or with the derive_device option. The input and output devices have to be of different types and compatible — the exact meaning of that is system-dependent, but typically it means they must refer to the same underlying hardware context (for example, refer to the same graphics card). iykyk 🫶

The following additional parameters are accepted:

derive_device type — rather than using the device supplied at initialisation, instead derive a new device of type type from the device the input frames exist on. ✨

hwupload_cuda 🦋

Uploads system memory frames to a CUDA device. Same energy, NVIDIA-flavoured.

It accepts the following optional parameters:

device — the number of the CUDA device to use. 🔥

hqx 🎀

Applies a high-quality magnification filter designed for pixel art. This filter was originally created by Maxim Stepin — absolute icon behavior, we stan. 👑

It accepts the following option:

n — set the scaling dimension: 2 for hq2x, 3 for hq3x and 4 for hq4x. Default is 3. 💜

hstack 🌈

Stacks input videos horizontally. Side by side, bestie, like a group photo.

All streams must be of same pixel format and of same height. Non-negotiable.

Note that this filter is faster than using overlay and pad filter to create same output. Efficiency queen. ⭐

The filter accepts the following option:

inputs — set number of input streams. Default is 2.
shortest — if set to 1, force the output to terminate when the shortest input terminates. Default value is 0.

hsvhold 🍓

Turns a certain HSV range into gray values.

This filter measures color difference between set HSV color in options and ones measured in video stream. Depending on options, output colors can be changed to be gray or not. It's giving selective desaturation main character energy. 💅

The filter accepts the following options:

hue — set the hue value which will be used in color difference calculation. Allowed range is from -360 to 360. Default value is 0.
sat — set the saturation value which will be used in color difference calculation. Allowed range is from -1 to 1. Default value is 0.
val — set the value which will be used in color difference calculation. Allowed range is from -1 to 1. Default value is 0.
similarity — set similarity percentage with the key color. Allowed range is from 0 to 1. Default value is 0.01.
- 0.00001 matches only the exact key color, while 1.0 matches everything. literally everything. 😭
blend — blend percentage. Allowed range is from 0 to 1. Default value is 0.
- 0.0 makes pixels either fully gray, or not gray at all.
- Higher values result in more gray pixels, with a higher gray pixel the more similar the pixels color is to the key color.

hsvkey 🧚

Turns a certain HSV range into transparency.

This filter measures color difference between set HSV color in options and ones measured in video stream. Depending on options, output colors can be changed to transparent by adding alpha channel. Yes she adds the alpha channel herself, we love a self-starter. ✨

The filter accepts the following options:

hue — set the hue value which will be used in color difference calculation. Allowed range is from -360 to 360. Default value is 0.
sat — set the saturation value which will be used in color difference calculation. Allowed range is from -1 to 1. Default value is 0.
val — set the value which will be used in color difference calculation. Allowed range is from -1 to 1. Default value is 0.
similarity — set similarity percentage with the key color. Allowed range is from 0 to 1. Default value is 0.01.
- 0.00001 matches only the exact key color, while 1.0 matches everything.
blend — blend percentage. Allowed range is from 0 to 1. Default value is 0.
- 0.0 makes pixels either fully transparent, or not transparent at all.
- Higher values result in semi-transparent pixels, with a higher transparency the more similar the pixels color is to the key color.

hue 🌸

Modifies the hue and/or the saturation of the input. The classic. The moment. The blueprint.

It accepts the following parameters:

h — specify the hue angle as a number of degrees. It accepts an expression, and defaults to "0".
s — specify the saturation in the [-10,10] range. It accepts an expression and defaults to "1".
H — specify the hue angle as a number of radians. It accepts an expression, and defaults to "0".
b — specify the brightness in the [-10,10] range. It accepts an expression and defaults to "0".

Bestie, h and H are mutually exclusive — you can't specify both at the same time. Pick a lane 💀

The b, h, H and s option values are expressions containing the following constants:

n — frame count of the input frame starting from 0
pts — presentation timestamp of the input frame expressed in time base units
r — frame rate of the input video, NAN if the input frame rate is unknown
t — timestamp expressed in seconds, NAN if the input timestamp is unknown
tb — time base of the input video

Examples 🎀

Set the hue to 90 degrees and the saturation to 1.0:

hue=h=90:s=1

Same command but expressing the hue in radians:

hue=H=PI/2:s=1

Rotate hue and make the saturation swing between 0 and 2 over a period of 1 second:

hue="H=2*PI*t: s=sin(2*PI*t)+1"

Apply a 3 seconds saturation fade-in effect starting at 0:

hue="s=min(t/3\,1)"

The general fade-in expression can be written as:

hue="s=min(0\, max((t-START)/DURATION\, 1))"

Apply a 3 seconds saturation fade-out effect starting at 5 seconds:

hue="s=max(0\, min(1\, (8-t)/3))"

The general fade-out expression can be written as:

hue="s=max(0\, min(1\, (START+DURATION-t)/DURATION))"

Commands 💅

This filter supports the following commands:

b
s
h
H

Modify the hue and/or the saturation and/or brightness of the input video. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value. No crash, no drama, just vibes. 🫶

huesaturation 🦋

Applies hue-saturation-intensity adjustments to input video stream.

This filter operates in RGB colorspace. Just so you know where you are, bestie.

This filter accepts the following options:

hue — set the hue shift in degrees to apply. Default is 0. Allowed range is from -180 to 180.
saturation — set the saturation shift. Default is 0. Allowed range is from -1 to 1.
intensity — set the intensity shift. Default is 0. Allowed range is from -1 to 1.
colors — set which primary and complementary colors are going to be adjusted. This options is set by providing one or multiple values. This can select multiple colors at once. By default all colors are selected.
- r — adjust reds.
- y — adjust yellows.
- g — adjust greens.
- c — adjust cyans.
- b — adjust blues.
- m — adjust magentas.
- a — adjust all colors.
strength — set strength of filtering. Allowed range is from 0 to 100. Default value is 1.
rw, gw, bw — set weight for each RGB component. Allowed range is from 0 to 1. By default is set to 0.333, 0.334, 0.333. Those options are used in saturation and lightess processing.
lightness — set preserving lightness, by default is disabled. Adjusting hues can change lightness from original RGB triplet, with this option enabled lightness is kept at same value. Protect the lightness at all costs 🔥

hysteresis ⭐

Grow first stream into second stream by connecting components. This makes it possible to build more robust edge masks. Lowkey the glow-up your edges have been waiting for.

This filter accepts the following options:

planes — set which planes will be processed as bitmap, unprocessed planes will be copied from first stream. By default value 0xf, all planes will be processed.
threshold — set threshold which is used in filtering. If pixel component value is higher than this value filter algorithm for connecting components is activated. By default value is 0.

The hysteresis filter also supports the framesync options. 💜

iccdetect 🧚

Detect the colorspace from an embedded ICC profile (if present), and update the frame's tags accordingly. Detective mode: activated. 🕵️‍♀️

This filter accepts the following options:

force — if true, the frame's existing colorspace tags will always be overridden by values detected from an ICC profile. Otherwise, they will only be assigned if they contain unknown. Enabled by default.

iccgen 🎀

Generate ICC profiles and attach them to frames. She's making them from scratch, we love an artisan. ✨

This filter accepts the following options:

color_primaries
color_trc

Configure the colorspace that the ICC profile will be generated for. The default value of "auto" infers the value from the input frame's metadata, defaulting to BT.709/sRGB as appropriate.

See the setparams filter for a list of possible values, but note that "unknown" are not valid values for this filter. 🙅‍♀️

force — if true, an ICC profile will be generated even if it would overwrite an already existing ICC profile. Disabled by default.

identity 😭

Obtains the identity score between two input videos. Basically a beauty contest, but for frames, and it's measuring how similar they are.

This filter takes two input videos.

Both input videos must have the same resolution and pixel format for this filter to work correctly. Also it assumes that both inputs have the same number of frames, which are compared one by one. No mixing and matching, bestie.

The obtained per component, average, min and max identity score is printed through the logging system.

The filter stores the calculated identity scores of each frame in frame metadata. Data girlie 💅

This filter also supports the framesync options.

In the below example the input file main.mpg being processed is compared with the reference file ref.mpg.

ffmpeg -i main.mpg -i ref.mpg -lavfi identity -f null -

idet 🔍

Detects video interlacing type.

This filter tries to detect if the input frames are interlaced, progressive, top or bottom field first. It will also try to detect fields that are repeated between adjacent frames (a sign of telecine). She's nosy and we love that for her. 🌸

Single frame detection considers only immediately adjacent frames when classifying each frame. Multiple frame detection incorporates the classification history of previous frames. The way that it remembers the past… iconic.

The filter will log these metadata values:

single.current_frame — detected type of current frame using single-frame detection. One of: `tff` (top field first), `bff` (bottom field first), `progressive`, or `undetermined`. The last one is the "idk she's giving mysterious" option. 💀
single.tff — cumulative count of frames the filter clocked as top field first using single-frame detection 🎀
multiple.tff — same thing but via multiple-frame detection, so the count of top-field-first frames the multi-frame way ✨
single.bff — cumulative count of frames detected as bottom field first, single-frame detection 🦋
multiple.current_frame — the vibe check on the current frame, determined with multiple-frame detection. It's one of: tff (top field first), bff (bottom field first), progressive, or undetermined 💅
multiple.bff — cumulative count of bottom-field-first frames found via multiple-frame detection 🌸
single.progressive — cumulative count of frames detected as progressive, single-frame detection 🫶
multiple.progressive — cumulative count of frames detected as progressive, multiple-frame detection 💖
single.undetermined — cumulative count of frames that single-frame detection just could NOT classify (mystery frames, we stan) 😭
multiple.undetermined — same energy, but the frames multiple-frame detection couldn't classify 💀
repeated.current_frame — which field in the current frame got repeated from the last one. One of neither, top, or bottom 👑
repeated.neither — cumulative count of frames with zero repeated fields ✨
repeated.top — cumulative count of frames where the top field is a repeat of the previous frame's top field 🌈
repeated.bottom — cumulative count of frames where the bottom field is repeated from the previous frame's bottom field ⭐

The filter accepts the following options:

intl_thres — sets the interlacing threshold 🎚️
prog_thres — sets the progressive threshold 🎚️
rep_thres — the threshold for repeated field detection 🔁
half_life — number of frames after which a given frame's contribution to the statistics gets halved (like, it only contributes 0.5 to its classification now). The default of 0 means all frames seen are given full weight of 1.0 forever, no decay, main character energy periodt 💅
analyze_interlaced_flag — when this isn't 0, idet uses the specified number of frames to figure out if the interlaced flag is actually accurate. It will not count undetermined frames. If the flag turns out accurate, it's used with no further computations; if it turns out inaccurate, it gets cleared with no further computations. This is the whole point: it lets you drop idet in as a low-computation way to clean up the interlaced flag 🧚

✨ Examples ✨

Wanna inspect the field order of the first 360 frames in a video, in verbose detail? Say less:

ffmpeg -i INPUT -filter:v idet,metadata=mode=print -frames:v 360 -an -f null -

The idet filter adds analysis metadata to each frame, which then gets discarded (rip metadata, you served your purpose 🙈). At the end, the filter also prints a final report with all the statistics. Slay.

🌀 `il`

Deinterleave or interleave fields. Iconic.

This filter lets you process interlaced image fields without deinterlacing them. Deinterleaving splits the input frame into 2 fields (the so-called half pictures). Odd lines get moved to the top half of the output image, even lines to the bottom half. Then you can process (filter) them independently and re-interleave them after 💖

The filter accepts the following options:

luma_mode, l
chroma_mode, c
alpha_mode, a

Available values for luma_mode, chroma_mode and alpha_mode are:

none — do nothing. Literally nothing. A vibe.
deinterleave, d — deinterleave fields, placing one above the other.
interleave, i — interleave fields. Reverses the effect of deinterleaving.

Default value is none.

luma_swap, ls
chroma_swap, cs
alpha_swap, as

Swap the luma/chroma/alpha fields. Exchanges even & odd lines. Default value is 0 🦋

Commands

This filter supports the all above options as commands 👑

💫 `inflate`

Apply the inflate effect to the video.

This filter replaces the pixel with the local (3x3) average, but only counting values higher than the pixel. So it puffs up the bright bits, basically ✨

It accepts the following options:

threshold0
threshold1
threshold2
threshold3

These limit the maximum change for each plane. Default is 65535. If 0, that plane remains unchanged, untouched, unbothered 🫶

Commands

This filter supports the all above options as commands.

🎀 `interlace`, `interlace_vulkan`

Simple interlacing filter from progressive contents. This interleaves upper (or lower) lines from odd frames with lower (or upper) lines from even frames — halving the frame rate and preserving image height. The way that math works out is genuinely so satisfying ngl 😭

     Original        Original             New Frame
     Frame 'j'      Frame 'j+1'             (tff)
    ==========      ===========       ==================
      Line 0  -------------------->    Frame 'j' Line 0
      Line 1          Line 1  ---->   Frame 'j+1' Line 1
      Line 2 --------------------->    Frame 'j' Line 2
      Line 3          Line 3  ---->   Frame 'j+1' Line 3
       ...             ...                   ...
  New Frame + 1 will be generated by Frame 'j+2' and Frame 'j+3' and so on

It accepts the following optional parameters:

scan — determines whether the interlaced frame is taken from the even (tff — default) or odd (bff) lines of the progressive frame 👀
lowpass — vertical lowpass filter to avoid twitter interlacing and reduce moire patterns 🌀
- 0, off — disable vertical lowpass filter
- 1, linear — enable linear filter (default)
- 2, complex — enable complex filter. This reduces twitter and moire slightly less, but better retains detail and the subjective impression of sharpness. Trade-offs, bestie 💅

🔥 `kerndeint`

Deinterlace input video by applying Donald Graft's adaptive kernel deinterling (that's the spelling in the original, we don't gatekeep 🫶). Works on the interlaced parts of a video to produce progressive frames.

The description of the accepted parameters follows.

thresh — set the threshold which affects the filter's tolerance when determining if a pixel line must be processed. It must be an integer in the range [0,255] and defaults to 10. A value of 0 will result in applying the process on every pixel. Every single one. No cap 💀
map — Paint pixels exceeding the threshold value to white if set to 1. Default is 0.
order — set the fields order. Swap fields if set to 1, leave fields alone if 0. Default is 0.
sharp — enable additional sharpening if set to 1. Default is 0.
twoway — enable twoway sharpening if set to 1. Default is 0.

Examples

Apply default values:

kerndeint=thresh=10:map=0:order=0:sharp=0:twoway=0

Enable additional sharpening:

kerndeint=sharp=1

Paint processed pixels in white:

kerndeint=map=1

🌈 `kirsch`

Apply the kirsch operator to the input video stream. Edge detection but make it fashion ✨

The filter accepts the following option:

planes — set which planes will be processed; unprocessed planes get copied. By default the value is 0xf, so all planes are processed 👑
scale — set the value which gets multiplied with the filtered result 🔢
delta — set the value which gets added to the filtered result ➕

Commands

This filter supports the all above options as commands.

💜 `lagfun`

Slowly update darker pixels.

This filter makes short flashes of light appear longer. It's giving afterimage, it's giving light trails, we're obsessed 🦋

This filter accepts the following options:

decay — set the factor for decaying. Default is .95. Allowed range is from 0 to 1 🍓
planes — set which planes to filter. Default is all. Allowed range is from 0 to 15.

Commands

This filter supports the all above options as commands.

🧚 `lenscorrection`

Correct radial lens distortion. Glow-up for your wide angle footage, lowkey.

This filter can be used to correct radial distortion as can result from the use of wide angle lenses, and thereby re-rectify the image. To find the right parameters one can use tools available for example as part of opencv or simply trial-and-error. To use opencv use the calibration sample (under samples/cpp) from the opencv sources and extract the k1 and k2 coefficients from the resulting matrix 📐

Note that effectively the same filter is available in the open-source tools Krita and Digikam from the KDE project. Snaps for them 💅

In contrast to the vignette filter, which can also be used to compensate lens errors, this filter corrects the distortion of the image, whereas vignette corrects the brightness distribution. So you may want to use both filters together in certain cases, though you'll have to take care of ordering, i.e. whether vignetting should be applied before or after lens correction. Order matters, bestie, always 👀

Options

The filter accepts the following options:

cx — relative x-coordinate of the focal point of the image, and thereby the center of the distortion. This value has a range [0,1] and is expressed as fractions of the image width. Default is 0.5 🎯
cy — relative y-coordinate of the focal point of the image, and thereby the center of the distortion. This value has a range [0,1] and is expressed as fractions of the image height. Default is 0.5 🎯
k1 — coefficient of the quadratic correction term. This value has a range [-1,1]. 0 means no correction. Default is 0.
k2 — coefficient of the double quadratic correction term. This value has a range [-1,1]. 0 means no correction. Default is 0.
i — set the interpolation type. Can be nearest or bilinear. Default is nearest.
fc — specify the color of the unmapped pixels. For the syntax of this option, check the Color section in the ffmpeg-utils manual. Default color is black@0 🖤

The formula that generates the correction is:

r_src = r_tgt * (1 + k1 * (r_tgt / r_0)^2 + k2 * (r_tgt / r_0)^4)

where r_0 is halve of the image diagonal and r_src and r_tgt are the distances from the focal point in the source and target images, respectively. Math ate, honestly 🧮

Commands

This filter supports the all above options as commands.

🌸 `lensfun`

Apply lens correction via the lensfun library (http://lensfun.sourceforge.net/).

The lensfun filter requires the camera make, camera model, and lens model to apply the lens correction. The filter will load the lensfun database and query it to find the corresponding camera and lens entries in the database. As long as these entries can be found with the given options, the filter can perform corrections on frames. Note that incomplete strings will result in the filter choosing the best match with the given options, and the filter will output the chosen camera and lens models (logged with level info). You must provide the make, camera model, and lens model as they are required — non-negotiable, bestie 🙈

Okay bestie, before we even get into the options: if you leave out one or both of make and model, the filter will just yeet the entire list of available makes and models into the log at level "INFO" — first column is the make, second column is the model, no cap. Want the lens list instead? Set any values you like for make and model and leave out lens_model, and the full list of lenses gets dropped into the log at level "INFO". Literally iconic. One giant heads-up though: the ffmpeg tool will exit after the list is printed, so don't be surprised when it dips out on you 💅

The filter accepts the following options:

make — the make of the camera (for example, "Canon"). This option is required, no skipping it 🫶
model — the model of the camera (for example, "Canon EOS 100D"). Also required, periodt.
lens_model — the model of the lens (for example, "Canon EF-S 18-55mm f/3.5-5.6 IS STM"). Required too, all three of them, the whole squad.
db_path — the full path to the lens database folder. If you don't set it, the filter tries to load the database from the install path it was built with at library-compile time. Default is unset.
mode — the type of correction to apply. The valid values are:
- vignetting — fixes lens vignetting ✨
- geometry — fixes lens geometry. This is the default, iykyk.
- subpixel — fixes chromatic aberrations.
- vig_geo — fixes lens vignetting and lens geometry together.
- vig_subpixel — fixes lens vignetting and chromatic aberrations.
- distortion — fixes both lens geometry and chromatic aberrations.
- all — enables all possible corrections. all of them. the whole buffet 🔥
focal_length — the focal length of the image/video (the zoom; expected to be constant for video). Like, an 18--55mm lens has a focal length range of [18--55], so pick a value inside that range when you're using that lens. Default 18.
aperture — the aperture of the image/video (again, expected constant for video). Note that aperture is only used for vignetting correction. Default 3.5.
focus_distance — the focus distance of the image/video (expected constant for video). Note that focus distance is only used for vignetting, and it only slightly affects the vignetting correction process. If you don't know it, just leave it at the default value, which is 1000.
scale — the scale factor applied after the transformation. After correction the video isn't necessarily rectangular anymore, so this parameter controls how much of the resulting image is visible. A value of 0 means a value gets chosen automatically so there's little or no unmapped area in the output image. 1.0 means no additional scaling at all. Lower values may show more of the corrected image, higher values may avoid unmapped areas in the output. Balance, bestie 🧚
target_geometry — the target geometry of the output image/video. Valid values:
- rectilinear (default)
- fisheye
- panoramic
- equirectangular
- fisheye_orthographic
- fisheye_stereographic
- fisheye_equisolid
- fisheye_thoby
reverse — apply the reverse of image correction, meaning instead of correcting distortion, you apply it. Going backwards, we love a plot twist 🙈
interpolation — the type of interpolation used when correcting distortion. Valid values:
- nearest
- linear (default)
- lanczos

✨ Examples ✨

Apply lens correction with make "Canon", camera model "Canon EOS 100D", and lens model "Canon EF-S 18-55mm f/3.5-5.6 IS STM", with a focal length of "18" and aperture of "8.0":

ffmpeg -i input.mov -vf lensfun=make=Canon:model="Canon EOS 100D":lens_model="Canon EF-S 18-55mm f/3.5-5.6 IS STM":focal_length=18:aperture=8 -c:v h264 -b:v 8000k output.mov

Same thing as before, but only for the first 5 seconds of video:

ffmpeg -i input.mov -vf lensfun=make=Canon:model="Canon EOS 100D":lens_model="Canon EF-S 18-55mm f/3.5-5.6 IS STM":focal_length=18:aperture=8:enable='lte(t\,5)' -c:v h264 -b:v 8000k output.mov

💜 `lcevc`

Low Complexity Enhancement Video Codec filter, built on liblcevc_dec (<https://github.com/v-novaltd/LCEVCdec>). That's the whole description, she's short and mysterious, main character energy fr.

🦋 `libplacebo`

Flexible GPU-accelerated processing filter, built on libplacebo (<https://code.videolan.org/videolan/libplacebo>). GPU go brrr.

✨ Options ✨

The options for this filter are split into the following sections:

🌈 Output mode 🌈

These options control the overall output mode. By default, libplacebo tries to preserve the source colorimetry and size as best as it can — but it will apply any embedded film grain, dolby vision metadata, or anamorphic SAR present in the source frames. Like, it's doing its best, okay?

inputs — set the number of inputs. Alongside the idx variable, this lets you place/blend multiple inputs inside the output frame. Effectively, it gives you functionality similar to hstack, overlay, etc. Slay 🎀
w, h — set the output video dimension expression. Default values are iw and ih. Allows for the same expressions as the scale filter, bestie.
crop_x, crop_y — set the input crop x/y expressions. Default values are (iw-cw)/2 and (ih-ch)/2.
crop_w, crop_h — set the input crop width/height expressions. Default values are iw and ih.
pos_x, pos_y — set the output placement x/y expressions. Default values are (ow-pw)/2 and (oh-ph)/2.
pos_w, pos_h — set the output placement width/height expressions. Default values are ow and oh.
rotate — rotate the input frame clockwise by the specified angle. The angles are:
- 0, 360
- 90
- 180
- 270
fps — set the output frame rate. This can be rational, e.g. 60000/1001. If you set it to the special string "none" (which is the default), input timestamps get passed through to the output unmodified instead. Otherwise, the input video frames are interpolated as necessary to rescale the video to the specified target framerate, in a manner determined by the frame_mixer option.
format — set the output format override. If unset (the default), frames are output in the same format as the respective input frames. Otherwise, format conversion is performed.
force_original_aspect_ratio, force_divisible_by — these work exactly the same as the identical scale filter options. Note that force_divisible_by also works with fit_sense=constraint 👑
reset_sar — if enabled, output frames always have a pixel aspect ratio of 1:1. If disabled (the default), any aspect ratio mismatches — including those coming from, like, anamorphic video sources — get forwarded to the output pixel aspect ratio.
normalize_sar — like reset_sar, except instead of stretching the video content to fill the new output aspect ratio, the content is padded or cropped as necessary. Mutually exclusive with fit_mode. Disabled by default.
pad_crop_ratio — specifies a ratio (between 0.0 and 1.0) between padding and cropping, for when the input aspect ratio doesn't match the output aspect ratio and normalize_sar is in effect. The default of 0.0 always pads the content with black borders, while 1.0 always crops off parts of the content. Intermediate values are possible and give you a mix of the two approaches. Yes queen, compromise 💅
fit_mode — specify the content fit strategy from a list of predefined modes. It determines how the input image is placed inside the destination crop rectangle (as defined by pos_x/y and pos_w/h). The names and their implementations are taken from the CSS 'object-fit' property. Note that this option is mutually exclusive with normalize_sar. Defaults to "fill". Valid values:
- fill — stretch the input to the output rectangle, ignoring aspect ratio mismatches. Note that unless reset_sar is also enabled, the output will still have the correct pixel aspect ratio tagged.
- contain — scale the input to fit inside the output, preserving aspect ratio by padding. Equivalent to normalize_sar with pad_crop_ratio set to 0.0.
- cover — scale the input to fill the output, preserving aspect ratio by cropping. Equivalent to normalize_sar with pad_crop_ratio set to 1.0.
- none, place — don't scale the input. It gets placed inside the output rectangle at its natural size, which may result in additional padding or cropping.
- scale_down — scale the input down as much as needed to fit inside the output. Equivalent to either "contain" or "none", depending on whether the input is larger than the output or not.
fit_sense — when fit_mode is in use, this option controls how the fit strategy is applied against the specified output resolution. Mutually exclusive with force_original_aspect_ratio. Valid values:
- target — the computed output resolution is taken as the exact size of the output frame. This is the default behavior.
- constraint — the computed output resolution is a size reference against which the fit mode is applied, enlarging or decreasing the true frame size as needed to fit the content.
fillcolor — set the color used to fill the output area not covered by the output image, for example as a result of normalize_sar. For the general syntax of this option, check the "Color" section in the ffmpeg-utils manual. Defaults to "black@0".
corner_rounding — render frames with rounded corners 🎀 The value, given as a float ranging from 0.0 to 1.0, indicates the relative degree of rounding, from fully square to fully circular. In other words, it gives the radius divided by half the smaller side length. Defaults to 0.0.
lut — specifies a custom LUT (in Adobe .cube format) to apply to the colors as part of color conversion. The exact interpretation depends on the value of lut_type.
lut_type — controls the interpretation of color values fed to and from the LUT specified as lut. Valid values:
- auto — chooses the interpretation of the LUT automatically from tagged metadata, and otherwise falls back to native. (Default)
- native — applied to raw image contents in its native RGB colorspace (non-linear light), before conversion to the output color space.
- normalized — applied to the normalized RGB image contents, in linear light, before conversion to the output color space.
conversion — this one fully replaces the conversion from the image color space to the output color space. Like, whole thing, swapped out. And if a LUT is present it has the highest priority — it overrides any ICC profiles, and it also overrides the options related to tone mapping and output colorimetry (color_primaries, color_trc). It's giving main character energy, it wins every argument 💅
extra_opts — pass extra libplacebo internal configuration options through here. You write them as a list of key=value pairs separated by :. The example below shows how to configure a custom filter kernel ("EWA LanczosSharp") and then use it to double the input image resolution:

-vf "libplacebo=w=iw*2:h=ih*2:extra_opts='upscaler=custom\:upscaler_preset=ewa_lanczos\:upscaler_blur=0.9812505644269356'"

shader_cache — file path of a cache directory that libplacebo will use to store and load cached shader objects. Important vibes: this cache is not cleaned up automatically, bestie. Nobody's coming to tidy it. Also, if the path does not end in a directory separator, the generated filenames will be effectively prefixed by the last path component. And all the directories must already exist — it will not make them for you.

-vf "libplacebo=shader_cache=/tmp/pl-shader-"

colorspace
color_primaries
color_trc
range
chroma_location — configure the colorspace that output frames will be delivered in. The default value of "auto" outputs frames in the same format as the input frames, which literally means no change at all. For any other value, conversion will be performed.

See the setparams filter for a list of possible values 👀

alpha_mode — choose the desired output alpha mode, for when the output format has an alpha channel. See the setparams filter for a list of possible values.
apply_filmgrain — apply film grain (like AV1 or H.274) if it's present in the source frames, and strip it from the output. Enabled by default ✨
apply_dolbyvision — apply Dolby Vision RPU metadata if it's present in the source frames, and strip it from the output. Also enabled by default. Heads up though: Dolby Vision will always output BT.2020+PQ, overriding the usual input frame metadata. Those get picked as the values of "auto" for the respective frame output options too.

So, on top of the expression constants documented for the scale filter, the crop_w, crop_h, crop_x, crop_y, pos_w, pos_h, pos_x and pos_y options can also contain these constants:

in_idx, idx — the (0-based) numeric index of the currently active input stream.
crop_w, cw — the computed values of crop_w and crop_h.
crop_h, ch — the computed values of crop_w and crop_h.
pos_w, pw — the computed values of pos_w and pos_h.
pos_h, ph — the computed values of pos_w and pos_h.
in_t, t — the input frame timestamp, in seconds. NAN if the input timestamp is unknown.
out_t, ot — the input frame timestamp, in seconds. NAN if the input timestamp is unknown.
n — the input frame number, starting with 0.

🌈 Scaling 🌈

The options in this section control how libplacebo performs upscaling and (if necessary) downscaling. And okay, super important detail so listen up 🫶 — libplacebo will always internally operate on 4:4:4 content, so any sub-sampled chroma formats like "yuv420p" will necessarily be upsampled and downsampled as part of the rendering process. Which means scaling might be in effect even if the source and destination resolution are the same 😭 that's just how it is.

upscaler
downscaler — configure the filter kernel used for upscaling and downscaling. The respective defaults are "spline36" and "mitchell". For a full list of possible values, pass "help" to these options. The most important values are:
- none — forces the use of built-in GPU texture sampling (typically bilinear). Extremely fast but poor quality, especially when downscaling. Fast and ugly, we've all been there 💀
- bilinear — bilinear interpolation. Can generally be done for free on GPUs, except when doing so would lead to aliasing. Fast and low quality.
- nearest — nearest-neighbour interpolation. Sharp but highly aliasing.
- oversample — an algorithm that looks visually similar to nearest-neighbour interpolation but tries to preserve pixel aspect ratio. Good for pixel art, since it results in minimal distortion of the artistic appearance 🎨 iconic for the pixel artists fr fr
- lanczos — standard sinc-sinc interpolation kernel.
- spline36 — cubic spline approximation of lanczos. No difference in performance, but has very slightly less ringing.
- ewa_lanczos — elliptically weighted average version of lanczos, based on a jinc-jinc kernel. This is also popularly referred to as just "Jinc scaling". Slow but very high quality. Queen behaviour 👑
- gaussian — gaussian kernel. Has certain ideal mathematical properties, but subjectively very blurry. Math says yes, eyes say no.
- mitchell — cubic BC spline with parameters recommended by Mitchell and Netravali. Very little ringing.
frame_mixer — controls the kernel used for mixing frames temporally. The default value is "none", which disables frame mixing. For a full list of possible values, pass "help" to this option. The most important values are:
- none — disables frame mixing, giving a result equivalent to "nearest neighbour" semantics.
- oversample — oversamples the input video to create a "Smooth Motion"-type effect: if an output frame would exactly fall on the transition between two video frames, it is blended according to the relative overlap. This is the recommended option whenever preserving the original subjective appearance is desired.
- mitchell_clamp — a larger filter kernel that smoothly interpolates multiple frames in a manner designed to eliminate ringing and other artefacts as much as possible. This is the recommended option wherever maximum visual smoothness is desired.
- linear — linear blend/fade between frames. Especially useful for constructing e.g. slideshows 🖼️
antiringing — enables anti-ringing (for non-EWA filters) — nay, slays ringing. The value (between 0.0 and 1.0) configures the strength of the anti-ringing algorithm. May increase aliasing if set too high, so don't get greedy. Disabled by default.
sigmoid — enable sigmoidal compression during upscaling. Reduces ringing slightly. Enabled by default.

🦋 Deinterlacing 🦋

Deinterlacing is automatically supported when frames are tagged as interlaced, however — and this is the sneaky part — frames are not deinterlaced unless a deinterlacing algorithm is chosen. So tagging alone gets you nothing, bestie.

deinterlace — the deinterlacing algorithm to use.
- weave — no deinterlacing, weave fields together into a single frame. This is the default.
- bob — naive bob deinterlacing, simply repeat each field line twice.
- yadif — yet another deinterlacing filter. See the yadif filter for more details.
- bwdif — bob weaver deinterlacing filter. See the bwdif filter for more details.
skip_spatial_check — skip the spatial deinterlacing check when using "yadif" deinterlacing.
send_fields — output a frame for each field, rather than for each frame. Note that this will always double the tagged output frame rate, even if the input does not contain any interlaced frames. Disabled by default.

💜 Debanding 💜

Libplacebo comes with a built-in debanding filter that is genuinely good at counteracting many common sources of banding and blocking. Turning this on is highly recommended whenever quality is desired — like, why wouldn't you 🔥

deband — enable the (fast) debanding algorithm. Disabled by default.
deband_iterations — number of deband iterations of the debanding algorithm. Each iteration is performed with progressively increased radius (and diminished threshold). Recommended values are in the range 1 to 4. Defaults to 1.
deband_threshold — debanding filter strength. Higher numbers lead to more aggressive debanding. Defaults to 4.0.
deband_radius — debanding filter radius. A higher radius is better for slow gradients, while a lower radius is better for steep gradients. Defaults to 16.0.
deband_grain — amount of extra output grain to add. Helps hide imperfections — a little texture never hurt anybody. Defaults to 6.0.

🎀 Color adjustment 🎀

A collection of subjective color controls. Not very rigorous, so the exact effect will vary somewhat depending on the input primaries and colorspace. Basically it's vibes-based color grading, and that's fine.

brightness — brightness boost, between -1.0 and 1.0. Defaults to 0.0.
contrast — contrast gain, between 0.0 and 16.0. Defaults to 1.0.
saturation — saturation gain, between 0.0 and 16.0. Defaults to 1.0.
hue — hue shift in radians, between -3.14 and 3.14. Defaults to 0.0. This will rotate the UV subvector, defaulting to BT.709 coefficients for RGB inputs.
gamma — gamma adjustment, between 0.0 and 16.0. Defaults to 1.0.
temperature — color temperature adjustment. Lower values make the output warmer/redder, down to 1667, while higher values make the output cooler/bluer, up to 25000. Defaults to 6500 (neutral white).
cones — cone model to use for color blindness simulation. Accepts any combination of "l", "m" and "s". Here are some examples:
- m — deuteranomaly / deuteranopia (affecting 3%-4% of the population)
- l — protanomaly / protanopia (affecting 1%-2% of the population)
- l+m — monochromacy (very rare)
- l+m+s — achromatopsy (complete loss of daytime vision, extremely rare)
cone-strength — gain factor for the cones specified by "cones", between 0.0 and 10.0. A value of 1.0 results in no change to color vision. A value of 0.0 (the default) simulates complete loss of those cones. Values above 1.0 result in exaggerating the differences between cones, which may help compensate for reduced color vision.

⭐ Peak detection ⭐

To help deal with sources that only have static HDR10 metadata (or no tagging whatsoever, we've all met those files 😭), libplacebo uses its own internal frame analysis compute shader to analyze source frames and adapt the tone mapping function in realtime. If this is too slow, or if you need exactly reproducible frame-perfect results, it's recommended to turn this feature off. Trade-offs, babe.

peak_detect — enable HDR peak detection. Ignores static MaxCLL/MaxFALL values in favor of dynamic detection from the input. Note that the detected values do not get written back to the output frames — they merely guide the internal tone mapping process. Enabled by default.
smoothing_period — the peak-detection smoothing period, and it lives between 0.0 and 1000.0 bestie. Higher values mean peak detection gets all chill and less responsive to changes in the input. Defaults to 20.0. 💅
scene_threshold_low, scene_threshold_high — the lower and upper thresholds for scene change detection, expressed on a logarithmic scale between 0.0 and 100.0. They default to 1.0 and 3.0 respectively. Set either one to a negative value and this whole feature just ✨ turns off ✨ — no cap.
percentile — which percentile of the frame brightness histogram gets used as the source peak for tone-mapping. Defaults to 99.995, which is a fairly conservative value (she's careful, she's sensible). Setting this to 100.0 disables frame histogram measurement entirely and instead uses the true peak brightness for tone-mapping. 🎀

🌈 Tone mapping 🌈

The options in this section control how libplacebo does tone-mapping and gamut-mapping when it's dealing with mismatches between wide-gamut or HDR content. Lowkey, libplacebo relies on accurate source tagging and mastering display gamut information to produce the best results — so like, tag your stuff properly, iykyk.

gamut_mode — how to handle out-of-gamut colors that can occur as a result of colorimetric gamut mapping. The values are:
- clip — do nothing, just clip out-of-range colors to the RGB volume. Low quality but extremely fast, it's giving "we have no time and we're fine with that".
- perceptual — perceptually soft-clip colors to the gamut volume. This is the default. 💖
- relative — relative colorimetric hard-clip. Similar to perceptual but without the soft knee.
- saturation — saturation mapping, maps primaries directly to primaries in RGB space. Not recommended except for artificial computer graphics where a bright, saturated display is what you want.
- absolute — absolute colorimetric hard-clip. Performs no adjustment of the white point.
- desaturate — hard-desaturates out-of-gamut colors towards white while preserving the luminance. Has a tendency to distort the visual appearance of bright objects 🫣.
- darken — linearly reduces content brightness to preserve saturated details, followed by clipping the remaining out-of-gamut colors.
- warn — highlight out-of-gamut pixels (by inverting/marking them). It's the "hey bestie look here" mode 👀.
- linear — linearly reduces chromaticity of the entire image to make it fit within the target color volume. Be careful when using this on BT.2020 sources without proper mastering metadata, because doing so will lead to excessive desaturation. That's a real warning, not a vibe.
tonemapping — the tone-mapping algorithm to use. Available values:
- auto — automatic selection based on internal heuristics. This is the default. 🧚
- clip — performs no tone-mapping, just clips out-of-range colors. Retains perfect color accuracy for in-range colors but completely destroys out-of-range information. Does not perform any black point adaptation. Not configurable.
- st2094-40 — EETF from SMPTE ST 2094-40 Annex B, which applies the Bezier curves from HDR10+ dynamic metadata based on Bezier curves to perform tone-mapping. The OOTF used is adjusted based on the ratio between the targeted and actual display peak luminances.
- st2094-10 — EETF from SMPTE ST 2094-10 Annex B.2, which takes into account the input signal average luminance in addition to the maximum/minimum. The configurable contrast parameter influences the slope of the linear output segment, defaulting to 1.0 for no increase/decrease in contrast. Note that this does not currently include the subjective gain/offset/gamma controls defined in Annex B.3 — so like, don't go looking for them, they're not here 🙈.
- bt.2390 — EETF from the ITU-R Report BT.2390, a hermite spline roll-off with linear segment. The knee point offset is configurable. Note that this parameter defaults to 1.0, rather than the value of 0.5 from the ITU-R spec.
- bt.2446a — EETF from ITU-R Report BT.2446, method A. Designed for well-mastered HDR sources. Can be used for both forward and inverse tone mapping. Not configurable.
- spline — simple spline consisting of two polynomials, joined by a single pivot point. The parameter gives the pivot point (in PQ space), defaulting to 0.30. Can be used for both forward and inverse tone mapping.
- reinhard — simple non-linear, global tone mapping algorithm. The parameter specifies the local contrast coefficient at the display peak. Essentially, a parameter of 0.5 implies that the reference white will be about half as bright as when clipping. Defaults to 0.5, which results in the simplest formulation of this function.
- mobius — generalization of the reinhard tone mapping algorithm to support an additional linear slope near black. The tone mapping parameter indicates the trade-off between the linear section and the non-linear section. Essentially, for a given parameter x, every color value below x will be mapped linearly, while higher values get non-linearly tone-mapped. Values near 1.0 make this curve behave like clip, while values near 0.0 make this curve behave like reinhard. The default value is 0.3, which provides a good balance between colorimetric accuracy and preserving out-of-gamut details. She's balanced, she's iconic 👑.
- hable — piece-wise, filmic tone-mapping algorithm developed by John Hable for use in Uncharted 2, inspired by a similar tone-mapping algorithm used by Kodak. Popularized by its use in video games with HDR rendering. Preserves both dark and bright details very well, but comes with the drawback of changing the average brightness quite significantly. This is sort of similar to reinhard with parameter 0.24.
- gamma — fits a gamma (power) function to transfer between the source and target color spaces, effectively resulting in a perceptual hard-knee joining two roughly linear sections. This preserves details at all scales fairly accurately but can result in an image with a muted or dull appearance. The parameter is used as the cutoff point, defaulting to 0.5.
- linear — linearly stretches the input range to the output range, in PQ space. This will preserve all details accurately but results in a significantly different average brightness. Can be used for inverse tone-mapping in addition to regular tone-mapping. The parameter can be used as an additional linear gain coefficient (defaulting to 1.0).
tonemapping_param — for tunable tone mapping functions, this parameter can be used to fine-tune the curve behavior. Refer to the documentation of tonemapping. The default value of 0.0 is replaced by the curve's preferred default setting.
inverse_tonemapping — if enabled, this filter will also attempt stretching SDR signals to fill HDR output color volumes. Disabled by default.
tonemapping_lut_size — size of the tone-mapping LUT, between 2 and 1024. Defaults to 256. Note that this figure is squared when combined with peak_detect — so like, do the math bestie, it grows fast 📈.
contrast_recovery — contrast recovery strength. If set to a value above 0.0, the source image will be divided into high-frequency and low-frequency components, and a portion of the high-frequency image is added back onto the tone-mapped output. May cause excessive ringing artifacts for some HDR sources, but can improve the subjective sharpness and detail left over in the image after tone-mapping. Defaults to 0.30.
contrast_smoothness — contrast recovery lowpass kernel size. Defaults to 3.5. Increasing or decreasing this will affect the visual appearance substantially. Has no effect when contrast_recovery is disabled. Fr fr, zero effect.

🦋 Dithering 🦋

By default, libplacebo will dither whenever necessary, which includes rendering to any integer format below 16-bit precision. It's recommended to always leave this on, since not doing so may result in visible banding in the output, even if the debanding filter is enabled. If maximum performance is needed, use ordered_fixed instead of disabling dithering. Leave it on, bestie, she's doing important work 💅

dithering — the dithering method to use. Accepts the following values:
- none — disables dithering completely. May result in visible banding 😭.
- blue — dither with pseudo-blue noise. This is the default. 💙
- ordered — tunable ordered dither pattern.
- ordered_fixed — faster ordered dither with a fixed size of 6. Texture-less.
- white — dither with white noise. Texture-less.
dither_lut_size — dither LUT size, as log base2 between 1 and 8. Defaults to 6, corresponding to a LUT size of 64x64.
dither_temporal — enables temporal dithering. Disabled by default.

✨ Custom shaders ✨

libplacebo supports a number of custom shaders based on the mpv .hook GLSL syntax. A collection of such shaders can be found here: <https://github.com/mpv-player/mpv/wiki/User-Scripts#user-shaders>

A full description of the mpv shader format is beyond the scope of this section, but a summary can be found here: <https://mpv.io/manual/master/#options-glsl-shader>

custom_shader_path — specifies a path to a custom shader file to load at runtime.
custom_shader_bin — specifies a complete custom shader as a raw string.

💀 Debugging / performance 💀

All of the options in this section default off. They may be of assistance when attempting to squeeze the maximum performance at the cost of quality — it's giving "speed over beauty" and honestly, sometimes that's the move.

skip_aa — disable anti-aliasing when downscaling.
disable_linear — disable linear light scaling.
disable_builtin — disable built-in GPU sampling (forces LUT).
disable_fbos — forcibly disable FBOs, resulting in loss of almost all functionality, but offering the maximum possible speed. Like yeah, it's fast, but at what cost 😭

🎀 Commands 🎀

This filter supports almost all of the above options as commands. Slay.

🌸 Examples 🌸

Tone-map input to standard gamut BT.709 output:

libplacebo=colorspace=bt709:color_primaries=bt709:color_trc=bt709:range=tv

Rescale input to fit into standard 1080p, with high quality scaling:

libplacebo=w=1920:h=1080:force_original_aspect_ratio=decrease:normalize_sar=true:upscaler=ewa_lanczos:downscaler=ewa_lanczos

Interpolate low FPS / VFR input to smoothed constant 60 fps output:

libplacebo=fps=60:frame_mixer=mitchell_clamp

Turn your input into a bog-standard sRGB JPEG — like, give it the whole yuv420p / bt470bg / bt709 / iec61966-2-1 / pc treatment:

libplacebo=format=yuv420p:colorspace=bt470bg:color_primaries=bt709:color_trc=iec61966-2-1:range=pc

Wanna deband like you actually mean it? crank those settings, bestie:

libplacebo=deband=true:deband_iterations=3:deband_radius=8:deband_threshold=6

Run this filter on the CPU — this is for systems with Mesa installed, and yeah you gotta turn off the most expensive options or it'll be tragic:

ffmpeg ... -init_hw_device vulkan:llvmpipe ... -vf libplacebo=upscaler=none:downscaler=none:peak_detect=false

Suppress that CPU-based AV1/H.274 film grain application in the decoder and let this filter handle it instead. But real talk: this is only a gain if the frames are already chillin' on the GPU, or if you're using libplacebo for other stuff anyway — otherwise the VRAM roundtrip will more than offset any speedup you were hoping for. 💀

ffmpeg -export_side_data +film_grain ... -vf libplacebo=apply_filmgrain=true

Interop with VAAPI hwdec so you don't have to round-trip through RAM (literally the worst commute):

ffmpeg -init_hw_device vulkan -hwaccel vaapi -hwaccel_output_format vaapi ... -vf libplacebo

🌈 libvmaf

Okay so this one calculates the VMAF (Video Multi-Method Assessment Fusion) score for a reference/distorted pair of input videos. Fancy quality math, we love to see it. ✨

Heads up on the order: the first input is the distorted video, and the second input is the reference video. Don't flip 'em, iykyk.

The VMAF score you get is printed through the logging system.

It needs Netflix's vmaf library (libvmaf) as a pre-requisite. Once you've installed the library you can enable it using: ./configure --enable-libvmaf.

The filter has the following options:

model — a | delimited list of vmaf models. Each model can be configured with a number of parameters. Default value: version=vmaf_v0.6.1
feature — a | delimited list of features. Each feature can be configured with a number of parameters.
log_path — set the file path to be used to store log files.
log_fmt — set the format of the log file (xml, json, csv, or sub).
pool — set the pool method to be used for computing vmaf. Options are min, harmonic_mean or mean (default).
n_threads — set number of threads to be used when initializing libvmaf. Default value: 0, no threads.
n_subsample — set frame subsampling interval to be used.

This filter also supports the framesync options.

Examples

In the examples below, a distorted video distorted.mpg is compared with a reference file reference.mpg.
Basic usage — the lil starter pack:

ffmpeg -i distorted.mpg -i reference.mpg -lavfi libvmaf=log_path=output.xml -f null -

Example with multiple models — more models, more opinions:

ffmpeg -i distorted.mpg -i reference.mpg -lavfi libvmaf='model=version=vmaf_v0.6.1\\:name=vmaf|version=vmaf_v0.6.1neg\\:name=vmaf_neg' -f null -

Example with multiple additional features — collector era:

ffmpeg -i distorted.mpg -i reference.mpg -lavfi libvmaf='feature=name=psnr|name=ciede' -f null -

Example with options and different containers — yes the second file is .mkv, that's the whole point, look at the filtergraph:

ffmpeg -i distorted.mpg -i reference.mkv -lavfi "[0:v]settb=AVTB,setpts=PTS-STARTPTS[main];[1:v]settb=AVTB,setpts=PTS-STARTPTS[ref];[main][ref]libvmaf=log_fmt=json:log_path=output.json" -f null -

🦋 libvmaf_cuda

This is the CUDA variant of the libvmaf filter. It only accepts CUDA frames — no mixing, no matching, no exceptions. 🙈

It requires Netflix's vmaf library (libvmaf) as a pre-requisite. After installing the library it can be enabled using: ./configure --enable-nonfree --enable-ffnvcodec --enable-libvmaf.

Examples

Basic usage showing CUVID hardware decoding and CUDA scaling with scale_cuda — the whole glow-up pipeline:

ffmpeg \
    -hwaccel cuda -hwaccel_output_format cuda -codec:v av1_cuvid -i dis.obu \
    -hwaccel cuda -hwaccel_output_format cuda -codec:v av1_cuvid -i ref.obu \
    -filter_complex "
        [0:v]scale_cuda=format=yuv420p[dis]; \
        [1:v]scale_cuda=format=yuv420p[ref]; \
        [dis][ref]libvmaf_cuda=log_fmt=json:log_path=output.json
    " \
    -f null -

💅 limitdiff

Apply a limited difference filter using the second and optionally third video stream. It's giving "compare and choose", honestly.

The filter accepts the following options:

threshold — set the threshold to use when allowing certain differences between video streams. Any absolute difference value lower or exact than this threshold will pick pixel components from the first video stream.
elasticity — set the elasticity of soft thresholding when processing video streams. This value multiplied with the first one sets the second threshold. Any absolute difference value greater or exact than the second threshold will pick pixel components from the second video stream. For values between those two thresholds, linear interpolation between the first and second video stream will be used.
reference — enable the reference (third) video stream processing. By default it's disabled. If set, this video stream will be used for calculating absolute difference with the first video stream.
planes — specify which planes will be processed. Defaults to all available.

Commands

This filter supports all of the above options as commands except the option reference. (Not everything gets to be a command, and that's okay.)

✨ limiter

Limits the pixel component values to the specified range [min, max]. Boundaries, but make it pixels. 🎀

The filter accepts the following options:

min — lower bound. Defaults to the lowest allowed value for the input.
max — upper bound. Defaults to the highest allowed value for the input.
planes — specify which planes will be processed. Defaults to all available.

Commands

This filter supports all of the above options as commands.

🌀 loop

Loop video frames. Round and round and round we go.

The filter accepts the following options:

loop — set the number of loops. Setting this value to -1 will result in infinite loops. Default is 0.
size — set maximal size in number of frames. Default is 0.
start — set first frame of loop. Default is 0.
time — set the time of loop start in seconds. Only used if the option named start is set to -1.

Examples

Loop the single first frame infinitely — the iconic eternal vibe:

loop=loop=-1:size=1:start=0

Loop the single first frame 10 times — a manageable amount of obsession:

loop=loop=10:size=1:start=0

Loop the first 10 frames 5 times — now we're getting somewhere:

loop=loop=5:size=10:start=0

🍓 lut1d

Apply a 1D LUT to an input video. Slap a whole lookup table on it. 💖

The filter accepts the following options:

file — set the 1D LUT file name.

Currently supported formats:

cube — Iridas
csp — cineSpace

interp — select interpolation mode.

Available values are:

nearest — use values from the nearest defined point.
linear — interpolate values using the linear interpolation.
cosine — interpolate values using the cosine interpolation.
cubic — interpolate values using the cubic interpolation.
spline — interpolate values using the spline interpolation.

Commands

This filter supports all of the above options as commands. 💫

🧚 lut3d

Apply a 3D LUT to an input video. The 3D one. It's giving depth, literally.

The filter accepts the following options:

file — set the 3D LUT file name.

Currently supported formats:

3dl — AfterEffects
cube — Iridas
dat — DaVinci
m3d — Pandora
csp — cineSpace

interp — select interpolation mode.

Available values are:

nearest — use values from the nearest defined point.
trilinear — interpolate values using the 8 points defining a cube.
tetrahedral — interpolate values using a tetrahedron.
pyramid — interpolate values using a pyramid.
prism — interpolate values using a prism.

Commands

This filter supports the interp option as commands. (Just the one, bestie — know your limits.)

🌸 lumakey

Turn certain luma values into transparency. Luma-based keying, hence the name, we're so clever. 👑

The filter accepts the following options:

threshold — set the luma which will be used as base for transparency. Default value is 0.
tolerance — set the range of luma values to be keyed out. Default value is 0.01.
softness — set the range of softness. Default value is 0. Use this to control the gradual transition from zero to full transparency.

Commands

This filter supports the same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value. (No crash, no drama — it just stays put.)

🎀 lut, lutrgb, lutyuv

Compute a look-up table for binding each pixel component input value to an output value, and apply it to the input video. Very "each one, teach one" but for pixels.

lutyuv applies a lookup table to a YUV input video, lutrgb to an RGB input video.

These filters accept the following parameters:

c0 — set first pixel component expression
c1 — set second pixel component expression
c2 — set third pixel component expression
c3 — set fourth pixel component expression, corresponds to the alpha component
r — set red component expression
g — set green component expression
b — set blue component expression
a — alpha component expression
y — set Y/luma component expression
u — set U/Cb component expression
v — set V/Cr component expression

Each of them specifies the expression to use for computing the lookup table for the corresponding pixel component values.

The exact component associated to each of the c* options depends on the format in input. Format-dependent, so read the room (the room being your pixel format).

The lut filter requires either YUV or RGB pixel formats in input, lutrgb requires RGB pixel formats in input, and lutyuv requires YUV. Do not come for lutrgb with YUV frames, it will not be cute.

The expressions can contain the following constants and functions:

w, h — the input width and height.
val — the input value for the pixel component.
clipval — the input value, clipped to the minval-maxval range.
maxval — the maximum value for the pixel component.
minval — the minimum value for the pixel component.
negval — the negated value for the pixel component value, clipped to the minval-maxval range; it corresponds to the expression maxval-clipval+minval.
clip(val) — the computed value in val, clipped to the minval-maxval range.
gammaval(gamma) — the computed gamma correction value of the pixel component value, clipped to the minval-maxval range. It corresponds to the expression pow((clipval-minval)/(maxval-minval)\,gamma)*(maxval-minval)+minval

All expressions default to clipval.

Commands

This filter supports the same commands as options.

Examples

Negate input video:

lutrgb="r=maxval+minval-val:g=maxval+minval-val:b=maxval+minval-val"
lutyuv="y=maxval+minval-val:u=maxval+minval-val:v=maxval+minval-val"

Which is literally, one hundred percent, the exact same thing as:

lutrgb="r=negval:g=negval:b=negval"
lutyuv="y=negval:u=negval:v=negval"

Negate luma — just the luma, bestie:

lutyuv=y=negval

Remove chroma components, turning the video into a graytone image — bye bye color 🌈:

lutyuv="u=128:v=128"

Apply a luma burning effect — 🔥:

lutyuv="y=2*val"

Remove green and blue components — red era only:

lutrgb="g=0:b=0"

Set a constant alpha channel value on input:

format=rgba,lutrgb=a="maxval-minval/2"

Correct luma gamma by a factor of 0.5:

lutyuv=y=gammaval(0.5)

Discard least significant bits of luma — she's letting go of the little details:

lutyuv=y='bitand(val, 128+64+32)'

Technicolor like effect — vintage movie moment ✨:

lutyuv=u='(val-maxval/2)*2+maxval/2':v='(val-maxval/2)*2+maxval/2'

lut2, tlut2

Okay so lut2 takes two input streams and gives you one output stream. Two in, one out, that's the whole vibe 🫶

And tlut2 — the time lut2 — takes two consecutive frames from a single stream instead. Same energy, different source.

This filter accepts the following parameters:

c0 — set first pixel component expression
c1 — set second pixel component expression
c2 — set third pixel component expression
c3 — set fourth pixel component expression, and yeah this one corresponds to the alpha component
d — set output bit depth, and heads up, it's only available for the lut2 filter. Default is 0, which means the bit depth is automatically picked from the first input format. autopilot mode ✨

Also, lut2 supports the framesync options too, iykyk.

Each one of those specifies the expression to use for computing the lookup table for the corresponding pixel component values. Literally what it says on the tin.

The exact component associated with each of the c* options depends on the format in the inputs — so check your format, bestie 💅

The expressions can contain the following constants:

w, h — the input width and height.
x — the first input value for the pixel component.
y — the second input value for the pixel component.
bdx — the first input video bit depth.
bdy — the second input video bit depth.

All expressions default to x.

Commands

This filter supports all of the above options as commands — except option d. Not d. Never d. 🚫

Examples

Highlight differences between two RGB video streams:

lut2='ifnot(x-y,0,pow(2,bdx)-1):ifnot(x-y,0,pow(2,bdx)-1):ifnot(x-y,0,pow(2,bdx)-1)'

Highlight differences between two YUV video streams:

lut2='ifnot(x-y,0,pow(2,bdx)-1):ifnot(x-y,pow(2,bdx-1),pow(2,bdx)-1):ifnot(x-y,pow(2,bdx-1),pow(2,bdx)-1)'

Show max difference between two video streams:

lut2='if(lt(x,y),0,if(gt(x,y),pow(2,bdx)-1,pow(2,bdx-1))):if(lt(x,y),0,if(gt(x,y),pow(2,bdx)-1,pow(2,bdx-1))):if(lt(x,y),0,if(gt(x,y),pow(2,bdx)-1,pow(2,bdx-1)))'

maskedclamp

Clamp the first input stream with the second input and third input stream. So that's three streams in the mix, bestie 🎀

It returns the value of the first stream, forced to be between second input stream minus undershoot and third input stream plus overshoot.

This filter accepts the following options:

undershoot — default value is 0.
overshoot — default value is 0.
planes — set which planes will be processed as bitmap; unprocessed planes will be copied from the first stream. Default value is 0xf, which means all planes get processed. the full set, no skips ✨

Commands

This filter supports all of the above options as commands.

maskedmax

Merge the second and third input stream into the output stream, using the absolute differences between the second input stream and the first input stream, and the absolute difference between the third input stream and the first input stream.

The picked value will come from the second input stream if the second absolute difference is greater than the first one — otherwise it comes from the third input stream. winner takes it 🏆

This filter accepts the following options:

planes — set which planes will be processed as bitmap; unprocessed planes will be copied from the first stream. Default value is 0xf, all planes processed.

Commands

This filter supports all of the above options as commands.

maskedmerge

Merge the first input stream with the second input stream, using per-pixel weights in the third input stream. Third stream's the bossy one here, lowkey 👑

A value of 0 in the third stream pixel component means the pixel component from the first stream comes back unchanged. Maximum value (eg. 255 for 8-bit videos) means the pixel component from the second stream comes back unchanged. Intermediate values define the amount of merging between both input stream's pixel components. It's a whole sliding scale, gorgeous.

This filter accepts the following options:

planes — set which planes will be processed as bitmap; unprocessed planes will be copied from the first stream. Default value is 0xf, all planes.

Commands

This filter supports all of the above options as commands.

maskedmin

The picked value comes from the second input stream if the second absolute difference is less than the first one — otherwise from the third input stream. Same as maskedmax but the other direction, twin flames fr fr 💜

This filter accepts the following options:

planes — set which planes will be processed as bitmap; unprocessed planes will be copied from the first stream. Default value is 0xf, all planes.

Commands

This filter supports all of the above options as commands.

maskedthreshold

Pick pixels by comparing the absolute difference of two video streams against a fixed threshold. Very demure, very threshold 🎀

If the absolute difference between the pixel component of the first and second video stream is equal to or lower than the user-supplied threshold, then the pixel component from the first video stream is picked — otherwise the pixel component from the second video stream is picked.

This filter accepts the following options:

threshold — set the threshold used when picking pixels from the absolute difference from two input video streams.
planes — set which planes will be processed as bitmap; unprocessed planes will be copied from the second stream (yes, second, not first — careful bestie). Default value is 0xf, all planes.
mode — set the mode of filter operation. Can be abs or diff. Default is abs.

Commands

This filter supports all of the above options as commands.

maskfun

Create a mask from the input video. Honestly iconic.

For example, it's useful for creating motion masks after the tblend filter. You're welcome 🫶

This filter accepts the following options:

low — set low threshold. Any pixel component lower than or exactly equal to this value will be set to 0.
high — set high threshold. Any pixel component higher than this value will be set to the max value allowed for the current pixel format.
planes — set planes to filter; by default all available planes are filtered.
fill — fill all frame pixels with this value.
sum — set max average pixel value for the frame. If the sum of all pixel components is higher than this average, the output frame will be completely filled with the value set by the fill option. Typically useful for scene changes when used in combination with the tblend filter. chef's kiss scenario detection 🔥

Commands

This filter supports all of the above options as commands.

mcdeint

Apply motion-compensation deinterlacing. Fancy stuff 💅

It needs one field per frame as input, and must therefore be used together with yadif=1/3 or equivalent. Don't forget that part, it literally won't work right 😭

This filter accepts the following options:

mode — set the deinterlacing mode.

It accepts one of the following values:

fast
medium
slow — use iterative motion estimation
extra_slow — like slow, but use multiple reference frames.

Default value is fast.

parity — set the picture field parity assumed for the input video. It must be one of the following values:
- 0, tff — assume top field first
- 1, bff — assume bottom field first

Default value is bff.

qp — set the per-block quantization parameter (QP) used by the internal encoder.

Higher values should result in a smoother motion vector field, but less optimal individual vectors. Default value is 1.

median

Pick the median pixel from a certain rectangle defined by radius. Very "middle child" of it 🌸

This filter accepts the following options:

radius — set horizontal radius size. Default value is 1. Allowed range is an integer from 1 to 127.
planes — set which planes to process. Default is 15, which is all available planes.
radiusV — set vertical radius size. Default value is 0. Allowed range is an integer from 0 to 127. If it's 0, the value will be picked from the horizontal radius option. inherit the vibe ✨
percentile — set median percentile. Default value is 0.5. A default value of 0.5 will always pick median values, while 0 picks minimum values, and 1 picks maximum values. choose your fighter.

Commands

This filter supports the same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it's kept at its current value. no crash, just vibes 🤷

mergeplanes

Merge color channel components from several video streams. Mixing and matching, we love to see it 💖

The filter accepts up to 4 input streams, and merges the selected input planes to the output video.

This filter accepts the following options:

mapping — set input to output plane mapping. Default is 0.

The mapping is specified as a bitmap. It should be specified as a hexadecimal number in the form 0xAa[Bb[Cc[Dd]]]. Aa describes the mapping for the first plane of the output stream. A sets the number of the input stream to use (from 0 to 3), and a sets the plane number of the corresponding input to use (from 0 to 3). The rest of the mappings are similar: Bb describes the mapping for the output stream's second plane, Cc describes the mapping for the output stream's third plane, and Dd describes the mapping for the output stream's fourth plane. hex-coded destiny, honestly 🌈

format — set output pixel format. Default is yuva444p.
map0s, map1s, map2s, map3s — set input to output stream mapping for output Nth plane. Default is 0.
map0p, map1p, map2p, map3p — set input to output plane mapping for output Nth plane. Default is 0.

Examples

Merge three gray video streams of same width and height into a single video stream:

[a0][a1][a2]mergeplanes=0x001020:yuv444p

Merge 1st yuv444p stream and 2nd gray video stream into yuva444p video stream:

[a0][a1]mergeplanes=0x00010210:yuva444p

Swap Y and A plane in yuva444p stream:
format=yuva444p,mergeplanes=0x03010200:yuva444p — swap the U and V planes in a yuv420p stream:

format=yuv420p,mergeplanes=0x000201:yuv420p

format=rgb24,mergeplanes=0x000102:yuv444p — cast an rgb24 clip to yuv444p:

format=rgb24,mergeplanes=0x000102:yuv444p

mestimate 🎯

Okay so this one estimates motion vectors using block matching algorithms and exports them. Where do they go? Into frame side data, bestie — so other filters can pick them up and do their thing. It's giving "behind the scenes hero" energy ngl.

This filter accepts the following options:

method — pick your motion estimation method. Accepted values are:
- esa — Exhaustive search algorithm.
- tss — Three step search algorithm.
- tdls — Two dimensional logarithmic search algorithm.
- ntss — New three step search algorithm.
- fss — Four step search algorithm.
- ds — Diamond search algorithm.
- hexbs — Hexagon-based search algorithm.
- epzs — Enhanced predictive zonal search algorithm.
- umh — Uneven multi-hexagon search algorithm.
- Default value is esa. 👑
mb_size — Macroblock size. Default 16.
search_param — Search parameter. Default 7.

mestimate_d3d12 🚀

Bestie. This is the GPU version and it's iconic. It estimates and exports motion vectors using D3D12 hardware-accelerated motion estimation, tapping into the GPU hardware motion estimation capabilities available in DirectX 12 Video APIs — so you get significant performance improvements over the software-based mestimate filter. Speed? Slayed.

Motion vectors are stored in frame side data to be used by other filters, just like its software sibling.

Heads up: this filter requires the input to be in d3d12 hardware pixel format. Also, it uses quarter-pixel precision for motion vector estimation. Don't forget or it will simply not work, no cap. 💅

This filter accepts the following options:

mb_size — Macroblock size. Only 8 and 16 are supported. Default 16.

Examples

Estimate motion vectors using D3D12 hardware acceleration with 16x16 blocks, and visualize them:

ffmpeg -hwaccel d3d12va -hwaccel_output_format d3d12 -i input.mp4 \
        -vf mestimate_d3d12=mb_size=16,hwdownload,format=nv12,codecview=mv=pf \
        -c:v libx264 -preset fast -b:v 5M output.mp4

midequalizer ⚖️

Apply the Midway Image Equalization effect using two video streams. What is that, you ask? It adjusts a pair of images to have the same histogram, while maintaining their dynamics as much as possible. Genuinely useful for stuff like matching exposures from a pair of stereo cameras — the way that just works. ✨

This filter has two inputs and one output, which must be of same pixel format, but may be of different sizes. The output of filter is first input adjusted with midway histogram of both inputs.

This filter accepts the following option:

planes — Set which planes to process. Default is 15, which is all available planes.

minterpolate 🌊

Convert the video to specified frame rate using motion interpolation. Smooth motion? We love to see it.

This filter accepts the following options:

fps — Specify the output frame rate. This can be rational e.g. "60000/1001". Frames are dropped if fps is lower than source fps. Default 60.
mi_mode — Motion interpolation mode. Following values are accepted:
- dup — Duplicate previous or next frame for interpolating new ones.
- blend — Blend source frames. Interpolated frame is mean of previous and next frames.
- mci — Motion compensated interpolation. Following options are effective when this mode is selected:
  - mc_mode — Motion compensation mode. Following values are accepted:
    - obmc — Overlapped block motion compensation.
    - aobmc — Adaptive overlapped block motion compensation. Window weighting coefficients are controlled adaptively according to the reliabilities of the neighboring motion vectors to reduce oversmoothing.
    - Default mode is obmc.
  - me_mode — Motion estimation mode. Following values are accepted:
    - bidir — Bidirectional motion estimation. Motion vectors are estimated for each source frame in both forward and backward directions.
    - bilat — Bilateral motion estimation. Motion vectors are estimated directly for interpolated frame.
    - Default mode is bilat.
  - me — The algorithm to be used for motion estimation. Following values are accepted:
    - esa — Exhaustive search algorithm.
    - tss — Three step search algorithm.
    - tdls — Two dimensional logarithmic search algorithm.
    - ntss — New three step search algorithm.
    - fss — Four step search algorithm.
    - ds — Diamond search algorithm.
    - hexbs — Hexagon-based search algorithm.
    - epzs — Enhanced predictive zonal search algorithm.
    - umh — Uneven multi-hexagon search algorithm.
    - Default algorithm is epzs.
  - mb_size — Macroblock size. Default 16.
  - search_param — Motion estimation search parameter. Default 32.
  - vsbmc — Enable variable-size block motion compensation. Motion estimation is applied with smaller block sizes at object boundaries in order to make them less blurry. Default is 0 (disabled).
scd — Scene change detection method. Scene change leads motion vectors to be in random direction. Scene change detection replace interpolated frames by duplicate ones. May not be needed for other modes. Following values are accepted:
- none — Disable scene change detection.
- fdiff — Frame difference. Corresponding pixel values are compared and if it satisfies scd_threshold scene change is detected.
- Default method is fdiff.
scd_threshold — Scene change detection threshold. Default is 10..

mix 🧃

Mix several video input streams into one video stream. A whole little blend session, honestly.

A description of the accepted options follows.

inputs — The number of inputs. If unspecified, it defaults to 2.
weights — Specify weight of each input video stream as sequence. Each weight is separated by space. If number of weights is smaller than number of frames last specified weight will be used for all remaining unset weights.
scale — Specify scale, if it is set it will be multiplied with sum of each weight multiplied with pixel values to give final destination pixel value. By default scale is auto scaled to sum of weights.
planes — Set which planes to filter. Default is all. Allowed range is from 0 to 15.
duration — Specify how end of stream is determined.
- longest — The duration of the longest input. (default)
- shortest — The duration of the shortest input.
- first — The duration of the first input.

Commands

This filter supports the following commands:

weights
scale
planes

Syntax is same as option with same name. 🫶

monochrome 🖤

Convert video to gray using custom color filter. Moody. Artistic. Very main character.

A description of the accepted options follows.

cb — Set the chroma blue spot. Allowed range is from -1 to 1. Default value is 0.
cr — Set the chroma red spot. Allowed range is from -1 to 1. Default value is 0.
size — Set the color filter size. Allowed range is from .1 to 10. Default value is 1.
high — Set the highlights strength. Allowed range is from 0 to 1. Default value is 0.

Commands

This filter supports the all above options as commands.

morpho 🦋

This filter allows to apply main morphological grayscale transforms, erode and dilate with arbitrary structures set in second input stream.

Unlike naive implementation and much slower performance in erosion and dilation filters, when speed is critical morpho filter should be used instead. Like, if you care about speed, this is the one. Periodt.

A description of accepted options follows,

mode — Set morphological transform to apply, can be:
- erode
- dilate
- open
- close
- gradient
- tophat
- blackhat

Default is erode.

planes — Set planes to filter, by default all planes except alpha are filtered.
structure — Set which structure video frames will be processed from second input stream, can be first or all. Default is all.

The morpho filter also supports the framesync options.

Commands

This filter supports same commands as options.

mpdecimate ✂️

Drop frames that do not differ greatly from the previous frame in order to reduce frame rate. Basically "if this frame is a copy of the last one, why are we keeping it" energy.

The main use of this filter is for very-low-bitrate encoding (e.g. streaming over dialup modem), but it could in theory be used for fixing movies that were inverse-telecined incorrectly.

A description of the accepted options follows.

max — Set the maximum number of consecutive frames which can be dropped (if positive), or the minimum interval between dropped frames (if negative). If the value is 0, the frame is dropped disregarding the number of previous sequentially dropped frames.

Default value is 0.

keep — Set the maximum number of consecutive similar frames to ignore before to start dropping them. If the value is 0, the frame is dropped disregarding the number of previous sequentially similar frames.

Default value is 0.

hi, lo, frac — Set the dropping threshold values.

Values for hi and lo are for 8x8 pixel blocks and represent actual pixel value differences, so a threshold of 64 corresponds to 1 unit of difference for each pixel, or the same spread out differently over the block.

A frame is a candidate for dropping if no 8x8 blocks differ by more than a threshold of hi, and if no more than frac blocks (1 meaning the whole image) differ by more than a threshold of lo.

Default value for hi is 64*12, default value for lo is 64*5, and default value for frac is 0.33.

msad 🔍

Obtain the MSAD (Mean Sum of Absolute Differences) between two input videos. A little comparison moment, we love a measurement.

This filter takes two input videos.

Both input videos must have the same resolution and pixel format for this filter to work correctly. Also it assumes that both inputs have the same number of frames, which are compared one by one. So yeah, match them up first or the math gets weird. 💀

The obtained per component, average, min and max MSAD is printed through the logging system.

The filter stores the calculated MSAD of each frame in frame metadata.

This filter also supports the framesync options.

okay so check it — in this example, the input file main.mpg that’s being processed gets compared with the reference file ref.mpg. 🫶

ffmpeg -i main.mpg -i ref.mpg -lavfi msad -f null -

multiply 💥

Multiply the pixel values of the first video stream with the pixel values of the second video stream. It’s giving math but make it video. ✨

The filter accepts the following options:

scale — set the scale applied to the second video stream. By default is 1. Allowed range is from 0 to 9. 🎚️
offset — set the offset applied to the second video stream. By default is 0.5. Allowed range is from -1 to 1. 🎛️
planes — specify planes from the input video stream that will be processed. By default all planes are processed. 🎞️

Commands 💬

This filter supports same commands as options. 💅

negate 🌗

Negate (invert) the input video. Literally flipping the vibes. 😭

It accepts the following option:

components — set components to negate.

Available values for components are:

negate_alpha — with value 1, it negates the alpha component, if present. Default value is 0. 🫥

Commands 💬

This filter supports same commands as options. 💅

nlmeans 🧼

Denoise frames using the Non-Local Means algorithm. Your video’s skin routine but for pixels. 💖

Each pixel is adjusted by looking for other pixels with similar contexts. This context similarity is defined by comparing their surrounding patches of size pxp. Patches are searched in an area of rxr around the pixel. 🔍

Note that the research area defines centers for patches, which means some patches will be made of pixels outside that research area. 👀

The filter accepts the following options.

s — set denoising strength. Default is 1.0. Must be in range [1.0, 30.0]. 💪
p — set patch size. Default is 7. Must be odd number in range [0, 99]. 🧩
pc — same as p but for chroma planes. The default value is 0 and means automatic. 🌈
r — set research size. Default is 15. Must be odd number in range [0, 99]. 🔎
rc — same as r but for chroma planes. The default value is 0 and means automatic. 🎨

nnedi 🧠

Deinterlace video using neural network edge directed interpolation. Big brain energy. 🤓

This filter accepts the following options:

weights — mandatory option, without the binary file the filter can not work. Currently the file can be found here: https://github.com/dubhater/vapoursynth-nnedi3/blob/master/src/nnedi3_weights.bin. ⚠️
deint — set which frames to deinterlace, by default it is "all". Can be "all" or "interlaced". 🎬
field — set mode of operation.

Can be one of the following:

af — use frame flags, both fields.
a — use frame flags, single field.
t — use top field only.
b — use bottom field only.
tf — use both fields, top first.
bf — use both fields, bottom first.

planes — set which planes to process, by default filter process all frames. 🖼️
nsize — set size of local neighborhood around each pixel, used by the predictor neural network.

Can be one of the following:

s8x6
s16x6
s32x6
s48x6
s8x4
s16x4
s32x4

nns — set the number of neurons in predictor neural network. Can be one of the following:
- n16
- n32
- n64
- n128
- n256
qual — controls the number of different neural network predictions that are blended together to compute the final output value. Can be "fast", default or "slow". ⚡
etype — set which set of weights to use in the predictor. Can be one of the following:
- a, abs — weights trained to minimize absolute error.
- s, mse — weights trained to minimize squared error.
pscrn — controls whether or not the prescreener neural network is used to decide which pixels should be processed by the predictor neural network and which can be handled by simple cubic interpolation. The prescreener is trained to know whether cubic interpolation will be sufficient for a pixel or whether it should be predicted by the predictor nn. The computational complexity of the prescreener nn is much less than that of the predictor nn. Since most pixels can be handled by cubic interpolation, using the prescreener generally results in much faster processing. The prescreener is pretty accurate, so the difference between using it and not using it is almost always unnoticeable.

Can be one of the following:

none
original
new
new2
new3

Default is "new".

Commands 💬

This filter supports same commands as options, excluding the weights option. 🚫

noformat 🚫

Force libavfilter not to use any of the specified pixel formats for the input to the next filter. It’s giving “not those pixels, bestie”. 💅

It accepts the following parameters:

pix_fmts — a '|'-separated list of pixel format names, such as pix_fmts=yuv420p|monow|rgb24". 📝

Examples ✨

Force libavfilter to use a format different from yuv420p for the input to the vflip filter:

noformat=pix_fmts=yuv420p,vflip

Convert the input video to any of the formats not contained in the list:

noformat=yuv420p|yuv444p|yuv410p

noise 🔊

Add noise on the video input frame. Chaos, but aesthetic. 🌪️

The filter accepts the following options:

all_seed, c0_seed, c1_seed, c2_seed, c3_seed — set noise seed for a specific pixel component, or all pixel components in case of all_seed. Default value is 123457. 🎲
all_strength, alls, c0_strength, c0s, c1_strength, c1s, c2_strength, c2s, c3_strength, c3s — set noise strength for a specific pixel component, or all pixel components in case all_strength. Default value is 0. Allowed range is [0, 100]. 💥
all_flags, allf, c0_flags, c0f, c1_flags, c1f, c2_flags, c2f, c3_flags, c3f — set pixel component flags, or set flags for all components if all_flags.

Available values for component flags are:

a — averaged temporal noise (smoother).
p — mix random noise with a (semi)regular pattern.
t — temporal noise (noise pattern changes between frames).
u — uniform noise (gaussian otherwise).

Examples ✨

Add temporal and uniform noise to input video:

noise=alls=20:allf=t+u

normalize 🌈

Normalize RGB video (aka histogram stretching, contrast stretching). See: https://en.wikipedia.org/wiki/Normalization_(image_processing). It’s giving glow-up for your contrast. ✨

For each channel of each frame, the filter computes the input range and maps it linearly to the user-specified output range. The output range defaults to the full dynamic range from pure black to pure white. 🖤🤍

Temporal smoothing can be used on the input range to reduce flickering (rapid changes in brightness) caused when small dark or bright objects enter or leave the scene. This is similar to the auto-exposure (automatic gain control) on a video camera, and, like a video camera, it may cause a period of over- or under-exposure of the video. 📹

The R,G,B channels can be normalized independently, which may cause some color shifting, or linked together as a single channel, which prevents color shifting. Linked normalization preserves hue. Independent normalization does not, so it can be used to remove some color casts. Independent and linked normalization can be combined in any ratio. 🎨

The normalize filter accepts the following options:

blackpt, whitept — colors which define the output range. The minimum input value is mapped to the blackpt. The maximum input value is mapped to the whitept. The defaults are black and white respectively. Specifying white for blackpt and black for whitept will give color-inverted, normalized video. Shades of grey can be used to reduce the dynamic range (contrast). Specifying saturated colors here can create some interesting effects. 🌈
smoothing — the number of previous frames to use for temporal smoothing. The input range of each channel is smoothed using a rolling average over the current frame and the smoothing previous frames. The default is 0 (no temporal smoothing). 🕰️
independence — controls the ratio of independent (color shifting) channel normalization to linked (color preserving) normalization. 0.0 is fully linked, 1.0 is fully independent. Defaults to 1.0 (fully independent). 🔗
strength — overall strength of the filter. 1.0 is full strength. 0.0 is a rather expensive no-op. Defaults to 1.0 (full strength). 💪

Commands 💬

This filter supports same commands as options, excluding the smoothing option. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value. ⚠️

Examples ✨

Stretch video contrast to use the full dynamic range, with no temporal smoothing; may flicker depending on the source content:

normalize=blackpt=black:whitept=white:smoothing=0

As above, but with 50 frames of temporal smoothing; flicker should be reduced, depending on the source content:

normalize=blackpt=black:whitept=white:smoothing=50

As above, but with hue-preserving linked channel normalization:

normalize=blackpt=black:whitept=white:smoothing=50:independence=0

As above, but with half strength:

normalize=blackpt=black:whitept=white:smoothing=50:independence=0:strength=0.5

Map the darkest input color to red, the brightest input color to cyan:

normalize=blackpt=red:whitept=cyan

null 🫥

Pass the video source unchanged to the output. Literally no thoughts, head empty. 💭

ocio 🎨

OpenColorIO library filter.

This filter allows you to do color management using the OpenColorIO library. See https://opencolorio.org/ for more details. To enable compilation of this filter, you need to configure FFmpeg with --enable-libopencolorio. 🎨

It accepts the following options:

config — by default the filter will use the OCIO config defined by the OCIO environment variable, but this parameter allows you to explicitly specify its location. If you are getting started, you can use config=ocio://studio-config-v1.0.0_aces-v1.3_ocio-v2.1 which specifies one of the built in defaults. ⚙️
input — set the input colorspace. 📥
output — set the output colorspace. 📤
display — set the display colorspace, used in combination with view. 🖥️
view — set the view colorspace, used in combination with display. 👀
inverse — when used in combination with display and view, this inverts the transform, so going from a display/view to the "input colorspace". 🔄
filetransform — lets you point at an external file-transform instead of the OCIO config file. Super handy when you just need to apply one single transform and you do not feel like dragging a whole OCIO config file along for the ride 💅
format — lets you set the output pix_fmt of the OCIO filter. And bestie, this has to be an RGB colorspace, so you are genuinely limited to rgb24, rgba, rgb48, rgba48, gbrp10, gbrp12, gbrpf32le, gbrapf32le. For most encoding we would recommend rgb48 ✨
context_params — lets you pass extra context parameters to the OCIO filter. It's a list of key=value pairs, separated by colons 🎀

✨ Examples ✨

Map from ACEScg to ACEScct — this assumes the OCIO file is defined with the OCIO environment variable.

input=ACEScg:output=ACEScct:format=rgb48

Map from ACEScg to an sRGB display using the "ACES 1.0 - SDR Video" view transform — this also assumes the OCIO file is defined with the OCIO environment variable. Heads up: you will need to wrap the argument in quotes so those spaces get interpreted correctly 🙈

input=ACEScg:display=sRGB - Display:view=ACES 1.0 - SDR Video:format=rgb48

Same thing as above, but using the OCIO file studio-config-v1.0.0_aces-v1.3_ocio-v2.1_ns.ocio instead of the OCIO environment variable:

config=studio-config-v1.0.0_aces-v1.3_ocio-v2.1_ns.ocio:input=ACEScg:display=sRGB - Display:view=ACES 1.0 - SDR Video:format=rgb48

If you are converting to YCrCb you still want to set the color matrix for the conversion. This one is a genuinely good example of combining the two 🫶

ffmpeg -y -i SOURCEFRAMES.%05d.exr -c:v libx265 -vf "ocio=input=ACEScg:output=ACEScct:format=rgb48,scale=in_color_matrix=bt709:out_color_matrix=bt709,format=yuv444p10"  OUTPUTFILE.mov

ocr

Optical Character Recognition 🔍

This filter uses Tesseract for optical character recognition. To enable compilation of this filter, you need to configure FFmpeg with --enable-libtesseract.

It accepts the following options:

datapath — set datapath to the tesseract data. Default behavior is to use whatever was set at installation.
language — set the language, default is "eng".
whitelist — set the character whitelist.
blacklist — set the character blacklist.

The filter exports the recognized text as the frame metadata lavfi.ocr.text 📝 The filter exports the confidence of recognized words as the frame metadata lavfi.ocr.confidence.

ocv

Apply a video transform using libopencv.

To enable this filter, install the libopencv library and headers and configure FFmpeg with --enable-libopencv.

It accepts the following parameters:

filter_name — the name of the libopencv filter to apply.
filter_params — the parameters to pass to the libopencv filter. If not specified, the default values are assumed.

Refer to the official libopencv documentation for more precise info: http://docs.opencv.org/master/modules/imgproc/doc/filtering.html

Several libopencv filters are supported — see the following subsections, bestie 👑

dilate

Dilate an image by using a specific structuring element. It corresponds to the libopencv function cvDilate.

It accepts the parameters: struct_el|nb_iterations.

struct_el represents a structuring element, and has the syntax: colsxrows+anchor_xxanchor_y/shape

cols and rows represent the number of columns and rows of the structuring element, anchor_x and anchor_y the anchor point, and shape the shape for the structuring element. shape must be "rect", "cross", "ellipse", or "custom".

If the value for shape is "custom", it must be followed by a string of the form "=filename". The file with name filename is assumed to represent a binary image, with each printable character corresponding to a bright pixel. When a custom shape is used, cols and rows are ignored, and the number of columns and rows of the read file are assumed instead.

The default value for struct_el is "3x3+0x0/rect".

nb_iterations specifies the number of times the transform is applied to the image, and defaults to 1.

Some examples:

# Use the default values
ocv=dilate

# Dilate using a structuring element with a 5x5 cross, iterating two times
ocv=filter_name=dilate:filter_params=5x5+2x2/cross|2

# Read the shape from the file diamond.shape, iterating two times.
# The file diamond.shape may contain a pattern of characters like this
#   *
#  ***
# *****
#  ***
#   *
# The specified columns and rows are ignored
# but the anchor point coordinates are not
ocv=dilate:0x0+2x2/custom=diamond.shape|2

erode

Erode an image by using a specific structuring element. It corresponds to the libopencv function cvErode.

It accepts the parameters: struct_el:nb_iterations, with the same syntax and semantics as the dilate filter.

smooth

Smooth the input video ✨

The filter takes the following parameters: type|param1|param2|param3|param4.

type is the type of smooth filter to apply, and must be one of these values: "blur", "blur_no_scale", "median", "gaussian", or "bilateral". The default value is "gaussian".

The meaning of param1, param2, param3, and param4 depends on the smooth type. param1 and param2 accept integer positive values or 0. param3 and param4 accept floating point values.

The default value for param1 is 3. The default value for the other parameters is 0.

These parameters correspond to the parameters assigned to the libopencv function cvSmooth.

oscilloscope

2D Video Oscilloscope 📈

Useful to measure spatial impulse, step responses, chroma delays, etc.

It accepts the following parameters:

x — set scope center x position.
y — set scope center y position.
s — set scope size, relative to frame diagonal.
t — set scope tilt/rotation.
o — set trace opacity.
tx — set trace center x position.
ty — set trace center y position.
tw — set trace width, relative to width of frame.
th — set trace height, relative to height of frame.
c — set which components to trace. By default it traces the first three components.
g — draw trace grid. By default it is enabled.
st — draw some statistics. By default it is enabled.
sc — draw scope. By default it is enabled.

Commands

This filter supports the same commands as options. The command accepts the same syntax as the corresponding option.

If the specified expression is not valid, it is kept at its current value.

Examples

Inspect the full first row of a video frame.

oscilloscope=x=0.5:y=0:s=1

Inspect the full last row of a video frame.

oscilloscope=x=0.5:y=1:s=1

Inspect the full 5th line of a video frame of height 1080.

oscilloscope=x=0.5:y=5/1080:s=1

Inspect the full last column of a video frame.

oscilloscope=x=1:y=0.5:s=1:t=1

overlay

Overlay one video on top of another 🎀

It takes two inputs and has one output. The first input is the "main" video on which the second input is overlaid.

It accepts the following parameters:

A description of the accepted options follows.

x
y — set the expression for the x and y coordinates of the overlaid video on the main video. Default value is "0" for both expressions. In case the expression is invalid, it is set to a huge value (meaning the overlay will not be displayed within the output visible area).
eof_action — see framesync.
eval — set when the expressions for x and y are evaluated. It accepts the following values:
- init — only evaluate expressions once during filter initialization or when a command is processed.
- frame — evaluate expressions for each incoming frame.

Default value is frame.

shortest — see framesync.
format — set the format for the output video. It accepts the following values:
- yuv420 — force YUV 4:2:0 8-bit planar output
- yuv420p10 — force YUV 4:2:0 10-bit planar output
- yuv422 — force YUV 4:2:2 8-bit planar output
- yuv422p10 — force YUV 4:2:2 10-bit planar output
- yuv444 — force YUV 4:4:4 8-bit planar output
- yuv444p10 — force YUV 4:4:4 10-bit planar output
- rgb — force RGB 8-bit packed output
- gbrp — force RGB 8-bit planar output
- auto — automatically pick format

Default value is yuv420.

repeatlast — see framesync.
alpha — set the format of the alpha of the overlaid video. It can be straight or premultiplied, or auto to choose the alpha mode automatically. Default is auto.

The x and y expressions can contain the following parameters.

main_w, W
main_h, H — the main input width and height.
overlay_w, w
overlay_h, h — the overlay input width and height.
x
y — the computed values for x and y. They are evaluated for each new frame.
hsub
vsub — horizontal and vertical chroma subsample values of the output format. For example, for the pixel format "yuv422p", hsub is 2 and vsub is 1.
n — the number of input frame, starting from 0.
pos — the position in the file of the input frame, NAN if unknown; deprecated, do not use 💀
t — the timestamp, expressed in seconds. It's NAN if the input timestamp is unknown.

This filter also supports the framesync options.

Note that the n and t variables are available only when evaluation is done per frame, and will evaluate to NAN when eval is set to init.

Be aware that frames are taken from each input video in timestamp order, so if their initial timestamps differ, it is a good idea to pass the two inputs through a setpts=PTS-STARTPTS filter to have them begin at the same zero timestamp, just like the example for the movie filter does 🫶

You can chain together more overlays, but you should test the efficiency of such an approach.

Commands

This filter supports the following commands:

x
y — modify the x and y of the overlay input. The command accepts the same syntax as the corresponding option.

If the specified expression is not valid, it is kept at its current value.

Examples

Draw the overlay at 10 pixels from the bottom right corner of the main video:

overlay=main_w-overlay_w-10:main_h-overlay_h-10

Using named options, the example above becomes:

overlay=x=main_w-overlay_w-10:y=main_h-overlay_h-10

Insert a transparent PNG logo in the bottom left corner of the input, using the ffmpeg tool with the -filter_complex option:
Next one in the overlay glow-up tour — drop a logo in using:

ffmpeg -i input -i logo -filter_complex 'overlay=10:main_h-overlay_h-10' output

Insert 2 different transparent PNG logos (second logo on bottom right corner) using the ffmpeg tool:

ffmpeg -i input -i logo1 -i logo2 -filter_complex 'overlay=x=10:y=H-h-10,overlay=x=W-w-10:y=H-h-10' output

Add a transparent color layer on top of the main video; "WxH" must specify the size of the main input to the overlay filter:

color=color=red@.3:size=WxH [over]; [in][over] overlay [out]

Play an original video and a filtered version (here with the deshake filter) side by side using the ffplay tool:

ffplay input.avi -vf 'split[a][b]; [a]pad=iw*2:ih[src]; [b]deshake[filt]; [src][filt]overlay=w'

The above command is the same as:

ffplay input.avi -vf 'split[b], pad=iw*2[src], [b]deshake, [src]overlay=w'

Make a sliding overlay appearing from the left to the right top part of the screen starting since time 2:

overlay=x='if(gte(t,2), -w+(t-2)*20, NAN)':y=0

Compose output by putting two input videos side to side:

ffmpeg -i left.avi -i right.avi -filter_complex "
nullsrc=size=200x100 [background];
[0:v] setpts=PTS-STARTPTS, scale=100x100 [left];
[1:v] setpts=PTS-STARTPTS, scale=100x100 [right];
[background][left]       overlay=shortest=1       [background+left];
[background+left][right] overlay=shortest=1:x=100 [left+right]
"

Mask 10-20 seconds of a video by applying the delogo filter to a section:

ffmpeg -i test.avi -codec:v:0 wmv2 -ar 11025 -b:v 9000k
-vf '[in]split[split_main][split_delogo];[split_delogo]trim=start=360:end=371,delogo=0:0:640:480[delogoed];[split_main][delogoed]overlay=eof_action=pass[out]'
masked.avi

Chain several overlays in cascade:

nullsrc=s=200x200 [bg];
testsrc=s=100x100, split=4 [in0][in1][in2][in3];
[in0] lutrgb=r=0, [bg]   overlay=0:0     [mid0];
[in1] lutrgb=g=0, [mid0] overlay=100:0   [mid1];
[in2] lutrgb=b=0, [mid1] overlay=0:100   [mid2];
[in3] null,       [mid2] overlay=100:100 [out0]

owdenoise

Apply Overcomplete Wavelet denoiser. fancy wavelet math, no thoughts head empty, just vibes and denoising ✨

The filter accepts the following options:

depth — Set depth.

Larger depth values will denoise lower frequency components more, but slow down filtering. trade-off city, bestie

Must be an int in the range 8-16, default is 8.

luma_strength, ls — Set luma strength.

Must be a double value in the range 0-1000, default is 1.0.

chroma_strength, cs — Set chroma strength.

Must be a double value in the range 0-1000, default is 1.0.

pad

Add paddings to the input image, and place the original input at the provided x, y coordinates. It's giving picture frame energy 🖼️

It accepts the following parameters:

width, w
height, h

Specify an expression for the size of the output image with the paddings added. If the value for width or height is 0, the corresponding input size is used for the output.

The width expression can reference the value set by the height expression, and vice versa. they talk to each other, it's cute

The default value of width and height is 0.

x
y

Specify the offsets to place the input image at within the padded area, with respect to the top/left border of the output image.

The x expression can reference the value set by the y expression, and vice versa.

The default value of x and y is 0.

If x or y evaluate to a negative number, they'll be changed so the input image is centered on the padded area. aww, it self-centers, growth mindset fr

color — Specify the color of the padded area. For the syntax of this option, check the "Color" section in the ffmpeg-utils manual.

The default value of color is "black".

eval — Specify when to evaluate width, height, x and y expression.

It accepts the following values:

init — Only evaluate expressions once during the filter initialization or when a command is processed.
frame — Evaluate expressions for each incoming frame.

Default value is init.

aspect — Pad to aspect instead to a resolution.

The value for the width, height, x, and y options are expressions containing the following constants:

in_w
in_h — The input video width and height.
iw
ih — These are the same as in_w and in_h.
out_w
out_h — The output width and height (the size of the padded area), as specified by the width and height expressions.
ow
oh — These are the same as out_w and out_h.
x
y — The x and y offsets as specified by the x and y expressions, or NAN if not yet specified.
a — same as iw / ih
sar — input sample aspect ratio
dar — input display aspect ratio, it is the same as (iw / ih) * sar
hsub
vsub — The horizontal and vertical chroma subsample values. For example for the pixel format "yuv422p" hsub is 2 and vsub is 1.

Examples

Add paddings with the color "violet" to the input video. The output video size is 640x480, and the top-left corner of the input video is placed at column 0, row 40:

pad=640:480:0:40:violet

The example above is equivalent to the following command:

pad=width=640:height=480:x=0:y=40:color=violet

Pad the input to get an output with dimensions increased by 3/2, and put the input video at the center of the padded area:

pad="3/2*iw:3/2*ih:(ow-iw)/2:(oh-ih)/2"

Pad the input to get a squared output with size equal to the maximum value between the input width and height, and put the input video at the center of the padded area:

pad="max(iw\,ih):ow:(ow-iw)/2:(oh-ih)/2"

Pad the input to get a final w/h ratio of 16:9:

pad="ih*16/9:ih:(ow-iw)/2:(oh-ih)/2"

In case of anamorphic video, in order to set the output display aspect correctly, it is necessary to use sar in the expression, according to the relation:

(ih * X / ih) * sar = output_dar
X = output_dar / sar

Thus the previous example needs to be modified to:

pad="ih*16/9/sar:ih:(ow-iw)/2:(oh-ih)/2"

Double the output size and put the input video in the bottom-right corner of the output padded area:

pad="2*iw:2*ih:ow-iw:oh-ih"

palettegen

Generate one palette for a whole video stream. COLOR TIME 🎨✨

It accepts the following options:

max_colors — Set the maximum number of colors to quantize in the palette. Note: the palette will still contain 256 colors; the unused palette entries will be black. so it's always a 256-slot fit, some slots just get the black fit, iykyk
reserve_transparent — Create a palette of 255 colors maximum and reserve the last one for transparency. Reserving the transparency color is useful for GIF optimization. If not set, the maximum of colors in the palette will be 256. You probably want to disable this option for a standalone image. Set by default.
transparency_color — Set the color that will be used as background for transparency.
stats_mode — Set statistics mode.

It accepts the following values:

full — Compute full frame histograms.
diff — Compute histograms only for the part that differs from previous frame. This might be relevant to give more importance to the moving part of your input if the background is static.
single — Compute new histogram for each frame.

Default value is full.

The filter also exports the frame metadata "lavfi.color_quant_ratio" ("nb_color_in / nb_color_out") which you can use to evaluate the degree of color quantization of the palette. This information is also visible at info logging level. receipts, available on demand

Examples

Generate a representative palette of a given video using ffmpeg:

ffmpeg -i input.mkv -vf palettegen palette.png

paletteuse

Use a palette to downsample an input video stream. palette goes in, cute lil limited-color video comes out 🦋

The filter takes two inputs: one video stream and a palette. The palette must be a 256 pixels image.

It accepts the following options:

dither — Select dithering mode. Available algorithms are:
- bayer — Ordered 8x8 bayer dithering (deterministic)
- heckbert — Dithering as defined by Paul Heckbert in 1982 (simple error diffusion). Note: this dithering is sometimes considered "wrong" and is included as a reference. elderly algorithm, still on the payroll for reference purposes only, no cap
- floyd_steinberg — Floyd and Steingberg dithering (error diffusion)
- sierra2 — Frankie Sierra dithering v2 (error diffusion)
- sierra2_4a — Frankie Sierra dithering v2 "Lite" (error diffusion)
- sierra3 — Frankie Sierra dithering v3 (error diffusion)
- burkes — Burkes dithering (error diffusion)
- atkinson — Atkinson dithering by Bill Atkinson at Apple Computer (error diffusion)
- none — Disable dithering.

Default is sierra2_4a.

bayer_scale — When bayer dithering is selected, this option defines the scale of the pattern (how much the crosshatch pattern is visible). A low value means more visible pattern for less banding, and higher value means less visible pattern at the cost of more banding.

The option must be an integer value in the range [0,5]. Default is 2.

diff_mode — If set, define the zone to process
- rectangle — Only the changing rectangle will be reprocessed. This is similar to GIF cropping/offsetting compression mechanism. This option can be useful for speed if only a part of the image is changing, and has use cases such as limiting the scope of the error diffusal dither to the rectangle that bounds the moving scene (it leads to more deterministic output if the scene doesn't change much, and as a result less moving noise and better GIF compression).

Default is none.

new — Take new palette for each output frame.
alpha_threshold — Sets the alpha threshold for transparency. Alpha values above this threshold will be treated as completely opaque, and values below this threshold will be treated as completely transparent. so dramatic, it's giving all-or-nothing

The option must be an integer value in the range [0,255]. Default is 128.

Examples

Use a palette (generated for example with palettegen) to encode a GIF using ffmpeg:

ffmpeg -i input.mkv -i palette.png -lavfi paletteuse output.gif

perspective

Correct perspective of video not recorded perpendicular to the screen. skewed footage? we fix. geometry slay. 💅

A description of the accepted parameters follows — buckle up bestie, there's a lil list 💅

x0, y0, x1, y1, x2, y2, x3, y3 — these set the coordinates expression for the top left, top right, bottom left and bottom right corners. The default values are "0:0:W:0:0:H:W:H", and with those the perspective just stays exactly the same, no cap. If the sense option is set to source, then the points you specified get sent to the corners of the destination. But if sense is set to destination, it flips — the corners of the source get sent to the coordinates you gave. 🔄

The expressions can use these variables, periodt:

W, H — the width and height of the video frame.
in — input frame count.
on — output frame count.

interpolation — sets the interpolation used for perspective correction ✨

It accepts these values:

linear
cubic

Default value is linear.

sense — sets how the coordinate options get interpreted 🧠

It accepts these values:

0, source — send the point in the source specified by the given coordinates to the corners of the destination.
1, destination — send the corners of the source to the point in the destination specified by the given coordinates.

Default value is source.

eval — sets when the expressions for coordinates x0,y0,...x3,y3 actually get evaluated ⏰

It accepts these values:

init — only evaluate the expressions once during filter initialization, or when a command is processed.
frame — evaluate the expressions for each incoming frame.

Default value is init.

🌀 phase

Delays interlaced video by one field time so that the field order flips. The intended use is fixing PAL movies that got captured with the opposite field order to the film-to-video transfer 💫

A description of the accepted parameters follows:

mode — sets the phase mode.

It accepts these values:

t — capture field order top-first, transfer bottom-first. The filter delays the bottom field.
b — capture field order bottom-first, transfer top-first. The filter delays the top field.
p — capture and transfer with the same field order. This mode literally only exists so the other options' docs have something to refer to — but if you actually select it, the filter will faithfully do nothing 💀
a — capture field order determined automatically by field flags, transfer opposite. The filter picks between t and b modes frame by frame using field flags. If no field information is available, then it works just like u.
u — capture unknown or varying, transfer opposite. The filter picks between t and b frame by frame by analyzing the images and choosing whichever alternative produces the best match between the fields.
T — capture top-first, transfer unknown or varying. The filter picks between t and p using image analysis.
B — capture bottom-first, transfer unknown or varying. The filter picks between b and p using image analysis.
A — capture determined by field flags, transfer unknown or varying. The filter picks between t, b and p using field flags and image analysis. If no field information is available, then it works just like U. This is the default mode 👑
U — both capture and transfer unknown or varying. The filter picks between t, b and p using image analysis only.

Commands 🫶

This filter supports all the above options as commands.

🩺 photosensitivity

Reduces the various flashes in video, so it can help users with epilepsy. Genuinely such a kind filter 💖

It accepts these options:

frames, f — set how many frames to use when filtering. Default is 30.
threshold, t — set the detection threshold factor. Default is 1. Lower is stricter.
skip — set how many pixels to skip when sampling frames. Default is 1. Allowed range is from 1 to 1024.
bypass — leave frames unchanged. Default is disabled.

🔬 pixdesctest

Pixel format descriptor test filter, mainly useful for internal testing. The output video should be equal to the input video.

For example:

format=monow, pixdesctest

can be used to test the monowhite pixel format descriptor definition.

🎛️ pixelize

Applies pixelization to the video stream. Big chunky vibes, honestly iconic.

The filter accepts these options:

width, w, height, h — set the block dimensions used for pixelization. Default value is 16.
mode, m — set the mode of pixelization used.

Possible values are:

avg
min
max

Default value is "avg".

planes, p — set what planes to filter. Default is to filter all planes.

Commands

This filter supports all options as commands.

🔍 pixscope

Displays sample values of color channels. Mainly useful for checking color and levels. Minimum supported resolution is 640x480 👀

The filters accept these options:

x — set scope X position, relative offset on the X axis.
y — set scope Y position, relative offset on the Y axis.
w — set scope width.
h — set scope height.
o — set window opacity. This window also holds statistics about the pixel area.
wx — set window X position, relative offset on the X axis.
wy — set window Y position, relative offset on the Y axis.

Commands

This filter supports the same commands as options.

🧼 pp7

Applies Postprocessing filter 7. It's a variant of the spp filter, similar to spp = 6 with 7 point DCT, where only the center sample is used after IDCT.

The filter accepts these options:

qp — force a constant quantization parameter. It accepts an integer in range 0 to 63. If it's not set, the filter will use the QP from the video stream (if available).
mode — set thresholding mode. Available modes are:
- hard — set hard thresholding.
- soft — set soft thresholding (better de-ringing effect, but likely blurrier).
- medium — set medium thresholding (good results, default).

✨ premultiply

Applies the alpha premultiply effect to the input video stream, using the first plane of the second stream as alpha.

Both streams must have the same dimensions and the same pixel format, no exceptions bestie.

The filter accepts this option:

planes — set which planes will be processed; unprocessed planes get copied. By default value is 0xf, so all planes get processed.
inplace — don't require a 2nd input for processing, instead use the alpha plane from the input stream.

🦋 premultiply_dynamic

Dynamically premultiplies or unpremultiplies the input video stream as needed, to match what the downstream filter graph requires. This is roughly equivalent to either "premultiply:inplace=yes" or "unpremultiply:inplace=yes", or otherwise a noop. It's giving adaptive queen 👑

The filter accepts this option:

planes — set which planes will be processed; unprocessed planes get copied. By default value is 0xf, all planes will be processed.

🌸 prewitt

Applies the prewitt operator to the input video stream.

The filter accepts these options:

planes — set which planes will be processed; unprocessed planes get copied. By default value is 0xf, all planes will be processed.
scale — set the value which will be multiplied with the filtered result.
delta — set the value which will be added to the filtered result.

Commands

This filter supports all the above options as commands.

🌈 pseudocolor

Alters frame colors in video with pseudocolors. Lowkey a whole aesthetic upgrade.

This filter accepts these options:

c0 — set pixel first component expression.
c1 — set pixel second component expression.
c2 — set pixel third component expression.
c3 — set pixel fourth component expression, corresponds to the alpha component.
index, i — set the component to use as the base for altering colors.
preset, p — pick one of the built-in LUTs. By default it's set to none.

Available LUTs:

magma
inferno
plasma
viridis
turbo
cividis
range1
range2
shadows
highlights
solar
nominal
preferred
total
spectral
cool
heat
fiery
blues
green
helix

opacity — set the opacity of output colors. Allowed range is from 0 to 1. Default value is 1.

Each of the expression options specifies which expression to use for computing the lookup table for the corresponding pixel component values.

The expressions can contain these constants and functions:

w, h — the input width and height.
val — the input value for the pixel component.
ymin, umin, vmin, amin — the minimum allowed component value.
ymax, umax, vmax, amax — the maximum allowed component value.

All expressions default to "val".

Commands

This filter supports all the above options as commands.

Examples

Change too high luma values to gradient:

pseudocolor="'if(between(val,ymax,amax),lerp(ymin,ymax,(val-ymax)/(amax-ymax)),-1):if(between(val,ymax,amax),lerp(umax,umin,(val-ymax)/(amax-ymax)),-1):if(between(val,ymax,amax),lerp(vmin,vmax,(val-ymax)/(amax-ymax)),-1):-1'"

📊 psnr

Obtains the average, maximum and minimum PSNR (Peak Signal to Noise Ratio) between two input videos.

This filter takes two input videos — the first input is considered the "main" source and gets passed unchanged to the output. The second input is used as a "reference" video for computing the PSNR. 👯

Both video inputs must have the same resolution and pixel format for this filter to work correctly. It also assumes both inputs have the same number of frames, and those get compared one by one.

The obtained average PSNR is printed through the logging system.

The filter stores the accumulated MSE (mean squared error) of each frame, and at the end of processing it's averaged across all frames equally, and this formula is applied to obtain the PSNR:

PSNR = 10*log10(MAX^2/MSE)

Where MAX is the average of the maximum values of each component of the image.

The description of the accepted parameters follows:

stats_file, f — if specified, the filter will use the named file to save the PSNR of each individual frame. When the filename equals "-", the data is sent to standard output.
stats_version — specifies which version of the stats file format to use. Details of each format are written below. Default value is 1.
output_max — controls whether the max value gets written to the stats log. Default is 0. Needs stats_version >= 2. And if you set this while stats_version < 2, the filter will ✨throw an error✨ at you, so don't do that.

This filter also supports the framesync options, btw.

The file that gets printed when stats_file is selected contains a sequence of key/value pairs of the form key:value for each compared couple of frames. Cute, right?

If you specify a stats_version greater than 1, a header line comes before the list of per-frame-pair stats, with key value pairs following the frame format, with these parameters:

psnr_log_version — the version of the log file format. It'll match stats_version. Slay, consistency.
fields — a comma separated list of the per-frame-pair parameters included in the log.

A description of each shown per-frame-pair parameter follows, bestie, don't scroll:

n — sequential number of the input frame, starting from 1.
mse_avg — Mean Square Error pixel-by-pixel average difference of the compared frames, averaged over all the image components.
mse_y, mse_u, mse_v, mse_r, mse_g, mse_b, mse_a — Mean Square Error pixel-by-pixel average difference of the compared frames for the component specified by the suffix.
psnr_y, psnr_u, psnr_v, psnr_r, psnr_g, psnr_b, psnr_a — Peak Signal to Noise ratio of the compared frames for the component specified by the suffix.
max_avg, max_y, max_u, max_v — Maximum allowed value for each channel, and average over all channels.

Examples

okay so for example:

movie=ref_movie.mpg, setpts=PTS-STARTPTS [main];
[main][ref] psnr="stats_file=stats.log" [out]

On this example the input file being processed is compared with the reference file ref_movie.mpg. The PSNR of each individual frame is stored in stats.log. Obsessed with the logs honestly.

another example with different containers:

ffmpeg -i main.mpg -i ref.mkv -lavfi  "[0:v]settb=AVTB,setpts=PTS-STARTPTS[main];[1:v]settb=AVTB,setpts=PTS-STARTPTS[ref];[main][ref]psnr" -f null -

pullup

Pulldown reversal (inverse telecine) filter, capable of handling mixed hard-telecine, 24000/1001 fps progressive, and 30000/1001 fps progressive content. It's giving "undo the telecine", fr fr 🎀

The pullup filter is designed to take advantage of future context in making its decisions. This filter is stateless in the sense that it does not lock onto a pattern to follow — instead it looks forward to the following fields in order to identify matches and rebuild progressive frames. Future context main character energy 💅

To produce content with an even framerate, insert the fps filter after pullup, use fps=24000/1001 if the input frame rate is 29.97fps, fps=24 for 30fps and the (rare) telecined 25fps input.

The filter accepts the following options:

jl, jr, jt, jb — these options set the amount of "junk" to ignore at the left, right, top, and bottom of the image, respectively. Left and right are in units of 8 pixels, while top and bottom are in units of 2 lines. The default is 8 pixels on each side. bye junk ✌️
sb — set the strict breaks. Setting this option to 1 will reduce the chances of filter generating an occasional mismatched frame, but it may also cause an excessive number of frames to be dropped during high motion sequences. Conversely, setting it to -1 will make filter match fields more easily. This may help processing of video where there is slight blurring between the fields, but may also cause there to be interlaced frames in the output. Default value is 0. It's a tradeoff, bestie, pick your fighter.
mp — set the metric plane to use. It accepts the following values:
- l — use luma plane.
- u — use chroma blue plane.
- v — use chroma red plane.

This option may be set to use chroma plane instead of the default luma plane for doing filter's computations. This may improve accuracy on very clean source material, but more likely will decrease accuracy, especially if there is chroma noise (rainbow effect) or any grayscale video. The main purpose of setting mp to a chroma plane is to reduce CPU load and make pullup usable in realtime on slow machines. Lowkey a whole personality trait of this filter.

For best results (without duplicated frames in the output file) it is necessary to change the output frame rate. Like, actually necessary. For example, to inverse telecine NTSC input:

ffmpeg -i input -vf pullup -r 24000/1001 ...

qp

Change video quantization parameters (QP). Little numbers, big mood.

The filter accepts the following option:

qp — set expression for quantization parameter.

The expression is evaluated through the eval API and can contain, among others, the following constants:

known — 1 if index is not 129, 0 otherwise. (yes, that's the whole rule, iconic)
qp — sequential index starting from -129 to 128.

Examples

some equation like:

qp=2+2*sin(PI*qp)

qrencode

Generate a QR code using the libqrencode library (see <https://fukuchi.org/works/qrencode/>), and overlay it on top of the current frame. QR codes on video, we love to see it 🦋

To enable the compilation of this filter, you need to configure FFmpeg with --enable-libqrencode. Don't forget or you'll be sad.

The QR code is generated from the provided text or text pattern. The corresponding QR code is scaled and overlaid into the video output according to the specified options.

In case no text is specified, no QR code is overlaied. (That's a typo in the og source and I'm keeping it, it's part of the vibe now 🙈 — the point: no text means no QR code at all.)

The filter accepts the following options:

qrcode_width, q and padded_qrcode_width, Q — specify an expression for the width of the rendered QR code, with and without padding. The qrcode_width expression can reference the value set by the padded_qrcode_width expression, and vice versa. By default padded_qrcode_width is set to qrcode_width, meaning that there is no padding.

These expressions are evaluated for each new frame.

See the qrencode Expressions section for details. 👀

x, y — specify an expression for positioning the padded QR code top-left corner. The x expression can reference the value set by the y expression, and vice.

By default x and y are set set to 0, meaning that the QR code is placed in the top left corner of the input. (yes, "set set", again, that's literally what the og says, love that for us 💀)

These expressions are evaluated for each new frame.

See the qrencode Expressions section for details.

case_sensitive, cs — instruct libqrencode to use case sensitive encoding. This is enabled by default. This can be disabled to reduce the QR encoding size. Smaller code, less case drama.
level, l — specify the QR encoding error correction level. With an higher correction level, the encoding size will increase but the code will be more robust to corruption. Lower level is L.

It accepts the following values:

expansion — select how the input text is expanded. Can be either none, or normal (default). See the qrencode Text expansion section below for details.
text, textfile — define the text to be rendered. In case neither is specified, no QR is encoded (just an empty colored frame).

In case expansion is enabled, the text is treated as a text template, using the qrencode expansion mechanism. See the qrencode Text expansion section below for details.

background_color, bc and foreground_color, fc — set the QR code and background color. The default value of foreground_color is black, the default value of background_color is white.

For the syntax of the color options, check the "Color" section in the ffmpeg-utils manual. Colors are a whole vibe, go wild 🌈

qrencode Expressions

The expressions set by the options contain the following constants and functions.

dar — input display aspect ratio, it is the same as (w / h) * sar
duration — the current frame's duration, in seconds
hsub, vsub — horizontal and vertical chroma subsample values. For example for the pixel format yuv422p hsub is 2 and vsub is 1.
main_h, H — the input height
main_w, W — the input width
n — the number of input frame, starting from 0
pict_type — a number representing the picture type
qr_w, w — the width of the encoded QR code
rendered_qr_w, q and rendered_padded_qr_w, Q — the width of the rendered QR code, without and without padding. (Yep, that's how the source reads, "without and without" — I'm not touching it, I just work here 💅 The next line is the part that actually matters:)

These parameters allow the q and Q expressions to refer to each other, so you can for example specify q=3/4*Q. Mutual reference, so cute.

rand(min, max) — return a random number included between min and max
sar — the input sample aspect ratio
t — timestamp expressed in seconds, NAN if the input timestamp is unknown
x, y — the x and y offset coordinates where the text is drawn.

These parameters allow the x and y expressions to refer to each other, so you can for example specify y=x/dar.

qrencode Text expansion

If expansion is set to none, the text is printed verbatim. No thoughts, just text.

If expansion is set to normal (which is the default), the following expansion mechanism is used.

The backslash character \, followed by any character, always expands to the second character.

Sequences of the form %{...} are expanded. The text between the braces is a function name, possibly followed by arguments separated by ':'. If the arguments contain special characters or delimiters (':' or '}'), they should be escaped.

Note that they probably must also be escaped as the value for the text option in the filter argument string and as the filter argument in the filtergraph description, and possibly also for the shell — that makes up to four levels of escaping 😭 using a text file with the textfile option avoids these problems. Bestie. Save yourself.

The following functions are available:

n, frame_num — return the frame number.
pts — Return the presentation timestamp of the current frame.

It can take up to two arguments.

The first argument is the format of the timestamp; it defaults to flt for seconds as a decimal number with microsecond accuracy; hms stands for a formatted [-]HH:MM:SS.mmm timestamp with millisecond accuracy. gmtime stands for the timestamp of the frame formatted as UTC time; localtime stands for the timestamp of the frame formatted as local time zone time. If the format is set to hms24hh, the time is formatted in 24h format (00-23).

The second argument is an offset added to the timestamp.

If the format is set to localtime or gmtime, a third argument may be supplied: a strftime C function format string. By default, YYYY-MM-DD HH:MM:SS format will be used. Timezones, we stan 🫶

expr, e — Evaluate the expression's value and output as a double.

It must take one argument specifying the expression to be evaluated, accepting the constants and functions defined in qrencode_expressions.

expr_formatted, ef — evaluates the expression's value and spits it back out as a formatted string ✨

First argument = the expression to evaluate, literally the same deal as the expr function. Second argument = the output format. Allowed values are x, X, d and u, and they're treated exactly as in the "printf" function. Third parameter is optional and sets the number of positions taken by the output — perfect for zero-padding from the left 💅

gmtime — the time at which the filter is running, expressed in UTC ⏰

It can take an argument: a "strftime" C function format string. That format string is extended to support the variable %[1-6]N, which prints fractions of the second with an optionally specified number of digits. fancy.

localtime — the time at which the filter is running, expressed in the local time zone 🌍

Same deal: it accepts an argument, a "strftime" C function format string, and that format string is extended to support %[1-6]N which prints fractions of the second with optionally specified number of digits.

metadata — frame metadata. Takes one or two arguments 📎

The first argument is mandatory and specifies the metadata key.

The second argument is optional and specifies a default value, used when the metadata key is not found or empty.

You can figure out what metadata is available by inspecting entries starting with TAG included within each frame section printed by running ffprobe -show_frames. String metadata generated in filters leading to the qrencode filter is also available 👀

rand(min, max) — returns a random number included between min and max 🎲

✨ Examples ✨

Generate a QR code encoding the specified text with the default size, overlaid in the top left corner of the input video, with the default size:

qrencode=text=www.ffmpeg.org

Same as below, but select blue on pink colors 🦋:

qrencode=text=www.ffmpeg.org:bc=pink@0.5:fc=blue

Place the QR code in the bottom right corner of the input video:

qrencode=text=www.ffmpeg.org:x=W-Q:y=H-Q

Generate a QR code with width of 200 pixels and padding, making the padded width 4/3 of the QR code width:

qrencode=text=www.ffmpeg.org:q=200:Q=4/3*q

Generate a QR code with padded width of 200 pixels and padding, making the QR code width 3/4 of the padded width:

qrencode=text=www.ffmpeg.org:Q=200:q=3/4*Q

Make the QR code a fraction of the input video width:

qrencode=text=www.ffmpeg.org:q=W/5

Generate a QR code encoding the frame number:

qrencode=text=%{n}

Generate a QR code encoding the GMT timestamp:

qrencode=text=%{gmtime}

Generate a QR code encoding the timestamp expressed as a float:

qrencode=text=%{pts}

🕵️ quirc

Identify and decode a QR code using the libquirc library (see <https://github.com/dlbeer/quirc/>), and print the identified QR codes positions and payload as metadata. it's giving detective mode fr fr 🔍

To enable the compilation of this filter, you need to configure FFmpeg with --enable-libquirc.

For each found QR code in the input video, some metadata entries are added with the prefix lavfi.quirc.N, where N is the index, starting from 0, associated to the QR code.

A description of each metadata value follows:

lavfi.quirc.count — the number of found QR codes; it is not set in case none was found 🙈
lavfi.quirc.N.corner.M.x
lavfi.quirc.N.coreer.M.y — the x/y positions of the four corners of the square containing the QR code, where M is the index of the corner starting from 0
lavfi.quirc.N.payload — the payload of the QR code

🎰 random

Flush video frames from internal cache of frames into a random order. No frame is discarded. Inspired by the frei0r nervous filter ✨

frames — set size in number of frames of internal cache, in range from 2 to 512. Default is 30.
seed — set seed for random number generator, must be an integer included between 0 and "UINT32_MAX". If not specified, or if explicitly set to less than 0, the filter will try to use a good random seed on a best effort basis 🎲

📺 readeia608

Read closed captioning (EIA-608) information from the top lines of a video frame.

This filter adds frame metadata for "lavfi.readeia608.X.cc" and "lavfi.readeia608.X.line", where "X" is the number of the identified line with EIA-608 data (starting from 0). A description of each metadata value follows:

lavfi.readeia608.X.cc — the two bytes stored as EIA-608 data (printed in hexadecimal)
lavfi.readeia608.X.line — the number of the line on which the EIA-608 data was identified and read

This filter accepts the following options:

scan_min — set the line to start scanning for EIA-608 data. Default is 0.
scan_max — set the line to end scanning for EIA-608 data. Default is 29.
spw — set the ratio of width reserved for sync code detection. Default is 0.27. Allowed range is "[0.1 - 0.7]".
chp — enable checking the parity bit. In the event of a parity error, the filter will output 0x00 for that character. Default is false.
lp — lowpass lines prior to further processing. Default is enabled.

Commands

This filter supports the all above options as commands 👑

✨ Examples ✨

Output a csv with presentation time and the first two lines of identified EIA-608 captioning data:

ffprobe -f lavfi -i movie=captioned_video.mov,readeia608 -show_entries frame=pts_time:frame_tags=lavfi.readeia608.0.cc,lavfi.readeia608.1.cc -of csv

⏱️ readvitc

Read vertical interval timecode (VITC) information from the top lines of a video frame.

The filter adds frame metadata key "lavfi.readvitc.tc_str" with the timecode value, if a valid timecode has been detected. Further metadata key "lavfi.readvitc.found" is set to 0/1 depending on whether timecode data has been found or not 💖

This filter accepts the following options:

scan_max — set the maximum number of lines to scan for VITC data. If the value is set to -1 the full video frame is scanned. Default is 45.
thr_b — set the luma threshold for black. Accepts float numbers in the range [0.0,1.0], default value is 0.2. The value must be equal or less than "thr_w".
thr_w — set the luma threshold for white. Accepts float numbers in the range [0.0,1.0], default value is 0.6. The value must be equal or greater than "thr_b".

✨ Examples ✨

Detect and draw VITC data onto the video frame; if no valid VITC is detected, draw --:--:--:-- as a placeholder:

ffmpeg -i input.avi -filter:v 'readvitc,drawtext=fontfile=FreeMono.ttf:text=%{metadata\\:lavfi.readvitc.tc_str\\:--\\\\\\:--\\\\\\:--\\\\\\:--}:x=(w-tw)/2:y=400-ascent'

🗺️ remap

Remap pixels using 2nd: Xmap and 3rd: Ymap input video stream.

Destination pixel at position (X, Y) will be picked from source (x, y) position where x = Xmap(X, Y) and y = Ymap(X, Y). If mapping values are out of range, zero value for pixel will be used for destination pixel.

Xmap and Ymap input video streams must be of same dimensions. Output video stream will have Xmap/Ymap video stream dimensions. Xmap and Ymap input video streams are 16bit depth, single channel 📐

format — specify pixel format of output from this filter. Can be "color" or "gray". Default is "color".
fill — specify the color of the unmapped pixels. For the syntax of this option, check the "Color" section in the ffmpeg-utils manual. Default color is "black".

🧼 removegrain

The removegrain filter is a spatial denoiser for progressive video. so clean, so fresh ✨

m0 — set mode for the first plane.
m1 — set mode for the second plane.
m2 — set mode for the third plane.
m3 — set mode for the fourth plane.

Range of mode is from 0 to 24. Description of each mode follows:

0 — leave input plane unchanged. Default.
1 — clips the pixel with the minimum and maximum of the 8 neighbour pixels.
2 — clips the pixel with the second minimum and maximum of the 8 neighbour pixels.
3 — clips the pixel with the third minimum and maximum of the 8 neighbour pixels.
4 — clips the pixel with the fourth minimum and maximum of the 8 neighbour pixels. This is equivalent to a median filter.
5 — line-sensitive clipping giving the minimal change.
6 — line-sensitive clipping, intermediate.
7 — line-sensitive clipping, intermediate.
8 — line-sensitive clipping, intermediate.
9 — line-sensitive clipping on a line where the neighbours pixels are the closest.
10 — replaces the target pixel with the closest neighbour.
11 — [1 2 1] horizontal and vertical kernel blur.
12 — same as mode 11.
13 — Bob mode, interpolates top field from the line where the neighbours pixels are the closest.
14 — Bob mode, interpolates bottom field from the line where the neighbours pixels are the closest.
15 — Bob mode, interpolates top field. Same as 13 but with a more complicated interpolation formula.
16 — Bob mode, interpolates bottom field. Same as 14 but with a more complicated interpolation formula.
17 — clips the pixel with the minimum and maximum of respectively the maximum and minimum of each pair of opposite neighbour pixels.
18 — line-sensitive clipping using opposite neighbours whose greatest distance from the current pixel is minimal.
19 — replaces the pixel with the average of its 8 neighbours.
20 — averages the 9 pixels ([1 1 1] horizontal and vertical blur).
21 — clips pixels using the averages of opposite neighbour.
22 — same as mode 21 but simpler and faster.
23 — small edge and halo removal, but reputed useless 💀
24 — similar as 23.

🚫 removelogo

Suppress a TV station logo, using an image file to determine which pixels comprise the logo. It works by filling in the pixels that comprise the logo with neighboring pixels. bye bye watermark 👋

The filter accepts the following options:

filename, f — set the filter bitmap file, which can be any image format supported by libavformat. The width and height of the image file must match those of the video stream being processed.

Pixels in the provided bitmap image with a value of zero are not considered part of the logo, non-zero pixels are considered part of the logo. If you use white (255) for the logo and black (0) for the rest, you will be safe 🫶 For making the filter bitmap, it is recommended to take a screen capture of a black frame with the logo visible, and then using a threshold filter followed by the erode filter once or twice.

Okay so real talk: sometimes a few little splotches sneak through and you gotta fix them by hand 💅. Remember that if logo pixels are not covered, the filter quality will be much reduced. Now — marking too many pixels as part of the logo doesn't hurt as much, but it will increase the amount of blurring needed to cover over the image and will destroy more information than necessary, and extra pixels will slow things down on a large logo. So like, cover the logo, don't cover the whole vibe check ✨

🌈 `repeatfields`

This filter uses the repeat_field flag from the Video ES headers and hard repeats fields based on its value. No cap, it just takes the flag's word for it 🫶

🦋 `reverse`

Reverse a video clip. Iconic for that "wait what just happened" moment 💫

Warning: this filter requires memory to buffer the entire clip, so trimming is suggested. Your RAM is gonna be working overtime 😭 trim first, bestie.

🌸 Examples

Take the first 5 seconds of a clip, and reverse it:

trim=end=5,reverse

🎀 `rgbashift`

Shift R/G/B/A pixels horizontally and/or vertically. Little nudge, big energy 💅

The filter accepts the following options:

rh — set amount to shift red horizontally.
rv — set amount to shift red vertically.
gh — set amount to shift green horizontally.
gv — set amount to shift green vertically.
bh — set amount to shift blue horizontally.
bv — set amount to shift blue vertically.
ah — set amount to shift alpha horizontally.
av — set amount to shift alpha vertically.
edge — set edge mode, can be smear, default, or warp.

⚡ Commands

This filter supports the all above options as commands. Yes, all of them 👑

⭐ `roberts`

Apply roberts cross operator to input video stream. Edge-detection slay 🔥

The filter accepts the following option:

planes — set which planes will be processed, unprocessed planes will be copied. By default value 0xf, all planes will be processed.
scale — set value which will be multiplied with filtered result.
delta — set value which will be added to filtered result.

⚡ Commands

This filter supports the all above options as commands. Same deal, bestie 💜

🌀 `rotate`

Rotate video by an arbitrary angle expressed in radians. Main character energy: the video spins 💫

The filter accepts the following options:

A description of the optional parameters follows.

angle, a — set an expression for the angle by which to rotate the input video clockwise, expressed as a number of radians. A negative value will result in a counter-clockwise rotation. By default it is set to "0".

This expression is evaluated for each frame. Every. Single. One 😤

out_w, ow — set the output width expression, default value is "iw". This expression is evaluated just once during configuration.
out_h, oh — set the output height expression, default value is "ih". This expression is evaluated just once during configuration.
bilinear — enable bilinear interpolation if set to 1, a value of 0 disables it. Default value is 1.
fillcolor, c — set the color used to fill the output area not covered by the rotated image. For the general syntax of this option, check the "Color" section in the ffmpeg-utils manual. If the special value "none" is selected then no background is printed (useful for example if the background is never shown).

Default value is "black".

The expressions for the angle and the output size can contain the following constants and functions:

n — sequential number of the input frame, starting from 0. It is always NAN before the first frame is filtered.
t — time in seconds of the input frame, it is set to 0 when the filter is configured. It is always NAN before the first frame is filtered.
hsub, vsub — horizontal and vertical chroma subsample values. For example for the pixel format "yuv422p" hsub is 2 and vsub is 1.
in_w, iw, in_h, ih — the input video width and height
out_w, ow, out_h, oh — the output width and height, that is the size of the padded area as specified by the width and height expressions
rotw(a), roth(a) — the minimal width/height required for completely containing the input video rotated by a radians.

These are only available when computing the out_w and out_h expressions. Like, only there, nowhere else 🚫

🌸 Examples

Rotate the input by PI/6 radians clockwise:

rotate=PI/6

Rotate the input by PI/6 radians counter-clockwise:

rotate=-PI/6

Rotate the input by 45 degrees clockwise:

rotate=45*PI/180

Apply a constant rotation with period T, starting from an angle of PI/3:

rotate=PI/3+2*PI*t/T

Make the input video rotation oscillating with a period of T seconds and an amplitude of A radians:

rotate=A*sin(2*PI/T*t)

Rotate the video, output size is chosen so that the whole rotating input video is always completely contained in the output:

rotate='2*PI*t:ow=hypot(iw,ih):oh=ow'

Rotate the video, reduce the output size so that no background is ever shown:

rotate=2*PI*t:ow='min(iw,ih)/sqrt(2)':oh=ow:c=none

⚡ Commands

The filter supports the following commands:

a, angle — set the angle expression. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value. No crash, no drama, it just stays put 🙈

🧚 `sab`

Apply Shape Adaptive Blur. Soft-focus filter, but make it smart ✨

The filter accepts the following options:

luma_radius, lr — set luma blur filter strength, must be a value in range 0.1-4.0, default value is 1.0. A greater value will result in a more blurred image, and in slower processing.
luma_pre_filter_radius, lpfr — set luma pre-filter radius, must be a value in the 0.1-2.0 range, default value is 1.0.
luma_strength, ls — set luma maximum difference between pixels to still be considered, must be a value in the 0.1-100.0 range, default value is 1.0.
chroma_radius, cr — set chroma blur filter strength, must be a value in range -0.9-4.0. A greater value will result in a more blurred image, and in slower processing.
chroma_pre_filter_radius, cpfr — set chroma pre-filter radius, must be a value in the -0.9-2.0 range.
chroma_strength, cs — set chroma maximum difference between pixels to still be considered, must be a value in the -0.9-100.0 range.

Each chroma option value, if not explicitly specified, is set to the corresponding luma option value. So chroma just copies luma's notes, honestly kind of a mood 🫶

👑 `scale`

Scale (resize) the input video, using the libswscale library. This is the filter, the icon, the legend 💅

The scale filter forces the output display aspect ratio to be the same of the input, by changing the output sample aspect ratio.

If the input image format is different from the format requested by the next filter, the scale filter will convert the input to the requested format. It's giving adaptable 🌈

🎀 Options

The filter accepts the following options, any of the options supported by the libswscale scaler, as well as any of the framesync options.

See the ffmpeg-scaler manual for the complete list of scaler options.

width, w / height, h — set the output video dimension expression. Default value is the input dimension.

If the width or w value is 0, the input width is used for the output. If the height or h value is 0, the input height is used for the output.

If one and only one of the values is -n with n >= 1, the scale filter will use a value that maintains the aspect ratio of the input image, calculated from the other specified dimension. After that it will, however, make sure that the calculated dimension is divisible by n and adjust the value if necessary.

If both values are -n with n >= 1, the behavior will be identical to both values being set to 0 as previously detailed.

See below for the list of accepted constants for use in the dimension expression.

eval — specify when to evaluate width and height expression. It accepts the following values:
- init — only evaluate expressions once during the filter initialization or when a command is processed.
- frame — evaluate expressions for each incoming frame.

Default value is init.

interl — set the interlacing mode. It accepts the following values:
- 1 — force interlaced aware scaling.
- 0 — do not apply interlaced scaling.
- -1 — select interlaced aware scaling depending on whether the source frames are flagged as interlaced or not.

Default value is 0.

flags — set libswscale scaling flags. See the ffmpeg-scaler manual for the complete list of values. If not explicitly specified the filter applies the default flags.
param0, param1 — set libswscale input parameters for scaling algorithms that need them. See the ffmpeg-scaler manual for the complete documentation. If not explicitly specified the filter applies empty parameters.
intent — set the ICC rendering intent to use when transforming between different color spaces. It accepts the following values:
- perceptual — use a perceptually guided tone and gamut mapping curve. The exact details of the mapping used may change at any time and should not be relied on as stable. This intent is recommended for final viewing of image/video content in typical viewing settings.
- relative_colorimetric — statically clip out-of-gamut colors using a colorimetric clipping curve which attempts to find the colorimetrically least dissimilar in-gamut color. This intent performs white point adaptation and black point adaptation. This is the default. This intent is recommended wherever faithful color reproduction is of the utmost importance, even at the cost of clipping.
- absolute_colorimetric — hard clip out-of-gamut colors with no attempt at white or black point reproduction. This intent will reproduce in-gamut colors 1:1 on the output display as they would appear on the reference display, assuming the output display is appropriately calibrated.
- saturation — performs saturation mapping - that is, stretches the input color volume directly onto the output color volume, in non-linear fashion that preserves the original signal appearance as much as possible. This intent is recommended for signal content evaluation, as it will not lead to any clipping. It is roughly analogous to not performing any color mapping, although it still takes into account the mastering display primaries and any differences in encoding TRC.
size, s — set the video size. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual.
in_color_matrix, out_color_matrix — set in/output YCbCr color space type.

This option lets you override the autodetected value and force a specific one for the output and the encoder. If you don’t specify it, the color space type depends on the pixel format. Possible values are literally:

auto — choose automatically.
bt709 — format conforming to International Telecommunication Union (ITU) Recommendation BT.709.
fcc — set color space conforming to the United States Federal Communications Commission (FCC) Code of Federal Regulations (CFR) Title 47 (2003) 73.682 (a).
bt601, bt470, smpte170m — set color space conforming to:
- ITU Radiocommunication Sector (ITU-R) Recommendation BT.601
- ITU-R Rec. BT.470-6 (1998) Systems B, B1, and G
- Society of Motion Picture and Television Engineers (SMPTE) ST 170:2004
smpte240m — set color space conforming to SMPTE ST 240:1999.
bt2020 — set color space conforming to ITU-R BT.2020 non-constant luminance system.

in_range, out_range — set in/output YCbCr sample range. This lets you override the autodetected value and force a specific one for the output and encoder. If not specified, the range depends on the pixel format. Possible values:
- auto/unknown — choose automatically.
- jpeg/full/pc — set full range (0-255 in case of 8-bit luma).
- mpeg/limited/tv — set "MPEG" range (16-235 in case of 8-bit luma).
in_chroma_loc, out_chroma_loc — set in/output chroma sample location. If not specified, center-sited chroma is used by default. Possible values:
- auto, unknown
- left
- center
- topleft
- top
- bottomleft
- bottom
in_primaries, out_primaries — set in/output RGB primaries. This lets you override the autodetected value and force a specific one for the output and encoder. Possible values:
- auto — choose automatically. This is the default.
- Other possible values are bt709, bt470m, bt470bg, smpte170m, smpte240m, film, bt2020, smpte428, smpte431, smpte432, jedec-p22, ebu3213.
in_transfer, out_transfer — set in/output transfer response curve (TRC). This lets you override the autodetected value and force a specific one for the output and encoder. Possible values:
- auto — choose automatically. This is the default.
- Other possible values are bt709, bt470m, gamma22, bt470bg, gamma28, smpte170m, smpte240m, linear, iec61966-2-1, srgb, iec61966-2-4, xvycc, bt1361e, bt2020-10, bt2020-12, smpte2084, smpte428, arib-std-b67.
force_original_aspect_ratio — enable decreasing or increasing output video width or height if necessary to keep the original aspect ratio. Possible values:
- disable — scale the video as specified and disable this feature.
- decrease — the output video dimensions will automatically be decreased if needed.
- increase — the output video dimensions will automatically be increased if needed.

One useful instance of this option is when you know a specific device’s maximum allowed resolution: you can use this to limit the output video to that while retaining the aspect ratio. For example, device A allows 1280x720 playback, and your video is 1920x800. Using this option (set it to decrease) and specifying 1280x720 to the command line makes the output 1280x533.

Please note that this is a different thing than specifying -1 for w or h — you still need to specify the output resolution for this option to work.

force_divisible_by — ensures that both the output dimensions, width and height, are divisible by the given integer when used together with force_original_aspect_ratio. This works similar to using "-n" in the w and h options.

This option respects the value set for force_original_aspect_ratio, increasing or decreasing the resolution accordingly. The video’s aspect ratio may be slightly modified.

This option can be handy if you need to have a video fit within or exceed a defined resolution using force_original_aspect_ratio but also have encoder restrictions on width or height divisibility.

reset_sar — enabling this option leads to the output SAR being reset to 1. Additionally, if the user requests proportional scaling either through the width or height expressions, e.g. "w=-4:h=360" or "w=iw/2:h=-1" or by enabling "force_original_aspect_ratio", then the input DAR is taken into account and the output is scaled to produce square pixels. Default is false.

The values of the w and h options are expressions containing the following constants:

in_w, in_h — the input width and height.
iw, ih — these are the same as in_w and in_h.
out_w, out_h — the output (scaled) width and height.
ow, oh — these are the same as out_w and out_h.
a — the same as iw / ih.
sar — input sample aspect ratio.
dar — the input display aspect ratio. Calculated from (iw / ih) * sar.
hsub, vsub — horizontal and vertical input chroma subsample values. For example, for the pixel format yuv422p, hsub is 2 and vsub is 1.
ohsub, ovsub — horizontal and vertical output chroma subsample values. For example, for the pixel format yuv422p, hsub is 2 and vsub is 1.
n — the (sequential) number of the input frame, starting from 0. Only available with eval=frame.
t — the presentation timestamp of the input frame, expressed as a number of seconds. Only available with eval=frame.
pos — the position (byte offset) of the frame in the input stream, or NaN if this information is unavailable and/or meaningless (for example, in case of synthetic video). Only available with eval=frame. 🚨 Deprecated, do not use.
ref_w, rw; ref_h, rh; ref_a; ref_dar, rdar; ref_n; ref_t; ref_pos — equivalent to the above, but for a second reference input. If any of these variables are present, this filter accepts two inputs.

Examples 🎬

Scale the input video to a size of 200x100:

scale=w=200:h=100

This is equivalent to:

scale=200:100

or:

scale=200x100

Specify a size abbreviation for the output size:

scale=qcif

which can also be written as:

scale=size=qcif

Scale the input to 2x:

scale=w=2*iw:h=2*ih

The above is the same as:

scale=2*in_w:2*in_h

Scale the input to 2x with forced interlaced scaling:

scale=2*iw:2*ih:interl=1

Scale the input to half size:

scale=w=iw/2:h=ih/2

Increase the width, and set the height to the same size:

scale=3/2*iw:ow

Seek Greek harmony:

scale=iw:1/PHI*iw
scale=ih*PHI:ih

Increase the height, and set the width to 3/2 of the height:

scale=w=3/2*oh:h=3/5*ih

Increase the size, making the size a multiple of the chroma subsample values:

scale="trunc(3/2*iw/hsub)*hsub:trunc(3/2*ih/vsub)*vsub"

Increase the width to a maximum of 500 pixels, keeping the same aspect ratio as the input:

scale=w='min(500\, iw*3/2):h=-1'

Make pixels square by combining scale and setsar:

scale='trunc(ih*dar):ih',setsar=1/1

Make pixels square using reset_sar, making sure the resulting resolution is even (required by some codecs):

scale='-2:ih-mod(ih,2):reset_sar=1'

Scale to target exactly, however reset SAR to 1:

scale='400:300:reset_sar=1'

Scale to even dimensions that fit within 400x300, preserving input SAR:

scale='400:300:force_original_aspect_ratio=decrease:force_divisible_by=2'

Scale to produce square pixels with even dimensions that fit within 400x300:

scale='400:300:force_original_aspect_ratio=decrease:force_divisible_by=2:reset_sar=1'

Scale a subtitle stream (sub) to match the main video (main) in size before overlaying. ("scale2ref")

'[main]split[a][b]; [ref][a]scale=rw:rh[c]; [b][c]overlay'

Scale a logo to 1/10th the height of a video, while preserving its display aspect ratio.

[logo-in][video-in]scale=w=oh*dar:h=rh/10[logo-out]

Commands 💬 — this filter supports:

width, w; height, h — set the output video dimension expression. The command accepts the same syntax of the corresponding option. If the specified expression is not valid, it is kept at its current value.

`scharr` 🌾

Apply scharr operator to input video stream. The filter accepts the following options:

planes — set which planes will be processed; unprocessed planes will be copied. By default value 0xf, all planes will be processed.
scale — set value which will be multiplied with filtered result.
delta — set value which will be added to filtered result.

Commands 💬 — this filter supports all the above options as commands.

`scroll` 🌀

Scroll input video horizontally and/or vertically by constant speed. The filter accepts the following options:

horizontal, h — set the horizontal scrolling speed. Default is 0. Allowed range is from -1 to 1. Negative values change scrolling direction.
vertical, v — set the vertical scrolling speed. Default is 0. Allowed range is from -1 to 1. Negative values change scrolling direction.
hpos — set the initial horizontal scrolling position. Default is 0. Allowed range is from 0 to 1.
vpos — set the initial vertical scrolling position. Default is 0. Allowed range is from 0 to 1.

Commands 💬 — this filter supports the following commands:

horizontal, h — set the horizontal scrolling speed.
vertical, v — set the vertical scrolling speed.

`scdet` 🎬

Detect video scene change. This filter sets frame metadata with mafd between frame, the scene score, and forwards the frame to the next filter, so they can use these metadata to detect scene change or others.

Also, this filter is a little chatterbox: it logs a message and sets frame metadata whenever it catches a scene change past your threshold. 🎬

The "lavfi.scd.mafd" metadata keys get set with mafd for every single frame.

The "lavfi.scd.score" metadata keys get set with the scene change score for every frame, so you can detect scene changes.

The "lavfi.scd.time" metadata keys get set with the current filtered frame time — the one that detects a scene change with the threshold.

This filter accepts these options, bestie:

threshold, t — sets the scene change detection threshold as a percentage of maximum change. The good-values zone is "[8.0, 14.0]", and the allowed range for threshold is "[0., 100.]". Default value is 10. 🎯
sc_pass, s — sets the flag to pass scene change frames to the next filter. Default value is 0. You can turn it on if you want to get a snapshot of the scene change frames only. 📸

💅 selectivecolor

This one adjusts cyan, magenta, yellow and black (CMYK) across certain ranges of colors (think "reds", "yellows", "greens", "cyans", ... and so on). The adjustment range is defined by the purity of the color — like, how saturated it already is. 🌈

It's basically the Adobe Photoshop Selective Color tool, but make it command line. ✨

The filter accepts the following options:

correction_method — select the color correction method. Available values are:
- absolute — the specified adjustments are applied "as-is" (added/subtracted to the original pixel component value).
- relative — the specified adjustments are relative to the original component value.

Default is "absolute".

reds — adjustments for red pixels (pixels where the red component is the max) ❤️
yellows — adjustments for yellow pixels (pixels where the blue component is the minimum) 💛
greens — adjustments for green pixels (pixels where the green component is the max) 💚
cyans — adjustments for cyan pixels (pixels where the red component is the minimum) 🩵
blues — adjustments for blue pixels (pixels where the blue component is the max) 💙
magentas — adjustments for magenta pixels (pixels where the green component is the minimum) 💜
whites — adjustments for white pixels (pixels where all components are greater than 128) 🤍
neutrals — adjustments for all pixels except pure black and pure white 🩶
blacks — adjustments for black pixels (pixels where all components are lesser than 128) 🖤
psfile — specify a Photoshop selective color file (".asv") to import the settings from. 📂

All the adjustment settings (reds, yellows, ...) accept up to 4 space-separated floating point adjustment values in the [-1,1] range — respectively to adjust how much cyan, magenta, yellow and black goes on the pixels of that range. no cap.

🌸 Examples

Increase cyan by 50% and reduce yellow by 33% in every green area, and increase magenta by 27% in blue areas:

selectivecolor=greens=.5 0 -.33 0:blues=0 .27

Use a Photoshop selective color preset:

selectivecolor=psfile=MySelectiveColorPresets/Misty.asv

✂️ separatefields

The "separatefields" filter takes a frame-based video input and splits each frame into its component fields, producing a new half-height clip with twice the frame rate and twice the frame count. It's giving mitosis, honestly. 🦋

It uses field-dominance info in the frame to decide which of each pair of fields goes first in the output. If it gets it wrong, use the setfield filter before the "separatefields" filter. 🔧

📐 setdar, setsar

The "setdar" filter sets the Display Aspect Ratio for the filter output video.

It does this by changing the specified Sample (aka Pixel) Aspect Ratio, according to this equation, periodt:

<DAR> = <HORIZONTAL_RESOLUTION> / <VERTICAL_RESOLUTION> * <SAR>

Keep in mind: "setdar" does not modify the pixel dimensions of the video frame. Also, the display aspect ratio set by this filter may get changed by later filters in the filterchain — like if there's scaling involved, or if another "setdar" or a "setsar" filter gets applied. 🙈

The "setsar" filter sets the Sample (aka Pixel) Aspect Ratio for the filter output video.

Note that, as a consequence of applying this filter, the output display aspect ratio will change according to the equation above. lowkey math.

And keep in mind that the sample aspect ratio set by the "setsar" filter may be changed by later filters in the filterchain too — e.g. if another "setsar" or a "setdar" filter gets applied.

It accepts these parameters:

r, ratio, dar ("setdar" only), sar ("setsar" only)` — set the aspect ratio used by the filter.

The parameter can be a floating point number string, or an expression. If the parameter isn't specified, the value "0" is assumed, which means the same input value is used. 🫶

max — set the maximum integer value to use for expressing the numerator and denominator when reducing the expressed aspect ratio to a rational. Default value is 100.

The sar parameter is an expression containing these constants:

w, h — the input width and height.
a — same as w / h.
sar — the input sample aspect ratio.
dar — the input display aspect ratio. Same as (w / h) * sar.
hsub, vsub — horizontal and vertical chroma subsample values. For example, for the pixel format "yuv422p", hsub is 2 and vsub is 1.

🌈 Examples

To change the display aspect ratio to 16:9, specify one of the following:

setdar=dar=1.77777
setdar=dar=16/9

To change the sample aspect ratio to 10:11, specify:

setsar=sar=10/11

To set a display aspect ratio of 16:9 and specify a maximum integer value of 1000 in the aspect ratio reduction, use the command:

setdar=ratio=16/9:max=1000

🎞️ setfield

Forces the field for the output video frame.

The "setfield" filter marks the interlace type field for the output frames. It doesn't change the input frame at all — it only sets the corresponding property, which affects how the frame is treated by the following filters (e.g. "fieldorder" or "yadif"). ✨

The filter accepts the following options:

mode — available values are:
- auto — keep the same field property.
- bff — mark the frame as bottom-field-first.
- tff — mark the frame as top-field-first.
- prog — mark the frame as progressive.

🎨 setparams

Forces the frame parameter for the output video frame.

The "setparams" filter marks interlace and color range for the output frames. Again, it doesn't change the input frame — it only sets the corresponding property, which affects how the frame is treated by filters/encoders. 💫

field_mode — available values are:
- auto — keep the same field property (default).
- bff — mark the frame as bottom-field-first.
- tff — mark the frame as top-field-first.
- prog — mark the frame as progressive.
range — available values are:
- auto — keep the same color range property (default).
- unspecified, unknown — mark the frame as unspecified color range.
- limited, tv, mpeg — mark the frame as limited range.
- full, pc, jpeg — mark the frame as full range.
color_primaries — set the color primaries. Available values are:
- auto — keep the same color primaries property (default).
- bt709
- unknown
- bt470m
- bt470bg
- smpte170m
- smpte240m
- film
- bt2020
- smpte428
- smpte431
- smpte432
- jedec-p22
color_trc — set the color transfer. Available values are:
- auto — keep the same color trc property (default).
- bt709
- unknown
- bt470m
- bt470bg
- smpte170m
- smpte240m
- linear
- log100
- log316
- iec61966-2-4
- bt1361e
- iec61966-2-1
- bt2020-10
- bt2020-12
- smpte2084
- smpte428
- arib-std-b67
colorspace — set the colorspace. Available values are:
- auto — keep the same colorspace property (default).
- gbr
- bt709
- unknown
- fcc
- bt470bg
- smpte170m
- smpte240m
- ycgco
- bt2020nc
- bt2020c
- smpte2085
- chroma-derived-nc
- chroma-derived-c
- ictcp
chroma_location — set the chroma sample location. Available values are:
- auto — keep the same chroma location (default).
- unspecified, unknown
- left
- center
- topleft
- top
- bottomleft
- bottom
alpha_mode — set the alpha moda (yes, "moda" — that's the manual's typo, it means mode 😭). Available values are:
- auto — keep the same alpha mode (default).
- unspecified, unknown
- premultiplied
- straight

🌀 shear

Applies a shear transform to the input video. it's giving slanted.

This filter supports the following options:

shx — shear factor in the X-direction. Default value is 0. Allowed range is from -2 to 2.
shy — shear factor in the Y-direction. Default value is 0. Allowed range is from -2 to 2.
fillcolor, c — set the color used to fill the output area not covered by the transformed video. For the general syntax of this option, check the "Color" section in the ffmpeg-utils manual. If the special value "none" is selected, then no background is printed (useful, for example, if the background is never shown). Default value is "black". 🖤
interp — set the interpolation type. Can be "bilinear" or "nearest". Default is "bilinear".

🎀 Commands

This filter supports the all above options as commands. yup, all of them, live, mid-stream. slay.

🔍 showinfo

Shows a line containing various information for each input video frame. The input video is not modified — purely a vibe check, no touchy. 👀

This filter supports the following options:

checksum — calculate checksums of each plane. Enabled by default.
udu_sei_as_ascii — try to print user data unregistered SEI as ascii characters when possible, in hex format otherwise.

The shown line contains a sequence of key/value pairs of the form key:value.

These values show up in the output:

n — the (sequential) number of the input frame, starting from 0.
pts — the Presentation TimeStamp of the input frame, expressed as a number of time base units. The time base unit depends on the filter input pad.
pts_time — the Presentation TimeStamp of the input frame, expressed as a number of seconds.
fmt — the pixel format name.
sar — the sample aspect ratio of the input frame, expressed in the form num/den.
s — the size of the input frame, bestie. For the exact syntax of this one, go peek at the "Video size" section in the ffmpeg-utils manual 📏
i — the type of interlaced mode. It's "P" for "progressive", "T" for top field first, "B" for bottom field first 🎞️
iskey — this is 1 if the frame is a key frame, 0 otherwise 🔑
type — the picture type of the input frame: "I" for an I-frame, "P" for a P-frame, "B" for a B-frame, or "?" for an unknown type. Also refer to the documentation of the "AVPictureType" enum and of the "av_get_picture_type_char" function defined in libavutil/avutil.h 💅
checksum — the Adler-32 checksum (printed in hexadecimal) of all the planes of the input frame 🧮
plane_checksum — the Adler-32 checksum (printed in hexadecimal) of each plane of the input frame, expressed in the form "[c0 c1 c2 c3]" ✨
mean — the mean value of pixels in each plane of the input frame, expressed in the form "[mean0 mean1 mean2 mean3]" 📊
stdev — the standard deviation of pixel values in each plane of the input frame, expressed in the form "[stdev0 stdev1 stdev2 stdev3]" 📈

🎨 showpalette

Displays the 256 colors palette of each frame. Heads up though — this filter is only relevant for pal8 pixel format frames, so if your frames aren't pal8 it's giving nothing, lowkey 🎀

It accepts the following option:

s — set the size of the box used to represent one palette color entry. Default is 30 (for a "30x30" pixel box) 🖼️

🔀 shuffleframes

Reorder and/or duplicate and/or drop video frames. Chaos, but make it intentional 💫

It accepts the following parameters:

mapping — set the destination indexes of input frames. This is a space or '|' separated list of indexes that maps input frames to output frames. Number of indexes also sets maximal value that each index may have. The '-1' index has special meaning and that is to drop the frame 🙈

The first frame has the index 0. The default is to keep the input unchanged.

Examples

Swap second and third frame of every three frames of the input:

ffmpeg -i INPUT -vf "shuffleframes=0 2 1" OUTPUT

Swap 10th and 1st frame of every ten frames of the input:

ffmpeg -i INPUT -vf "shuffleframes=9 1 2 3 4 5 6 7 8 0" OUTPUT

🧩 shufflepixels

Reorder pixels in video frames. Yes, actually. Pixel chaos, on purpose 🦋

This filter accepts the following options:

direction, d — set shuffle direction. Can be forward or inverse direction. Default direction is forward ⬅️➡️
mode, m — set shuffle mode. Can be horizontal, vertical or block mode 🧱
width, w and height, h — set shuffle block_size. In case of horizontal shuffle mode only the width part of size is used, and in case of vertical shuffle mode only the height part of size is used 📐
seed, s — set random seed used with shuffling pixels. Mainly useful to set to be able to reverse the filtering process to get the original input back. For example, to reverse forward shuffle you need to use the same parameters and the exact same seed and to set direction to inverse 🌱

🎭 shuffleplanes

Reorder and/or duplicate video planes. Plane remix, no cap 💅

It accepts the following parameters:

map0 — the index of the input plane to be used as the first output plane.
map1 — the index of the input plane to be used as the second output plane.
map2 — the index of the input plane to be used as the third output plane.
map3 — the index of the input plane to be used as the fourth output plane.

The first plane has the index 0. The default is to keep the input unchanged.

Examples

Swap the second and third planes of the input:

ffmpeg -i INPUT -vf shuffleplanes=0:2:1:3 OUTPUT

📡 signalstats

Evaluate various visual metrics that assist in determining issues associated with the digitization of analog video media. Basically a lil detective for your tape transfers 🕵️‍♀️

By default the filter will log these metadata values:

YMIN — display the minimal Y value contained within the input frame. Expressed in range of [0-255].
YLOW — display the Y value at the 10% percentile within the input frame. Expressed in range of [0-255].
YAVG — display the average Y value within the input frame. Expressed in range of [0-255].
YHIGH — display the Y value at the 90% percentile within the input frame. Expressed in range of [0-255].
YMAX — display the maximum Y value contained within the input frame. Expressed in range of [0-255].
UMIN — display the minimal U value contained within the input frame. Expressed in range of [0-255].
ULOW — display the U value at the 10% percentile within the input frame. Expressed in range of [0-255].
UAVG — display the average U value within the input frame. Expressed in range of [0-255].
UHIGH — display the U value at the 90% percentile within the input frame. Expressed in range of [0-255].
UMAX — display the maximum U value contained within the input frame. Expressed in range of [0-255].
VMIN — display the minimal V value contained within the input frame. Expressed in range of [0-255].
VLOW — display the V value at the 10% percentile within the input frame. Expressed in range of [0-255].
VAVG — display the average V value within the input frame. Expressed in range of [0-255].
VHIGH — display the V value at the 90% percentile within the input frame. Expressed in range of [0-255].
VMAX — display the maximum V value contained within the input frame. Expressed in range of [0-255].
SATMIN — display the minimal saturation value contained within the input frame. Expressed in range of [0-~181.02].
SATLOW — display the saturation value at the 10% percentile within the input frame. Expressed in range of [0-~181.02].
SATAVG — display the average saturation value within the input frame. Expressed in range of [0-~181.02].
SATHIGH — display the saturation value at the 90% percentile within the input frame. Expressed in range of [0-~181.02].
SATMAX — display the maximum saturation value contained within the input frame. Expressed in range of [0-~181.02].
HUEMED — display the median value for hue within the input frame. Expressed in range of [0-360].
HUEAVG — display the average value for hue within the input frame. Expressed in range of [0-360].
YDIF — display the average of sample value difference between all values of the Y plane in the current frame and corresponding values of the previous input frame. Expressed in range of [0-255].
UDIF — display the average of sample value difference between all values of the U plane in the current frame and corresponding values of the previous input frame. Expressed in range of [0-255].
VDIF — display the average of sample value difference between all values of the V plane in the current frame and corresponding values of the previous input frame. Expressed in range of [0-255].
YBITDEPTH — display bit depth of Y plane in current frame. Expressed in range of [0-16].
UBITDEPTH — display bit depth of U plane in current frame. Expressed in range of [0-16].
VBITDEPTH — display bit depth of V plane in current frame. Expressed in range of [0-16].

The filter accepts the following options:

stat, out — specify an additional form of image analysis. out outputs video with the specified type of pixel highlighted ✨

Both options accept the following values:

tout — identify temporal outliers pixels. A temporal outlier is a pixel unlike the neighboring pixels of the same field. Examples of temporal outliers include the results of video dropouts, head clogs, or tape tracking issues 💀
vrep — identify vertical line repetition. Vertical line repetition includes similar rows of pixels within a frame. In born-digital video vertical line repetition is common, but this pattern is uncommon in video digitized from an analog source. When it occurs in video that results from the digitization of an analog source it can indicate concealment from a dropout compensator 👀
brng — identify pixels that fall outside of legal broadcast range 🚨
color, c — set the highlight color for the out option. The default color is yellow 💛

Examples

Output data of various video metrics:

ffprobe -f lavfi movie=example.mov,signalstats="stat=tout+vrep+brng" -show_frames

Output specific data about the minimum and maximum values of the Y plane per frame:

ffprobe -f lavfi movie=example.mov,signalstats -show_entries frame_tags=lavfi.signalstats.YMAX,lavfi.signalstats.YMIN

Playback video while highlighting pixels that are outside of broadcast range in red:

ffplay example.mov -vf signalstats="out=brng:color=red"

Playback video with signalstats metadata drawn over the frame:

ffplay example.mov -vf signalstats=stat=brng+vrep+tout,drawtext=fontfile=FreeSerif.ttf:textfile=signalstat_drawtext.txt

The contents of signalstat_drawtext.txt used in the command are:

time %{pts:hms}
Y (%{metadata:lavfi.signalstats.YMIN}-%{metadata:lavfi.signalstats.YMAX})
U (%{metadata:lavfi.signalstats.UMIN}-%{metadata:lavfi.signalstats.UMAX})
V (%{metadata:lavfi.signalstats.VMIN}-%{metadata:lavfi.signalstats.VMAX})
saturation maximum: %{metadata:lavfi.signalstats.SATMAX}

🖋️ signature

Calculates the MPEG-7 Video Signature. The filter can handle more than one input. In this case the matching between the inputs can be calculated additionally. The filter always passes through the first input. The signature of each stream can be written into a file. Very main character of it, honestly 👑

It accepts the following options:

detectmode — enable or disable the matching process 🔍

Available values are:

off — disable the calculation of a matching (default) 🚫
full — calculate the matching for the whole video and output whether the whole video matches or only parts 🎯
fast — calculate only until a matching is found or the video ends. Should be faster in some cases ⚡
nb_inputs — set the number of inputs. The option value must be a non negative integer. Default value is 1 🔢
filename — sets the path where the output gets written. Okay so real talk: if you've got more than one input, this path has to be a prototype, meaning it must contain %d or %0nd (where n is a positive integer) — that's the slot the input number gets swapped into. And if you don't specify a filename at all? Nothing gets written. That's the default, babes. 💅
format — picks the output format. Your options are:
- binary — use the specified binary representation (this is the default, so it's giving lowkey reliable 🎀).
- xml — use the specified xml representation.
th_d — sets the threshold for detecting one word as similar. Value has to be an integer greater than zero. Default is 9000. ✨
th_dc — sets the threshold for detecting all words as similar. Integer greater than zero, bestie. Default is 60000.
th_xh — sets the threshold for detecting frames as similar. Integer greater than zero. Default is 116.
th_di — sets the minimum length of a sequence, in frames, for it to count as a matching sequence. Gotta be a non-negative integer value. Default is 0. 🦋
th_it — sets the minimum relation that matching frames to all frames must have. The value must be a double between 0 and 1. Default is 0.5. 🌸

Examples 💖

Wanna calculate the signature of an input video and stash it in signature.bin? Say less:

ffmpeg -i input.mkv -vf signature=filename=signature.bin -map 0:v -f null -

Wanna detect whether two videos match and store the signatures in XML format in signature0.xml and signature1.xml? Iconic, do this:

ffmpeg -i input1.mkv -i input2.mkv -filter_complex "[0:v][1:v] signature=nb_inputs=2:detectmode=full:format=xml:filename=signature%d.xml" -map :v -f null -

🧚 `siti` 🧚

Calculates Spatial Information (SI) and Temporal Information (TI) scores for a video, as defined in ITU-T Rec. P.910 (11/21): Subjective video quality assessment methods for multimedia applications. The PDF lives at <https://www.itu.int/rec/T-REC-P.910-202111-S/en> if you wanna read the whole thing (respect if you do, ngl).

Heads up bestie — this is a legacy implementation that corresponds to a superseded recommendation. If you want the latest and greatest, refer to ITU-T Rec. P.910 (07/22): <https://www.itu.int/rec/T-REC-P.910-202207-I/en> 💀

It accepts the following option:

print_summary — if set to 1, summary statistics get printed to the console. Default is 0.

Examples ✨

To calculate SI/TI metrics and print summary:

ffmpeg -i input.mp4 -vf siti=print_summary=1 -f null -

🌈 `smartblur` 🌈

Blurs the input video without impacting the outlines. The way it keeps your edges crisp while smoothing everything else? Obsessed. 😭

It accepts the following options:

luma_radius, lr — sets the luma radius. Value must be a float in the range [0.1,5.0] that specifies the variance of the gaussian filter used to blur the image (slower if larger, obviously). Default is 1.0.
luma_strength, ls — sets the luma strength. Float in the range [-1.0,1.0] that configures the blurring. A value in [0.0,1.0] will blur the image, whereas a value in [-1.0,0.0] will sharpen it. Default is 1.0.
luma_threshold, lt — sets the luma threshold used as a coefficient to decide whether a pixel should be blurred or not. Integer in the range [-30,30]. A value of 0 filters all the image, a value in [0,30] filters flat areas, and a value in [-30,0] filters edges. Default is 0.
chroma_radius, cr — sets the chroma radius. Float in [0.1,5.0], same gaussian-variance situation (slower if larger). Default value is luma_radius — it literally inherits, cute 🎀.
chroma_strength, cs — sets the chroma strength. Float in [-1.0,1.0]. Blurs in [0.0,1.0], sharpens in [-1.0,0.0]. Default value is luma_strength.
chroma_threshold, ct — sets the chroma threshold coefficient for the blur-or-not decision. Integer in [-30,30]: 0 filters all the image, [0,30] filters flat areas, [-30,0] filters edges. Default value is luma_threshold.
alpha_radius, ar — sets the alpha radius. Float in [0.1,5.0], gaussian variance again (slower if larger). Default value is luma_radius.
alpha_strength, as — sets the alpha strength. Float in [-1.0,1.0]. Blur in [0.0,1.0], sharpen in [-1.0,0.0]. Default value is luma_strength.
alpha_threshold, at — sets the alpha threshold coefficient for the blur-or-not decision. Integer in [-30,30]: 0 filters all the image, [0,30] filters flat areas, [-30,0] filters edges. Default value is luma_threshold.

If a chroma or alpha option is not explicitly set, the corresponding luma value gets used instead. So like, set nothing and everything just vibes off luma. 💜

⭐ `sobel` ⭐

Applies the sobel operator to the input video stream. Edge detection goes brrr.

The filter accepts the following option:

planes — sets which planes get processed; unprocessed planes are copied over untouched. Default value 0xf means all planes get processed.
scale — sets the value that gets multiplied with the filtered result.
delta — sets the value that gets added to the filtered result.

Commands 🫶

This filter supports the all above options as commands. (Yes, all of them, periodt.)

🍓 `spp` 🍓

Applies a simple postprocessing filter that compresses and decompresses the image at several (or — in the case of quality level 6 — all) shifts, then averages the results. It's giving "brute force but make it fashion". 💅

The filter accepts the following options:

quality — sets quality. This option defines the number of levels for averaging. Accepts an integer in the range 0-6. If set to 0, the filter has no effect at all. A value of 6 means the higher quality. For each increment of that value, the speed drops by a factor of approximately 2. Default value is 3. 😭 the math is mathing
qp — force a constant quantization parameter. If not set, the filter will use the QP from the video stream (if available).
mode — sets the thresholding mode. Available modes are:
- hard — set hard thresholding (default).
- soft — set soft thresholding (better de-ringing effect, but likely blurrier).
use_bframe_qp — enables the use of the QP from the B-Frames if set to 1. Careful bestie — using this option may cause flicker, since the B-Frames often have larger QP. Default is 0 (not enabled). 🙈

Commands 👑

This filter supports the following commands:

quality, level — sets the quality level. The value "max" can be used to set the maximum level, currently 6.

🦋 `sr` 🦋

Scales the input by applying one of the super-resolution methods based on convolutional neural networks. Supported models:

Super-Resolution Convolutional Neural Network model (SRCNN). See <https://arxiv.org/abs/1501.00092>.
Efficient Sub-Pixel Convolutional Neural Network model (ESPCN). See <https://arxiv.org/abs/1609.05158>.

Training scripts as well as scripts for model file (.pb) saving can be found at <https://github.com/XueweiMeng/sr/tree/sr_dnn_native>. Original repository is at <https://github.com/HighVoltageRocknRoll/sr.git>. Go give 'em a star, they earned it 🌟

The filter accepts the following options:

dnn_backend — specifies which DNN backend to use for model loading and execution. Accepts the following values:
- tensorflow — TensorFlow backend. To enable this backend you need to install the TensorFlow for C library (see <https://www.tensorflow.org/install/lang_c>) and configure FFmpeg with --enable-libtensorflow. No cap, you actually have to do both. 🔥
model — sets the path to the model file specifying the network architecture and its parameters. Note that different backends use different file formats. TensorFlow and OpenVINO backend can load files for only its format.
scale_factor — sets the scale factor for the SRCNN model. Allowed values are 2, 3 and 4. Default value is 2. The scale factor is necessary for the SRCNN model because it accepts input upscaled using bicubic upscaling with the proper scale factor. ✨

Want full functionality (such as async execution)? Please use the dnn_processing filter instead. iykyk 💖

🎀 `sr_amf` 🎀

Upscales (size increasing) the input video using the AMD Advanced Media Framework library for hardware acceleration. Uses advanced algorithms for upscaling with higher output quality. Setting the output width and height works in the same way as for the scale filter — so if you already know that one, you're chillin'. 🧚

The filter accepts the following options:

w, h — set the output video dimension expression. Default value is the input dimension. Allows for the same expressions as the scale filter.
algorithm — sets the algorithm used for scaling:
- bilinear — Bilinear
- bicubic — Bicubic
- sr1-0 — Video SR1.0. This is a default value 👑
- point — Point
- sr1-1 — Video SR1.1
sharpness — controls hq scaler sharpening. The value is a float in the range of [0.0, 2.0].
format — controls the output pixel format. By default, or if none is specified, the input pixel format is used.
keep-ratio — forces the scaler to keep the aspect ratio of the input image when the output size has a different aspect ratio. Default value is false. Bestie do NOT let your video get squished 😭
fill — specifies whether the output image outside the region of interest, which doesn't fill the entire output surface, should be filled with a solid color.

Examples 🌸

Scale input to 720p, keeping aspect ratio and ensuring the output is yuv420p:

sr_amf=-2:720:format=yuv420p

Upscale to 4K with algorithm video SR1.1:

sr_amf=4096:2160:algorithm=sr1-1

💜 `ssim` 💜

Obtains the SSIM (Structural SImilarity Metric) between two input videos. Comparison girlie behavior fr fr.

This filter takes in input two input videos; the first input is considered the "main" source and is passed unchanged to the output. The second input is used as a "reference" video for computing the SSIM.

Okay bestie, so both video inputs have got to match on resolution AND pixel format or this filter simply will not do its thing 💅. It also lowkey assumes both inputs have the same number of frames, and it compares them one by one.

The filter stores the calculated SSIM of each frame.

Here's what the accepted parameters look like:

stats_file, f — if you set this, the filter uses the named file to save the SSIM of each individual frame. And when the filename equals -, the data gets sent to standard output. 🫶

The file printed when stats_file is selected contains a sequence of key:value pairs, one of the form key:value for each compared couple of frames.

A description of each shown parameter follows:

n — sequential number of the input frame, starting from 1.
Y, U, V, R, G, B — SSIM of the compared frames for the component specified by the suffix.
All — SSIM of the compared frames for the whole frame.
dB — same as above but in dB representation.

This filter also supports the framesync options.

✨ Examples

For example:

movie=ref_movie.mpg, setpts=PTS-STARTPTS [main];
[main][ref] ssim="stats_file=stats.log" [out]

On this example the input file being processed is compared with the reference file ref_movie.mpg. The SSIM of each individual frame is stored in stats.log.

Another example with both psnr and ssim at same time:

ffmpeg -i main.mpg -i ref.mpg -lavfi  "ssim;[0:v][1:v]psnr" -f null -

Another example with different containers:

ffmpeg -i main.mpg -i ref.mkv -lavfi  "[0:v]settb=AVTB,setpts=PTS-STARTPTS[main];[1:v]settb=AVTB,setpts=PTS-STARTPTS[ref];[main][ref]ssim" -f null -

🕶️ stereo3d

Convert between different stereoscopic image formats. Literally 3D glasses energy, we love to see it 🌈.

The filters accept the following options:

in — Set stereoscopic image format of input.

Available values for input image formats are:

sbsl — side by side parallel (left eye left, right eye right)
sbsr — side by side crosseye (right eye left, left eye right)
sbs2l — side by side parallel with half width resolution (left eye left, right eye right)
sbs2r — side by side crosseye with half width resolution (right eye left, left eye right)
abl, tbl — above-below (left eye above, right eye below)
abr, tbr — above-below (right eye above, left eye below)
ab2l, tb2l — above-below with half height resolution (left eye above, right eye below)
ab2r, tb2r — above-below with half height resolution (right eye above, left eye below)
al — alternating frames (left eye first, right eye second)
ar — alternating frames (right eye first, left eye second)
irl — interleaved rows (left eye has top row, right eye starts on next row)
irr — interleaved rows (right eye has top row, left eye starts on next row)
icl — interleaved columns, left eye first
icr — interleaved columns, right eye first

Default value is sbsl.

out — Set stereoscopic image format of output.
- sbsl — side by side parallel (left eye left, right eye right)
- sbsr — side by side crosseye (right eye left, left eye right)
- sbs2l — side by side parallel with half width resolution (left eye left, right eye right)
- sbs2r — side by side crosseye with half width resolution (right eye left, left eye right)
- abl, tbl — above-below (left eye above, right eye below)
- abr, tbr — above-below (right eye above, left eye below)
- ab2l, tb2l — above-below with half height resolution (left eye above, right eye below)
- ab2r, tb2r — above-below with half height resolution (right eye above, left eye below)
- al — alternating frames (left eye first, right eye second)
- ar — alternating frames (right eye first, left eye second)
- irl — interleaved rows (left eye has top row, right eye starts on next row)
- irr — interleaved rows (right eye has top row, left eye starts on next row)
- arbg — anaglyph red/blue gray (red filter on left eye, blue filter on right eye)
- argg — anaglyph red/green gray (red filter on left eye, green filter on right eye)
- arcg — anaglyph red/cyan gray (red filter on left eye, cyan filter on right eye)
- arch — anaglyph red/cyan half colored (red filter on left eye, cyan filter on right eye)
- arcc — anaglyph red/cyan color (red filter on left eye, cyan filter on right eye)
- arcd — anaglyph red/cyan color optimized with the least squares projection of dubois (red filter on left eye, cyan filter on right eye)
- agmg — anaglyph green/magenta gray (green filter on left eye, magenta filter on right eye)
- agmh — anaglyph green/magenta half colored (green filter on left eye, magenta filter on right eye)
- agmc — anaglyph green/magenta colored (green filter on left eye, magenta filter on right eye)
- agmd — anaglyph green/magenta color optimized with the least squares projection of dubois (green filter on left eye, magenta filter on right eye)
- aybg — anaglyph yellow/blue gray (yellow filter on left eye, blue filter on right eye)
- aybh — anaglyph yellow/blue half colored (yellow filter on left eye, blue filter on right eye)
- aybc — anaglyph yellow/blue colored (yellow filter on left eye, blue filter on right eye)
- aybd — anaglyph yellow/blue color optimized with the least squares projection of dubois (yellow filter on left eye, blue filter on right eye)
- ml — mono output (left eye only)
- mr — mono output (right eye only)
- chl — checkerboard, left eye first
- chr — checkerboard, right eye first
- icl — interleaved columns, left eye first
- icr — interleaved columns, right eye first
- hdmi — HDMI frame pack

Default value is arcd.

🎀 Examples

Convert input video from side by side parallel to anaglyph yellow/blue dubois:

stereo3d=sbsl:aybd

Convert input video from above below (left eye above, right eye below) to side by side crosseye.

stereo3d=abl:sbsr

🔀 streamselect, astreamselect

Select video or audio streams. Pick-your-fighter but for streams 👑

The filter accepts the following options:

inputs — Set number of inputs. Default is 2.
map — Set input indexes to remap to outputs.

Commands

The streamselect and astreamselect filter supports the following commands:

map — Set input indexes to remap to outputs.

🌸 Examples

Select first 5 seconds 1st stream and rest of time 2nd stream:

sendcmd='5.0 streamselect map 1',streamselect=inputs=2:map=0

Same as above, but for audio:

asendcmd='5.0 astreamselect map 1',astreamselect=inputs=2:map=0

💬 subtitles

Draw subtitles on top of input video using the libass library. The way this filter just pops your captions on screen, obsessed 💖.

To enable compilation of this filter you need to configure FFmpeg with --enable-libass. This filter also requires a build with libavcodec and libavformat to convert the passed subtitles file to ASS (Advanced Substation Alpha) subtitles format.

The filter accepts the following options:

filename, f — Set the filename of the subtitle file to read. It must be specified.
original_size — Specify the size of the original video, the video for which the ASS file was composed. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Due to a misdesign in ASS aspect ratio arithmetic, this is necessary to correctly scale the fonts if the aspect ratio has been changed.
fontsdir — Set a directory path containing fonts that can be used by the filter. These fonts will be used in addition to whatever the font provider uses.
alpha — Process alpha channel, by default alpha channel is untouched.
charenc — Set subtitles input character encoding. subtitles filter only. Only useful if not UTF-8.
stream_index, si — Set subtitles stream index. subtitles filter only.
force_style — Override default style or script info parameters of the subtitles. It accepts a string containing ASS style format KEY=VALUE couples separated by ,.
wrap_unicode — Break lines according to the Unicode Line Breaking Algorithm. Availability requires at least libass release 0.17.0 (or LIBASS_VERSION 0x01600010), and libass must have been built with libunibreak.

The option is enabled by default except for native ASS.

If the first key is not specified, it is assumed that the first value specifies the filename.

For example, to render the file sub.srt on top of the input video, use the command:

subtitles=sub.srt

which is equivalent to:

subtitles=filename=sub.srt

To render the default subtitles stream from file video.mkv, use:

subtitles=video.mkv

To render the second subtitles stream from that file, use:

subtitles=video.mkv:si=1

To make the subtitles stream from sub.srt appear in 80% transparent blue DejaVu Serif, use:

subtitles=sub.srt:force_style='Fontname=DejaVu Serif,PrimaryColour=&HCCFF0000'

🔍 super2xsai

Scale the input by 2x and smooth using the Super2xSaI (Scale and Interpolate) pixel art scaling algorithm. Pixel art girlies stay winning 🎨✨.

Useful for enlarging pixel art images without reducing sharpness.

🔁 swaprect

Swap two rectangular objects in video. It's giving "trading places" but make it rectangles 😭.

This filter accepts the following options:

w — Set object width.
h — Set object height.
x1 — Set 1st rect x coordinate.
y1 — Set 1st rect y coordinate.
x2 — Set 2nd rect x coordinate.
y2 — Set 2nd rect y coordinate.

All expressions are evaluated once for each frame.

The all options are expressions containing the following constants:

w, h — The input width and height.
a — same as w / h
sar — input sample aspect ratio
dar — input display aspect ratio, it is the same as (w / h) * sar
n — The number of the input frame, starting from 0.
t — The timestamp expressed in seconds. It's NAN if the input timestamp is unknown.
pos — the position in the file of the input frame, NAN if unknown; deprecated, do not use

Commands

This filter supports the all above options as commands.

🌀 swapuv

Swaps the U and V planes. That's it, that's the whole filter — a little chroma flip, no cap. 💅

🌀 tblend

Blends successive video frames together, one after another.

See blend for the full deets, bestie — this one's just the simple cousin. ✨

🌀 telecine

Applies the telecine process to your video. Iconic film-to-video pipeline behavior, lowkey.

It takes these options:

first_field — picks which field comes first:
- top, t — top field first
- bottom, b — bottom field first The default value is top.
pattern — a string of numbers representing the pulldown pattern you want to apply. The default value is 23.

Some typical patterns, for the girlies who need a reference 💖:

NTSC output (30i):

27.5p: 32222
24p: 23 (classic)
24p: 2332 (preferred)
20p: 33
18p: 334
16p: 3444

PAL output (25i):

27.5p: 12222
24p: 222222222223 ("Euro pulldown")
16.67p: 33
16p: 33333334

🌀 thistogram

Computes AND draws a color distribution histogram for your input video — across time. Obsessed with the temporal glow-up honestly 🦋

Unlike the histogram video filter, which only shows the histogram of a single input frame at a certain time, this filter also shows past histograms for a number of frames defined by the width option. Basically it's the memories-having version of histogram, and we respect her for that.

The computed histogram is a representation of the color component distribution in an image.

It accepts these options:

width, w — sets the width of a single color component output. Default value is 0. A value of 0 means the width will be picked from the input video. This also sets the number of passed histograms to keep. Allowed range is [0, 8192].
display_mode, d — sets the display mode. Accepted values:
- stack — per-color-component graphs are placed below each other.
- parade — per-color-component graphs are placed side by side.
- overlay — presents information identical to that in parade, except that the graphs representing color components are superimposed directly over one another. Default is stack.
levels_mode, m — sets the mode. Can be either linear or logarithmic. Default is linear.
components, c — sets what color components to display. Default is 7.
bgopacity, b — sets the background opacity. Default is 0.9.
envelope, e — show the envelope. Default is disabled.
ecolor, ec — sets the envelope color. Default is gold.
slide — sets the slide mode. Available values:
- frame — draw a new frame when the right border is reached.
- replace — replace old columns with new ones.
- scroll — scroll from right to left.
- rscroll — scroll from left to right.
- picture — draw a single picture. Default is replace.

🌀 threshold

Applies a threshold effect to your video stream. Fr fr a whole little math moment 🧮

This filter needs four video streams to do its thresholding:

The first stream is the stream we're filtering.
The second stream holds the threshold values.
The third stream holds the min values.
And the last, fourth stream holds the max values.

It accepts this option:

planes — sets which planes will be processed; unprocessed planes get copied. By default the value is 0xf, meaning all planes are processed.

So like, for example: if the first stream's pixel component value is less than the threshold value of the pixel component from the 2nd (threshold) stream, then the third stream value gets picked — otherwise the fourth stream's pixel component value gets picked. Slay, conditional logic 💅

Using the color source filter, you can perform various types of thresholding:

📜 Commands

This filter supports all options as commands.

📜 Examples

Binary threshold, using gray color as threshold:

ffmpeg -i 320x240.avi -f lavfi -i color=gray -f lavfi -i color=black -f lavfi -i color=white -lavfi threshold output.avi

Inverted binary threshold, using gray color as threshold:

ffmpeg -i 320x240.avi -f lavfi -i color=gray -f lavfi -i color=white -f lavfi -i color=black -lavfi threshold output.avi

Truncate binary threshold, using gray color as threshold:

ffmpeg -i 320x240.avi -f lavfi -i color=gray -i 320x240.avi -f lavfi -i color=gray -lavfi threshold output.avi

Threshold to zero, using gray color as threshold:

ffmpeg -i 320x240.avi -f lavfi -i color=gray -f lavfi -i color=white -i 320x240.avi -lavfi threshold output.avi

Inverted threshold to zero, using gray color as threshold:

ffmpeg -i 320x240.avi -f lavfi -i color=gray -i 320x240.avi -f lavfi -i color=white -lavfi threshold output.avi

🌀 thumbnail

Selects the most representative frame out of a given sequence of consecutive frames. She's the group-photo editor of filters — picks the one where everybody looks good 📸✨

Options:

n — sets the frames batch size to analyze. In a set of n frames, the filter picks one of them, then handles the next batch of n frames until the end. Default is 100.
log — sets the log level to display picked frame stats. Default is info.

⚠️ Heads up bestie: since the filter keeps track of the whole frames sequence, a bigger n value means higher memory usage — so a high value is not recommended. Don't be out here setting n=99999 and then acting surprised, I'm begging 😭

📜 Examples

Extract one picture every 50 frames:

thumbnail=50

Complete example of a thumbnail creation with ffmpeg:

ffmpeg -i in.avi -vf thumbnail,scale=300:200 -frames:v 1 out.png

🌀 tile

Tiles several successive frames together into one grid. Very collage, very mood board 🎀

The untile filter can do the reverse — iykyk.

Options:

layout — sets the grid size in the form COLUMNSxROWS. Range is up to UINT_MAX cells. Default is 6x5.
nb_frames — sets the maximum number of frames to render in the given area. It must be less than or equal to wxh. Default value is 0, meaning the whole area will be used.
margin — sets the outer border margin in pixels. Range is 0 to 1024. Default is 0.
padding — sets the inner border thickness (i.e. the number of pixels between frames). For more advanced padding options (like different values for the edges), refer to the pad video filter. Range is 0 to 1024. Default is 0.
color — specifies the color of the unused area. For the syntax of this option, check the Color section in the ffmpeg-utils manual. The default value of color is black.
overlap — sets the number of frames to overlap when tiling several successive frames together. The value must be between 0 and nb_frames - 1. Default is 0.
init_padding — sets the number of frames to initially be empty before displaying the first output frame. This controls how soon you get that first output frame. The value must be between 0 and nb_frames - 1. Default is 0.

📜 Examples

Produce 8x8 PNG tiles of all keyframes (-skip_frame nokey) in a movie:

ffmpeg -skip_frame nokey -i file.avi -vf 'scale=128:72,tile=8x8' -an -vsync 0 keyframes%03d.png

The -vsync 0 is necessary to prevent ffmpeg from duplicating each output frame to accommodate the originally detected frame rate. Do not skip it, bestie, I'm serious 💀

Display 5 pictures in an area of 3x2 frames, with 7 pixels between them, and 2 pixels of initial margin, using mixed flat and named options:

tile=3x2:nb_frames=5:padding=7:margin=2

🌀 tiltandshift

Applies a tilt-and-shift effect. Okay so this one is genuinely a philosophical moment, buckle up 🌈

What happens when you invert time and space?

Normally a video is composed of several frames that represent a different instant of time and shows a scene that evolves in the space captured by the frame. This filter is the antipode of that concept — it's taking inspiration from tilt and shift photography.

A filtered frame contains the whole timeline of events composing the sequence, and this is obtained by placing a slice of pixels from each frame into a single one. However — since there are no infinite-width frames — this is done up to the width of the input frame, and a video is recomposed by shifting away one column for each subsequent frame. In order to map space to time, the filter tilts each input frame as well, so that motion is preserved. This is accomplished by progressively selecting a different column from each input frame.

The end result is a sort of inverted parallax, so that far away objects move much faster than the ones in the front. The ideal conditions for this video effect are when there's either very little motion and the background is static, or when there's a lot of motion and a very wide depth of field (e.g. wide panorama, while moving on a train). 🚂✨

It accepts these parameters:

tilt — tilt video while shifting (default). When unset, the video will be sliding a static image, composed of the first column of each frame.
start — what to do at the start of filtering (see below).
end — what to do at the end of filtering (see below).
hold — how many columns should pass through before start of filtering.
pad — how many columns should be inserted before end of filtering.

Normally the filter shifts and tilts from the very first frame, and stops when the last one is received. However, before filtering starts, normal video may be preserved, so that the effect is slowly shifted into its place. Similarly, the last video frame may be reconstructed at the end. Alternatively, it's possible to just start and end with black.

none — filtering starts immediately and ends when the last frame is received.
frame — the first frames or the very last frame are kept intact during processing.
black — black is padded at the beginning or at the end of filtering.

🌀 tinterlace

Performs various types of temporal field interlacing. Retro TV energy, respectfully 🫶

Frames are counted starting from 1, so the first input frame is considered odd.

Options:

mode — specifies the mode of the interlacing. This option can also be specified as a value alone. See below for the list of values.

Available values:

merge, 0 — move odd frames into the upper field, even into the lower field, generating a double height frame at half frame rate.

         ------> time
        Input:
        Frame 1         Frame 2         Frame 3         Frame 4

        11111           22222           33333           44444
        11111           22222           33333           44444
        11111           22222           33333           44444
        11111           22222           33333           44444

…aaaand the output for that one looks like:

11111                           33333
22222                           44444
11111                           33333
22222                           44444
11111                           33333
22222                           44444
11111                           33333
22222                           44444

`drop_even`, `1`

Only outputs the odd frames, bestie — even frames get yeeted, and you get a frame with unchanged height at half frame rate. 💅

         ------> time
        Input:
        Frame 1         Frame 2         Frame 3         Frame 4

        11111           22222           33333           44444
        11111           22222           33333           44444
        11111           22222           33333           44444
        11111           22222           33333           44444

        Output:
        11111                           33333
        11111                           33333
        11111                           33333
        11111                           33333

`drop_odd`, `2`

Opposite energy: only the even frames make it out, odd frames are dropped, frame height stays the same, half frame rate. 🦋

         ------> time
        Input:
        Frame 1         Frame 2         Frame 3         Frame 4

        11111           22222           33333           44444
        11111           22222           33333           44444
        11111           22222           33333           44444
        11111           22222           33333           44444

        Output:
                        22222                           44444
                        22222                           44444
                        22222                           44444
                        22222                           44444

`pad`, `3`

Each frame gets expanded to full height, but alternating lines get padded with black — so you get double height at the same input frame rate. 🎀

         ------> time
        Input:
        Frame 1         Frame 2         Frame 3         Frame 4

        11111           22222           33333           44444
        11111           22222           33333           44444
        11111           22222           33333           44444
        11111           22222           33333           44444

        Output:
        11111           .....           33333           .....
        .....           22222           .....           44444
        11111           .....           33333           .....
        .....           22222           .....           44444
        11111           .....           33333           .....
        .....           22222           .....           44444
        11111           .....           33333           .....
        .....           22222           .....           44444

`interleave_top`, `4`

Takes the upper field from odd frames and interleaves it with the lower field from even frames — unchanged height, half frame rate. 🧚

         ------> time
        Input:
        Frame 1         Frame 2         Frame 3         Frame 4

        11111<-         22222           33333<-         44444
        11111           22222<-         33333           44444<-
        11111<-         22222           33333<-         44444
        11111           22222<-         33333           44444<-

        Output:
        11111                           33333
        22222                           44444
        11111                           33333
        22222                           44444

`interleave_bottom`, `5`

The flip side, literally: the lower field from odd frames gets interleaved with the upper field from even frames. Unchanged height, half frame rate. 🌸

         ------> time
        Input:
        Frame 1         Frame 2         Frame 3         Frame 4

        11111           22222<-         33333           44444<-
        11111<-         22222           33333<-         44444
        11111           22222<-         33333           44444<-
        11111<-         22222           33333<-         44444

        Output:
        22222                           44444
        11111                           33333
        22222                           44444
        11111                           33333

`interlacex2`, `6`

Double frame rate, unchanged height. New frames get inserted, each one holding the second temporal field from the previous input frame plus the first temporal field from the next input frame. This mode relies on the top_field_first flag, okay? It's for interlaced video displays with no field synchronisation. ⭐

         ------> time
        Input:
        Frame 1         Frame 2         Frame 3         Frame 4

        11111           22222           33333           44444
         11111           22222           33333           44444
        11111           22222           33333           44444
         11111           22222           33333           44444

        Output:
        11111   22222   22222   33333   33333   44444   44444
         11111   11111   22222   22222   33333   33333   44444
        11111   22222   22222   33333   33333   44444   44444
         11111   11111   22222   22222   33333   33333   44444

`mergex2`, `7`

Odd frames get moved into the upper field, even ones into the lower field — so you get a double height frame at the same frame rate. 👑

         ------> time
        Input:
        Frame 1         Frame 2         Frame 3         Frame 4

        11111           22222           33333           44444
        11111           22222           33333           44444
        11111           22222           33333           44444
        11111           22222           33333           44444

        Output:
        11111           33333           33333           55555
        22222           22222           44444           44444
        11111           33333           33333           55555
        22222           22222           44444           44444
        11111           33333           33333           55555
        22222           22222           44444           44444
        11111           33333           33333           55555
        22222           22222           44444           44444

Heads up bestie: numeric values are deprecated but still accepted for backward compatibility reasons. 🙈

Default mode is merge. 💖

`flags`

Set flags that influence the filter process. ✨

Available value for flags is:

low_pass_filter, vlpf — enables linear vertical low-pass filtering in the filter. You need vertical low-pass filtering when you're making an interlaced destination from a progressive source that's got high-frequency vertical detail. The filtering reduces interlace 'twitter' and Moire patterning. 🌈
complex_filter, cvlpf — enables complex vertical low-pass filtering. This slightly less reduces interlace 'twitter' and Moire patterning, but keeps detail and that subjective sharpness impression better. 🔥
bypass_il — bypasses already interlaced frames, and only adjusts the frame rate.

Real talk: vertical low-pass filtering and bypassing already interlaced frames can only be enabled for mode interleave_top and interleave_bottom. 😭

🌟 tmedian

Picks median pixels out of several successive input video frames. Iconic little filter ngl.

The filter accepts the following options:

radius — sets the radius of the median filter. Default is 1. Allowed range is from 1 to 127.
planes — sets which planes to filter. Default value is 15, and by that all planes are processed.
percentile — sets the median percentile. Default value is 0.5. The default 0.5 will always pick median values, while 0 picks minimum values, and 1 picks maximum values.

Commands

This filter supports all the above options as commands, excluding the option radius. 💀

💜 tmidequalizer

Applies the Temporal Midway Video Equalization effect. 🦋

Midway Video Equalization adjusts a sequence of video frames so they all have the same histograms, while keeping their dynamics as much as possible. It's useful for stuff like matching exposures from a video frames sequence.

This filter accepts the following option:

radius — sets the filtering radius. Default is 5. Allowed range is from 1 to 127.
sigma — sets the filtering sigma. Default is 0.5. This controls the strength of filtering. Setting this option to 0 effectively does nothing. 🙈
planes — sets which planes to process. Default is 15, which is all available planes.

🎀 tmix

Mixes successive video frames. It's giving temporal blender energy fr fr.

A description of the accepted options follows.

frames — the number of successive frames to mix. If unspecified, it defaults to 3.
weights — specifies the weight of each input video frame. Each weight is separated by space. If the number of weights is smaller than the number of frames, the last specified weight gets used for all the remaining unset weights.
scale — specifies the scale. If it's set, it gets multiplied with the sum of each weight multiplied with pixel values to give the final destination pixel value. By default, scale is auto scaled to the sum of the weights.
planes — sets which planes to filter. Default is all. Allowed range is from 0 to 15.

Examples

Average 7 successive frames:

tmix=frames=7:weights="1 1 1 1 1 1 1"

Apply simple temporal convolution:

tmix=frames=3:weights="-1 3 -1"

Similar as above but only showing temporal differences:

tmix=frames=3:weights="-1 2 -1":scale=1

Commands

This filter supports the following commands:

weights
scale
planes

Syntax is the same as the option with the same name. 🫶

🌈 tonemap

okay so this one's a whole vibe remapper — it tone maps colors from different dynamic ranges 💅

Real talk: this filter expects data in single precision floating point, because it literally needs to operate on (and can output) out-of-range values. So you can't just feed it whatever — you need another filter, like zscale, to convert the resulting frame back into a usable format. It's giving "I do my job, you handle the cleanup" energy 😭

Also — the tonemapping algorithms implemented here only work on linear light, so your input data should be linearized beforehand (and possibly correctly tagged, bestie, don't skip that part).

ffmpeg -i INPUT -vf zscale=transfer=linear,tonemap=clip,zscale=transfer=bt709,format=yuv420p OUTPUT

Options

The filter accepts the following options. 🎀

tonemap — Set the tone map algorithm to use. Possible values are:
- none — Do not apply any tone map, only desaturate overbright pixels.
- clip — Hard-clip any out-of-range values. Use it for perfect color accuracy for in-range values, while distorting out-of-range values.
- linear — Stretch the entire reference gamut to a linear multiple of the display.
- gamma — Fit a logarithmic transfer between the tone curves.
- reinhard — Preserve overall image brightness with a simple curve, using nonlinear contrast, which results in flattening details and degrading color accuracy.
- hable — Preserve both dark and bright details better than reinhard, at the cost of slightly darkening everything. Use it when detail preservation is more important than color and brightness accuracy.
- mobius — Smoothly map out-of-range values, while retaining contrast and colors for in-range material as much as possible. Use it when color accuracy is more important than detail preservation.

Default is none. ✨

param — Tune the tone mapping algorithm.

This affects the following algorithms:

none — Ignored. (ngl it does nothing, bestie)
linear — Specifies the scale factor to use while stretching. Default to 1.0.
gamma — Specifies the exponent of the function. Default to 1.8.
clip — Specify an extra linear coefficient to multiply into the signal before clipping. Default to 1.0.
reinhard — Specify the local contrast coefficient at the display peak. Default to 0.5, which means that in-gamut values will be about half as bright as when clipping.
hable — Ignored.
mobius — Specify the transition point from linear to mobius transform. Every value below this point is guaranteed to be mapped 1:1. The higher the value, the more accurate the result will be, at the cost of losing bright details. Default to 0.3, which due to the steep initial slope still preserves in-range colors fairly accurately.

desat — Apply desaturation for highlights that exceed this level of brightness. The higher the parameter, the more color information will be preserved. This setting helps prevent unnaturally blown-out colors for super-highlights, by (smoothly) turning into white instead. This makes images feel more natural, at the cost of reducing information about out-of-range colors.

The default of 2.0 is somewhat conservative and will mostly just apply to skies or directly sunlit surfaces. A setting of 0.0 disables this option.

Heads up 🫶 — this option works only if the input frame has a supported color tag.

peak — Override signal/nominal/reference peak with this value. Useful when the embedded peak information in display metadata is not reliable or when tone mapping from a lower range to a higher range.

🦋 tpad

Temporarily pad video frames. Like, literally add blank frames at the start or end. Iconic for syncing things up 💅

The filter accepts the following options:

start — Specify number of delay frames before input video stream. Default is 0.
stop — Specify number of padding frames after input video stream. Set to -1 to pad indefinitely. Default is 0.
start_mode — Set kind of frames added to beginning of stream. Can be either add or clone. With add frames of solid-color are added. With clone frames are clones of first frame. Default is add.
stop_mode — Set kind of frames added to end of stream. Can be either add or clone. With add frames of solid-color are added. With clone frames are clones of last frame. Default is add.
start_duration, stop_duration — Specify the duration of the start/stop delay. See the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. These options override start and stop. Default is 0.
color — Specify the color of the padded area. For the syntax of this option, check the "Color" section in the ffmpeg-utils manual.

The default value of color is "black".

💀 transpose

Transpose rows with columns in the input video and optionally flip it. It's the "turn your video sideways" filter, and yes it does exactly what it says 🔥

It accepts the following parameters:

dir — Specify the transposition direction.

Can assume the following values:

0, 4, cclock_flip — Rotate by 90 degrees counterclockwise and vertically flip (default), that is:

L.R     L.l
. . ->  . .
l.r     R.r

1, 5, clock — Rotate by 90 degrees clockwise, that is:

L.R     l.L
. . ->  . .
l.r     r.R

2, 6, cclock — Rotate by 90 degrees counterclockwise, that is:

L.R     R.r
. . ->  . .
l.r     L.l

3, 7, clock_flip — Rotate by 90 degrees clockwise and vertically flip, that is:

L.R     r.R
. . ->  . .
l.r     l.L

For values between 4-7, the transposition is only done if the input video geometry is portrait and not landscape. These values are deprecated, the "passthrough" option should be used instead. (fr fr, don't use them)

Numerical values are deprecated too, and should be dropped in favor of symbolic constants. 💅

passthrough — Do not apply the transposition if the input geometry matches the one specified by the specified value. It accepts the following values:
- none — Always apply transposition.
- portrait — Preserve portrait geometry (when height >= width).
- landscape — Preserve landscape geometry (when width >= height).

Default value is "none".

For example to rotate by 90 degrees clockwise and preserve portrait layout:

transpose=dir=1:passthrough=portrait

The command above can also be specified as:

transpose=1:portrait

✨ trim

Trim the input so that the output contains one continuous subpart of the input. It's giving "cut to the good part" 😌

It accepts the following parameters:

start — Specify the time of the start of the kept section, i.e. the frame with the timestamp start will be the first frame in the output.
end — Specify the time of the first frame that will be dropped, i.e. the frame immediately preceding the one with the timestamp end will be the last frame in the output.
start_pts — This is the same as start, except this option sets the start timestamp in timebase units instead of seconds.
end_pts — This is the same as end, except this option sets the end timestamp in timebase units instead of seconds.
duration — The maximum duration of the output in seconds.
start_frame — The number of the first frame that should be passed to the output.
end_frame — The number of the first frame that should be dropped.

start, end, and duration are expressed as time duration specifications; see the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax.

Bestie, listen 👑 — note that the first two sets of the start/end options and the duration option look at the frame timestamp, while the _frame variants simply count the frames that pass through the filter. Also note that this filter does not modify the timestamps. If you wish for the output timestamps to start at zero, insert a setpts filter after the trim filter. iykyk.

If multiple start or end options are set, this filter tries to be greedy and keep all the frames that match at least one of the specified constraints. To keep only the part that matches all the constraints at once, chain multiple trim filters.

The defaults are such that all the input is kept. So it is possible to set e.g. just the end values to keep everything before the specified time. 🫶

Examples:

Drop everything except the second minute of input:

ffmpeg -i INPUT -vf trim=60:120

Keep only the first second:

ffmpeg -i INPUT -vf trim=duration=1

🎀 unpremultiply

Apply alpha unpremultiply effect to input video stream using first plane of second stream as alpha.

Both streams must have same dimensions and same pixel format. No exceptions, sorry 😭

The filter accepts the following option:

planes — Set which planes will be processed, unprocessed planes will be copied. By default value 0xf, all planes will be processed.

If the format has 1 or 2 components, then luma is bit 0. If the format has 3 or 4 components: for RGB formats bit 0 is green, bit 1 is blue and bit 2 is red; for YUV formats bit 0 is luma, bit 1 is chroma-U and bit 2 is chroma-V. If present, the alpha channel is always the last bit.

inplace — Do not require 2nd input for processing, instead use alpha plane from input stream.

🌸 unsharp

Sharpen or blur the input video. Yes bestie, one filter, both directions — it's versatile like that 💅

It accepts the following parameters:

luma_msize_x, lx — Set the luma matrix horizontal size. It must be an odd integer between 3 and 23. The default value is 5.
luma_msize_y, ly — Set the luma matrix vertical size. It must be an odd integer between 3 and 23. The default value is 5.
luma_amount, la — Set the luma effect strength. It must be a floating point number, reasonable values lay between -1.5 and 1.5.

Negative values will blur the input video, while positive values will sharpen it, a value of zero will disable the effect.

Default value is 1.0.

chroma_msize_x, cx — Set the chroma matrix horizontal size. It must be an odd integer between 3 and 23. The default value is 5.
chroma_msize_y, cy — Set the chroma matrix vertical size. It must be an odd integer between 3 and 23. The default value is 5.
chroma_amount, ca — Set the chroma effect strength. It must be a floating point number, reasonable values lay between -1.5 and 1.5.
alpha_msize_x, ax — sets the alpha matrix horizontal size 💖 it has to be an odd integer between 3 and 23, and the default is 5.
alpha_msize_y, ay — same deal but vertical: odd integer between 3 and 23, default 5.
alpha_amount, aa — the alpha effect strength, a floating point number, and the reasonable zone is between -1.5 and 1.5.

Negative values will blur the input video, while positive values will sharpen it — a value of zero will disable the effect. ✨

Default value is 0.0.

All parameters are optional and default to the equivalent of the string '5:5:1.0:5:5:0.0'.

Examples 🎀

Apply strong luma sharpen effect:

unsharp=luma_msize_x=7:luma_msize_y=7:luma_amount=2.5

Apply a strong blur of both luma and chroma parameters:

unsharp=7:7:-2:7:7:-2

untile

Decomposes a video made of tiled images back into the individual images 🦋 it's literally the un-tiling of it all.

The frame rate of the output video is the frame rate of the input video multiplied by the number of tiles. math! iconic.

This filter does the reverse of tile.

The filter accepts the following options:

layout — sets the grid size (i.e. the number of lines and columns). For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. 👑

Examples 🌸

Produce a 1-second video from a still image file made of 25 frames stacked vertically, like an analogic film reel:

ffmpeg -r 1 -i image.jpg -vf untile=1x25 movie.mkv

uspp

Applies the ultra slow/simple postprocessing filter that compresses and decompresses the image at several (or — in the case of quality level 8 — all) shifts and averages the results 💅

The way this differs from the behavior of spp is that uspp actually encodes & decodes each case with libavcodec Snow, whereas spp uses a simplified intra only 8x8 DCT similar to MJPEG. no cap, different machinery fr fr.

Heads up bestie 💀 This filter is not available in ffmpeg versions between 5.0 and 6.0. like it just dipped for a bit.

The filter accepts the following options:

quality — sets quality. This option defines the number of levels for averaging. It accepts an integer in the range 0-8. If set to 0, the filter will have no effect. A value of 8 means the higher quality 🔥 For each increment of that value the speed drops by a factor of approximately 2. Default value is 3.
qp — force a constant quantization parameter. If not set, the filter will use the QP from the video stream (if available).
codec — use specified codec instead of snow.

v360

Converts 360 videos between various formats 🌈 like a glow-up but for projections.

The filter accepts the following options:

input, output — set the format of the input/output video.

Available formats:

e, equirect — equirectangular projection.
c3x2, c6x1, c1x6 — cubemap with 3x2/6x1/1x6 layout.
- Format specific options:
- in_pad, out_pad — set padding proportion for the input/output cubemap. Values in decimals. Example values:
  - 0 — no padding.
  - 0.01 — 1% of face is padding. For example, with 1920x1280 resolution face size would be 640x640 and padding would be 3 pixels from each side. (640 * 0.01 = 6 pixels) Default value is 0. Maximum value is 0.1.
- fin_pad, fout_pad — set fixed padding for the input/output cubemap. Values in pixels. Default value is 0. If greater than zero it overrides other padding options.
- in_forder, out_forder — set order of faces for the input/output cubemap. Choose one direction for each position. Designation of directions:
  - r — right
  - l — left
  - u — up
  - d — down
  - f — forward
  - b — back Default value is rludfb.
- in_frot, out_frot — set rotation of faces for the input/output cubemap. Choose one angle for each position. Designation of angles:
  - 0 — 0 degrees clockwise
  - 1 — 90 degrees clockwise
  - 2 — 180 degrees clockwise
  - 3 — 270 degrees clockwise Default value is 000000.
eac — Equi-Angular Cubemap.
flat, gnomonic, rectilinear — regular video.
- Format specific options:
- h_fov, v_fov, d_fov — set output horizontal/vertical/diagonal field of view. Values in degrees. If diagonal field of view is set it overrides horizontal and vertical field of view.
- ih_fov, iv_fov, id_fov — set input horizontal/vertical/diagonal field of view. Values in degrees. If diagonal field of view is set it overrides horizontal and vertical field of view.
dfisheye — dual fisheye.
- Format specific options:
- h_fov, v_fov, d_fov — set output horizontal/vertical/diagonal field of view. Values in degrees. If diagonal field of view is set it overrides horizontal and vertical field of view.
- ih_fov, iv_fov, id_fov — set input horizontal/vertical/diagonal field of view. Values in degrees. If diagonal field of view is set it overrides horizontal and vertical field of view.
barrel, fb, barrelsplit — Facebook's 360 formats.
sg — stereographic format.
- Format specific options:
- h_fov, v_fov, d_fov — set output horizontal/vertical/diagonal field of view. Values in degrees. If diagonal field of view is set it overrides horizontal and vertical field of view.
- ih_fov, iv_fov, id_fov — set input horizontal/vertical/diagonal field of view. Values in degrees. If diagonal field of view is set it overrides horizontal and vertical field of view.
mercator — Mercator format.
ball — Ball format, gives significant distortion toward the back.
hammer — Hammer-Aitoff map projection format.
sinusoidal — Sinusoidal map projection format.
fisheye — fisheye projection.
- Format specific options:
- h_fov, v_fov, d_fov — set output horizontal/vertical/diagonal field of view. Values in degrees. If diagonal field of view is set it overrides horizontal and vertical field of view.
- ih_fov, iv_fov, id_fov — set input horizontal/vertical/diagonal field of view. Values in degrees. If diagonal field of view is set it overrides horizontal and vertical field of view.
pannini — Pannini projection.
- Format specific options:
- h_fov — set output pannini parameter.
- ih_fov — set input pannini parameter.
cylindrical — cylindrical projection.
- Format specific options:
- h_fov, v_fov, d_fov — set output horizontal/vertical/diagonal field of view. Values in degrees. If diagonal field of view is set it overrides horizontal and vertical field of view.
- ih_fov, iv_fov, id_fov — set input horizontal/vertical/diagonal field of view. Values in degrees. If diagonal field of view is set it overrides horizontal and vertical field of view.
perspective — perspective projection. (output only 👀)
- Format specific options:
- v_fov — set perspective parameter.
tetrahedron — Tetrahedron projection.
tsp — Truncated square pyramid projection.
he, hequirect — half equirectangular projection.
equisolid — equisolid format.
- Format specific options:
- h_fov, v_fov, d_fov — set output horizontal/vertical/diagonal field of view. Values in degrees. If diagonal field of view is set it overrides horizontal and vertical field of view.
- ih_fov, iv_fov, id_fov — set input horizontal/vertical/diagonal field of view. Values in degrees. If diagonal field of view is set it overrides horizontal and vertical field of view.
og — orthographic format.
- Format specific options:
- h_fov, v_fov, d_fov — set output horizontal/vertical/diagonal field of view. Values in degrees. If diagonal field of view is set it overrides horizontal and vertical field of view.
- ih_fov, iv_fov, id_fov — set input horizontal/vertical/diagonal field of view. Values in degrees. If diagonal field of view is set it overrides horizontal and vertical field of view.
octahedron — Octahedron projection.
cylindricalea — Cylindrical Equal Area projection.
interp — sets the interpolation method. Note: more complex interpolation methods require much more memory to run 🫶 like your RAM is gonna feel it.

Available methods:

near, nearest — nearest neighbour.
line, linear — bilinear interpolation.
lagrange9 — Lagrange9 interpolation.
cube, cubic — bicubic interpolation.
lanc, lanczos — Lanczos interpolation.
sp16, spline16 — Spline16 interpolation.
gauss, gaussian — Gaussian interpolation.
mitchell — Mitchell interpolation.

Default? line. That's the whole vibe, bestie, nothing else to say. 💅

w, h — set the output video resolution, obviously. The default resolution depends on formats — like, it's not gonna magically pick for you, check your formats. 🫶
in_stereo, out_stereo — set the input/output stereo format. The menu is:
- 2d — 2D mono
- sbs — side by side
- tb — top bottom

Default value is 2d for both input and output format. Classic. ✨

yaw, pitch, roll — set rotation for the output video. Values are in degrees. Spin it, tilt it, roll it, go off. 🌀
rorder — sets the rotation order for the output video. You pick one item for each position, so three picks total, no skipping:
- y, Y — yaw
- p, P — pitch
- r, R — roll

Default value is ypr. iykyk.

h_flip, v_flip, d_flip — flip the output video horizontally (swaps left↔right) / vertically (swaps up↔down) / in-depth (swaps back↔forward). Boolean values, so yes/no energy only. 🔄
ih_flip, iv_flip — set if the input video is flipped horizontally/vertically. Boolean values, keep it truthful.
in_trans — set if the input video is transposed. Boolean value, by default disabled.
out_trans — set if the output video needs to be transposed. Boolean value, by default disabled.
h_offset, v_offset — set output horizontal/vertical off-axis offset. Default is 0, and the allowed range is from -1 to 1. Nudge responsibly. 🎯
alpha_mask — builds a mask in the alpha plane for all unmapped pixels by marking them fully transparent. Boolean value, by default disabled.
reset_rot — reset the rotation of the output video. Boolean value, by default disabled. Un-spin era. 🙈

📸 Examples 📸

Convert equirectangular video to cubemap with 3x2 layout and 1% padding using bicubic interpolation:

ffmpeg -i input.mkv -vf v360=e:c3x2:cubic:out_pad=0.01 output.mkv

Extract back view of Equi-Angular Cubemap:

ffmpeg -i input.mkv -vf v360=eac:flat:yaw=180 output.mkv

Convert transposed and horizontally flipped Equi-Angular Cubemap in side-by-side stereo format to equirectangular top-bottom stereo format:

v360=eac:equirect:in_stereo=sbs:in_trans=1:ih_flip=1:out_stereo=tb

🕹️ Commands 🕹️

This filter supports a subset of the above options as commands. Not all of them, just a subset — don't come for me. 💅

🧚 vaguedenoiser

Apply a wavelet based denoiser. Yes queen, denoise. ✨

Here's the tea: it transforms each frame from the video input into the wavelet domain using Cohen-Daubechies-Feauveau 9/7. Then it applies some filtering to the obtained coefficients. Then it does an inverse wavelet transform after. And because wavelet properties are iconic like that, it should give you a nice smoothed result with reduced noise, without blurring picture features. Periodt. 💖

This filter accepts the following options:

threshold — the filtering strength. Higher = more filtered the video will be. Ngl, hard thresholding can use a higher threshold than soft thresholding before the video starts looking overfiltered. Default value is 2.
method — the filtering method the filter will use. It accepts the following values:
- hard — all values under the threshold will be zeroed. Just gone. 💀
- soft — all values under the threshold will be zeroed, and all values above will be reduced by the threshold.
- garrote — scales or nullifies coefficients, so it's the intermediary between (more) soft and (less) hard thresholding. Main character of the trio, honestly.

Default is garrote.

nsteps — number of times the wavelet will decompose the picture. Plot twist: the picture can't be decomposed beyond a particular point (typically, 8 for a 640x480 frame — as 2^9 = 512 > 480). Valid values are integers between 1 and 32. Default value is 6.
percent — partial of full denoising (limited coefficients shrinking), from 0 to 100. Default value is 85.
planes — a list of the planes to process. By default all planes are processed.
type — the threshold type the filter will use. It accepts the following values:
- universal — the threshold used is the same for all decompositions.
- bayes — the threshold used also depends on each decomposition's coefficients.

Default is universal.

🌈 varblur

Apply a variable blur filter by using the 2nd video stream to set the blur radius. The 2nd stream must have the same dimensions — non-negotiable, bestie. 🫶

This filter accepts the following options:

min_r — set min allowed radius. Allowed range is from 0 to 254. Default is 0.
max_r — set max allowed radius. Allowed range is from 1 to 255. Default is 8.
planes — set which planes to process. By default, all are used.

The varblur filter also supports the framesync options. Slay. ✨

🕹️ Commands 🕹️

This filter supports all the above options as commands. All of them. Generous queen.

🦋 vectorscope

Display 2 color component values in the two dimensional graph (which is called a vectorscope). It's giving data visualization but make it color theory. 🎨

This filter accepts the following options:

mode, m — set vectorscope mode. It accepts the following values:
- gray, tint — gray values are displayed on the graph, and higher brightness means more pixels have the same component color value at that location in the graph. This is the default mode.
- color — gray values are displayed on the graph. Surrounding pixel values which are not present in the video frame are drawn in a gradient of the 2 color components which are set by option x and y. The 3rd color component is static.
- color2 — the actual color component values present in the video frame are displayed on the graph.
- color3 — similar to color2, but a higher frequency of the same values x and y on the graph increases the value of another color component, which is luminance by default values of x and y.
- color4 — the actual colors present in the video frame are displayed on the graph. If two different colors map to the same position on the graph, then the color with the higher value of the component not present in the graph is picked.
- color5 — gray values are displayed on the graph. Similar to color but with the 3rd color component picked from a radial gradient.
x — set which color component will be represented on the X-axis. Default is 1.
y — set which color component will be represented on the Y-axis. Default is 2.
intensity, i — set intensity, used by modes gray, color, color3 and color5 for increasing the brightness of the color component which represents the frequency of the (X, Y) location in the graph.
envelope, e —
- none — no envelope, this is the default.
- instant — instant envelope, so even the darkest single pixel will be clearly highlighted. Not a single pixel left unseen. 👑
- peak — hold the maximum and minimum values presented in the graph over time. This way you can still spot out-of-range values without constantly staring at the vectorscope like it owes you money.
- peak+instant — peak and instant envelope combined together. Both. At once. We love a collab. 💅
graticule, g — set what kind of graticule to draw:
- none
- green
- color
- invert
opacity, o — set graticule opacity.
flags, f — set graticule flags:
- white — draw graticule for the white point.
- black — draw graticule for the black point.
- name — draw color point short names.
bgopacity, b — set background opacity.
lthreshold, l — set low threshold for the color component not represented on the X or Y axis. Values lower than this value will be ignored. Default is 0.

Heads up: this value is multiplied with the actual max possible value one pixel component can have. So for 8-bit input and a low threshold value of 0.1, the actual threshold is 0.1 * 255 = 25. Math is math, bestie. 🧮

hthreshold, h — set high threshold for the color component not represented on the X or Y axis. Values higher than this value will be ignored. Default is 1.

Same deal: this value is multiplied with the actual max possible value one pixel component can have. So for 8-bit input and a high threshold value of 0.9, the actual threshold is 0.9 * 255 = 230.

colorspace, c — set what kind of colorspace to use when drawing the graticule:
- auto
- 601
- 709

Default is auto.

tint0, t0, tint1, t1 — set the color tint for gray/tint vectorscope mode. By default both options are zero. That means no tint, and the output will remain gray. Respectfully monochrome. 🤍

💜 vidstabdetect

Analyze video stabilization/deshaking. This is perform pass 1 of 2, so see vidstabtransform for pass 2. Two-part saga, we love a sequel. ✨

This filter generates a file with relative translation and rotation transform information about subsequent frames, which is then used by the vidstabtransform filter.

To enable compilation of this filter you need to configure FFmpeg with --enable-libvidstab. No lib, no filter, that's just how it is. 😭

This filter accepts the following options:

result — set the path to the file used to write the transforms information. Default value is transforms.trf.
shakiness — set how shaky the video is and how quick the camera is. It accepts an integer in the range 1-10; a value of 1 means little shakiness, a value of 10 means strong shakiness. Default value is 5.
accuracy — set the accuracy of the detection process. It must be a value in the range 1-15. A value of 1 means low accuracy, a value of 15 means high accuracy. Default value is 15.
stepsize — set stepsize of the search process. The region around the minimum is scanned with 1 pixel resolution. Default value is 6.
mincontrast — set minimum contrast. Below this value a local measurement field is discarded. Must be a floating point value in the range 0-1. Default value is 0.3.
tripod — set reference frame number for tripod mode.

If enabled, the motion of the frames is compared to a reference frame in the filtered stream, identified by the specified number. The idea is to compensate all movements in a more-or-less static scene and keep the camera view absolutely still. Tripod mode is literally the "hold still, bestie" of filters. 🎀

If set to 0, it is disabled. The frames are counted starting from 1.

show — show fields and transforms in the resulting frames. It accepts an integer in the range 0-2. Default value is 0, which disables any visualization.
fileformat — format for the transforms data file to be written. Acceptable values are:
- ascii — human-readable plain text
- binary — binary format, roughly 40% smaller than ascii. (default)

📸 Examples 📸

Use default values:

vidstabdetect

Analyze a strongly shaky movie and put the results in file mytransforms.trf:

vidstabdetect=shakiness=10:accuracy=15:result="mytransforms.trf"

Visualize the result of internal transformations in the resulting video:

vidstabdetect=show=1

Analyze a video with medium shakiness using ffmpeg:

ffmpeg -i input -vf vidstabdetect=shakiness=5:show=1 dummy.avi

vidstabtransform 🎬✨

Video stabilization/deshaking: this is pass 2 of 2, and you peep vidstabdetect for pass 1. It reads a file that has transform info for each frame and applies/compensates those transforms. Team it up with the vidstabdetect filter and boom, you can deshake videos. See also http://public.hronopik.de/vid.stab.

It is important to also use the unsharp filter, see below. fr fr, don’t skip it.

To enable compilation of this filter you need to configure FFmpeg with --enable-libvidstab.

Options

input — sets the path to the file used to read the transforms. Default value is transforms.trf.
smoothing — sets the number of frames (value*2 + 1) used for lowpass filtering the camera movements. Default value is 10.

For example, a number of 10 means 21 frames are used (10 in the past and 10 in the future) to smoothen the motion in the video. A larger value leads to a smoother video, but limits the acceleration of the camera (pan/tilt movements). 0 is a special case where a static camera is simulated. iconic little zero behavior, ngl.

optalgo — sets the camera path optimization algorithm.

Accepted values are:

gauss — gaussian kernel low-pass filter on camera motion (default)
avg — averaging on transformations

maxshift — sets the maximal number of pixels to translate frames. Default value is -1, meaning no limit.
maxangle — sets the maximal angle in radians (degree*PI/180) to rotate frames. Default value is -1, meaning no limit.
crop — specifies how to deal with borders that may be visible due to movement compensation.

Available values are:

keep — keep image information from previous frame (default)
black — fill the border black

invert — invert transforms if set to 1. Default value is 0.
relative — consider transforms as relative to previous frame if set to 1, absolute if set to 0. Default value is 0.
zoom — sets percentage to zoom. A positive value gives a zoom-in effect, a negative value gives a zoom-out effect. Default value is 0 (no zoom).
optzoom — sets optimal zooming to avoid borders.

Accepted values are:

0 — disabled
1 — optimal static zoom value is determined (only very strong movements will lead to visible borders) (default)
2 — optimal adaptive zoom value is determined (no borders will be visible), see zoomspeed

Note that the value given at zoom is added to the one calculated here.

zoomspeed — sets percent to zoom maximally each frame (enabled when optzoom is set to 2). Range is from 0 to 5, default value is 0.25.
interpol — specifies type of interpolation.

Available values are:

no — no interpolation
linear — linear only horizontal
bilinear — linear in both directions (default)
bicubic — cubic in both directions (slow)

tripod — enable virtual tripod mode if set to 1, which is equivalent to relative=0:smoothing=0. Default value is 0.

Use also tripod option of vidstabdetect.

debug — increase log verbosity if set to 1. Also the detected global motions are written to the temporary file global_motions.trf. Default value is 0.

Examples

Use ffmpeg for a typical stabilization with default values:

ffmpeg -i inp.mpeg -vf vidstabtransform,unsharp=5:5:0.8:3:3:0.4 inp_stabilized.mpeg

Note the use of the unsharp filter which is always recommended. like always always.

Zoom in a bit more and load transform data from a given file:

vidstabtransform=zoom=5:input="mytransforms.trf"

Smoothen the video even more:

vidstabtransform=smoothing=30

vflip 🔄💅

Flip the input video vertically. simple, clean, iconic.

For example, to vertically flip a video with ffmpeg:

ffmpeg -i in.avi -vf "vflip" out.avi

vfrdet 🕵️‍♀️🎞️

Detect variable frame rate video.

This filter tries to detect if the input is variable or constant frame rate.

At end it will output number of frames detected as having variable delta pts, and ones with constant delta pts. If there was frames with variable delta, then it will also show min, max and average delta encountered. numbers girlies rise.

vibrance 🌈💖

Boost or alter saturation.

The filter accepts the following options:

intensity — set strength of boost if positive value, or strength of alter if negative value. Default is 0. Allowed range is from -2 to 2.
rbal — set the red balance. Default is 1. Allowed range is from -10 to 10.
gbal — set the green balance. Default is 1. Allowed range is from -10 to 10.
bbal — set the blue balance. Default is 1. Allowed range is from -10 to 10.
rlum — set the red luma coefficient.
glum — set the green luma coefficient.
blum — set the blue luma coefficient.
alternate — if intensity is negative and this is set to 1, colors will change, otherwise colors will be less saturated, more towards gray.

Commands

All the options above also work as commands. slay.

vif 👀📊

Obtain the average VIF (Visual Information Fidelity) between two input videos.

This filter takes two input videos.

Both input videos must have the same resolution and pixel format for this filter to work correctly. Also it assumes that both inputs have the same number of frames, which are compared one by one.

The obtained average VIF score is printed through the logging system.

The filter stores the calculated VIF score of each frame.

This filter also supports the framesync options.

In the below example the input file main.mpg being processed is compared with the reference file ref.mpg.

ffmpeg -i main.mpg -i ref.mpg -lavfi vif -f null -

vignette 🌸🕯️

Make or reverse a natural vignetting effect.

The filter accepts the following options:

angle, a — set lens angle expression as a number of radians.

The value is clipped in the [0,PI/2] range.

Default value: PI/5

x0, y0 — set center coordinates expressions. Respectively w/2 and h/2 by default.
mode — set forward/backward mode.

Available modes are:

forward — the larger the distance from the central point, the darker the image becomes.
backward — the larger the distance from the central point, the brighter the image becomes. This can be used to reverse a vignette effect, though there is no automatic detection to extract the lens angle and other settings (yet). It can also be used to create a burning effect.

Default value is forward.

eval — set evaluation mode for the expressions (angle, x0, y0).

It accepts the following values:

init — evaluate expressions only once during the filter initialization.
frame — evaluate expressions for each incoming frame. This is way slower than the init mode since it requires all the scalers to be re-computed, but it allows advanced dynamic expressions.

Default value is init.

dither — set dithering to reduce the circular banding effects. Default is 1 (enabled).
aspect — set vignette aspect. This setting allows one to adjust the shape of the vignette. Setting this value to the SAR of the input will make a rectangular vignetting following the dimensions of the video.

Default is 1/1.

Expressions

The alpha, x0 and y0 expressions can contain the following parameters.

w, h — input width and height
n — the number of input frame, starting from 0
pts — the PTS (Presentation TimeStamp) time of the filtered video frame, expressed in TB units, NAN if undefined
r — frame rate of the input video, NAN if the input frame rate is unknown
t — the PTS (Presentation TimeStamp) of the filtered video frame, expressed in seconds, NAN if undefined
tb — time base of the input video

Examples

Apply simple strong vignetting effect:

vignette=PI/4

Make a flickering vignetting:

vignette='PI/4+random(1)*PI/50':eval=frame

vmafmotion 🎥📈

Obtain the average VMAF motion score of a video. It is one of the component metrics of VMAF.

The obtained average motion score is printed through the logging system.

The filter accepts the following options:

stats_file — if specified, the filter will use the named file to save the motion score of each frame with respect to the previous frame. When filename equals - the data is sent to standard output.

Example:

ffmpeg -i ref.mpg -vf vmafmotion -f null -

vpp_amf ⚡🎛️

Scale (resize) and convert colorspace, transfer characteristics or color primaries for the input video, using AMD Advanced Media Framework library for hardware acceleration. Setting the output width and height works in the same way as for the scale filter.

The filter accepts the following options:

w, h — set the output video dimension expression. Default value is the input dimension.

Allows for the same expressions as the scale filter.

scale_type — sets the algorithm used for scaling:
- bilinear — Bilinear

This is the default.

bicubic — Bicubic

format — controls the output pixel format. By default, or if none is specified, the input pixel format is used.
force_original_aspect_ratio, force_divisible_by — work the same as the identical scale filter options.
reset_sar — works the same as the identical scale filter option.
color_profile — specify all color properties at once.

The accepted values are:

bt601 — BT.601
bt709 — BT.709
bt2020 — BT.2020

trc — specify output transfer characteristics.

The accepted values are:

bt709 — BT.709
gamma22 — Constant gamma of 2.2
gamma28 — Constant gamma of 2.8
smpte170m — SMPTE-170M
smpte240m — SMPTE-240M
linear — Linear
log — LOG
log-sqrt — LOG_SQRT
iec61966-2-4 — iec61966-2-4
bt1361-ecg — BT1361_ECG
iec61966-2-1 — iec61966-2-1
bt2020-10 — BT.2020 for 10-bits content
bt2020-12 — BT.2020 for 12-bits content

smpte2084 — SMPTE2084 💅
smpte428 — SMPTE428 ✨
arib-std-b67 — ARIB_STD_B67 🦋

🎨 primaries 🎨

This one sets the output color primaries, bestie — like, what physical colors your RGB points actually are. The accepted values are:

bt709 — BT.709 👑
bt470m — BT.470M
bt470bg — BT.470BG or BT.601-6 625 🌸
smpte170m — SMPTE-170M or BT.601-6 525
smpte240m — SMPTE-240M
film — film 🎞️ (yes, literally just "film", iconic)
bt2020 — BT.2020 🔥
smpte428 — SMPTE-428
smpte431 — SMPTE-431
smpte432 — SMPTE-432
jedec-p22 — JEDEC P22 phosphors ⭐

💫 Examples 💫

Scale input to 720p, keeping aspect ratio and ensuring the output is yuv420p. No cap, this is the classic:

vpp_amf=-2:720:format=yuv420p

Upscale to 4K and change color profile to bt2020. Big screen energy 🌈:

vpp_amf=4096:2160:color_profile=bt2020

📚 vstack 📚

Stacks your input videos vertically — one on top of the other, like a lil video sandwich 🥪.

All streams must be of same pixel format and of same width. Like, that's non-negotiable bestie, they gotta match.

Note that this filter is faster than using overlay and pad filter to create same output — so like, stop doing it the hard way 💅.

The filter accepts the following options:

inputs — Set number of input streams. Default is 2. 🫶
shortest — If set to 1, force the output to terminate when the shortest input terminates. Default value is 0.

🧚 w3fdif 🧚

Deinterlace the input video (w3fdif stands for "Weston 3 Field Deinterlacing Filter") — say that five times fast, I dare you 😭.

Based on the process described by Martin Weston for BBC R&D, and implemented based on the de-interlace algorithm written by Jim Easterbrook for BBC R&D, the Weston 3 field deinterlacing filter uses filter coefficients calculated by BBC R&D. Shoutout the BBC, fr fr 👑

This filter uses field-dominance information in frame to decide which of each pair of fields to place first in the output. If it gets it wrong use setfield filter before w3fdif filter. That's your lil fix-it moment 🩹.

There are two sets of filter coefficients, so called "simple" and "complex". Which set of filter coefficients is used can be set by passing an optional parameter:

filter — Set the interlacing filter coefficients. Accepts one of the following values:
- simple — Simple filter coefficient set.
- complex — More-complex filter coefficient set.

Default value is complex.

mode — The interlacing mode to adopt. It accepts one of the following values:
- frame — Output one frame for each frame.
- field — Output one frame for each field.

The default value is "field".

parity — The picture field parity assumed for the input interlaced video. It accepts one of the following values:
- tff — Assume the top field is first.
- bff — Assume the bottom field is first.
- auto — Enable automatic detection of field parity.

The default value is "auto". If the interlacing is unknown or the decoder does not export this information, top field first will be assumed.

deint — Specify which frames to deinterlace. Accepts one of the following values:
- all — Deinterlace all frames,
- interlaced — Only deinterlace frames marked as interlaced.

Default value is all.

🎀 Commands 🎀

This filter supports same commands as options. Slay, consistent API 💅.

🌊 waveform 🌊

Video waveform monitor — it's giving oscilloscope but make it aesthetic 📈.

The waveform monitor plots color component intensity. By default luma only. Each column of the waveform corresponds to a column of pixels in the source video. Cute, right?

It accepts the following options:

mode, m — Can be either row, or column. Default is column. In row mode, the graph on the left side represents color component value 0 and the right side represents value = 255. In column mode, the top side represents color component value = 0 and bottom side represents value = 255. 🦋
intensity, i — Set intensity. Smaller values are useful to find out how many values of the same luminance are distributed across input rows/columns. Default value is 0.04. Allowed range is [0, 1]. ✨
mirror, r — Set mirroring mode. 0 means unmirrored, 1 means mirrored. In mirrored mode, higher values will be represented on the left side for row mode and at the top for column mode. Default is 1 (mirrored). 🪞
display, d — Set display mode. It accepts the following values:
- overlay — Presents information identical to that in the parade, except that the graphs representing color components are superimposed directly over one another.

This display mode makes it easier to spot relative differences or similarities in overlapping areas of the color components that are supposed to be identical, such as neutral whites, grays, or blacks. 👀

stack — Display separate graph for the color components side by side in row mode or one below the other in column mode.
parade — Display separate graph for the color components side by side in column mode or one below the other in row mode.

Using this display mode makes it easy to spot color casts in the highlights and shadows of an image, by comparing the contours of the top and the bottom graphs of each waveform. Since whites, grays, and blacks are characterized by exactly equal amounts of red, green, and blue, neutral areas of the picture should display three waveforms of roughly equal width/height. If not, the correction is easy to perform by making level adjustments the three waveforms. 🎨

Default is stack.

components, c — Set which color components to display. Default is 1, which means only luma or red color component if input is in RGB colorspace. If is set for example to 7 it will display all 3 (if) available color components. 🔴🟢🔵
envelope, e —
- none — No envelope, this is default.
- instant — Instant envelope, minimum and maximum values presented in graph will be easily visible even with small step value.
- peak — Hold minimum and maximum values presented in graph across time. This way you can still spot out of range values without constantly looking at waveforms. 👑
- peak+instant — Peak and instant envelope combined together. Best of both worlds, periodt.
filter, f —
- lowpass — No filtering, this is default.
- flat — Luma and chroma combined together.
- aflat — Similar as above, but shows difference between blue and red chroma.
- xflat — Similar as above, but use different colors.
- yflat — Similar as above, but again with different colors.
- chroma — Displays only chroma.
- color — Displays actual color value on waveform. 🌈
- acolor — Similar as above, but with luma showing frequency of chroma values.
graticule, g — Set which graticule to display.
- none — Do not display graticule.
- green — Display green graticule showing legal broadcast ranges. 💚
- orange — Display orange graticule showing legal broadcast ranges. 🧡
- invert — Display invert graticule showing legal broadcast ranges.
opacity, o — Set graticule opacity.
flags, fl — Set graticule flags.
- numbers — Draw numbers above lines. By default enabled.
- dots — Draw dots instead of lines. ⚪
scale, s — Set scale used for displaying graticule.
- digital
- millivolts
- ire

Default is digital.

bgopacity, b — Set background opacity.
tint0, t0
tint1, t1 — Set tint for output. Only used with lowpass filter and when display is not overlay and input pixel formats are not RGB. 💜
fitmode, fm — Set sample aspect ratio of video output frames. Can be used to configure waveform so it is not stretched too much in one of directions.
- none — Set sample aspect ration to 1/1.
- size — Set sample aspect ratio to match input size of video

Default is none.

input — Set input formats for filter to pick from. Can be all, for selecting from all available formats, or first, for selecting first available format. Default is first. 🫶

🧵 weave, doubleweave 🧵

The weave takes a field-based video input and join each two sequential fields into single frame, producing a new double height clip with half the frame rate and half the frame count. Two fields walk into a bar and become one frame, iykyk 😌.

The doubleweave works same as weave but without halving frame rate and frame count. Same vibes, no rate cut 💅.

It accepts the following option:

first_field — Set first field. Available values are:
- top, t — Set the frame as top-field-first.
- bottom, b — Set the frame as bottom-field-first.

🌸 Examples 🌸

Interlace video using select and separatefields filter. Look at this lil pipeline go:

separatefields,select=eq(mod(n,4),0)+eq(mod(n,4),3),weave

🔍 xbr 🔍

Apply the xBR high-quality magnification filter which is designed for pixel art. It follows a set of edge-detection rules, see https://forums.libretro.com/t/xbr-algorithm-tutorial/123. Pixel art besties, this one's for you 🎮✨.

It accepts the following option:

n — Set the scaling dimension: 2 for "2xBR", 3 for "3xBR" and 4 for "4xBR". Default is 3.

🧮 xcorrelate 🧮

Apply normalized cross-correlation between first and second input video stream. Math girlies rise 🙌.

Second input video stream dimensions must be lower than first input video stream. Like, second place stays second, no cap.

The filter accepts the following options:

planes — Set which planes to process.
secondary — Set which secondary video frames will be processed from second input video stream, can be first or all. Default is all.

The xcorrelate filter also supports the framesync options. 🔗

🌅 xfade 🌅

Apply cross fade from one input video stream to another input video stream. The cross fade is applied for specified duration. It's giving dramatic scene transition, honestly iconic 🎬.

Both inputs must be constant frame-rate and have the same resolution, pixel format, frame rate and timebase. Yeah, they gotta be twins or this whole thing falls apart 😭.

The filter accepts the following options:

transition — Set one of available transition effects:
custom
fade
wipeleft
wiperight
wipeup
wipedown
slideleft
slideright
slideup
slidedown
circlecrop
rectcrop
distance
fadeblack
fadewhite
radial
smoothleft
smoothright
smoothup
smoothdown
circleopen
circleclose
vertopen
vertclose
horzopen
horzclose
dissolve
pixelize
diagtl
diagtr
diagbl
diagbr
hlslice
hrslice
vuslice
vdslice
hblur
fadegrays
wipetl
wipetr
wipebl
wipebr
squeezeh
squeezev
zoomin
fadefast
fadeslow
hlwind
hrwind
vuwind
vdwind
coverleft
coverright
coverup
coverdown
revealleft
revealright
revealup
revealdown

The default transition effect is fade, bestie. 🎀

duration — sets the cross fade duration in seconds. The range is 0 to 60 seconds, and the default duration is 1 second. ⏱️
offset — sets where the cross fade starts, relative to the first input stream, in seconds. Default offset is 0.
expr — sets the expression for a custom transition effect. 💅

The expressions can use the following variables and functions:

X, Y — the coordinates of the current sample.
W, H — the width and height of the image.
P — progress of the transition effect.
PLANE — the currently processed plane.
A — the return value of the first input at the current location and plane.
B — the return value of the second input at the current location and plane.
a0(x, y), a1(x, y), a2(x, y), a3(x, y) — return the value of the pixel at location (x,y) of the first/second/third/fourth component of the first input.
b0(x, y), b1(x, y), b2(x, y), b3(x, y) — return the value of the pixel at location (x,y) of the first/second/third/fourth component of the second input.

Examples ✨

Cross fade from one input video to another input video, with fade transition and a transition duration of 2 seconds starting at an offset of 5 seconds:

ffmpeg -i first.mp4 -i second.mp4 -filter_complex xfade=transition=fade:duration=2:offset=5 output.mp4

`xmedian` 👑

Picks the median pixels out of several input videos. So extra, so iconic.

The filter accepts the following options:

inputs — sets the number of inputs. Default is 3, and the allowed range is 3 to 255. Heads up: if the number of inputs is an even number, the result will be the mean value between the two median values. 🧮
planes — sets which planes to filter. Default value is 15, which means all planes get processed.
percentile — sets the median percentile. Default value is 0.5. A default of 0.5 will always pick median values, while 0 picks minimum values and 1 picks maximum values.

Commands

This filter supports all the above options as commands, except the inputs option. Not me leaving inputs out of the group chat. 🙈

`xpsnr` 📊

Gets you the average (across all input frames) and minimum (across all color plane averages) eXtended Perceptually weighted peak Signal-to-Noise Ratio — XPSNR — between two input videos. Literally the glow-up metric.

The XPSNR is a low-complexity psychovisually motivated distortion measurement algorithm for assessing the difference between two video streams or images. This is especially useful for objectively quantifying the distortions caused by video and image codecs, as an alternative to a formal subjective test. The logarithmic XPSNR output values are in a similar range as those of traditional psnr assessments but better reflect human impressions of visual coding quality. More details on the XPSNR measure, which essentially represents a blockwise weighted variant of the PSNR measure, can be found in the following freely available papers:

C. R. Helmrich, M. Siekmann, S. Becker, S. Bosse, D. Marpe, and T. Wiegand, "XPSNR: A Low-Complexity Extension of the Perceptually Weighted Peak Signal-to-Noise Ratio for High-Resolution Video Quality Assessment," in Proc. IEEE Int. Conf. Acoustics, Speech, Sig. Process. (ICASSP), virt./online, May 2020. <www.ecodis.de/xpsnr.htm>
C. R. Helmrich, S. Bosse, H. Schwarz, D. Marpe, and T. Wiegand, "A Study of the Extended Perceptually Weighted Peak Signal-to-Noise Ratio (XPSNR) for Video Compression with Different Resolutions and Bit Depths," ITU Journal: ICT Discoveries, vol. 3, no. 1, pp. 65 - 72, May 2020. <http://handle.itu.int/11.1002/pub/8153d78b-en>

When publishing the results of XPSNR assessments obtained using, e.g., this FFmpeg filter, a reference to the above papers as a means of documentation is strongly encouraged. 🫶

The filter requires two input videos. The first input is considered a (usually not distorted) reference source and is passed unchanged to the output, whereas the second input is a (distorted) test signal. Except for the bit depth, these two video inputs must have the same pixel format. In addition, for best performance, both compared input videos should be in YCbCr color format. 💜

The obtained overall XPSNR values mentioned above are printed through the logging system. In case of input with multiple color planes, we suggest reporting of the minimum XPSNR average.

The following parameter, which behaves like the one for the psnr filter, is accepted:

stats_file, f — if specified, the filter will use the named file to save the XPSNR value of each individual frame and color plane. When the file name equals -, that data is sent to standard output. 🍓

This filter also supports the framesync options.

Examples

XPSNR analysis of two 1080p HD videos, ref_source.yuv and test_video.yuv, both at 24 frames per second, with color format 4:2:0, bit depth 8, and output of a logfile named xpsnr.log:

ffmpeg -s 1920x1080 -framerate 24 -pix_fmt yuv420p -i ref_source.yuv -s 1920x1080 -framerate
24 -pix_fmt yuv420p -i test_video.yuv -lavfi xpsnr="stats_file=xpsnr.log" -f null -

XPSNR analysis of two 2160p UHD videos, ref_source.yuv with bit depth 8 and test_video.yuv with bit depth 10, both at 60 frames per second with color format 4:2:0, no logfile output:

ffmpeg -s 3840x2160 -framerate 60 -pix_fmt yuv420p -i ref_source.yuv -s 3840x2160 -framerate
60 -pix_fmt yuv420p10le -i test_video.yuv -lavfi xpsnr="stats_file=-" -f null -

`xstack` 🧱

Stacks video inputs into a custom layout. Main character collage energy.

All streams must be of the same pixel format. No mixing, that's the rule. ✋

The filter accepts the following options:

inputs — sets the number of input streams. Default is 2.
layout — specifies the layout of the inputs. This option requires the desired layout configuration to be explicitly set by the user. It sets the position of each video input in the output. Each input is separated by '|'. The first number represents the column, and the second number represents the row. Numbers start at 0 and are separated by '_'. Optionally one can use wX and hX, where X is the video input from which to take the width or height. Multiple values can be used when separated by '+' — in that case the values are summed together. ➕

Note that if inputs are of different sizes gaps may appear, as not all of the output video frame will be filled. Similarly, videos can overlap each other if their position doesn't leave enough space for the full frame of adjoining videos. It's giving chaotic scrapbook. 😭

For 2 inputs, a default layout of "0_0|w0_0" (equivalent to "grid=2x1") is set. In all other cases, a layout or a grid must be set by the user. Either grid or layout can be specified at a time. Specifying both will result in an error. 🚨

grid — specifies a fixed size grid of inputs. This option is used to create a fixed size grid of the input streams. Set the grid size in the form COLUMNSxROWS. There must be ROWS * COLUMNS input streams and they will be arranged as a grid with ROWS rows and COLUMNS columns. When using this option, each input stream within a row must have the same height and all the rows must have the same width.

If grid is set, then the inputs option is ignored and is implicitly set to ROWS * COLUMNS.

For 2 inputs, a default grid of "2x1" (equivalent to "layout=0_0|w0_0") is set. In all other cases, a layout or a grid must be set by the user. Either grid or layout can be specified at a time. Specifying both will result in an error.

shortest — if set to 1, forces the output to terminate when the shortest input terminates. Default value is 0.
fill — if set to a valid color, all unused pixels will be filled with that color. By default fill is set to none, so it is disabled. 🎨

Examples

Display 4 inputs into a 2x2 grid.

Layout:

input1(0, 0)  | input3(w0, 0)
input2(0, h0) | input4(w0, h0)

xstack=inputs=4:layout=0_0|0_h0|w0_0|w0_h0

Note that if inputs are of different sizes, gaps or overlaps may occur.

Display 4 inputs into a 1x4 grid.

Layout:

input1(0, 0)
input2(0, h0)
input3(0, h0+h1)
input4(0, h0+h1+h2)

xstack=inputs=4:layout=0_0|0_h0|0_h0+h1|0_h0+h1+h2

Note that if inputs are of different widths, unused space will appear.

Display 9 inputs into a 3x3 grid.

Layout:

input1(0, 0)       | input4(w0, 0)      | input7(w0+w3, 0)
input2(0, h0)      | input5(w0, h0)     | input8(w0+w3, h0)
input3(0, h0+h1)   | input6(w0, h0+h1)  | input9(w0+w3, h0+h1)

xstack=inputs=9:layout=0_0|0_h0|0_h0+h1|w0_0|w0_h0|w0_h0+h1|w0+w3_0|w0+w3_h0|w0+w3_h0+h1

Note that if inputs are of different sizes, gaps or overlaps may occur.

Display 16 inputs into a 4x4 grid.

Layout:

input1(0, 0)       | input5(w0, 0)       | input9 (w0+w4, 0)       | input13(w0+w4+w8, 0)
input2(0, h0)      | input6(w0, h0)      | input10(w0+w4, h0)      | input14(w0+w4+w8, h0)
input3(0, h0+h1)   | input7(w0, h0+h1)   | input11(w0+w4, h0+h1)   | input15(w0+w4+w8, h0+h1)
input4(0, h0+h1+h2)| input8(w0, h0+h1+h2)| input12(w0+w4, h0+h1+h2)| input16(w0+w4+w8, h0+h1+h2)

xstack=inputs=16:layout=0_0|0_h0|0_h0+h1|0_h0+h1+h2|w0_0|w0_h0|w0_h0+h1|w0_h0+h1+h2|w0+w4_0|
w0+w4_h0|w0+w4_h0+h1|w0+w4_h0+h1+h2|w0+w4+w8_0|w0+w4+w8_h0|w0+w4+w8_h0+h1|w0+w4+w8_h0+h1+h2

Heads up bestie: this bit starts mid-thought from the filter before it, so don't panic, nothing's missing 💖

Note that if inputs are of different sizes, gaps or overlaps may occur.

🌀 yadif

Deinterlace the input video — and yes, yadif literally stands for "yet another deinterlacing filter" 😭 these devs really said "we've done this before, we'll do it again" and honestly? iconic.

It accepts the following parameters:

mode — the interlacing mode to adopt. Pick one of these:
- 0, send_frame — output one frame for each frame.
- 1, send_field — output one frame for each field.
- 2, send_frame_nospatial — like send_frame, but it skips the spatial interlacing check. Lowkey speedrunning it 🏃‍♀️
- 3, send_field_nospatial — like send_field, but it skips the spatial interlacing check.

The default value is send_frame.

parity — the picture field parity assumed for the input interlaced video. It accepts one of the following values:
- 0, tff — assume the top field is first. Top first, main character energy 👑
- 1, bff — assume the bottom field is first.
- -1, auto — enable automatic detection of field parity. Let the filter figure it out, she's smart ✨

The default value is auto. If the interlacing is unknown or the decoder does not export this information, top field first will be assumed.

deint — specify which frames to deinterlace. Accepts one of the following values:
- 0, all — deinterlace all frames.
- 1, interlaced — only deinterlace frames marked as interlaced.

The default value is all.

🌸 yaepblur

Apply blur filter while preserving edges — and yaepblur means "yet another edge preserving blur filter" 🫠 the naming department was clearly running on fumes but the algorithm slaps. It's described in J. S. Lee, Digital image enhancement and noise filtering by use of local statistics, IEEE Trans. Pattern Anal. Mach. Intell. PAMI-2, 1980. — yes that's a real citation, she's academic 💅

It accepts the following parameters:

radius, r — set the window radius. Default value is 3.
planes, p — set which planes to filter. Default is only the first plane.
sigma, s — set blur strength. Default value is 128.

Commands

This filter supports same commands as options. Simple as that, no notes 🎀

🔍 zoompan

Apply Zoom & Pan effect. Cinema. Peak. 🎬

This filter accepts the following options:

zoom, z — set the zoom expression. Range is 1-10. Default is 1.
x
y — set the x and y expression. Default is 0.
d — set the duration expression in number of frames. This sets for how many number of frames effect will last for single input image. Default is 90.
s — set the output image size, default is hd720.
fps — set the output frame rate, default is 25.

Each expression can contain the following constants:

in_w, iw — input width.
in_h, ih — input height.
out_w, ow — output width.
out_h, oh — output height.
in — input frame count.
on — output frame count.
in_time, it — the input timestamp expressed in seconds. It's NAN if the input timestamp is unknown. 🫥
out_time, time, ot — the output timestamp expressed in seconds.
x
y — last calculated x and y position from x and y expression for current input frame.
px
py — x and y of last output frame of previous input frame or 0 when there was not yet such frame (first input frame).
zoom — last calculated zoom from z expression for current input frame.
pzoom — last calculated zoom of last output frame of previous input frame.
duration — number of output frames for current input frame. Calculated from d expression for each input frame.
pduration — number of output frames created for previous input frame.
a — rational number: input width / input height.
sar — sample aspect ratio.
dar — display aspect ratio.

Examples

Zoom in up to 1.5x and pan at same time to some spot near center of picture:

zoompan=z='min(zoom+0.0015,1.5)':d=700:x='if(gte(zoom,1.5),x,x+1/a)':y='if(gte(zoom,1.5),y,y+1)':s=640x360

Zoom in up to 1.5x and pan always at center of picture:

zoompan=z='min(zoom+0.0015,1.5)':d=700:x='iw/2-(iw/zoom/2)':y='ih/2-(ih/zoom/2)'

Same as above but without pausing:

zoompan=z='min(max(zoom,pzoom)+0.0015,1.5)':d=1:x='iw/2-(iw/zoom/2)':y='ih/2-(ih/zoom/2)'

Zoom in 2x into center of picture only for the first second of the input video:

zoompan=z='if(between(in_time,0,1),2,1)':d=1:x='iw/2-(iw/zoom/2)':y='ih/2-(ih/zoom/2)'

🦋 zscale

Scale (resize) the input video, using the z.lib library: <https://github.com/sekrit-twc/zimg>. To enable compilation of this filter, you need to configure FFmpeg with --enable-libzimg.

The zscale filter forces the output display aspect ratio to be the same as the input, by changing the output sample aspect ratio. No stretching, no distortion, we respect the aspect ratio in this house 🫶

If the input image format is different from the format requested by the next filter, the zscale filter will convert the input to the requested format. Helpful queen behavior.

Options

The filter accepts the following options.

width, w
height, h — set the output video dimension expression. Default value is the input dimension.

If the width or w value is 0, the input width is used for the output. If the height or h value is 0, the input height is used for the output.

If one and only one of the values is -n with n >= 1, the zscale filter will use a value that maintains the aspect ratio of the input image, calculated from the other specified dimension. After that it will, however, make sure that the calculated dimension is divisible by n and adjust the value if necessary. Math girlies stay winning 📐✨

If both values are -n with n >= 1, the behavior will be identical to both values being set to 0 as previously detailed.

See below for the list of accepted constants for use in the dimension expression.

size, s — set the video size. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual.
dither, d — set the dither type.

Possible values are:

none
ordered
random
error_diffusion

Default is none.

filter, f — set the resize filter type.

Possible values are:

point
bilinear
bicubic
spline16
spline36
spline64
lanczos

Default is bilinear.

range, r — set the color range.

Possible values are:

input
limited
full

Default is same as input.

primaries, p — set the color primaries.

Possible values are:

input
709
unspecified
170m
240m
2020

Default is same as input.

transfer, t — set the transfer characteristics.

Possible values are:

input
709
unspecified
601
linear
2020_10
2020_12
smpte2084
iec61966-2-1
arib-std-b67

Default is same as input.

matrix, m — set the colorspace matrix.

Possible value are:

input
709
unspecified
470bg
170m
2020_ncl
2020_cl

Default is same as input.

rangein, rin — set the input color range.

Possible values are:

input
limited
full

Default is same as input.

primariesin, pin — set the input color primaries.

Possible values are:

input
709
unspecified
170m
240m
2020

Default is same as input.

transferin, tin — set the input transfer characteristics.

Possible values are:

input
709
unspecified
601
linear
2020_10
2020_12

Default is same as input.

matrixin, min — set the input colorspace matrix.

Possible value are:

input
709
unspecified
470bg
170m
2020_ncl
2020_cl

chromal, c — set the output chroma location.

Possible values are:

input
left
center
topleft
top
bottomleft
bottom

chromalin, cin — set the input chroma location.

Possible values are:

input
left
center
topleft
top
bottomleft
bottom

npl — set the nominal peak luminance.
param_a — parameter A for scaling filters. Parameter b for bicubic, and the number of filter taps for lanczos.
param_b — parameter B for scaling filters. Parameter c for bicubic.

The values of the w and h options are expressions containing the following constants:

in_w
in_h — the input width and height.
iw
ih — these are the same as in_w and in_h.
out_w
out_h — the output (scaled) width and height.
ow
oh — these are the same as out_w and out_h.
a — the same as iw / ih.
sar — input sample aspect ratio.
dar — the input display aspect ratio. Calculated from (iw / ih) * sar.
hsub
vsub — horizontal and vertical input chroma subsample values. For example for the pixel format yuv422p hsub is 2 and vsub is 1.
ohsub
ovsub — horizontal and vertical output chroma subsample values. For example for the pixel format yuv422p hsub is 2 and vsub is 1.

Commands

This filter supports the following commands:

width, w
height, h — set the output video dimension expression. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value. No crash, no drama, she just keeps it 💅

🌈 CUDA VIDEO FILTERS

To enable CUDA and/or NPP filters please refer to configuration guidelines for CUDA and for CUDA NPP filters. (Yes, the docs are a vibe-check-yourself situation here, go read those 🫶)

Running CUDA filters requires you to initialize a hardware device and to pass that device to all filters in any filter graph. Like introducing your friend to the whole group chat — everyone needs the intro or nobody talks to them.

-init_hw_device cuda[=name][:device[,key=value...]] — initialise a new hardware device of type cuda called name, using the given device parameters.
-filter_hw_device name — pass the hardware device called name to all filters in any filter graph.

For more detailed information see <https://www.ffmpeg.org/ffmpeg.html#Advanced-Video-options>

Example of initializing second CUDA device on the system and running scale_cuda and bilateral_cuda filters.

./ffmpeg -hwaccel cuda -hwaccel_output_format cuda -i input.mp4 -init_hw_device cuda:1 -filter_complex \
"[0:v]scale_cuda=format=yuv444p[scaled_video];[scaled_video]bilateral_cuda=window_size=9:sigmaS=3.0:sigmaR=50.0" \
-an -sn -c:v h264_nvenc -cq 20 out.mp4

Fr fr copy that one exactly — the cuda:1 is doing the whole "use my second GPU" thing and the \ line breaks matter 💀

Quick vibe check before the filters: CUDA filters operate exclusively on GPU memory, so frame data sometimes has to be uploaded (hwupload) to hardware surfaces associated with the appropriate CUDA device before processing — and then downloaded (hwdownload) back to normal memory afterward, if you need it. Whether hwupload or hwdownload is actually necessary totally depends on your workflow, bestie:

If your input frames are already in GPU memory (like, say, when you're using -hwaccel cuda or -hwaccel_output_format cuda), you do not need to explicitly use hwupload — the data is already chilling in the right memory space. 💅
If your input frames are in CPU memory (software-decoded frames, or frames that got processed by CPU-based filters, iykyk), then you do need hwupload to move that data over to GPU memory for the CUDA processing. 🚚
If the output of your CUDA filters needs further processing by software-based filters, or needs saving in a format the GPU-based encoders don't support, then hwdownload is required to ferry the data back to CPU memory. 🛟

Heads up, queen: hwupload uploads data to a surface with the same layout as the software frame, so you may need to drop a format filter immediately before hwupload to make sure the input is in the right format. Same energy the other way — hwdownload may not support all output formats, so you might need to slot an extra format filter in immediately after hwdownload in the filter graph for compatibility. 🌸

✨ CUDA ✨

Okay so here's the rundown on the currently available Nvidia CUDA video filters. 🎀

Prerequisites:

Install Nvidia CUDA Toolkit 👑

Note: if FFmpeg detects the Nvidia CUDA Toolkit during configuration, it'll just enable CUDA filters automatically, no extra flags required. But if you wanna explicitly enable them, use these: - Configure FFmpeg with --enable-cuda-nvcc --enable-nonfree. - Configure FFmpeg with --enable-cuda-llvm. Additional requirement: the llvm lib must be installed.

bilateral_cuda

CUDA accelerated bilateral filter — an edge-preserving filter, bestie. 🦋 This one is mathematically accurate thanks to the GPU acceleration glow-up. For best output quality, use one-to-one chroma subsampling, i.e. yuv444p format.

The filter accepts the following options:

sigmaS — set the sigma of the gaussian function used to calculate the spatial weight, also called sigma space. Allowed range is 0.1 to 512. Default is 0.1.
sigmaR — set the sigma of the gaussian function used to calculate the color range weight, also called sigma color. Allowed range is 0.1 to 512. Default is 0.1.
window_size — set the window size of the bilateral function, which determines the number of neighbours to loop on. If the number you enter is even, one will be added automatically (the filter simply refuses to be weird about it ✨). Allowed range is 1 to 255. Default is 1.

Examples

Apply the bilateral filter on a video:

./ffmpeg -v verbose \
-hwaccel cuda -hwaccel_output_format cuda -i input.mp4  \
-init_hw_device cuda \
-filter_complex \
" \
[0:v]scale_cuda=format=yuv444p[scaled_video];
[scaled_video]bilateral_cuda=window_size=9:sigmaS=3.0:sigmaR=50.0" \
-an -sn -c:v h264_nvenc -cq 20 out.mp4

bwdif_cuda

Deinterlace the input video using the bwdif algorithm, but implemented in CUDA so it can run as part of a GPU accelerated pipeline with nvdec and/or nvenc. 🔥

It accepts the following parameters:

mode — the interlacing mode to adopt. It accepts one of the following values:
- 0, send_frame — output one frame for each frame.
- 1, send_field — output one frame for each field.

The default value is send_field.

parity — the picture field parity assumed for the input interlaced video. It accepts one of the following values:
- 0, tff — assume the top field is first.
- 1, bff — assume the bottom field is first.
- -1, auto — enable automatic detection of field parity.

The default value is auto. If the interlacing is unknown or the decoder does not export this information, top field first will be assumed.

deint — specify which frames to deinterlace. Accepts one of the following values:
- 0, all — deinterlace all frames.
- 1, interlaced — only deinterlace frames marked as interlaced.

The default value is all.

chromakey_cuda

CUDA accelerated YUV colorspace color/chroma keying. 💚

This filter works like the normal chromakey filter but operates on CUDA frames. For more details and parameters see chromakey.

Examples

Make all the green pixels in the input video transparent and use it as an overlay for another video:

./ffmpeg \
    -hwaccel cuda -hwaccel_output_format cuda -i input_green.mp4  \
    -hwaccel cuda -hwaccel_output_format cuda -i base_video.mp4 \
    -init_hw_device cuda \
    -filter_complex \
    " \
        [0:v]chromakey_cuda=0x25302D:0.1:0.12:1[overlay_video]; \
        [1:v]scale_cuda=format=yuv420p[base]; \
        [base][overlay_video]overlay_cuda" \
    -an -sn -c:v h264_nvenc -cq 20 output.mp4

Process two software sources, explicitly uploading the frames:

./ffmpeg -init_hw_device cuda=cuda -filter_hw_device cuda \
    -f lavfi -i color=size=800x600:color=white,format=yuv420p \
    -f lavfi -i yuvtestsrc=size=200x200,format=yuv420p \
    -filter_complex \
    " \
        [0]hwupload[under]; \
        [1]hwupload,chromakey_cuda=green:0.1:0.12[over]; \
        [under][over]overlay_cuda" \
    -c:v hevc_nvenc -cq 18 -preset slow output.mp4

colorspace_cuda

CUDA accelerated implementation of the colorspace filter. 🌈

Ngl, it is by no means feature complete compared to the software colorspace filter — at the current time it only supports color range conversion between jpeg/full and mpeg/limited range. Don't come for her, she's doing her best. 💀

The filter accepts the following options:

range — specify output color range.

The accepted values are:

tv — TV (restricted) range
mpeg — MPEG (restricted) range
pc — PC (full) range
jpeg — JPEG (full) range

overlay_cuda

Overlay one video on top of another. 👑

This is the CUDA variant of the overlay filter. It only accepts CUDA frames, and the underlying input pixel formats have to match.

It takes two inputs and has one output. The first input is the "main" video, and the second input gets overlaid on top of it.

It accepts the following parameters:

x, y — set expressions for the x and y coordinates of the overlaid video on the main video.

They can contain the following parameters:

main_w, W — the main input width and height.
main_h, H — the main input width and height.
overlay_w, w — the overlay input width and height.
overlay_h, h — the overlay input width and height.
x, y — the computed values for x and y. They are evaluated for each new frame.
n — the ordinal index of the main input frame, starting from 0.
pos — the byte offset position in the file of the main input frame, NAN if unknown. Deprecated, do not use. 🙈
t — the timestamp of the main input frame, expressed in seconds, NAN if unknown.

Default value is 0 for both expressions.

eval — set when the expressions for x and y are evaluated.

It accepts the following values:

init — evaluate expressions once during filter initialization or when a command is processed.
frame — evaluate expressions for each incoming frame

Default value is frame.

eof_action — see framesync.
shortest — see framesync.
repeatlast — see framesync.

This filter also supports the framesync options. 🫶

pad_cuda

Add paddings to an input video stream using CUDA. 🎀

This filter is the CUDA-accelerated version of the pad filter. It accepts the same options and expressions and provides the same core functionality. For a detailed description of available options, please see the documentation for the pad filter.

Examples

Add a 200-pixel black border to all sides of a video frame:

ffmpeg -hwaccel cuda -hwaccel_output_format cuda -i input.mp4 -vf "pad_cuda=w=iw+400:h=ih+400:x=200:y=200" -c:v h264_nvenc out.mp4

Pad the input video to a 16:9 aspect ratio, filling with the color "blue":

ffmpeg -hwaccel cuda -hwaccel_output_format cuda -i input.mp4 -vf "pad_cuda=w=ih*16/9/sar:h=ih:x=(ow-iw)/2:y=(oh-ih)/2:color=blue" -c:v h264_nvenc out.mp4

scale_cuda

Scale (resize) and convert (pixel format) the input video, using accelerated CUDA kernels. ✨ Setting the output width and height works in the same way as for the scale filter.

The filter accepts the following options:

w, h — set the output video dimension expression. Default value is the input dimension.

Allows for the same expressions as the scale filter.

interp_algo — sets the algorithm used for scaling:
- nearest — nearest neighbour. Used by default if input parameters match the desired output.
- bilinear — bilinear.
- bicubic — bicubic. This is the default.
- lanczos — lanczos.
format — controls the output pixel format. By default, or if none is specified, the input pixel format is used.

The filter does not support converting between YUV and RGB pixel formats. 😭

passthrough — if set to 0, every frame is processed, even if no conversion is necessary. This mode can be useful to use the filter as a buffer for a downstream frame-consumer that exhausts the limited decoder frame pool.

If set to 1, frames are passed through as-is if they match the desired output parameters. This is the default behaviour.

use_filters — if set to 1, filter with a generic weight LUT instead of using fixed-function shader kernels. May be faster or slower depending on the hardware. A value of auto (the default) enables this automatically when required for correct anti-aliasing when downscaling.
param — algorithm-specific parameter.

Affects the curves of the bicubic algorithm.

force_original_aspect_ratio, force_divisible_by — work the same as the identical scale filter options.
reset_sar — works the same as the identical scale filter option.

Examples

Scale it to 720p while keeping the aspect ratio and making sure the output is yuv420p 🎀

scale_cuda=-2:720:format=yuv420p

Upscale to 4K with the nearest neighbour algorithm — crunchy pixels, on purpose, no cap

scale_cuda=4096:2160:interp_algo=nearest

Do literally zero conversion or scaling, just copy every input frame into freshly allocated ones ✨ This one is a lifesaver when your filter-and-encode chain would otherwise drain the decoder's frame pool dry 💀

scale_cuda=passthrough=0

thumbnail_cuda 🖼️

Picks the most representative frame out of a run of consecutive frames, using CUDA. It's giving "choose the main character of this batch" energy 👑

Options it takes:

n — sets the batch size of frames to analyze. Out of a set of n frames it picks one, then moves on to the next batch of n frames, all the way to the end. Default is 100.

Heads up bestie 🙈 — since this filter keeps track of the whole frame sequence, a bigger n means higher memory usage, so a high value is not recommended. Be reasonable, ok?

Example 🌸

Thumbnails get pulled from every n=150-frame batch, one pick per batch. The chosen frames are then scaled with scale_cuda.

./ffmpeg  -hwaccel cuda -hwaccel_output_format cuda  -i ./input.mp4 -vf "thumbnail_cuda=150,scale_cuda=1920:1080,hwdownload,format=nv12" ./output/out%03d.png

yadif_cuda 💫

Deinterlaces your input video with the yadif algorithm, except it's implemented in CUDA so it can slot right into a GPU-accelerated pipeline alongside nvdec and/or nvenc. Slay 🔥

Parameters:

mode — the interlacing mode to adopt. Accepts one of these:
- 0, send_frame — output one frame for each frame.
- 1, send_field — output one frame for each field.
- 2, send_frame_nospatial — like send_frame, but it skips the spatial interlacing check.
- 3, send_field_nospatial — like send_field, but it skips the spatial interlacing check.

Default value is send_frame.

parity — the picture field parity assumed for the input interlaced video. Accepts one of these:
- 0, tff — assume the top field is first.
- 1, bff — assume the bottom field is first.
- -1, auto — enable automatic detection of field parity.

Default value is auto. If the interlacing is unknown, or the decoder doesn't export that info, top field first gets assumed. 😌

deint — which frames to deinterlace. Accepts one of these:
- 0, all — deinterlace all frames.
- 1, interlaced — only deinterlace frames marked as interlaced.

Default value is all.

✨ CUDA NPP ✨

Here's the rundown on the NVIDIA Performance Primitives (libnpp) video filters that are currently available 💜

Prerequisites, bestie, don't skip these:

Install Nvidia CUDA Toolkit
Install libnpp

To enable CUDA NPP filters:

Configure FFmpeg with --enable-nonfree --enable-libnpp.

scale_npp 🌀

Uses the NVIDIA Performance Primitives (libnpp) to do scaling and/or pixel format conversion on CUDA video frames. Setting the output width and height works exactly the same way as it does for the scale filter.

Extra options it accepts:

format — the pixel format of the output CUDA frames. If you set it to the string same (which is the default), the input format is kept. Note that automatic format negotiation and conversion is not yet supported for hardware frames 🫠
interp_algo — the interpolation algorithm used for resizing. One of these:
- nn — nearest neighbour.
- linear
- cubic
- cubic2p_bspline — 2-parameter cubic (B=1, C=0)
- cubic2p_catmullrom — 2-parameter cubic (B=0, C=1/2)
- cubic2p_b05c03 — 2-parameter cubic (B=1/2, C=3/10)
- super — supersampling
- lanczos
force_original_aspect_ratio — lets the output video width or height decrease or increase when necessary, so the original aspect ratio stays intact. Possible values:
- disable — scale the video as specified and turn this feature off.
- decrease — the output video dimensions get automatically decreased if needed.
- increase — the output video dimensions get automatically increased if needed.

A genuinely useful case for this: when you know a specific device's maximum allowed resolution, you can use this to cap the output video to it while retaining the aspect ratio. Example — device A allows 1280x720 playback and your video is 1920x800. Using this option (set to decrease) and specifying 1280x720 on the command line makes the output 1280x533. Iconic 🎀

Please note this is a different thing than specifying -1 for w or h — you still need to specify the output resolution for this option to actually work. Don't come for me, that's just how it is 😭

force_divisible_by — makes sure both output dimensions, width and height, are divisible by the given integer when used together with force_original_aspect_ratio. Works similarly to using -n in the w and h options.

This option respects whatever value you set for force_original_aspect_ratio, increasing or decreasing the resolution accordingly. Warning: the video's aspect ratio may get slightly modified. 🫶

This one's handy if you need a video to fit within or exceed a defined resolution using force_original_aspect_ratio but you also have encoder restrictions on width or height divisibility.

reset_sar — works the same as the identical scale filter option.
eval — specify when to evaluate the width and height expression. Accepts:
- init — only evaluate expressions once during filter initialization, or when a command is processed.
- frame — evaluate expressions for every incoming frame.

The values of the w and h options are expressions containing these constants:

in_w, in_h — the input width and height
iw, ih — same as in_w and in_h.
out_w, out_h — the output (scaled) width and height
ow, oh — same as out_w and out_h
a — same as iw / ih
sar — input sample aspect ratio
dar — the input display aspect ratio. Calculated from (iw / ih) * sar.
n — the (sequential) number of the input frame, starting from 0. Only available with eval=frame.
t — the presentation timestamp of the input frame, expressed as a number of seconds. Only available with eval=frame.
pos — the position (byte offset) of the frame in the input stream, or NaN if that info is unavailable and/or meaningless (for example, in the case of synthetic video). Only available with eval=frame. Deprecated, do not use. 💀

scale2ref_npp 🔁

Uses the NVIDIA Performance Primitives (libnpp) to scale (resize) the input video based on a reference video.

Check out the scale_npp filter for the available options — scale2ref_npp supports the same ones, but uses the reference video instead of the main input as its basis. It also supports these extra constants for the w and h options:

main_w, main_h — the main input video's width and height
main_a — same as main_w / main_h
main_sar — the main input video's sample aspect ratio
main_dar, mdar — the main input video's display aspect ratio. Calculated from (main_w / main_h) * main_sar.
main_n — the (sequential) number of the main input frame, starting from 0. Only available with eval=frame.
main_t — the presentation timestamp of the main input frame, expressed as a number of seconds. Only available with eval=frame.
main_pos — the position (byte offset) of the frame in the main input stream, or NaN if that info is unavailable and/or meaningless (for example, in the case of synthetic video). Only available with eval=frame.

Examples 🌈

Scale a subtitle stream (b) so it matches the main video (a) in size before overlaying:

'scale2ref_npp[b][a];[a][b]overlay_cuda'

Scale a logo to 1/10th the height of a video while preserving its display aspect ratio:

[logo-in][video-in]scale2ref_npp=w=oh*mdar:h=ih/10[logo-out][video-out]

sharpen_npp 🔪✨

Uses the NVIDIA Performance Primitives (libnpp) to do image sharpening with border control.

Extra options it accepts:

border_type — the type of sampling used at frame borders. One of these:
- replicate — replicate pixel values.

transpose_npp 🦋

Transposes rows with columns in the input video and optionally flips it. For more in-depth examples, peep the transpose video filter — it shares mostly the same options.

Parameters:

dir — specify the transposition direction.

Can be one of these:

cclock_flip — rotate 90 degrees counterclockwise and vertically flip. (default)
clock — rotate 90 degrees clockwise.
cclock — rotate 90 degrees counterclockwise.
clock_flip — rotate 90 degrees clockwise and vertically flip.

passthrough — don't apply the transposition if the input geometry already matches the one specified by this value. Accepts:
- none — always apply transposition. (default)
- portrait — preserve portrait geometry (when height >= width).
- landscape — preserve landscape geometry (when width >= height).

🌊 OPENCL VIDEO FILTERS 🌊

And here's the breakdown of the OpenCL video filters that are currently available 💙

To enable compilation of these filters you need to configure FFmpeg with --enable-opencl.

Running OpenCL filters means you gotta initialize a hardware device and pass that device to every filter in any filter graph. Periodt.

-init_hw_device opencl[=name][:device[,key=value...]] — initialise a new hardware device of type opencl called name, using the given device parameters.
-filter_hw_device name — pass the hardware device called name to all filters in any filter graph.

For more detailed info, see <https://www.ffmpeg.org/ffmpeg.html#Advanced-Video-options>

Example of choosing the first device on the second platform and running the avgblur_opencl filter with default parameters on it 🎀

-init_hw_device opencl=gpu:1.0 -filter_hw_device gpu -i INPUT -vf "hwupload, avgblur_opencl, hwdownload" OUTPUT

okay so real talk, bestie: OpenCL filters can't just reach into normal memory and grab frame data. All frame data needs to be uploaded first — that's hwupload — to hardware surfaces connected to the appropriate device before being used, and then downloaded back to normal memory with hwdownload. 💅 Also note that hwupload will upload to a surface with the same layout as the software frame, so it may be necessary to add a format filter immediately before to get the input into the right format. And hwdownload does not support all formats on the output — it may be necessary to insert an additional format filter immediately following in the graph to get the output in a supported format. No cap.

avgblur_opencl 💖

Apply average blur filter. It's giving soft-focus main character energy.

The filter accepts the following options:

sizeX — Set horizontal radius size. Range is [1, 1024] and default value is 1.
planes — Set which planes to filter. Default value is 0xf, by which all planes are processed.
sizeY — Set vertical radius size. Range is [1, 1024] and default value is 0. If zero, the sizeX value will be used.

Example ✨

Apply average blur filter with horizontal and vertical size of 3, setting each pixel of the output to the average value of the 7x7 region centered on it in the input. For pixels on the edges of the image, the region does not extend beyond the image boundaries, and so out-of-range coordinates are not used in the calculations.

-i INPUT -vf "hwupload, avgblur_opencl=3, hwdownload" OUTPUT

boxblur_opencl 🎀

Apply a boxblur algorithm to the input video.

It accepts the following parameters:

luma_radius, lr
luma_power, lp
chroma_radius, cr
chroma_power, cp
alpha_radius, ar
alpha_power, ap

A description of the accepted options follows.

luma_radius, lr, chroma_radius, cr, alpha_radius, ar — Set an expression for the box radius in pixels used for blurring the corresponding input plane.

The radius value must be a non-negative number, and must not be greater than the value of the expression min(w,h)/2 for the luma and alpha planes, and of min(cw,ch)/2 for the chroma planes.

Default value for luma_radius is 2. If not specified, chroma_radius and alpha_radius default to the corresponding value set for luma_radius.

The expressions can contain the following constants:

w, h — The input width and height in pixels.
cw, ch — The input chroma image width and height in pixels.
hsub, vsub — The horizontal and vertical chroma subsample values. For example, for the pixel format yuv422p, hsub is 2 and vsub is 1.

luma_power, lp, chroma_power, cp, alpha_power, ap — Specify how many times the boxblur filter is applied to the corresponding plane.

Default value for luma_power is 2. If not specified, chroma_power and alpha_power default to the corresponding value set for luma_power.

A value of 0 will disable the effect. 💀

Examples 🦋

Apply boxblur filter, setting each pixel of the output to the average value of box-radiuses luma_radius, chroma_radius, alpha_radius for each plane respectively. The filter will apply luma_power, chroma_power, alpha_power times onto the corresponding plane. For pixels on the edges of the image, the radius does not extend beyond the image boundaries, and so out-of-range coordinates are not used in the calculations.

Apply a boxblur filter with the luma, chroma, and alpha radius set to 2 and luma, chroma, and alpha power set to 3. The filter will run 3 times with box-radius set to 2 for every plane of the image.

-i INPUT -vf "hwupload, boxblur_opencl=luma_radius=2:luma_power=3, hwdownload" OUTPUT
-i INPUT -vf "hwupload, boxblur_opencl=2:3, hwdownload" OUTPUT

Apply a boxblur filter with luma radius set to 2, luma_power to 1, chroma_radius to 4, chroma_power to 5, alpha_radius to 3 and alpha_power to 7.

For the luma plane, a 2x2 box radius will be run once.

For the chroma plane, a 4x4 box radius will be run 5 times.

For the alpha plane, a 3x3 box radius will be run 7 times.

-i INPUT -vf "hwupload, boxblur_opencl=2:1:4:5:3:7, hwdownload" OUTPUT

colorkey_opencl 🍓

RGB colorspace color keying.

The filter accepts the following options:

color — The color which will be replaced with transparency.
similarity — Similarity percentage with the key color.

0.01 matches only the exact key color, while 1.0 matches everything.

blend — Blend percentage.

0.0 makes pixels either fully transparent, or not transparent at all.

Higher values result in semi-transparent pixels, with a higher transparency the more similar the pixels color is to the key color.

Examples 🌸

Make every semi-green pixel in the input transparent with some slight blending:

-i INPUT -vf "hwupload, colorkey_opencl=green:0.3:0.1, hwdownload" OUTPUT

convolution_opencl 🔥

Apply convolution of 3x3, 5x5, 7x7 matrix.

The filter accepts the following options:

0m, 1m, 2m, 3m — Set matrix for each plane. Matrix is sequence of 9, 25 or 49 signed numbers. Default value for each plane is 0 0 0 0 1 0 0 0 0.
0rdiv, 1rdiv, 2rdiv, 3rdiv — Set multiplier for calculated value for each plane. If unset or 0, it will be sum of all matrix elements. The option value must be a float number greater or equal to 0.0. Default value is 1.0.
0bias, 1bias, 2bias, 3bias — Set bias for each plane. This value is added to the result of the multiplication. Useful for making the overall image brighter or darker. The option value must be a float number greater or equal to 0.0. Default value is 0.0.

Examples 👑

Apply sharpen:

-i INPUT -vf "hwupload, convolution_opencl=0 -1 0 -1 5 -1 0 -1 0:0 -1 0 -1 5 -1 0 -1 0:0 -1 0 -1 5 -1 0 -1 0:0 -1 0 -1 5 -1 0 -1 0, hwdownload" OUTPUT

Apply blur:

-i INPUT -vf "hwupload, convolution_opencl=1 1 1 1 1 1 1 1 1:1 1 1 1 1 1 1 1 1:1 1 1 1 1 1 1 1 1:1 1 1 1 1 1 1 1 1:1/9:1/9:1/9:1/9, hwdownload" OUTPUT

Apply edge enhance:

-i INPUT -vf "hwupload, convolution_opencl=0 0 0 -1 1 0 0 0 0:0 0 0 -1 1 0 0 0 0:0 0 0 -1 1 0 0 0 0:0 0 0 -1 1 0 0 0 0:5:1:1:1:0:128:128:128, hwdownload" OUTPUT

Apply edge detect:

-i INPUT -vf "hwupload, convolution_opencl=0 1 0 1 -4 1 0 1 0:0 1 0 1 -4 1 0 1 0:0 1 0 1 -4 1 0 1 0:0 1 0 1 -4 1 0 1 0:5:5:5:1:0:128:128:128, hwdownload" OUTPUT

Apply laplacian edge detector which includes diagonals:

-i INPUT -vf "hwupload, convolution_opencl=1 1 1 1 -8 1 1 1 1:1 1 1 1 -8 1 1 1 1:1 1 1 1 -8 1 1 1 1:1 1 1 1 -8 1 1 1 1:5:5:5:1:0:128:128:0, hwdownload" OUTPUT

Apply emboss:

-i INPUT -vf "hwupload, convolution_opencl=-2 -1 0 -1 1 1 0 1 2:-2 -1 0 -1 1 1 0 1 2:-2 -1 0 -1 1 1 0 1 2:-2 -1 0 -1 1 1 0 1 2, hwdownload" OUTPUT

erosion_opencl 🌈

Apply erosion effect to the video.

This filter replaces the pixel by the local(3x3) minimum.

It accepts the following options:

threshold0, threshold1, threshold2, threshold3 — Limit the maximum change for each plane. Range is [0, 65535] and default value is 65535. If 0, plane will remain unchanged.
coordinates — Flag which specifies the pixel to refer to. Range is [0, 255] and default value is 255, i.e. all eight pixels are used.

Flags to local 3x3 coordinates region centered on x:

Example 🫶

Apply erosion filter with threshold0 set to 30, threshold1 set 40, threshold2 set to 50 and coordinates set to 231, setting each pixel of the output to the local minimum between pixels: 1, 2, 3, 6, 7, 8 of the 3x3 region centered on it in the input. If the difference between input pixel and local minimum is more then threshold of the corresponding plane, output pixel will be set to input pixel - threshold of corresponding plane.

-i INPUT -vf "hwupload, erosion_opencl=30:40:50:coordinates=231, hwdownload" OUTPUT

deshake_opencl 🦋

Feature-point based video stabilization filter.

The filter accepts the following options:

tripod — Simulates a tripod by preventing any camera movement whatsoever from the original frame. Defaults to 0.
debug — Whether or not additional debug info should be displayed, both in the processed output and in the console.

Note that in order to see console debug output you will also need to pass -v verbose to ffmpeg.

Viewing point matches in the output video is only supported for RGB input.

Defaults to 0.

adaptive_crop — Whether or not to do a tiny bit of cropping at the borders to cut down on the amount of mirrored pixels.

Defaults to 1.

refine_features — Whether or not feature points should be refined at a sub-pixel level.

This can be turned off for a slight performance gain at the cost of precision.

Defaults to 1.

smooth_strength — The strength of the smoothing applied to the camera path from 0.0 to 1.0.

1.0 is the maximum smoothing strength while values less than that result in less smoothing.

0.0 causes the filter to adaptively choose a smoothing strength on a per-frame basis.

Defaults to 0.0.

smooth_window_multiplier — Controls the size of the smoothing window (the number of frames buffered to determine motion information from).

The size of the smoothing window is determined by multiplying the framerate of the video by this number.

Acceptable values range from 0.1 to 10.0.

Larger values increase the amount of motion data available for determining how to smooth the camera path, potentially improving smoothness, but also increase latency and memory usage.

Defaults to 2.0.

Examples ⭐

Stabilize a video with a fixed, medium smoothing strength:

-i INPUT -vf "hwupload, deshake_opencl=smooth_strength=0.5, hwdownload" OUTPUT

Stabilize a video with debugging (both in console and in rendered video):

-i INPUT -filter_complex "[0:v]format=rgba, hwupload, deshake_opencl=debug=1, hwdownload, format=rgba, format=yuv420p" -v verbose OUTPUT

dilation_opencl 👑

Apply dilation effect to the video.

This filter replaces the pixel by the local(3x3) maximum.

It accepts the following options:

okay so the last dilation_opencl dials are these:

threshold0 — caps the max change for plane 0. Range is [0, 65535], default is 65535. If it's 0, that plane stays unchanged, periodt.
threshold1 — same thing but for plane 1. Range is [0, 65535], default is 65535. If it's 0, plane 1 remains untouched.
threshold2 — plane 2's max-change cap. Range is [0, 65535], default is 65535. Set it to 0 and plane 2 will remain unchanged.
threshold3 — plane 3's max-change cap. Range is [0, 65535], default is 65535. 0 means plane 3 remains unchanged.
coordinates — tells it which pixel to refer to. Range is [0, 255], default is 255, which means all eight pixels are used. Flags to local 3x3 coordinates region centered on x:

Example

Apply dilation filter with threshold0 set to 30, threshold1 set 40, threshold2 set to 50 and coordinates set to 231, setting each pixel of the output to the local maximum between pixels: 1, 2, 3, 6, 7, 8 of the 3x3 region centered on it in the input. If the difference between input pixel and local maximum is more than threshold of the corresponding plane, output pixel will be set to input pixel + threshold of corresponding plane.

-i INPUT -vf "hwupload, dilation_opencl=30:40:50:coordinates=231, hwdownload" OUTPUT

`nlmeans_opencl`

Non-local Means denoise filter through OpenCL, this filter accepts same options as nlmeans. easy win ✨

`overlay_opencl`

Overlay one video on top of another.

It takes two inputs and has one output. The first input is the "main" video on which the second input is overlaid. This filter requires same memory layout for all the inputs. So, format conversion may be needed.

The filter accepts the following options:

x — Set the x coordinate of the overlaid video on the main video. Default value is 0.
y — Set the y coordinate of the overlaid video on the main video. Default value is 0.

Examples

Overlay an image LOGO at the top-left corner of the INPUT video. Both inputs are yuv420p format.

-i INPUT -i LOGO -filter_complex "[0:v]hwupload[a], [1:v]format=yuv420p, hwupload[b], [a][b]overlay_opencl, hwdownload" OUTPUT

The inputs have same memory layout for color channels, the overlay has additional alpha plane, like INPUT is yuv420p, and the LOGO is yuva420p.

-i INPUT -i LOGO -filter_complex "[0:v]hwupload[a], [1:v]format=yuva420p, hwupload[b], [a][b]overlay_opencl, hwdownload" OUTPUT

`pad_opencl`

Add paddings to the input image, and place the original input at the provided x, y coordinates.

It accepts the following options:

width, w, height, h — Specify an expression for the size of the output image with the paddings added. If the value for width or height is 0, the corresponding input size is used for the output. The width expression can reference the value set by the height expression, and vice versa. The default value of width and height is 0.
x, y — Specify the offsets to place the input image at within the padded area, with respect to the top/left border of the output image. The x expression can reference the value set by the y expression, and vice versa. The default value of x and y is 0. If x or y evaluate to a negative number, they'll be changed so the input image is centered on the padded area.
color — Specify the color of the padded area. For the syntax of this option, check the Color section in the ffmpeg-utils manual.
aspect — Pad to an aspect instead to a resolution.

The value for the width, height, x, and y options are expressions containing the following constants:

in_w, in_h — The input video width and height.
iw, ih — These are the same as in_w and in_h.
out_w, out_h — The output width and height (the size of the padded area), as specified by the width and height expressions.
ow, oh — These are the same as out_w and out_h.
x, y — The x and y offsets as specified by the x and y expressions, or NAN if not yet specified.
a — same as iw / ih
sar — input sample aspect ratio
dar — input display aspect ratio, it is the same as (iw / ih) * sar

`prewitt_opencl`

Apply the Prewitt operator (https://en.wikipedia.org/wiki/Prewitt_operator) to input video stream.

The filter accepts the following option:

planes — Set which planes to filter. Default value is 0xf, by which all planes are processed.
scale — Set value which will be multiplied with filtered result. Range is [0.0, 65535] and default value is 1.0.
delta — Set value which will be added to filtered result. Range is [-65535, 65535] and default value is 0.0.

Example

Apply the Prewitt operator with scale set to 2 and delta set to 10.

-i INPUT -vf "hwupload, prewitt_opencl=scale=2:delta=10, hwdownload" OUTPUT

`program_opencl`

Filter video using an OpenCL program.

source — OpenCL program source file.
kernel — Kernel name in program.
inputs — Number of inputs to the filter. Defaults to 1.
size, s — Size of output frames. Defaults to the same as the first input.

The program_opencl filter also supports the framesync options.

The program source file must contain a kernel function with the given name, which will be run once for each plane of the output. Each run on a plane gets enqueued as a separate 2D global NDRange with one work-item for each pixel to be generated. The global ID offset for each work-item is therefore the coordinates of a pixel in the destination image.

The kernel function needs to take the following arguments:

Destination image, __write_only image2d_t.

This image will become the output; the kernel should write all of it.

Frame index, unsigned int.

This is a counter starting from zero and increasing by one for each frame.

Source images, __read_only image2d_t.

These are the most recent images on each input. The kernel may read from them to generate the output, but they can't be written to.

Example programs:

Copy the input to the output (output must be the same size as the input).

__kernel void copy(__write_only image2d_t destination,
                   unsigned int index,
                   __read_only  image2d_t source)
{
    const sampler_t sampler = CLK_NORMALIZED_COORDS_FALSE;

    int2 location = (int2)(get_global_id(0), get_global_id(1));

    float4 value = read_imagef(source, sampler, location);

    write_imagef(destination, location, value);
}

Apply a simple transformation, rotating the input by an amount increasing with the index counter. Pixel values are linearly interpolated by the sampler, and the output need not have the same dimensions as the input.

__kernel void rotate_image(__write_only image2d_t dst,
                           unsigned int index,
                           __read_only  image2d_t src)
{
    const sampler_t sampler = (CLK_NORMALIZED_COORDS_FALSE |
                               CLK_FILTER_LINEAR);

    float angle = (float)index / 100.0f;

    float2 dst_dim = convert_float2(get_image_dim(dst));
    float2 src_dim = convert_float2(get_image_dim(src));

    float2 dst_cen = dst_dim / 2.0f;
    float2 src_cen = src_dim / 2.0f;

    int2   dst_loc = (int2)(get_global_id(0), get_global_id(1));

    float2 dst_pos = convert_float2(dst_loc) - dst_cen;
    float2 src_pos = {
        cos(angle) * dst_pos.x - sin(angle) * dst_pos.y,
        sin(angle) * dst_pos.x + cos(angle) * dst_pos.y
    };
    src_pos = src_pos * src_dim / dst_dim;

    float2 src_loc = src_pos + src_cen;

    if (src_loc.x < 0.0f      || src_loc.y < 0.0f ||
        src_loc.x > src_dim.x || src_loc.y > src_dim.y)
        write_imagef(dst, dst_loc, 0.5f);
    else
        write_imagef(dst, dst_loc, read_imagef(src, sampler, src_loc));
}

Blend two inputs together, with the amount of each input used varying with the index counter.

__kernel void blend_images(__write_only image2d_t dst,
                           unsigned int index,
                           __read_only  image2d_t src1,
                           __read_only  image2d_t src2)
{
    const sampler_t sampler = (CLK_NORMALIZED_COORDS_FALSE |
                               CLK_FILTER_LINEAR);

    float blend = (cos((float)index / 50.0f) + 1.0f) / 2.0f;

    int2  dst_loc = (int2)(get_global_id(0), get_global_id(1));
    int2 src1_loc = dst_loc * get_image_dim(src1) / get_image_dim(dst);
    int2 src2_loc = dst_loc * get_image_dim(src2) / get_image_dim(dst);

    float4 val1 = read_imagef(src1, sampler, src1_loc);
    float4 val2 = read_imagef(src2, sampler, src2_loc);

    write_imagef(dst, dst_loc, val1 * blend + val2 * (1.0f - blend));
}

`remap_opencl`

Remap pixels using 2nd: Xmap and 3rd: Ymap input video stream.

Destination pixel at position (X, Y) will be picked from source (x, y) position where x = Xmap(X, Y) and y = Ymap(X, Y). If mapping values are out of range, zero value for pixel will be used for destination pixel.

Xmap and Ymap input video streams must be of same dimensions. Output video stream will have Xmap/Ymap video stream dimensions. Xmap and Ymap input video streams are 32bit float pixel format, single channel.

interp — Specify interpolation used for remapping of pixels. Allowed values are near and linear. Default value is linear.
fill — Specify the color of the unmapped pixels. For the syntax of this option, check the Color section in the ffmpeg-utils manual. Default color is black.

`roberts_opencl`

Apply the Roberts cross operator (https://en.wikipedia.org/wiki/Roberts_cross) to input video stream.

The filter accepts the following option:

planes — Set which planes to filter. Default value is 0xf, by which all planes are processed.
scale — Set value which will be multiplied with filtered result. Range is [0.0, 65535] and default value is 1.0.
delta — Set value which will be added to filtered result. Range is [-65535, 65535] and default value is 0.0.

Example

Apply the Roberts cross operator with scale set to 2 and delta set to 10

-i INPUT -vf "hwupload, roberts_opencl=scale=2:delta=10, hwdownload" OUTPUT

🌀 sobel_opencl

This one runs the Sobel operator over your input video stream — edge-detection classic, iykyk. (The Wikipedia link, if you wanna read up: https://en.wikipedia.org/wiki/Sobel_operator.)

It takes these options:

planes — picks which planes get filtered. Default is 0xf, which means every plane gets the treatment 💅
scale — the value that gets multiplied with the filtered result. Range is [0.0, 65535], default is 1.0.
delta — the value added onto the filtered result. Range is [-65535, 65535], default is 0.0.

✨ Example

So here's the vibe: sobel operator with scale set to 2 and delta set to 10, slay:

-i INPUT -vf "hwupload, sobel_opencl=scale=2:delta=10, hwdownload" OUTPUT

🌈 tonemap_opencl

HDR (PQ/HLG) → SDR conversion with tone-mapping. Basically your overly-bright video gets a little reality check ✨

It accepts these params:

tonemap — picks the tone-mapping operator. Same as the tonemap option in tonemap.
param — tunes the tone-mapping algorithm. Same as the param option in tonemap.
desat — desaturates highlights that go past this brightness level. The higher you set it, the more color info gets preserved. It's the thing that stops super-highlights from looking like a blown-out crime scene, by (smoothly) fading them into white instead. Makes images feel way more natural, at the cost of losing info about out-of-range colors.

Default is 0.5, and ngl the algorithm here is a lil different from the CPU version of tonemap at the moment. Setting it to 0.0 disables the option entirely.

threshold — the tonemapping algorithm gets fine-tuned per scene, and this threshold is how it detects whether the scene changed. If the distance between the current frame's average brightness and the current running average exceeds the threshold, it re-calculates the scene average and peak brightness. Default is 0.2.
format — sets the output pixel format.

Formats supported rn are:

p010
nv12

range, r — sets the output color range.

Possible values:

tv/mpeg
pc/jpeg

Default is same as input.

primaries, p — sets the output color primaries.

Possible values:

bt709
bt2020

Default is same as input.

transfer, t — sets the output transfer characteristics.

Possible values:

bt709
bt2020

Default is bt709.

matrix, m — sets the output colorspace matrix.

Possible values:

bt709
bt2020

Default is same as input.

✨ Example

Convert an HDR (PQ/HLG) video to bt2020-transfer-characteristic p010 format using the linear operator. Obsessed with this one:

-i INPUT -vf "format=p010,hwupload,tonemap_opencl=t=bt2020:tonemap=linear:format=p010,hwdownload,format=p010" OUTPUT

🦋 unsharp_opencl

Sharpen or blur your input video. Pick your fighter 💅

It accepts these parameters:

luma_msize_x, lx — luma matrix horizontal size. Range [1, 23], default 5.
luma_msize_y, ly — luma matrix vertical size. Range [1, 23], default 5.
luma_amount, la — luma effect strength. Range [-10, 10], default 1.0.

Negative values blur the input, positive values sharpen it, and zero disables the effect completely.

chroma_msize_x, cx — chroma matrix horizontal size. Range [1, 23], default 5.
chroma_msize_y, cy — chroma matrix vertical size. Range [1, 23], default 5.
chroma_amount, ca — chroma effect strength. Range [-10, 10], default 0.0.

Same deal: negative blurs, positive sharpens, zero = effect off.

Every single parameter is optional and defaults to the equivalent of the string '5:5:1.0:5:5:0.0'. Cute, right?

✨ Examples

Strong luma sharpen, go off:

-i INPUT -vf "hwupload, unsharp_opencl=luma_msize_x=7:luma_msize_y=7:luma_amount=2.5, hwdownload" OUTPUT

Strong blur on both luma and chroma:

-i INPUT -vf "hwupload, unsharp_opencl=7:7:-2:7:7:-2, hwdownload" OUTPUT

🧚 xfade_opencl

Cross-fades two videos with a custom transition effect, using OpenCL. Main character energy for your cuts 🔥

It accepts these options:

transition — sets one of the possible transition effects:
- custom — selects the custom transition effect; the actual transition description gets picked from the source and kernel options.
- fade
- wipeleft
- wiperight
- wipeup
- wipedown
- slideleft
- slideright
- slideup
- slidedown

Default transition is fade.

source — the OpenCL program source file for a custom transition.
kernel — sets the name of the kernel to use for the custom transition from the program source file.
duration — sets the duration of the video transition.
offset — sets the time the transition starts, relative to the first video.

Heads up bestie: the program source file must contain a kernel function with the given name, and that function gets run once for each plane of the output. Each run on a plane is enqueued as a separate 2D global NDRange with one work-item for each pixel to be generated. So the global ID offset for each work-item is literally the coordinates of a pixel in the destination image. No cap.

The kernel function needs to take these arguments:

Destination image, __write_only image2d_t.

This image becomes the output — the kernel should write all of it.

First Source image, __read_only image2d_t. Second Source image, __read_only image2d_t.

These are the most recent images on each input. The kernel can read from them to generate the output, but writing to them? Not happening, sorry.

Transition progress, float. This value is always between 0 and 1 inclusive.

✨ Example programs

Apply a dots curtain transition effect — this is the iconic one 🎀

__kernel void blend_images(__write_only image2d_t dst,
                           __read_only  image2d_t src1,
                           __read_only  image2d_t src2,
                           float progress)
{
    const sampler_t sampler = (CLK_NORMALIZED_COORDS_FALSE |
                               CLK_FILTER_LINEAR);
    int2  p = (int2)(get_global_id(0), get_global_id(1));
    float2 rp = (float2)(get_global_id(0), get_global_id(1));
    float2 dim = (float2)(get_image_dim(src1).x, get_image_dim(src1).y);
    rp = rp / dim;

    float2 dots = (float2)(20.0, 20.0);
    float2 center = (float2)(0,0);
    float2 unused;

    float4 val1 = read_imagef(src1, sampler, p);
    float4 val2 = read_imagef(src2, sampler, p);
    bool next = distance(fract(rp * dots, &unused), (float2)(0.5, 0.5)) < (progress / distance(rp, center));

    write_imagef(dst, p, next ? val1 : val2);
}

✨ VAAPI VIDEO FILTERS ✨

Okay bestie, so VAAPI video filters are these iconic little guys that usually team up with a VAAPI decoder and a VAAPI encoder. Like, they're a whole squad that shows up together. 💅 Below is the full rundown of every VAAPI video filter, no gatekeeping.

To get these filters to even compile, you gotta configure FFmpeg with --enable-vaapi. That's the vibe check for the build step, fr.

And to actually use vaapi filters, you need to set up the vaapi device correctly first. No device, no slay. For more info, please read https://trac.ffmpeg.org/wiki/Hardware/VAAPI — go read it, iykyk. 🦋

🎀 overlay_vaapi

Overlay one video right on top of another. Main character on top of main character, literally.

It takes two inputs and has exactly one output. The first input is the "main" video — the one getting overlaid — and the second input is the one doing the overlaying. 💖

The filter accepts the following options:

x, y — Set expressions for the x and y coordinates of the overlaid video on the main video. Default value is 0 for both expressions.
w, h — Set expressions for the width and height of the overlaid video on the main video. Default values are overlay_iw for w and overlay_ih*w/overlay_iw for h.

The expressions can contain the following parameters:

main_w, W / main_h, H — The main input width and height.
overlay_iw / overlay_ih — The overlay input width and height.
overlay_w, w / overlay_h, h — The overlay output width and height.
overlay_x, x / overlay_y, y — Position of the overlay layer inside of main.

More options, because of course there are:

alpha — Set transparency of overlaid video. Allowed range is 0.0 to 1.0. Higher value means lower transparency. Default value is 1.0. (Yes, that's backwards from what your brain wants it to be — higher alpha = less see-through. Note it and move on. 💀)
eof_action — See framesync.
shortest — See framesync.
repeatlast — See framesync.

This filter also supports the framesync options. 🫶

🌸 Examples

Overlay an image LOGO at the top-left corner of the INPUT video. Both inputs for this filter are yuv420p format.

-i INPUT -i LOGO -filter_complex "[0:v]hwupload[a], [1:v]format=yuv420p, hwupload[b], [a][b]overlay_vaapi" OUTPUT

Overlay an image LOGO at the offset (200, 100) from the top-left corner of the INPUT video. The inputs have same memory layout for color channels, the overlay has additional alpha plane, like INPUT is yuv420p, and the LOGO is yuva420p.

-i INPUT -i LOGO -filter_complex "[0:v]hwupload[a], [1:v]format=yuva420p, hwupload[b], [a][b]overlay_vaapi=x=200:y=100:w=400:h=300:alpha=1.0, hwdownload, format=nv12" OUTPUT

🦋 tonemap_vaapi

Perform HDR-to-SDR or HDR-to-HDR tone-mapping. It currently only accepts HDR10 as input — so if you roll up with something else, it's not gonna work, bestie. 😭

It accepts the following parameters:

format — Specify the output pixel format. Default is nv12 for HDR-to-SDR tone-mapping and p010 for HDR-to-HDR tone-mapping.
primaries, p — Set the output color primaries. Default is bt709 for HDR-to-SDR tone-mapping and same as input for HDR-to-HDR tone-mapping.
transfer, t — Set the output transfer characteristics. Default is bt709 for HDR-to-SDR tone-mapping and same as input for HDR-to-HDR tone-mapping.
matrix, m — Set the output colorspace matrix. Default is bt709 for HDR-to-SDR tone-mapping and same as input for HDR-to-HDR tone-mapping.
display — Set the output mastering display colour volume. It is given by a |-separated list of two values, two values are space separated. It set display primaries x & y in G, B, R order, then white point x & y, the nominal minimum & maximum display luminances. HDR-to-HDR tone-mapping will be performed when this option is set. ✨
light — Set the output content light level information. It accepts 2 space-separated values, the first input is the maximum light level and the second input is the maximum average light level. It is ignored for HDR-to-SDR tone-mapping, and optional for HDR-to-HDR tone-mapping.

💖 Example

Convert HDR(HDR10) video to bt2020-transfer-characteristic p010 format

tonemap_vaapi=format=p010:t=bt2020-10

Convert HDR video to HDR video

tonemap_vaapi=display=7500\ 3000|34000\ 16000|13250\ 34500|15635\ 16450|500\ 10000000

🧚 hstack_vaapi

Stack input videos horizontally. Side by side, bestie. 👑

This is the VA-API variant of the hstack filter, and here's the cute part — each input stream may have different height, and this filter will scale down/up each input stream while keeping the original aspect. Flexible and polite. 🎀

It accepts the following options:

inputs — See hstack.
shortest — See hstack.
height — Set height of output. If set to 0, this filter will set height of output to height of the first input stream. Default value is 0.

🌈 vstack_vaapi

Stack input videos vertically. Same energy as above but stacked up-and-down instead. 🫶

This is the VA-API variant of the vstack filter, each input stream may have different width, this filter will scale down/up each input stream while keeping the original aspect.

It accepts the following options:

inputs — See vstack.
shortest — See vstack.
width — Set width of output. If set to 0, this filter will set width of output to width of the first input stream. Default value is 0.

⭐ xstack_vaapi

Stack video inputs into a custom layout. This one is the artsy one — you decide where everything goes. 💅

This is the VA-API variant of the xstack filter, each input stream may have different size, this filter will scale down/up each input stream to the given output size, or the size of the first input stream.

It accepts the following options:

inputs — See xstack.
shortest — See xstack.
layout — See xstack. Moreover, this permits the user to supply output size for each input stream.

xstack_vaapi=inputs=4:layout=0_0_1920x1080|0_h0_1920x1080|w0_0_1920x1080|w0_h0_1920x1080

grid — See xstack.
grid_tile_size — Set output size for each input stream when grid is set. If this option is not set, this filter will set output size by default to the size of the first input stream. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual.
fill — See xstack.

🍓 pad_vaapi

Add paddings to the input image, and place the original input at the provided x, y coordinates. It's giving scrapbook page with a cute border. 💖

It accepts the following options:

width, w / height, h — Specify an expression for the size of the output image with the paddings added. If the value for width or height is 0, the corresponding input size is used for the output.

The width expression can reference the value set by the height expression, and vice versa.

The default value of width and height is 0.

x, y — Specify the offsets to place the input image at within the padded area, with respect to the top/left border of the output image.

The x expression can reference the value set by the y expression, and vice versa.

The default value of x and y is 0.

If x or y evaluate to a negative number, they'll be changed so the input image is centered on the padded area. 🎯

color — Specify the color of the padded area. For the syntax of this option, check the "Color" section in the ffmpeg-utils manual.
aspect — Pad to an aspect instead to a resolution.

The value for the width, height, x, and y options are expressions containing the following constants:

in_w / in_h — The input video width and height.
iw / ih — These are the same as in_w and in_h.
out_w / out_h — The output width and height (the size of the padded area), as specified by the width and height expressions.
ow / oh — These are the same as out_w and out_h.
x / y — The x and y offsets as specified by the x and y expressions, or NAN if not yet specified.
a — same as iw / ih
sar — input sample aspect ratio
dar — input display aspect ratio, it is the same as (iw / ih) * sar

💜 drawbox_vaapi

Draw a colored box on the input image. Arts and crafts era, fr fr. 🎨

It accepts the following parameters:

x, y — The expressions which specify the top left corner coordinates of the box. It defaults to 0.
width, w / height, h — The expressions which specify the width and height of the box; if 0 they are interpreted as the input width and height. It defaults to 0.
color, c — Specify the color of the box to write. For the general syntax of this option, check the "Color" section in the ffmpeg-utils manual.
thickness, t — The expression which sets the thickness of the box edge. A value of fill will create a filled box. Default value is 3. See below for the list of accepted constants.
replace — With value 1, the pixels of the painted box will overwrite the video's color and alpha pixels. Default is 0, which composites the box onto the input video.

The parameters for x, y, w and h and t are expressions containing the following constants:

in_h, ih / in_w, iw — The input width and height.
x / y — The x and y offset coordinates where the box is drawn.
w / h — The width and height of the drawn box.
t — The thickness of the drawn box.

🌟 Examples

Draw a black box around the edge of the input image:

drawbox

Draw a box with color red and an opacity of 50%:

drawbox=10:20:200:60:red@0.5

The previous example can be specified as:

drawbox=x=10:y=20:w=200:h=60:color=red@0.5

Fill the box with pink color:

drawbox=x=10:y=10:w=100:h=100:color=pink@0.5:t=fill

Draw a 2-pixel red 2.40:1 mask:

drawbox=x=-t:y=0.5*(ih-iw/2.4)-t:w=iw+t*2:h=iw/2.4+t*2:t=2:c=red

VIDEOTOOLBOX VIDEO FILTERS 🍎✨

Okay bestie, here's the tea on the VideoToolbox video filters that are currently available. 👑

VideoToolbox filter depends on VideoToolbox framework, and it's auto detected when you build ffmpeg for an Apple platform such as macOS. If autodetect is off for some reason, just add --enable-videotoolbox to configure and you're so back. 🫶

scale_vt 🎀

It scales and converts the color parameters using VTPixelTransferSession. Two jobs, one filter, main character energy fr.

The filter accepts the following options:

w
h — Set the output video dimension expression. Default value is the input dimension. (so if you don't say anything, it just matches what came in 💅)
color_matrix — Set the output colorspace matrix.
color_primaries — Set the output color primaries.
color_transfer — Set the output transfer characteristics.

Examples 🌸

Perform HDR to SDR conversion, and scale to half size of input 🔥

ffmpeg -hwaccel videotoolbox \
        -hwaccel_output_format videotoolbox_vld \
        -i hdr.mov \
        -c:v hevc_videotoolbox \
        -profile:v main \
        -b:v 3M \
        -vf scale_vt=w=iw/2:h=ih/2:color_matrix=bt709:color_primaries=bt709:color_transfer=bt709 \
        -c:a copy \
        -tag:v hvc1 \
        sdr.mp4

transpose_vt 🦋

This one transposes rows with columns in the input video and optionally flips it. For more in depth examples, go peek at the transpose video filter — it shares mostly the same options, iykyk. 👀

It accepts the following parameters:

dir — Specify the transposition direction.

Can assume the following values:

cclock_flip — Rotate by 90 degrees counterclockwise and vertically flip. (default) ⭐
clock — Rotate by 90 degrees clockwise.
cclock — Rotate by 90 degrees counterclockwise.
clock_flip — Rotate by 90 degrees clockwise and vertically flip.
hflip — Flip the input video horizontally.
vflip — Flip the input video vertically.

passthrough — Do not apply the transposition if the input geometry matches the one specified by the specified value. It accepts the following values:
- none — Always apply transposition. (default)
- portrait — Preserve portrait geometry (when height >= width).
- landscape — Preserve landscape geometry (when width >= height).

ngl, passthrough is lowkey the icon here — it's literally the "read the room before you rotate" option 💀🫶

🌋 VULKAN VIDEO FILTERS 🌋

okay so this is the Vulkan filter fam — here's the whole rundown of what's currently available 💜

real talk first: to even compile these filters you gotta configure FFmpeg with --enable-vulkan and either --enable-libglslang or --enable-libshaderc. no shortcuts, bestie 🫶

and running Vulkan filters isn't a plug-and-play vibe — you need to initialize a hardware device and then hand that device to every filter in the filter graph ✨

-init_hw_device vulkan[=name][:device[,key=value...]] — inits a brand-new hardware device of type vulkan called name, using the device params and options you pass in key=value. The supported options are:
- debug — flips validation layers ON if you set it to 1.
- linear_images — allocates linear images. Does NOT apply to decoding, jsyk.
- disable_multiplane — disables multiplane images. Also does NOT apply to decoding.
- avoid_host_import — skips the dynamic host memory imports and instead goes for a regular memcpy() into a buffer that was mapped earlier.
-filter_hw_device name — passes the hardware device called name to all the filters in any filter graph. literally all of them 💅

for the deep lore, see https://www.ffmpeg.org/ffmpeg.html#Advanced-Video-options

🎀 example of grabbing the first device and running the nlmeans_vulkan filter on it with default params:

-init_hw_device vulkan=vk:0 -filter_hw_device vk -i INPUT -vf "hwupload,nlmeans_vulkan,hwdownload" OUTPUT

⚠️ important vibe check: Vulkan filters can't touch frame data sitting in normal memory 😭 so all frame data has to be uploaded (hwupload) to hardware surfaces attached to the right device before use, and then downloaded (hwdownload) back into normal memory afterwards. Also note that hwupload uploads to a frame with the same layout as the software frame — so you might need to slap a format filter immediately before it to get the input into the right format. And hwdownload doesn't support all formats on the output — usually you gotta insert another format filter right after it in the graph so the output lands in a supported format. plan accordingly bestie 🦋

💫 avgblur_vulkan

applies an average blur filter, implemented on the GPU using Vulkan ✨

it accepts these options:

sizeX — horizontal radius size. Range is [1, 32], default value is 3.
sizeY — vertical radius size. Range is [1, 32], default value is 3.
planes — which planes get filtered. Default value is 0xf, which means all planes get processed 💅

🌸 blend_vulkan

blends two Vulkan frames into each other, iconic behavior honestly.

The blend filter takes two input streams and spits out one stream — the first input is the "top" layer, the second input is the "bottom" layer. By default, the output terminates when the longest input terminates.

here's the option rundown:

c0_mode
c1_mode
c2_mode
c3_mode
all_mode

— these set the blend mode for a specific pixel component, or for all pixel components in the case of all_mode. Default value is "normal".

the available values for component modes are:

normal
multiply

🦋 bwdif_vulkan

a deinterlacer using bwdif — that's the "Bob Weaver Deinterlacing Filter" algorithm — implemented on the GPU using Vulkan 🔥

it accepts these parameters:

mode — the interlacing mode to adopt. takes one of:
- 0, send_frame — output one frame for each frame.
- 1, send_field — output one frame for each field.

default value is "send_field".

parity — the picture field parity assumed for the input interlaced video. takes one of:
- 0, tff — assume the top field is first.
- 1, bff — assume the bottom field is first.
- -1, auto — enable automatic detection of field parity.

default value is "auto". If the interlacing is unknown or the decoder doesn't export that info, top field first gets assumed. no cap 💀

deint — which frames to deinterlace. accepts one of:
- 0, all — deinterlace all frames.
- 1, interlaced — only deinterlace frames marked as interlaced.

default value is "all".

🎀 chromaber_vulkan

applies an effect that emulates chromatic aberration ✨ works best with RGB inputs, but still gives you a similar effect with YCbCr inputs too, so no worries.

dist_x — horizontal displacement multiplier. Each chroma pixel's position gets multiplied by this amount, starting from the center of the image. Default is 0.
dist_y — same deal but vertical: sets the vertical displacement multiplier. Default is 0.

👑 color_vulkan

a video source that creates a Vulkan frame of a solid color. super useful for benchmarking, or for overlaying. main character energy for a solid rectangle honestly 🫶

it accepts these parameters:

color — the color to use. either a name, or a hexadecimal value. default value is "black".
size — the size of the output frame. default value is "1920x1080".
rate — the framerate to output at. default value is 60 frames per second.
duration — the video duration. default value is -0.000001.
sar — the video signal aspect ratio. default value is "1/1".
format — the pixel format of the output Vulkan frames. default value is "yuv444p".
out_range — set the output YCbCr sample range.

this lets you override the autodetected value and force a specific value used for the output and encoder. If you don't specify it, the range depends on the pixel format. possible values:

auto/unknown — choose automatically.
jpeg/full/pc — set full range (0-255 in case of 8-bit luma).
mpeg/limited/tv — set "MPEG" range (16-235 in case of 8-bit luma).

⭐ vflip_vulkan

flips an image vertically. that's it, that's the whole filter 💅

⭐ hflip_vulkan

flips an image horizontally. equally unbothered, equally iconic.

⭐ flip_vulkan

flips an image along both the vertical and horizontal axis. double flip, double slay 🌈

✨ gblur_vulkan

applies a Gaussian blur filter on Vulkan frames 🦋

the filter accepts these options:

sigma — horizontal sigma, the standard deviation of the Gaussian blur. Default is 0.5.
sigmaV — vertical sigma. if it's negative it'll be the same as sigma. Default is -1.
planes — which planes to filter. by default all planes get filtered.
size — the kernel size along the horizontal axis. Default is 19.
sizeV — the kernel size along the vertical axis. Default is 0, which means "just use the same value as size".

💜 nlmeans_vulkan

denoises frames using the Non-Local Means algorithm, implemented on the GPU using Vulkan ✨ it supports more pixel formats than nlmeans or nlmeans_opencl, including alpha channel support. we love an upgrade 👑

the filter accepts these options:

s — denoising strength for all components. Default is 1.0. Must be in range [0.0, 100.0].
p — patch size for all planes. Default is 7. Must be an odd number in range [0, 99].
r — research size. Default is 15. Must be an odd number in range [0, 99].
t — parallelism. Default is 8. Must be a number in range [1, 64]. Larger values use more VRAM but might not actually be faster — the optimal value depends on your hardware and your input, ngl.
s0
s1
s2
s3

— denoising strength for a specific component. Default is 1.0, equal to s. Must be in range [0.0, 100.0]. Setting 0.0 disables denoising in that component entirely.

p0
p1
p2
p3

— patch size for a specific component. Default is 7, equal to p. Must be an odd number in range [0, 99].

🍓 overlay_vulkan

overlays one video on top of another, classic.

it takes two inputs and has one output. the first input is the "main" video, and the second input gets overlaid on top of it. heads up: this filter requires all inputs to use the same pixel format, so you might need some format conversion first 🙈

the filter accepts these options:

x — the x coordinate of the overlaid video on the main video. Default value is 0.
y — the y coordinate of the overlaid video on the main video. Default value is 0.

🧚 transpose_vulkan

transposes rows with columns in the input video and optionally flips it. for more in-depth examples go check the transpose video filter, which shares mostly the same options 💖

it accepts these parameters:

dir — specify the transposition direction.

can assume these values:

cclock_flip — rotate by 90 degrees counterclockwise and vertically flip. (default)
clock — rotate by 90 degrees clockwise.
cclock — rotate by 90 degrees counterclockwise.
clock_flip — rotate by 90 degrees clockwise and vertically flip.

passthrough — do NOT apply the transposition if the input geometry matches the one specified by the given value. accepts these values:
- none — always apply transposition. (default)
- portrait — preserve portrait geometry (when height >= width).
- landscape — preserve landscape geometry (when width >= height).

💅 QSV VIDEO FILTERS 💅

Okay bestie, here's the rundown on the QSV filters that are available rn. ✨

To even get these filters to compile, you gotta configure FFmpeg with --enable-libmfx or --enable-libvpl. No cap, that's the price of admission. 🎟️

And to actually use QSV filters, you need to set up the QSV device correctly first — like, properly properly. For the full tea, go read <https://trac.ffmpeg.org/wiki/Hardware/QuickSync> 🫶

hstack_qsv

Stack your input videos horizontally, horizontally, horizontally. 🎀

This is the QSV glow-up of the hstack filter, and the cute thing about it is each input stream can have a different height — the filter will scale each input stream down/up while keeping the original aspect. Iconic behavior honestly. 🦋

It accepts the following options:

inputs — see hstack.
shortest — see hstack.
height — sets the height of the output. If you set it to 0, this filter will set the height of the output to the height of the first input stream. Default value is 0.

vstack_qsv

Stack your input videos vertically. ⬆️⬇️

This is the QSV variant of the vstack filter, each input stream may have different width, and this filter will scale down/up each input stream while keeping the original aspect. Same energy, different axis. 💜

It accepts the following options:

inputs — see vstack.
shortest — see vstack.
width — set the width of the output. If set to 0, this filter will set the width of the output to the width of the first input stream. Default value is 0.

xstack_qsv

Stack video inputs into a custom layout. Like interior design but for pixels. 🛋️✨

This is the QSV variant of the xstack filter.

It accepts the following options:

inputs — see xstack.
shortest — see xstack.
layout — see xstack. Moreover, this one lets you supply the output size for each input stream. Yes bestie, per-input sizing, we love to see it:

xstack_qsv=inputs=4:layout=0_0_1920x1080|0_h0_1920x1080|w0_0_1920x1080|w0_h0_1920x1080

grid — see xstack.
grid_tile_size — sets the output size for each input stream when grid is set. If this option is not set, this filter will set the output size by default to the size of the first input stream. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. 📐
fill — see xstack.

🌈 VIDEO SOURCES 🌈

Below is a description of the currently available video sources. Yes, we're going to the source of things. 💀

buffer

Buffer video frames, and make them available to the filter chain. It's literally a little holding pen for frames. 🧺

This source is mainly intended for a programmatic use, in particular through the interface defined in libavfilter/buffersrc.h.

It accepts the following parameters:

video_size — specify the size (width and height) of the buffered video frames. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual.
width — the input video width.
height — the input video height.
pix_fmt — a string representing the pixel format of the buffered video frames. It may be a number corresponding to a pixel format, or a pixel format name. Either hits. 💅
time_base — specify the timebase assumed by the timestamps of the buffered frames.
frame_rate — specify the frame rate expected for the video stream.
colorspace — a string representing the color space of the buffered video frames. It may be a number corresponding to a color space, or a color space name.
range — a string representing the color range of the buffered video frames. It may be a number corresponding to a color range, or a color range name.
alpha_mode — a string representing the alpha mode of the buffered video frames. It may be a number corresponding to an alpha mode, or an alpha mode name.
pixel_aspect, sar — the sample (pixel) aspect ratio of the input video.
hw_frames_ctx — when using a hardware pixel format, this should be a reference to an AVHWFramesContext describing input frames.

For example:

buffer=width=320:height=240:pix_fmt=yuv410p:time_base=1/24:sar=1

will instruct the source to accept video frames with size 320x240 and with format yuv410p, assuming 1/24 as the timestamps timebase and square pixels (1:1 sample aspect ratio). And since the pixel format with name yuv410p corresponds to the number 6 (check the enum AVPixelFormat definition in libavutil/pixfmt.h), this example corresponds to:

buffer=size=320x240:pixfmt=6:time_base=1/24:pixel_aspect=1/1

Alternatively, the options can be specified as a flat string, but heads up bestie:

This syntax is deprecated. ⚠️ Like, it still works, but it's giving legacy.

width:height:pix_fmt:time_base.num:time_base.den:pixel_aspect.num:pixel_aspect.den

cellauto

Create a pattern generated by an elementary cellular automaton. Yes, math girlies rise up, this one's for you. 🧚‍♀️✨

The initial state of the cellular automaton can be defined through the filename and pattern options. If such options are not specified, an initial state is created randomly. Chaos mode: enabled. 🎲

At each new frame a new row in the video is filled with the result of the cellular automaton next generation. The behavior when the whole frame is filled is defined by the scroll option.

This source accepts the following options:

filename, f — read the initial cellular automaton state, i.e. the starting row, from the specified file. In the file, each non-whitespace character is considered an alive cell, a newline will terminate the row, and further characters in the file will be ignored.
pattern, p — read the initial cellular automaton state, i.e. the starting row, from the specified string. Each non-whitespace character in the string is considered an alive cell, a newline will terminate the row, and further characters in the string will be ignored.
rate, r — set the video rate, that is the number of frames generated per second. Default is 25.
random_fill_ratio, ratio — set the random fill ratio for the initial cellular automaton row. It is a floating point number value ranging from 0 to 1, defaults to 1/PHI. This option is ignored when a file or a pattern is specified.
random_seed, seed — set the seed for filling randomly the initial row, must be an integer included between 0 and UINT32_MAX. If not specified, or if explicitly set to -1, the filter will try to use a good random seed on a best effort basis. She'll try her best. 🫡
rule — set the cellular automaton rule, it is a number ranging from 0 to 255. Default value is 110.
size, s — set the size of the output video. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. If filename or pattern is specified, the size is set by default to the width of the specified initial state row, and the height is set to width * PHI. If size is set, it must contain the width of the specified pattern string, and the specified pattern will be centered in the larger row. If a filename or a pattern string is not specified, the size value defaults to 320x518 (used for a randomly generated initial state).
scroll — if set to 1, scroll the output upward when all the rows in the output have been already filled. If set to 0, the new generated row will be written over the top row just after the bottom row is filled. Defaults to 1.
start_full, full — if set to 1, completely fill the output with generated rows before outputting the first frame. This is the default behavior, for disabling set the value to 0.
stitch — if set to 1, stitch the left and right row edges together. This is the default behavior, for disabling set the value to 0.

Examples

Read the initial state from pattern, and specify an output of size 200x400.

cellauto=f=pattern:s=200x400

Generate a random initial row with a width of 200 cells, with a fill ratio of 2/3:

cellauto=ratio=2/3:s=200x200

Create a pattern generated by rule 18 starting by a single alive cell centered on an initial row with width 100:

cellauto=p=@s=100x400:full=0:rule=18

Specify a more elaborated initial pattern:

cellauto=p='@@ @ @@':s=100x400:full=0:rule=18

coreimagesrc

Video source generated on GPU using Apple's CoreImage API on OSX. 🍎✨

This video source is a specialized version of the coreimage video filter. Use a core image generator at the beginning of the applied filterchain to generate the content.

The coreimagesrc video source accepts the following options:

list_generators — list all available generators along with all their respective options as well as possible minimum and maximum values along with the default values.

list_generators=true

size, s — specify the size of the sourced video. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. The default value is 320x240.
rate, r — specify the frame rate of the sourced video, as the number of frames generated per second. It has to be a string in the format frame_rate_num/frame_rate_den, an integer number, a floating point number or a valid video frame rate abbreviation. The default value is 25.
sar — set the sample aspect ratio of the sourced video.
duration, d — set the duration of the sourced video. See the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. If not specified, or the expressed duration is negative, the video is supposed to be generated forever. ♾️

Additionally, all options of the coreimage video filter are accepted. A complete filterchain can be used for further processing of the generated input without CPU-HOST transfer. See coreimage documentation and examples for details.

Examples

Use CIQRCodeGenerator to create a QR code for the FFmpeg homepage, given as complete and escaped command-line for Apple's standard bash shell:

ffmpeg -f lavfi -i coreimagesrc=s=100x100:filter=CIQRCodeGenerator@inputMessage=https\\\\\://FFmpeg.org/@inputCorrectionLevel=H -frames:v 1 QRCode.png

This example is equivalent to the QRCode example of coreimage without the need for a nullsrc video source. Slay, no extra source needed. 🎀

ddagrab

Captures the Windows Desktop via Desktop Duplication API. Screen-record girlie hours. 🪟💅

The filter exclusively returns D3D11 Hardware Frames, for on-gpu encoding or processing. So an explicit hwdownload is needed for any kind of software processing. Don't forget it or you'll be confused and sad. 😭

It accepts the following options:

output_idx — DXGI Output Index to capture. Usually corresponds to the index Windows has given the screen minus one, so it's starting at 0. Defaults to output 0.
draw_mouse — whether to draw the mouse cursor. Defaults to true. Only affects hardware cursors. If a game or application renders its own cursor, it'll always be captured. 🖱️ (can't hide that one, sorry bestie)
framerate — the max framerate your desktop gets captured at, bestie ✨ the gap between frames is never smaller than 1/framerate. When dup_frames is true (and it is, by default) and your desktop just isn't updating fast enough, the filter goes ahead and duplicates a previous frame to keep the vibe going. Heads up though 💀 there's no background buffering happening, so if the filter isn't polled often enough, the actual inter-frame interval can end up way bigger than you'd think.

Defaults to 30 FPS.

video_size — sets the size of the captured video 📐

Defaults to the full size of the screen.

Cropped from the bottom/right if it's smaller than the screen size.

offset_x — horizontal offset of the captured video.
offset_y — vertical offset of the captured video.
output_fmt — the output format you want, queen 💅 defaults to 8 Bit BGRA.

It accepts these values:

auto — passes all supported output formats to DDA and just returns whatever DDA decides to use. letting DDA cook 🔥
8bit, bgra — 8 Bit formats always work, and DDA converts to them if necessary. no notes, iconic.
10bit, x2bgr10 — filter initialization will fail if a 10 bit format is requested but unavailable 😭 no cap, that's a hard fail.

dup_frames — when set to true (the default), the filter duplicates frames whenever the desktop hasn't been updated, so it can keep an approximately constant target framerate 🎯 When set to false, the filter just waits for the desktop to update — and in that case the inter-frame intervals may vary significantly.

Examples

Capture the primary screen and encode using nvenc:

ffmpeg -f lavfi -i ddagrab -c:v h264_nvenc -cq 18 output.mp4

You can also skip the lavfi device and directly use the filter ✨ This one also demonstrates downloading the frame and encoding with libx264. Explicit output format specification is required in this case:

ffmpeg -filter_complex ddagrab=output_idx=1:framerate=60,hwdownload,format=bgra -c:v libx264 -crf 18 output.mp4

If you want to capture only a subsection of the desktop, you can do that by specifying a smaller size plus its offsets into the screen 🫶

ddagrab=video_size=800x600:offset_x=100:offset_y=100

`gfxcapture`

Capture windows or monitors using the Windows.Graphics.Capture API 🪟✨

This source gives you low overhead capture of application windows or entire monitors. The filter outputs hardware frames in d3d11 format — so use hwdownload,format= if you need system memory frames, mkay?

Which window gets captured? You can pick it via regular expressions on its title, class name or backing executable name, by explicit native handles, or by monitor index or explicit native handle. A window has to match all the expressions you provide to be selected 👑 The first matching window gets picked, in whatever order Windows happens to return them (chaotic, I know 💀).

Explicit handles (hwnd, hmonitor) override pattern or index based selection. If neither handles nor a monitor index are given, the first window matching the provided regular expressions is captured.

This source does NOT hold a stable FPS, bestie 😭 It returns frames at whatever rate the compositor provides them, only capped by max_framerate. If you need a stable rate, you need to add an fps filter to drop/duplicate frames as needed.

If the capture source disappears mid-capture (window closed, monitor disconnected), the filter will return EOF.

This source accepts the following options:

window_title — ECMAScript regular expression matched against the window title. Supports a PCRE style (?i) prefix for case-insensitive matching 🌸
window_class — same as window_title, but matched against the window class name.
window_exe — same as window_title, but matched against the executable file name of the window's process.
monitor_idx — zero-based index of the monitor to capture.

Can also be set to window to capture the monitor the selected window is displayed on at filter initialization time 🖥️

hwnd — explicit native window handle (HWND).
hmonitor — explicit native monitor handle (HMONITOR).
capture_cursor — capture the mouse cursor. Enabled by default.
capture_border — capture the full area of the window, including its window decorarions/border. Disabled by default.
display_border — draw a yellow highlight border around the captured window 💛 Disabled by default.
max_framerate — maximum capture frame rate. Accepts a video rate (e.g. 30, "60/1", "24000/1001"). The default is 60 FPS. The actual rate is the rate at which the compositor renders the window/monitor, capped by this option.
width — force the output canvas width. If zero (default) the initial captured source width is used. If a negative number is provided, the width will be rounded down to the next multiple of that number.

See resize_mode.

height — force the output canvas height. If zero (default) the initial captured source height is used. If a negative number is provided, the height will be rounded down to the next multiple of that number.

See resize_mode.

crop_left — crop this many pixels from the left side of the captured frames.
crop_top — crop this many pixels from the left side of the captured frames.
crop_right — crop this many pixels from the left side of the captured frames.
crop_bottom — crop this many pixels from the left side of the captured frames.

okay so yes, the docs literally say "left side" for all four of these 🙈 that's exactly what the page says, I'm just the messenger bestie 💅

premultiplied — if set to 1, return frames with premultiplied alpha. Default is 0 (straight alpha).
resize_mode — defines how the captured content is fitted into the output canvas size. Possible values:
- crop — crop (or pad with black) to the canvas size. (default)
- scale — scale the source to fill the canvas, potentially altering aspect ratio 👀
- scale_aspect — scale the source to fit inside the canvas while preserving aspect ratio. Remaining area is filled with black.
scale_mode — scaling algorithm used when resizing is required. Possible values:
- point — nearest neighbour (pixelated) scaling.
- bilinear — bilinear filtering. (default)
- bicubic — bicubic filtering. Potentially more blurry, but fewer scaling artifacts depending on contents.
output_fmt — desired output pixel format inside the D3D11 hardware frames. Possible values:
- bgra, 8bit — 8 bit BGRA output (default)
- x2bgr10, 10bit — 10 bit BGR output
- rgbaf16, 16bit — 16bit float RGBA output

Examples

Capture a window by title (case-insensitive) at a maximum of 60 fps:

ffmpeg -filter_complex gfxcapture=window_title='(?i)My Application':max_framerate=60,hwdownload,format=bgra,format=yuv420p -c:v libx264 -crf 15 capture.mp4

Capture monitor 1 at native refresh, 10bit color depth, scale to 1920x1080 preserving aspect:

ffmpeg -filter_complex gfxcapture=monitor_idx=1:width=1920:height=1080:resize_mode=scale_aspect:output_fmt=10bit -c:v hevc_nvenc -cq 15 capture.mp4

Capture a window by executable name, draw border, force point scaling, fixed 60 fps:

ffmpeg -filter_complex gfxcapture=window_exe='^firefox.exe$':display_border=1:scale_mode=point,fps=60 -rc qvbr -qvbr_quality_level 15 -c:v h264_amf capture.mp4

`gradients`

Generate several gradients 🌈✨ aesthetic af, honestly.

size, s — set frame size. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Default value is 640x480.
rate, r — set frame rate, expressed as number of frames per second. Default value is 25.
c0, c1, c2, c3, c4, c5, c6, c7 — set 8 colors. The default behavior for colors is to pick a random one 🎲
x0, y0, y0, y1 — set gradient line source and destination points. If negative or out of range, random ones are picked.
nb_colors, n — set number of colors to use at once. Allowed range is from 2 to 8. Default value is 2.
seed — set seed for picking gradient line points.
duration, d — set the duration of the sourced video. See the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax.

If not specified, or the expressed duration is negative, the video is supposed to be generated forever ♾️

speed — set speed of gradients rotation.
type, t — set type of gradients. Available values are:
- linear
- radial
- circular
- spiral
- square

Default type is linear.

Commands

This source supports the some above options as commands 💅

`mandelbrot`

Generate a Mandelbrot set fractal, and progressively zoom towards the point specified with start_x and start_y 🌀 lowkey mesmerizing, ngl.

This source accepts the following options:

end_pts — set the terminal pts value. Default value is 400.
end_scale — set the terminal scale value. Must be a floating point value. Default value is 0.3.
inner — set the inner coloring mode, that is the algorithm used to draw the Mandelbrot fractal internal region.

It shall assume one of the following values:

black — set black mode.
convergence — show time until convergence.
mincol — set color based on point closest to the origin of the iterations.
period — set period mode.

Default value is mincol.

bailout — set the bailout value. Default value is 10.0.
maxiter — set the maximum of iterations performed by the rendering algorithm. Default value is 7189.
outer — set outer coloring mode. It shall assume one of following values:
- iteration_count — set iteration count mode.
- normalized_iteration_count — set normalized iteration count mode.

Default value is normalized_iteration_count.

rate, r — set frame rate, expressed as number of frames per second. Default value is 25.
size, s — set frame size. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Default value is 640x480.
start_scale — set the initial scale value. Default value is 3.0.
start_x — set the initial x position. Must be a floating point value between -100 and 100. Default value is -0.743643887037158704752191506114774.
start_y — set the initial y position. Must be a floating point value between -100 and 100. Default value is -0.131825904205311970493132056385139.

mptestsrc

Okay bestie, this one serves up a whole buffet of test patterns — the same ones the MPlayer test filter makes. The size is fixed and non-negotiable: 512x512. It's highkey the go-to source for testing encoding features, so if you're doing a vibe check on your encoder, this is your girl. ✨

It takes these options:

rate, r — the frame rate of the sourced video, i.e. how many frames get generated per second. Gotta be a string shaped like frame_rate_num/frame_rate_den, or a plain integer, or a floating point number, or a legit video frame rate abbreviation. Default is "25".
duration, d — how long the sourced video runs. Peep the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. If you don't set it, or the duration you express is negative, the video is assumed to be generated forever (main character energy, no end in sight 😭).
test, t — the number or name of the test to run. Supported tests are:
- dc_luma
- dc_chroma
- freq_luma
- freq_chroma
- amp_luma
- amp_chroma
- cbp
- mv
- ring1
- ring2
- all
- max_frames, m — the max number of frames generated for each test, default value is 30.

Default value is "all", which cycles through the whole list of tests. 🎠

Some examples, because we love a demo:

mptestsrc=t=dc_luma

That one generates a "dc_luma" test pattern. Slay. 💅

frei0r_src

Provides a frei0r source. 🧚

Heads up: to enable compilation of this filter you need to install the frei0r header and configure FFmpeg with --enable-frei0r. No header, no filter, that's the deal.

It takes these parameters:

size — the size of the video to generate. For the syntax, check the "Video size" section in the ffmpeg-utils manual.
framerate — the framerate of the generated video. Can be a string of the form num/den or a frame rate abbreviation.
filter_name — the name of the frei0r source to load. Want more info on frei0r and how to set the parameters? Read the frei0r section in the video filters documentation.
filter_params — a '|'-separated list of parameters to pass to the frei0r source.

For example, to generate a frei0r partik0l source with size 200x200 and frame rate 10, overlaid on the overlay filter's main input:

frei0r_src=size=200x200:framerate=10:filter_name=partik0l:filter_params=1234 [overlay]; [in][overlay] overlay

life

Generates a life pattern, bestie. 🦋

This source is based on a generalization of John Conway's life game.

The sourced input represents a life grid — every pixel is a cell that can be in one of two states: alive or dead. Every cell interacts with its eight neighbours, meaning the cells that are horizontally, vertically, or diagonally adjacent.

At each interaction the grid evolves according to whichever rule is adopted, and the rule specifies how many alive neighbor cells will make a cell stay alive or be born. The rule option is literally how you pick that rule. 👑

This source accepts these options:

filename, f — the file to read the initial grid state from. In that file, every non-whitespace character counts as an alive cell, and a newline delimits the end of each row. If you don't set this option, the initial grid is generated randomly. 🎲
rate, r — the video rate, i.e. the number of frames generated per second. Default is 25.
random_fill_ratio, ratio — the random fill ratio for the initial random grid. It's a floating point number from 0 to 1, defaults to 1/PHI. It is ignored when a file is specified (so don't @ me when it does nothing).
random_seed, seed — the seed for filling the initial random grid, must be an integer between 0 and UINT32_MAX. If not specified, or explicitly set to -1, the filter will try to use a good random seed on a best effort basis.
rule — the life rule. You can set a rule with a code of the kind "SNS/BNB", where NS and NB are sequences of numbers in the range 0-8. NS specifies the number of alive neighbor cells which make a live cell stay alive, and NB the number of alive neighbor cells which make a dead cell become alive (i.e. to be "born"). "s" and "b" can be used in place of "S" and "B" respectively.

Alternatively, a rule can be specified by an 18-bits integer. The 9 high order bits encode the next cell state if it's alive, for each number of neighbor alive cells; the low order bits specify the rule for "borning" new cells. Higher order bits encode for a higher number of neighbor cells. For example the number 6153 = (12<<9)+9 specifies a stay alive rule of 12 and a born rule of 9, which corresponds to "S23/B03".

Default value is "S23/B3", which is the original Conway's game of life rule — keeps a cell alive if it has 2 or 3 neighbor alive cells, and borns a new cell if there are three alive cells around a dead cell. iconic. 💀

size, s — the size of the output video. For the syntax, check the "Video size" section in the ffmpeg-utils manual.

If filename is specified, the size defaults to the same size as the input file. If size is set, it must contain the size specified in the input file, and the initial grid defined in that file gets centered in the larger resulting area.

If a filename is not specified, the size value defaults to "320x240" (used for a randomly generated initial grid).

stitch — if set to 1, stitch the left and right grid edges together, and the top and bottom edges too. Defaults to 1.
mold — cell mold speed. If set, a dead cell goes from death_color to mold_color with a step of mold. mold can have a value from 0 to 255. 🍓
life_color — the color of living (or new born) cells.
death_color — the color of dead cells. If mold is set, this is the first color used to represent a dead cell.
mold_color — the mold color, for definitely dead and moldy cells.

For the syntax of these 3 color options, check the "Color" section in the ffmpeg-utils manual.

Examples

Read a grid from pattern, and center it on a grid of size 300x300 pixels:

life=f=pattern:s=300x300

Generate a random grid of size 200x200, with a fill ratio of 2/3:

life=ratio=2/3:s=200x200

Specify a custom rule for evolving a randomly generated grid:

life=rule=S14/B34

Full example with slow death effect (mold) using ffplay:

ffplay -f lavfi life=s=300x200:mold=10:r=60:ratio=0.1:death_color=#C83232:life_color=#00ff00,scale=1200:800:flags=16

perlin

Generates Perlin noise. 🌫️

Perlin noise is a kind of noise with local continuity in space. It's useful for generating patterns with continuity in space and time — think simulating smoke, fluids, or terrain. Obsessed honestly.

If more than one octave is specified through the octaves option, Perlin noise gets generated as a sum of components, each one with doubled frequency. In that case the persistence option specifies the ratio of the amplitude with respect to the previous component. More octave components let you specify more high frequency details in the generated noise (like small size variations due to boulders in a generated terrain 🪨).

Options

size, s — the size (width and height) of the buffered video frames. For the syntax, check the "Video size" section in the ffmpeg-utils manual. Default value is "320x240".
rate, r — the frame rate expected for the video stream, expressed as a number of frames per second. Default value is 25.
octaves — the total number of components making up the noise, each one with doubled frequency. Default value is 1.
persistence — the ratio used to compute the amplitude of the next octave component with respect to the previous component amplitude. Default value is 1.
xscale, yscale — define a scale factor used to multiple the x, y coordinates. Useful for an effect with a pattern stretched along the x or y axis. Default value is 1.
tscale — define a scale factor used to multiple the time coordinate. Useful to change the time variation speed. Default value is 1.
random_mode — the random mode used to compute the initial pattern. Supported values are:
- random — compute and use random seed.
- ken — use the predefined initial pattern defined by Ken Perlin in the original article, can be useful to compare the output with other sources.
- seed — use the value specified by the random_seed option.

Default value is "random".

random_seed, seed — when random_mode is set to random_seed, use this value to compute the initial pattern. Default value is 0.

Examples

Generate single component:

perlin

Use Perlin noise with 7 components, each one with a halved contribution to total amplitude:

perlin=octaves=7:persistence=0.5

Chain Perlin noise with the lutyuv to generate a black&white effect:

perlin=octaves=3:tscale=0.3,lutyuv=y='if(lt(val\,128)\,255\,0)'

Stretch noise along the y axis, and convert gray level to red-only signal:

perlin=octaves=7:tscale=0.4:yscale=0.3,lutrgb=r=val:b=0:g=0

qrencodesrc

Generates a QR code using the libqrencode library (see <https://fukuchi.org/works/qrencode/>). 🎀

To enable the compilation of this source, you need to configure FFmpeg with --enable-libqrencode. No configure flag, no QR codes, sorry bestie.

The QR code gets generated from the provided text or text pattern. The corresponding QR code is scaled and put in the video output according to the specified output size options.

In case no text is specified, the QR code is not generated — instead you get an empty colored output back. 🙈

This source accepts these options:

qrcode_width, q — specify an expression for the width of the rendered QR code, with padding.
padded_qrcode_width, Q — specify an expression for the width of the rendered QR code, without padding.

The qrcode_width expression can reference the value set by the padded_qrcode_width expression, and vice versa. By default padded_qrcode_width is set to qrcode_width, meaning there is no padding.

These expressions are evaluated only once, when initializing the source. See the qrencode Expressions section for details. ✨

okay so real talk bestie: some of the constants are missing for the source (for example x or t or ¸n), because they only make sense when you evaluate the expression for each frame rather than at initialization time. no cap. 💅

rate, r — set the frame rate of the sourced video, as the number of frames generated per second. It has to be a string in the format frame_rate_num/frame_rate_den, an integer number, a floating point number or a valid video frame rate abbreviation. Default is "25". ✨
case_sensitive, cs — tell libqrencode to use case sensitive encoding. It’s enabled by default. You can disable it to reduce the QR encoding size.
level, l — set the QR encoding error correction level. Higher correction level means bigger encoding size, but the code is more robust to corruption. The lower level is L. It accepts L, M, Q, H. 🎀
expansion — choose how the input text is expanded. Can be none or normal (default). See the qrencode Text expansion section for details.
text, textfile — define the text to render. If neither is specified, no QR is encoded (just an empty colored frame). If expansion is enabled, the text is treated as a text template using the qrencode expansion mechanism. See the qrencode Text expansion section for details. 🫶
background_color, bc and foreground_color, fc — set the QR code and background color. The default for foreground_color is black, and the default for background_color is white. For the syntax of the color options, check the Color section in the ffmpeg-utils manual. 🌈

✨ Examples ✨

Generate a QR code encoding the specified text with the default size:

qrencodesrc=text=www.ffmpeg.org

Same as below, but select blue on pink colors:

qrencodesrc=text=www.ffmpeg.org:bc=pink:fc=blue

Generate a QR code with width of 200 pixels and padding, making the padded width 4/3 of the QR code width:

qrencodesrc=text=www.ffmpeg.org:q=200:Q=4/3*q

Generate a QR code with padded width of 200 pixels and padding, making the QR code width 3/4 of the padded width:

qrencodesrc=text=www.ffmpeg.org:Q=200:q=3/4*Q

Generate a QR code encoding the frame number:

qrencodesrc=text=%{n}

Generate a QR code encoding the GMT timestamp:

qrencodesrc=text=%{gmtime}

Generate a QR code encoding the timestamp expressed as a float:

qrencodesrc=text=%{pts}

✨ allrgb, allyuv, color, colorchart, colorspectrum, haldclutsrc, nullsrc, pal75bars, pal100bars, rgbtestsrc, smptebars, smptehdbars, testsrc, testsrc2, yuvtestsrc ✨

allrgb — returns frames of size 4096x4096 of all rgb colors. 👑
allyuv — returns frames of size 4096x4096 of all yuv colors.
color — provides a uniformly colored input.
colorchart — provides a colors checker chart. 🎨
colorspectrum — provides a color spectrum input.
haldclutsrc — provides an identity Hald CLUT. See also the haldclut filter.
nullsrc — returns unprocessed video frames. It’s mainly useful in analysis / debugging tools, or as the source for filters which ignore the input data. 💀
pal75bars — generates a color bars pattern based on EBU PAL recommendations with 75% color levels.
pal100bars — generates a color bars pattern based on EBU PAL recommendations with 100% color levels.
rgbtestsrc — generates an RGB test pattern useful for detecting RGB vs BGR issues. You should see a red, green and blue stripe from top to bottom. 🌈
smptebars — generates a color bars pattern based on the SMPTE Engineering Guideline EG 1-1990.
smptehdbars — generates a color bars pattern based on the SMPTE RP 219-2002.
testsrc — generates a test video pattern showing a color pattern, a scrolling gradient and a timestamp. This is mainly intended for testing purposes. ✨
testsrc2 — similar to testsrc, but supports more pixel formats instead of just rgb24. This lets you use it as an input for other tests without requiring a format conversion.
yuvtestsrc — generates a YUV test pattern. You should see a y, cb and cr stripe from top to bottom.

The sources accept the following parameters:

level — set the level of the Hald CLUT, only available in the haldclutsrc source. A level of N generates a picture of N*N*N by N*N*N pixels to be used as identity matrix for 3D lookup tables. Each component is coded on a 1/(N*N) scale.
color, c — set the color of the source, only available in the color source. For the syntax of this option, check the Color section in the ffmpeg-utils manual. 💖
size, s — set the size of the sourced video. For the syntax of this option, check the Video size section in the ffmpeg-utils manual. Default is 320x240. This option is not available with the allrgb, allyuv, and haldclutsrc filters.
rate, r — set the frame rate of the sourced video, as the number of frames generated per second. It has to be a string in the format frame_rate_num/frame_rate_den, an integer number, a floating point number or a valid video frame rate abbreviation. Default is "25". 🌸
duration, d — set the duration of the sourced video. See the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. If not specified, or if the expressed duration is negative, the video is supposed to be generated forever. Since the frame rate is used as time base, all frames including the last one will have their full duration. If the specified duration is not a multiple of the frame duration, it will be rounded up. 🕰️
sar — set the sample aspect ratio of the sourced video.
alpha — set the alpha (opacity) of the background, only available in the testsrc2 source. The value must be between 0 (fully transparent) and 255 (fully opaque, the default).
decimals, n — set the number of decimals to show in the timestamp, only available in the testsrc source. The displayed timestamp value corresponds to the original timestamp value multiplied by the power of 10 of the specified value. Default is 0.
type — set the type of the color spectrum, only available in the colorspectrum source. Can be one of black, white, all.
patch_size — set the patch size of a single color patch, only available in the colorchart source. Default is 64x64.
preset — set the colorchecker colors preset, only available in the colorchart source. Available values are reference, skintones. Default is reference. 🎀

✨ Examples ✨

Generate a video with a duration of 5.3 seconds, with size 176x144 and a frame rate of 10 frames per second:

testsrc=duration=5.3:size=qcif:rate=10

The following graph description will generate a red source with an opacity of 0.2, with size qcif and a frame rate of 10 frames per second:

color=c=red@0.2:s=qcif:r=10

If the input content is to be ignored, nullsrc can be used. The following command generates noise in the luma plane by employing the geq filter:

nullsrc=s=256x256, geq=random(1)*255:128:128

✨ Commands ✨

The color source supports the following commands:

c, color — set the color of the created image. Accepts the same syntax as the corresponding color option.

✨ openclsrc ✨

Generate video using an OpenCL program. 🧚

source — OpenCL program source file.
kernel — kernel name in program.
size, s — size of frames to generate. This must be set.
format — pixel format to use for the generated frames. This must be set.
rate, r — number of frames generated every second. Default is '25'.

For details of how the program loading works, see the program_opencl filter.

Example programs:

Generate a colour ramp by setting pixel values from the position of the pixel in the output image. (Note that this will work with all pixel formats, but the generated output will not be the same.)

__kernel void ramp(__write_only image2d_t dst,
                   unsigned int index)
{
    int2 loc = (int2)(get_global_id(0), get_global_id(1));

    float4 val;
    val.xy = val.zw = convert_float2(loc) / convert_float2(get_image_dim(dst));

    write_imagef(dst, loc, val);
}

Generate a Sierpinski carpet pattern, panning by a single pixel each frame.

__kernel void sierpinski_carpet(__write_only image2d_t dst,
                                unsigned int index)
{
    int2 loc = (int2)(get_global_id(0), get_global_id(1));

    float4 value = 0.0f;
    int x = loc.x + index;
    int y = loc.y + index;
    while (x > 0 || y > 0) {
        if (x % 3 == 1 && y % 3 == 1) {
            value = 1.0f;
            break;
        }
        x /= 3;
        y /= 3;
    }

    write_imagef(dst, loc, value);
}

✨ sierpinski ✨

Generate a Sierpinski carpet/triangle fractal, and randomly pan around. 🌀

This source accepts the following options:

size, s — set frame size. For the syntax of this option, check the Video size section in the ffmpeg-utils manual. Default is 640x480.
rate, r — set frame rate, expressed as number of frames per second. Default is "25".
seed — set seed which is used for random panning.
jump — set max jump for single pan destination. Allowed range is from 1 to 10000.
type — set fractal type, can be default carpet or triangle. 🔥

✨ zoneplate ✨

Generate a zoneplate test video pattern. ⭐

This source accepts the following options:

size, s — set frame size. For the syntax of this option, check the Video size section in the ffmpeg-utils manual. Default is 320x240.
rate, r — sets the frame rate, as a number of frames per second. default is 25 💫
duration, d — sets how long the sourced video lasts. the syntax is the Time duration section in the ffmpeg-utils(1) manual, go check it 🫶

if you don't specify it, or the duration you give is negative, the video is just gonna generate forever. infinite content era 😌

sar — sets the sample aspect ratio of the sourced video ✨
precision — sets precision in bits for the look-up table used for the sine calculations. default is 10, and the allowed range is 4 to 16 🎀
xo — sets the horizontal axis offset for the output signal. default is 0
yo — sets the vertical axis offset for the output signal. default is 0
to — sets the time axis offset for the output signal. default is 0
k0 — sets the 0-order constant added to the signal phase. default is 0
kx — sets the 1-order phase factor multiplier for the horizontal axis. default is 0
ky — sets the 1-order phase factor multiplier for the vertical axis. default is 0
kt — sets the 1-order phase factor multiplier for the time axis. default is 0
kxt, kyt, kxy — set the phase factor multipliers for combinations of spatial and temporal axes. default is 0 🌸
kx2 — sets the 2-order phase factor multiplier for the horizontal axis. default is 0
ky2 — sets the 2-order phase factor multiplier for the vertical axis. default is 0
kt2 — sets the 2-order phase factor multiplier for the time axis. default is 0
ku — sets the constant added to the final phase to produce the chroma-blue component of the signal. default is 0 💙
kv — sets the constant added to the final phase to produce the chroma-red component of the signal. default is 0

🎮 Commands

this source supports some of the options above as commands, iykyk 💅

🌈 Examples

generate a horizontal color sine sweep:

zoneplate=ku=512:kv=0:kt2=0:kx2=256:s=wvga:xo=-426:kt=11

generate a vertical color sine sweep:

zoneplate=ku=512:kv=0:kt2=0:ky2=156:s=wvga:yo=-240:kt=11

generate a circular zone-plate:

zoneplate=ku=512:kv=100:kt2=0:ky2=256:kx2=556:s=wvga:yo=0:kt=11

🎬 VIDEO SINKS 🎬

okay bestie here's the rundown of the video sinks that are currently available. these are the endpoints — where the pixels go to rest 🦋

`buffersink`

buffers video frames and makes them available to the end of the filter graph. literally the finish line 🏁

this sink is mainly intended for programmatic use — especially through the interface defined in libavfilter/buffersink.h or the options system.

it accepts a pointer to an AVBufferSinkContext structure, which defines the incoming buffers' formats, and that pointer gets passed as the opaque parameter to avfilter_init_filter for initialization. yes that's very specific, yes you need all of it 🎀

`nullsink`

a null video sink: it does absolutely nothing with the input video. zero notes. it just sits there 💀

it's mainly useful as a template, and for use in analysis / debugging tools. king of doing nothing but doing it flawlessly 👑

💖 MULTIMEDIA FILTERS 💖

and now the multimedia filters that are currently available — these ones ate and left no crumbs ✨

`a3dscope`

converts the input audio into 3d scope video output. it's giving main character visualization 🌟

the filter accepts the following options:

rate, r — sets the frame rate, as a number of frames per second. default is 25
size, s — specify the video size for the output. for the syntax of this option, check the Video size section in the ffmpeg-utils manual. default is hd720 🖼️
fov — sets the camera field of view. default is 90 degrees, and the allowed range is from 40 to 150 👀
roll — sets the camera roll
pitch — sets the camera pitch
yaw — sets the camera yaw
xzoom — sets the camera zoom on the X-axis
yzoom — sets the camera zoom on the Y-axis
zzoom — sets the camera zoom on the Z-axis
xpos — sets the camera position on the X-axis
ypos — sets the camera position on the Y-axis
zpos — sets the camera position on the Z-axis
length — sets the length of the displayed audio waves, in number of frames 💫

🎮 Commands

the filter supports some of the options above as commands 💅

`abitscope`

converts the input audio into a video output, displaying the audio bit scope. it's literally the audio version of a glow-up reveal 😭

the filter accepts the following options:

rate, r — sets the frame rate, as a number of frames per second. default is 25
size, s — specify the video size for the output. for the syntax of this option, check the Video size section in the ffmpeg-utils manual. default is 1024x256 🎀
colors — specify a list of colors, separated by a space or by |, that'll be used to draw the channels. any unrecognized or missing colors get replaced by white. no cap 🌈
mode, m — sets the output mode. can be bars or trace. default is bars

`adrawgraph`

draws a graph using the input audio metadata. the audio is the moment, the graph is the aesthetic 📈

see drawgraph for the details 🫶

`agraphmonitor`

see graphmonitor — that's the whole vibe check, go read that one 💜

`ahistogram`

converts the input audio into a video output, displaying the volume histogram. it's giving data visualization slay 🍓

the filter accepts the following options:

dmode — specify how the histogram is calculated. it accepts the following values:
- single — use a single histogram for all channels
- separate — use a separate histogram for each channel

default is single ✨

rate, r — sets the frame rate, as a number of frames per second. default is 25
size, s — specify the video size for the output. for the syntax of this option, check the Video size section in the ffmpeg-utils manual. default is hd720
scale — sets the display scale. it accepts the following values:
- log — logarithmic
- sqrt — square root
- cbrt — cubic root
- lin — linear
- rlog — reverse logarithmic

default is log 🌸

ascale — sets the amplitude scale. it accepts the following values:
- log — logarithmic
- lin — linear

default is log

acount — sets how many frames to accumulate in the histogram. default is 1. set it to -1 and it accumulates all the frames. hoarder behavior, we respect it 💀
rheight — sets the histogram ratio of window height
slide — sets the sonogram sliding. it accepts the following values:
- replace — replace old rows with new ones
- scroll — scroll from top to bottom

default is replace 🦋

hmode — sets the histogram mode. it accepts the following values:
- abs — use absolute values of samples
- sign — use untouched values of samples

default is abs

`aphasemeter`

measures the phase of the input audio, which gets exported as metadata lavfi.aphasemeter.phase, representing the mean phase of the current audio frame. a video output can also be produced, and it's enabled by default. the audio is passed through as the first output 🔊

the audio will be rematrixed to stereo if it has a different channel layout. the phase value is in range [-1, 1], where -1 means the left and right channels are completely out of phase, and 1 means the channels are in phase. it's giving relationship status ngl 💔💖

the filter accepts the following options, all related to its video output:

rate, r — sets the output frame rate. default value is 25
size, s — sets the video size for the output. for the syntax of this option, check the Video size section in the ffmpeg-utils manual. default value is 800x400 🎀
rc, gc, bc — specify the red, green and blue contrast. default values are 2, 7 and 1. the allowed range is [0, 255]
mpc — sets the color that'll be used for drawing the median phase. if the color is none, which is the default, no median phase value gets drawn 🙈
video — enables the video output. default is enabled

🔍 phasing detection 🔍

the filter also detects out of phase and mono sequences in stereo streams. it logs the sequence start, end and duration when it lasts longer than — or as long as — the minimum you set. it's literally a detective, obsessed 🕵️‍♀️

the filter accepts the following options for this detection:

phasing — enables mono and out of phase detection. default is disabled
tolerance, t — sets the phase tolerance for mono detection, in amplitude ratio. default is 0. the allowed range is [0, 1]
angle, a — sets the angle threshold for out of phase detection, in degrees. default is 170. the allowed range is [90, 180] 📐
duration, d — sets the mono or out of phase duration until notification, expressed in seconds. default is 2

✨ Examples ✨

a complete example with ffmpeg to detect 1 second of mono with 0.001 phase tolerance:

ffmpeg -i stereo.wav -af aphasemeter=video=0:phasing=1:duration=1:tolerance=0.001 -f null -

`avectorscope`

converts the input audio into a video output, representing the audio vector scope. this one's a whole aesthetic 🌈

the filter is used to measure the difference between the channels of a stereo audio stream. a monaural signal — meaning identical left and right signals — results in a straight vertical line. any stereo separation shows up as a deviation from that line, creating a Lissajous figure. and if the straight line (or a deviation from it) shows up horizontal, that means the left and right channels are out of phase. literal visual proof, no cap 😭

the filter accepts the following options:

mode, m — sets the vectorscope mode. available values are:
- lissajous — Lissajous rotated by 45 degrees
- lissajous_xy — same as above but not rotated
- polar — a shape resembling half of a circle

default value is lissajous 💫

size, s — sets the video size for the output. for the syntax of this option, check the Video size section in the ffmpeg-utils manual. default value is 400x400 🎀
rate, r — sets the output frame rate. default value is 25
rc, gc, bc, ac — specify the red, green, blue and alpha contrast. default values are 40, 160, 80 and 255. the allowed range is [0, 255]
rf, gf, bf, af — specify the red, green, blue and alpha fade. default values are 15, 10, 5 and 5. the allowed range is [0, 255] 🌸
zoom — sets the zoom factor. default value is 1, and the allowed range is [0, 10]. values lower than 1 will auto adjust the zoom factor to the maximal possible value. it fixes itself, self-care icon 💅
draw — sets the vectorscope drawing mode. available values are:
- dot — draw a dot for each sample
- line — draw a line between the previous and current sample
- aaline — draw an anti-aliased line between the previous and current sample

default value is dot ✨

scale — specify the amplitude scale of the audio samples. available values are:
- lin — linear
- sqrt — square root
- cbrt — cubic root
- log — logarithmic
swap — swaps the left channel axis with the right channel axis 🔄
mirror — mirrors the axis. available values are:
- none — no mirror

okay so the mirror options are short and iconic:

x — mirror only the x axis.
y — mirror only the y axis.
xy — mirror both axis.

🎬 Examples

Here’s the complete example using ffplay — copy it exactly, bestie:

ffplay -f lavfi 'amovie=input.mp3, asplit [a][out1];
             [a] avectorscope=zoom=1.3:rc=2:gc=200:bc=10:rf=1:gf=8:bf=7 [out0]'

🕹️ Commands

This filter supports the all above options as commands except options size and rate.

🧚 bench, abench

Benchmark part of a filtergraph. It’s literally giving stopwatch energy for your filtergraph’s little runway walk. 💅

The filter accepts the following options:

action — start or stop a timer.

Available values are:

start — get the current time, set it as frame metadata (using the key lavfi.bench.start_time), and forward the frame to the next filter.
stop — get the current time and fetch the lavfi.bench.start_time metadata from the input frame metadata to get the time difference. Time difference, average, maximum and minimum time (respectively t, avg, max and min) are then printed. The timestamps are expressed in seconds.

🌸 Examples

Benchmark selectivecolor filter:

bench=start,selectivecolor=reds=-.2 .12 -.49,bench=stop

🧵 concat

Concatenate audio and video streams, joining them together one after the other. It’s the “stitch these clips together” filter, and yes, it ate. ✨

The filter works on segments of synchronized video and audio streams. All segments must have the same number of streams of each type, and that will also be the number of streams at output.

The filter accepts the following options:

n — set the number of segments. Default is 2.
v — set the number of output video streams, that is also the number of video streams in each segment. Default is 1.
a — set the number of output audio streams, that is also the number of audio streams in each segment. Default is 0.
unsafe — activate unsafe mode: do not fail if segments have a different format.

The filter has v+a outputs: first v video outputs, then a audio outputs.

There are n × (v+a) inputs: first the inputs for the first segment, in the same order as the outputs, then the inputs for the second segment, etc.

Related streams do not always have exactly the same duration, for various reasons including codec frame size or sloppy authoring. For that reason, related synchronized streams (e.g. a video and its audio track) should be concatenated at once. The concat filter will use the duration of the longest stream in each segment (except the last one), and if necessary pad shorter audio streams with silence.

🚨 For this filter to work correctly, all segments must start at timestamp 0.

All corresponding streams must have the same parameters in all segments; the filtering system will automatically select a common pixel format for video streams, and a common sample format, sample rate and channel layout for audio streams, but other settings, such as resolution, must be converted explicitly by the user.

Different frame rates are acceptable but will result in variable frame rate at output; be sure to configure the output file to handle it.

🎬 Examples

Concatenate an opening, an episode and an ending, all in bilingual version (video in stream 0, audio in streams 1 and 2):

ffmpeg -i opening.mkv -i episode.mkv -i ending.mkv -filter_complex \
  '[0:0] [0:1] [0:2] [1:0] [1:1] [1:2] [2:0] [2:1] [2:2]
   concat=n=3:v=1:a=2 [v] [a1] [a2]' \
  -map '[v]' -map '[a1]' -map '[a2]' output.mkv

Concatenate two parts, handling audio and video separately, using the (a)movie sources, and adjusting the resolution:

movie=part1.mp4, scale=512:288 [v1] ; amovie=part1.mp4 [a1] ;
movie=part2.mp4, scale=512:288 [v2] ; amovie=part2.mp4 [a2] ;
[v1] [v2] concat [outv] ; [a1] [a2] concat=v=0:a=1 [outa]

🚨 Note that a desync will happen at the stitch if the audio and video streams do not have exactly the same duration in the first file.

🕹️ Commands

This filter supports the following commands:

next — close the current segment and step to the next one.

🎧 ebur128

EBU R128 scanner filter. This filter takes an audio stream and analyzes its loudness level. By default, it logs a message at a frequency of 10Hz with the Momentary loudness (identified by M), Short-term loudness (S), Integrated loudness (I) and Loudness Range (LRA). Loudness girlie is doing a full vibe check on your audio. 👑✨

The filter can only analyze streams which have sample format is double-precision floating point. The input stream will be converted to this specification, if needed. Users may need to insert aformat and/or aresample filters after this filter to obtain the original parameters.

The filter also has a video output (see the video option) with a real time graph to observe the loudness evolution. The graphic contains the logged message mentioned above, so it is not printed anymore when this option is set, unless the verbose logging is set. The main graphing area contains the short-term loudness (3 seconds of analysis), and the gauge on the right is for the momentary loudness (400 milliseconds), but can optionally be configured to instead display short-term loudness (see gauge).

The green area marks a +/- 1LU target range around the target loudness (-23LUFS by default, unless modified through target).

More info about the Loudness Recommendation EBU R128 is at http://tech.ebu.ch/loudness.

The filter accepts the following options:

video — activate the video output. The audio stream is passed unchanged whether this option is set or no. The video stream will be the first output stream if activated. Default is 0.
size — set the video size. This option is for video only. For the syntax of this option, check the “Video size” section in the ffmpeg-utils manual. Default and minimum resolution is 640x480.
meter — set the EBU scale meter. Default is 9. Common values are 9 and 18, respectively for EBU scale meter +9 and EBU scale meter +18. Any other integer value between this range is allowed.
metadata — set metadata injection. If set to 1, the audio input will be segmented into 100ms output frames, each of them containing various loudness information in metadata. All the metadata keys are prefixed with lavfi.r128..

Default is 0.

framelog — force the frame logging level.

Available values are:

quiet — logging disabled.
info — information logging level.
verbose — verbose logging level.

By default, the logging level is set to info. If the video or the metadata options are set, it switches to verbose.

peak — set peak mode(s).

Available modes can be cumulated (the option is a flag type). Possible values are:

none — disable any peak mode (default).
sample — enable sample-peak mode. Simple peak mode looking for the higher sample value. It logs a message for sample-peak (identified by SPK).
true — enable true-peak mode. If enabled, the peak lookup is done on an over-sampled version of the input stream for better peak accuracy. It logs a message for true-peak (identified by TPK) and true-peak per frame (identified by FTPK). This mode requires a build with libswresample.

dualmono — treat mono input files as “dual mono”. If a mono file is intended for playback on a stereo system, its EBU R128 measurement will be perceptually incorrect. If set to true, this option will compensate for this effect. Multi-channel input files are not affected by this option.
panlaw — set a specific pan law to be used for the measurement of dual mono files. This parameter is optional, and has a default value of -3.01dB.
target — set a specific target level (in LUFS) used as relative zero in the visualization. This parameter is optional and has a default value of -23LUFS as specified by EBU R128. However, material published online may prefer a level of -16LUFS (e.g. for use with podcasts or video platforms).
gauge — set the value displayed by the gauge. Valid values are momentary and shortterm. By default the momentary value will be used, but in certain scenarios it may be more useful to observe the short term value instead (e.g. live mixing).
scale — sets the display scale for the loudness. Valid parameters are absolute (in LUFS) or relative (LU) relative to the target. This only affects the video output, not the summary or continuous log output.
integrated — read-only exported value for measured integrated loudness, in LUFS.
range — read-only exported value for measured loudness range, in LU.
lra_low — read-only exported value for measured LRA low, in LUFS.
lra_high — read-only exported value for measured LRA high, in LUFS.
sample_peak — read-only exported value for measured sample peak, in dBFS.
true_peak — read-only exported value for measured true peak, in dBFS.

🎬 Examples

Real-time graph using ffplay, with a EBU scale meter +18:

ffplay -f lavfi -i "amovie=input.mp3,ebur128=video=1:meter=18 [out0][out1]"

Run an analysis with ffmpeg:

ffmpeg -nostats -i input.mp3 -filter_complex ebur128 -f null -

🦋 interleave, ainterleave

Temporally interleave frames from several inputs. It’s giving “everyone gets a turn” energy, and honestly? Obsessed. 💖

interleave works with video inputs, ainterleave with audio.

These filters read frames from several inputs and send the oldest queued frame to the output.

Input streams must have well defined, monotonically increasing frame timestamp values.

In order to submit one frame to output, these filters need to enqueue at least one frame for each input, so they cannot work in case one input is not yet terminated and will not receive incoming frames.

For example consider the case when one input is a select filter which always drops input frames. The interleave filter will keep reading from that input, but it will never be able to send new frames to output until the input sends an end-of-stream signal.

Also, depending on inputs synchronization, the filters will drop frames in case one input receives more frames than the other ones, and the queue is already filled.

These filters accept the following options:

nb_inputs, n — set the number of different inputs, it is 2 by default.
duration — how to determine the end-of-stream.
- longest — the duration of the longest input. (default)
- shortest — the duration of the shortest input.
first — how long the first input runs for.

✨ Examples ✨

Interleave frames from different streams using ffmpeg — bestie, this is the whole vibe:

ffmpeg -i bambi.avi -i pr0n.mkv -filter_complex "[0:v][1:v] interleave" out.avi

Add a flickering blur effect (yes it's chaotic, yes it eats):

select='if(gt(random(0), 0.2), 1, 2)':n=2 [tmp], boxblur=2:2, [tmp] interleave

💅 `latency`, `alatency`

Measure how long filtering is taking. Basically a stopwatch for your filtergraph 🫶

It reports the previous filter's filtering latency — that's a delay measured in number of audio samples for audio filters, or number of video frames for video filters.

When the input stream ends, the filter reports the min and max measured latency for the previous running filter in the filtergraph. Receipts, periodt.

🎀 `metadata`, `ametadata`

Manipulate frame metadata. Little key-value gremlins living inside your frames 💖

This filter accepts the following options:

mode — sets the mode of operation of the filter.

It can be one of the following:

select — if both value and key are set, it selects frames that have such metadata. If only key is set, it selects every frame that has such key in metadata.
add — add new metadata key and value. If the key is already available, do nothing.
modify — modify the value of an already present key.
delete — if value is set, delete only keys that have such value. Otherwise, delete the key. If key is not set, delete all metadata values in the frame. Gone, deleted, blocked 💀
print — print the key and its value if metadata was found. If key is not set, print all metadata values available in the frame.

key — set the key used with all modes. Mandatory for all modes except print and delete.
value — set the metadata value that will be used. This option is mandatory for modify and add mode.
function — which function to use when comparing metadata value and value.

Can be one of the following:

same_str — values are interpreted as strings, returns true if the metadata value is the same as value.
starts_with — values are interpreted as strings, returns true if the metadata value starts with the value option string.
less — values are interpreted as floats, returns true if the metadata value is less than value.
equal — values are interpreted as floats, returns true if value is equal to the metadata value.
greater — values are interpreted as floats, returns true if the metadata value is greater than value.
expr — values are interpreted as floats, returns true if the expression from option expr evaluates to true.
ends_with — values are interpreted as strings, returns true if the metadata value ends with the value option string.

expr — set the expression used when function is set to expr. The expression is evaluated through the eval API and can contain the following constants:
- VALUE1, FRAMEVAL — float representation of value from the metadata key.
- VALUE2, USERVAL — float representation of value as supplied by the user in the value option.
file — if specified in print mode, output is written to the named file. Instead of a plain filename, any writable url can be specified. The filename - is a shorthand for standard output. If the file option is not set, output is written to the log with the AV_LOG_INFO loglevel.
direct — reduces buffering in print mode when output is written to a URL set using file.

✨ Examples ✨

Print all metadata values for frames with key lavfi.signalstats.YDIF with values between 0 and 1:

signalstats,metadata=print:key=lavfi.signalstats.YDIF:value=0:function=expr:expr='between(VALUE1,0,1)'

Print silencedetect output to the file metadata.txt:

silencedetect,ametadata=mode=print:file=metadata.txt

Direct all metadata to a pipe with file descriptor 4:

metadata=mode=print:file='pipe\:4'

🌸 `perms`, `aperms`

Set read/write permissions for the output frames. Lowkey playing god with your buffers 😭

These filters are mainly aimed at developers to test the direct path in the following filter in the filtergraph.

The filters accept the following options:

mode — select the permissions mode.

It accepts the following values:

none — do nothing. This is the default.
ro — set all the output frames read-only.
rw — set all the output frames directly writable.
toggle — make the frame read-only if writable, and writable if read-only.
random — set each output frame read-only or writable randomly.

seed — set the seed for the random mode; must be an integer included between 0 and UINT32_MAX. If not specified, or if explicitly set to -1, the filter will try to use a good random seed on a best effort basis.

Note: in case of an auto-inserted filter between the permission filter and the following one, the permission might not be received as expected in that following filter. Inserting a format or aformat filter before the perms/aperms filter can avoid this problem. Fix it bestie 🫶

🦋 `realtime`, `arealtime`

Slow down filtering to match real time approximately. It's giving "chill, we have time" energy ✨

These filters will pause the filtering for a variable amount of time to match the output rate with the input timestamps. They are similar to the re option to ffmpeg.

They accept the following options:

limit — time limit for the pauses. Any pause longer than that will be considered a timestamp discontinuity and reset the timer. Default is 2 seconds.
speed — speed factor for processing. The value must be a float larger than zero. Values larger than 1.0 will result in faster than realtime processing, smaller will slow processing down. The limit is automatically adapted accordingly. Default is 1.0.

A processing speed faster than what is possible without these filters cannot be achieved. No cap, but also, cap.

✨ Commands ✨

Both filters support all the above options as commands.

🍓 `segment`, `asegment`

Split a single input stream into multiple streams. The opposite of the concat filters, fr fr.

segment works on video frames, asegment on audio samples.

This filter accepts the following options:

timestamps — timestamps of output segments separated by |. The first segment will run from the beginning of the input stream. The last segment will run until the end of the input stream.
frames, samples — exact frame/sample count to split the segments.

In all cases, prefixing each segment with + will make it relative to the previous segment.

✨ Examples ✨

Split the input audio stream into three output audio streams — starting at the start of the input audio stream and storing that in the 1st output audio stream, then following at the 60th second and storing that in the 2nd output audio stream, and last after the 150th second of the input audio stream stored in the 3rd output audio stream:

asegment=timestamps="60|150"

💜 `select`, `aselect`

Select frames to pass in the output. The bouncer of your filtergraph 👑

This filter accepts the following options:

expr, e — set the expression, which is evaluated for each input frame.

If the expression is evaluated to zero, the frame is discarded.

If the evaluation result is negative or NaN, the frame is sent to the first output; otherwise it is sent to the output with index ceil(val)-1, assuming that the input index starts from 0.

For example a value of 1.2 corresponds to the output with index ceil(1.2)-1 = 2-1 = 1, that is the second output.

outputs, n — set the number of outputs. The output to which to send the selected frame is based on the result of the evaluation. Default value is 1.

The expression can contain the following constants:

n — the (sequential) number of the filtered frame, starting from 0.
selected_n — the (sequential) number of the selected frame, starting from 0.
prev_selected_n — the sequential number of the last selected frame. It's NAN if undefined.
TB — the timebase of the input timestamps.
pts — the PTS (Presentation TimeStamp) of the filtered frame, expressed in TB units. It's NAN if undefined.
t — the PTS of the filtered frame, expressed in seconds. It's NAN if undefined.
prev_pts — the PTS of the previously filtered frame. It's NAN if undefined.
prev_selected_pts — the PTS of the last previously filtered frame. It's NAN if undefined.
prev_selected_t — the PTS of the last previously selected frame, expressed in seconds. It's NAN if undefined.
start_pts — the first PTS in the stream which is not NAN. It remains NAN if not found.
start_t — the first PTS, in seconds, in the stream which is not NAN. It remains NAN if not found.
pict_type (video only) — the type of the filtered frame. It can assume one of the following values:
- I
- P
- B
- S
- SI
- SP
- BI
interlace_type (video only) — the frame interlace type. It can assume one of the following values:
- PROGRESSIVE — the frame is progressive (not interlaced).
- TOPFIRST — the frame is top-field-first.
- BOTTOMFIRST — the frame is bottom-field-first.
consumed_sample_n (audio only) — the number of selected samples before the current frame.
samples_n (audio only) — the number of samples in the current frame.
sample_rate (audio only) — the input sample rate.
key — this is 1 if the filtered frame is a key-frame, 0 otherwise.
pos — the position in the file of the filtered frame, -1 if the information is not available (e.g. for synthetic video); deprecated, do not use 🚩
scene (video only) — value between 0 and 1 to indicate a new scene; a low value reflects a low probability for the current frame to introduce a new scene, while a higher value means the current frame is more likely to be one (see the example below).
concatdec_select — the concat demuxer can select only part of a concat input file by setting an inpoint and an outpoint, but the output packets may not be entirely contained in the selected interval. By using this variable, it is possible to skip frames generated by the concat demuxer which are not exactly contained in the selected interval.

This works by comparing the frame pts against the lavf.concat.start_time and the lavf.concat.duration packet metadata values which are also present in the decoded frames.

concatdec_select is -1 when the frame pts is at least start_time AND either the duration metadata is missing or the frame pts is less than start_time + duration. It's 0 otherwise, and NaN if the start_time metadata is missing. 🫶

Which honestly just means: a frame gets selected if its pts lands inside the interval the concat demuxer set up. That's it, that's the whole vibe check.

iw (video only) — the width of the input video frame, bestie. 📐
ih (video only) — the height of the input video frame. 📏
view (video only) — the View ID, for multi-view video. 👀

The default value of the select expression is literally "1" — so with no expression at all you're grabbing everything. 🎀

Examples

Select all the frames in the input:

select

The example above is exactly the same as:

select=1

Skip all frames (bold move, but okay):

select=0

Select only I-frames:

select='eq(pict_type\,I)'

Select one frame every 100:

select='not(mod(n\,100))'

Select only the frames inside the 10-20 time interval:

select=between(t\,10\,20)

Select only the I-frames inside the 10-20 time interval:

select=between(t\,10\,20)*eq(pict_type\,I)

Select frames with a minimum distance of 10 seconds between them:

select='isnan(prev_selected_t)+gte(t-prev_selected_t\,10)'

Use aselect to grab only audio frames with a samples_n greater than 100:

aselect='gt(samples_n\,100)'

Create a mosaic of the first scenes, iconic behaviour honestly:

ffmpeg -i video.avi -vf select='gt(scene\,0.4)',scale=160:120,tile -frames:v 1 preview.png

Comparing scene against a value somewhere between 0.3 and 0.5 is generally a sane choice. ✨

Send the even and odd frames to separate outputs, then compose them back together:

select=n=2:e='mod(n, 2)+1' [odd][even]; [odd] pad=h=2*ih [tmp]; [tmp][even] overlay=y=h

Select the useful frames out of an ffconcat file that's using inpoints and outpoints, but where the source files are not intra-frame only:

ffmpeg -copyts -vsync 0 -segment_time_metadata 1 -i input.ffconcat -vf select=concatdec_select -af aselect=concatdec_select output.avi

sendcmd, asendcmd

Okay so these two literally just send commands to the other filters chilling in the filtergraph. 📨

They read commands and pass them along to other filters in the filtergraph. sendcmd has to be wedged between two video filters, and asendcmd has to be wedged between two audio filters — but other than that, they act exactly the same. Twins. 💅

You can hand over the commands in the filter arguments via the commands option, or in a file named by the filename option.

These filters accept the following options:

commands, c — sets the commands that get read and sent to the other filters.
filename, f — sets the filename of the commands that get read and sent to the other filters.

Commands syntax

A commands description is basically a sequence of interval specs, each one holding a list of commands to fire when a particular event related to that interval goes down. The event is typically the current frame time entering or leaving a given time interval. 🕰️

An interval is specified with this syntax:

<START>[-<END>] <COMMANDS>;

The time interval is defined by the START and END times. END is optional and defaults to the maximum time.

The current frame time counts as inside the specified interval if it's included in the interval [START, END) — meaning the time is greater or equal to START and less than END.

COMMANDS is a sequence of one or more command specs separated by ",", all relating to that interval. The syntax of a single command spec is:

[<FLAGS>] <TARGET> <COMMAND> <ARG>

FLAGS is optional and specifies which type of events relating to the time interval actually enable sending that command. It has to be a non-null sequence of identifier flags separated by "+" or "|" and wrapped in "[" and "]".

The following flags are recognized: 🚩

enter — the command is sent when the current frame timestamp enters the specified interval. In other words: the previous frame timestamp was not in the interval and the current one is.
leave — the command is sent when the current frame timestamp leaves the specified interval. In other words: the previous frame timestamp was in the interval and the current one is not.
expr — the command ARG is interpreted as an expression, and the result of that expression gets passed as ARG.

The expression is evaluated through the eval API and can contain these constants: 🧮

POS — original position in the file of the frame, or undefined if it's undefined for the current frame. Deprecated, do not use. 🚫
PTS — the presentation timestamp in input.
N — the count of the input frame for video or audio, starting from 0.
T — the time in seconds of the current frame.
TS — the start time in seconds of the current command interval.
TE — the end time in seconds of the current command interval.
TI — the interpolated time of the current command interval, TI = (T - TS) / (TE - TS).
W — the video frame width.
H — the video frame height.

If FLAGS isn't specified, a default value of "[enter]" is assumed.

TARGET specifies the target of the command — usually the name of the filter class or a specific filter instance name.

COMMAND specifies the name of the command for the target filter.

ARG is optional and specifies the optional list of arguments for the given COMMAND.

Between one interval spec and the next, whitespace — or sequences of characters starting with "#" all the way to the end of the line — is ignored, and you can absolutely use it to annotate comments. 📝

A simplified BNF description of the commands specification syntax, for the girlies who love grammar:

<COMMAND_FLAG>  ::= "enter" | "leave"
<COMMAND_FLAGS> ::= <COMMAND_FLAG> [(+|"|")<COMMAND_FLAG>]
<COMMAND>       ::= ["[" <COMMAND_FLAGS> "]"] <TARGET> <COMMAND> [<ARG>]
<COMMANDS>      ::= <COMMAND> [,<COMMANDS>]
<INTERVAL>      ::= <START>[-<END>] <COMMANDS>
<INTERVALS>     ::= <INTERVAL>[;<INTERVALS>]

Examples

Specify an audio tempo change at second 4:

asendcmd=c='4.0 atempo tempo 1.5',atempo

Target a specific filter instance:

asendcmd=c='4.0 atempo@my tempo 1.5',atempo@my

Specify a list of drawtext and hue commands in a file.

# show text in the interval 5-10
5.0-10.0 [enter] drawtext reinit 'fontfile=FreeSerif.ttf:text=hello world',
         [leave] drawtext reinit 'fontfile=FreeSerif.ttf:text=';

# desaturate the image in the interval 15-20
15.0-20.0 [enter] hue s 0,
          [enter] drawtext reinit 'fontfile=FreeSerif.ttf:text=nocolor',
          [leave] hue s 1,
          [leave] drawtext reinit 'fontfile=FreeSerif.ttf:text=color';

# apply an exponential saturation fade-out effect, starting from time 25
25 [enter] hue s exp(25-t)

A filtergraph that lets you read and process the above command list stored in a file test.cmd can be specified like this:

sendcmd=f=test.cmd,drawtext=fontfile=FreeSerif.ttf:text='',hue

setpts, asetpts

Change the PTS (presentation timestamp) of the input frames. ⏱️

setpts does its thing on video frames, asetpts does it on audio frames.

This filter accepts the following options:

expr — the expression that gets evaluated for each frame to construct its timestamp.
strip_fps (video only) — boolean option that decides whether the original framerate and frame duration metadata gets unset. If you set it to true, be advised that you really should explicitly specify a sane frame rate if the output is being sent to a constant frame rate muxer. Default is "false". ⚠️

The expression is evaluated through the eval API and can contain these constants:

FRAME_RATE, FR — frame rate, only defined for constant frame-rate video.
PTS — the presentation timestamp in input.
N — the count of the input frame for video, or the number of consumed samples not including the current frame for audio, starting from 0.
NB_CONSUMED_SAMPLES — the number of consumed samples, not including the current frame (only audio).
NB_SAMPLES, S — the number of samples in the current frame (only audio).
SAMPLE_RATE, SR — the audio sample rate.
STARTPTS — the PTS of the first frame.
STARTT — the time in seconds of the first frame.
INTERLACED — state whether the current frame is interlaced.
T — the time in seconds of the current frame.
POS — original position in the file of the frame, or undefined if it's undefined for the current frame; deprecated, do not use. 🚫
PREV_INPTS — the previous input PTS.
PREV_INT — previous input time in seconds.
PREV_OUTPTS — the previous output PTS.
PREV_OUTT — previous output time in seconds.
RTCTIME — the wallclock (RTC) time in microseconds. This is deprecated, use time(0) instead.
RTCSTART — the wallclock (RTC) time at the start of the movie, in microseconds.
TB — the timebase of the input timestamps.
T_CHANGE — time of the first frame after the command was applied, or the time of the first frame if there were no commands.

Examples

Start counting PTS from zero:

setpts=PTS-STARTPTS

Apply a fast motion effect:

setpts=0.5*PTS

Apply a slow motion effect:

setpts=2.0*PTS

Set a fixed rate of 25 frames per second:

setpts=N/(25*TB)

Apply a random jitter effect of +/-100 TB units:

setpts=PTS+randomi(0, -100\,100)

Set a fixed rate of 25 fps with some jitter:

setpts='1/(25*TB) * (N + 0.05 * sin(N*2*PI/25))'

Apply an offset of 10 seconds to the input PTS:

setpts=PTS+10/TB

Generate timestamps from a "live source" and rebase them onto the current timebase:

setpts='(RTCTIME - RTCSTART) / (TB * 1000000)'

Generate timestamps by counting samples:

asetpts=N/SR/TB

Commands

Both filters support all of the options above as commands. 💅

setrange

Force the color range of the output video frame. 🌈

The setrange filter just marks the color range property on the output frames. It does not change the input frame at all — it only sets the corresponding property, which then affects how the frame gets treated by the filters that come after it.

The filter accepts the following options:

range — sets the color range property, and the available values are:
- auto — keep the same color range property.
- unspecified, unknown — set the color range as unspecified.
- limited, tv, mpeg — set the color range as limited.
- full, pc, jpeg — set the color range as full.

settb, asettb

Timebase queen behavior — this sets the timebase used for output frame timestamps. It’s mainly useful for testing timebase configuration, iykyk.

It accepts these parameters:

expr, tb — the expression that gets evaluated into the output timebase.

The value for tb is an arithmetic expression representing a rational. The expression can contain the constants AVTB (the default timebase), intb (the input timebase), and sr (the sample rate, audio only). Default value is "intb".

Examples

Set the timebase to 1/25:

settb=expr=1/25

Set the timebase to 1/10:

settb=expr=0.1

Set the timebase to 1001/1000:

settb=1+0.001

Set the timebase to 2*intb:

settb=2*intb

Set the default timebase value:

settb=AVTB

showcqt

Bestie, it’s giving spectrum visualizer — this converts input audio into a video output that represents the frequency spectrum logarithmically using the Brown-Puckette constant Q transform algorithm with direct frequency domain coefficient calculation (but the transform itself is not really constant Q, instead the Q factor is actually variable/clamped), with musical tone scale, from E0 to D#10. 💖

The filter accepts the following options:

size, s — specify the video size for the output. It must be even. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Default value is "1920x1080".
fps, rate, r — set the output frame rate. Default value is 25.
bar_h — set the bargraph height. It must be even. Default value is -1, which computes the bargraph height automatically.
axis_h — set the axis height. It must be even. Default value is -1, which computes the axis height automatically.
sono_h — set the sonogram height. It must be even. Default value is -1, which computes the sonogram height automatically.
fullhd — set the fullhd resolution. This option is deprecated, use size, s instead. Default value is 1.
sono_v, volume — specify the sonogram volume expression. It can contain variables:
- bar_v — the bar_v evaluated expression.
- frequency, freq, f — the frequency where it is evaluated.
- timeclamp, tc — the value of the timeclamp option.

And functions:

a_weighting(f) — A-weighting of equal loudness.
b_weighting(f) — B-weighting of equal loudness.
c_weighting(f) — C-weighting of equal loudness.

Default value is 16.

bar_v, volume2 — specify the bargraph volume expression. It can contain variables:
- sono_v — the sono_v evaluated expression.
- frequency, freq, f — the frequency where it is evaluated.
- timeclamp, tc — the value of the timeclamp option.

And functions:

a_weighting(f) — A-weighting of equal loudness.
b_weighting(f) — B-weighting of equal loudness.
c_weighting(f) — C-weighting of equal loudness.

Default value is "sono_v".

sono_g, gamma — specify the sonogram gamma. Lower gamma makes the spectrum more contrast, higher gamma makes the spectrum have more range. Default value is 3. Acceptable range is [1, 7].
bar_g, gamma2 — specify the bargraph gamma. Default value is 1. Acceptable range is [1, 7].
bar_t — specify the bargraph transparency level. Lower value makes the bargraph sharper. Default value is 1. Acceptable range is [0, 1].
timeclamp, tc — specify the transform timeclamp. At low frequency, there is a trade-off between accuracy in time domain and frequency domain. If timeclamp is lower, an event in time domain is represented more accurately (such as fast bass drum), otherwise an event in frequency domain is represented more accurately (such as bass guitar). Acceptable range is [0.002, 1]. Default value is 0.17.
attack — set attack time in seconds. The default is 0 (disabled). Otherwise, it limits future samples by applying asymmetric windowing in time domain, useful when low latency is required. Accepted range is [0, 1].
basefreq — specify the transform base frequency. Default value is 20.01523126408007475, which is frequency 50 cents below E0. Acceptable range is [10, 100000].
endfreq — specify the transform end frequency. Default value is 20495.59681441799654, which is frequency 50 cents above D#10. Acceptable range is [10, 100000].
coeffclamp — this option is deprecated and ignored.
tlength — specify the transform length in time domain. Use this option to control the accuracy trade-off between time domain and frequency domain at every frequency sample. It can contain variables:
- frequency, freq, f — the frequency where it is evaluated.
- timeclamp, tc — the value of the timeclamp option.

Default value is "384*tc/(384+tc*f)".

count — specify the transform count for every video frame. Default value is 6. Acceptable range is [1, 30].
fcount — specify the transform count for every single pixel. Default value is 0, which makes it computed automatically. Acceptable range is [0, 10].
fontfile — specify the font file for use with freetype to draw the axis. If not specified, use embedded font. Note that drawing with font file or embedded font is not implemented with custom basefreq and endfreq, use axisfile option instead.
font — specify fontconfig pattern. This has lower priority than fontfile. The ":" in the pattern may be replaced by "|" to avoid unnecessary escaping.
fontcolor — specify font color expression. This is an arithmetic expression that should return integer value 0xRRGGBB. It can contain variables:
- frequency, freq, f — the frequency where it is evaluated.
- timeclamp, tc — the value of the timeclamp option.

And functions:

midi(f) — midi number of frequency f, some midi numbers: E0(16), C1(24), C2(36), A4(69).
r(x), g(x), b(x) — red, green, and blue value of intensity x.

Default value is "st(0, (midi(f)-59.5)/12); st(1, if(between(ld(0),0,1), 0.5-0.5*cos(2*PI*ld(0)), 0)); r(1-ld(1)) + b(ld(1))".

axisfile — specify image file to draw the axis. This option overrides the fontfile and fontcolor option.
axis, text — enable/disable drawing text to the axis. If it is set to 0, drawing to the axis is disabled, ignoring the fontfile and axisfile option. Default value is 1.
csp — set colorspace. The accepted values are:
- unspecified — Unspecified (default).
- bt709 — BT.709.
- fcc — FCC.
- bt470bg — BT.470BG or BT.601-6 625.
- smpte170m — SMPTE-170M or BT.601-6 525.
- smpte240m — SMPTE-240M.
- bt2020ncl — BT.2020 with non-constant luminance.
cscheme — set spectrogram color scheme. This is a list of floating point values with format "left_r|left_g|left_b|right_r|right_g|right_b". The default is "1|0.5|0|0|0.5|1".

Examples

Playing audio while showing the spectrum:

ffplay -f lavfi 'amovie=a.mp3, asplit [a][out1]; [a] showcqt [out0]'

Same as above, but with frame rate 30 fps:

ffplay -f lavfi 'amovie=a.mp3, asplit [a][out1]; [a] showcqt=fps=30:count=5 [out0]'

Playing at 1280x720:

ffplay -f lavfi 'amovie=a.mp3, asplit [a][out1]; [a] showcqt=s=1280x720:count=4 [out0]'

Disable sonogram display:

sono_h=0

A1 and its harmonics: A1, A2, (near)E3, A3:

ffplay -f lavfi 'aevalsrc=0.1*sin(2*PI*55*t)+0.1*sin(4*PI*55*t)+0.1*sin(6*PI*55*t)+0.1*sin(8*PI*55*t),
                                    asplit[a][out1]; [a] showcqt [out0]'

Same as above, but with more accuracy in frequency domain:

ffplay -f lavfi 'aevalsrc=0.1*sin(2*PI*55*t)+0.1*sin(4*PI*55*t)+0.1*sin(6*PI*55*t)+0.1*sin(8*PI*55*t),
                                    asplit[a][out1]; [a] showcqt=timeclamp=0.5 [out0]'

Custom volume:

bar_v=10:sono_v=bar_v*a_weighting(f)

Custom gamma, now spectrum is linear to the amplitude:

bar_g=2:sono_g=2

Custom tlength equation:

tc=0.33:tlength='st(0,0.17); 384*tc / (384 / ld(0) + tc*f /(1-ld(0))) + 384*tc / (tc*f / ld(0) + 384 /(1-ld(0)))'

Custom fontcolor and fontfile, C-note is colored green, others are colored blue:

fontcolor='if(mod(floor(midi(f)+0.5),12), 0x0000FF, g(1))':fontfile=myfont.ttf

Custom font using fontconfig:

font='Courier New,Monospace,mono|bold'

Custom frequency range with custom axis using image file:

axisfile=myaxis.png:basefreq=40:endfreq=10000

showcwt

Wavelet era — this converts input audio to video output representing the frequency spectrum using Continuous Wavelet Transform and Morlet wavelet. 🦋

The filter accepts the following options:

size, s — specify the video size for the output. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Default value is "640x512".
rate, r — set the output frame rate. Default value is 25.
scale — set the frequency scale used. Allowed values are:
- linear
- log
- bark
- mel
- erbs
- sqrt
- cbrt
- qdrt
- fm

Default value is "linear".

iscale — set the intensity scale used. Allowed values are:
- linear
- log
- sqrt
- cbrt
- qdrt

Default value is "log".

min — set the minimum frequency that will be used in output. Default is 20 Hz.
max — set the maximum frequency that will be used in output. Default is 20000 Hz. The real frequency upper limit depends on input audio’s sample rate, and such will be enforced on this value when it is set to a value greater than Nyquist frequency.
imin — set the minimum intensity that will be used in output.
imax — set the maximum intensity that will be used in output.
logb — sets the logarithmic basis for brightness strength when those calculated magnitude values get mapped to pixel values ✨ allowed range is 0 to 1, and the default is 0.0001 (yes, tiny, that's the point).
deviation — sets the frequency deviation 🌀 lower values than 1 are more frequency oriented, while higher values than 1 are more time oriented. Allowed range is 0 to 10, default 1.
pps — sets the number of pixel output per each second in one row 🎀 allowed range is 1 to 1024, default 64.
mode — sets the output visual mode, bestie. allowed values are:
- magnitude — show magnitude.
- phase — show only phase.
- magphase — show combination of magnitude and phase. Magnitude is mapped to brightness and phase to color. (iconic combo tbh 💅)
- channel — show unique color per channel magnitude.
- stereo — show unique color per stereo difference.

Default value is "magnitude".

slide — sets the output slide method. allowed values are:
- replace
- scroll
- frame
direction — sets the direction method for the output slide method. allowed values are:
- lr — direction from left to right.
- rl — direction from right to left.
- ud — direction from up to down.
- du — direction from down to up.
bar — sets the ratio of bargraph display to display size. Default is 0.
rotation — sets color rotation, must be in [-1.0, 1.0] range. Default value is 0.

showfreqs</p> 💖

Converts input audio into a video output representing the audio power spectrum. Audio amplitude is on the Y-axis while frequency is on the X-axis — like a little graph of your sound's whole personality 📊✨

The filter accepts the following options:

size, s — specify size of video. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Default is "1024x512".
rate, r — set video rate. Default is 25.
mode — set display mode. This sets how each frequency bin will be represented. It accepts the following values:
- line
- bar
- dot

Default is "bar".

ascale — set amplitude scale. It accepts the following values:
- lin — linear scale.
- sqrt — square root scale.
- cbrt — cubic root scale.
- log — logarithmic scale.

Default is "log".

fscale — set frequency scale. It accepts the following values:
- lin — linear scale.
- log — logarithmic scale.
- rlog — reverse logarithmic scale.

Default is "lin".

win_size — set window size. Allowed range is from 16 to 65536. Default is 2048.
win_func — set windowing function. It accepts the following values:

rect, bartlett, hanning, hamming, blackman, welch, flattop, bharris, bnuttall, bhann, sine, nuttall, lanczos, gauss, tukey, dolph, cauchy, parzen, poisson, bohman, kaiser

Default is "hanning".

overlap — set window overlap. In range [0, 1]. Default is 1, which means optimal overlap for the selected window function will be picked 🎯
averaging — set time averaging. Setting this to 0 will display current maximal peaks. Default is 1, which means time averaging is disabled.
colors — specify list of colors separated by space or by '|' which will be used to draw channel frequencies. Unrecognized or missing colors will be replaced by white color. (rude but ok 🙈)
cmode — set channel display mode. It accepts the following values:
- combined
- separate

Default is "combined".

minamp — set minimum amplitude used in "log" amplitude scaler.
data — set data display mode. It accepts the following values:
- magnitude
- phase
- delay

Default is "magnitude".

channels — set channels to use when processing audio. By default all are processed.

showspatial 🦋

Converts stereo input audio into a video output representing the spatial relationship between two channels 🌈✨ it's giving "I can literally see the stereo field" energy.

The filter accepts the following options:

size, s — specify the video size for the output. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Default value is "512x512".
win_size — set window size. Allowed range is from 1024 to 65536. Default size is 4096.
win_func — set window function. It accepts the following values:

rect, bartlett, hann, hanning, hamming, blackman, welch, flattop, bharris, bnuttall, bhann, sine, nuttall, lanczos, gauss, tukey, dolph, cauchy, parzen, poisson, bohman, kaiser

Default value is "hann".

rate, r — set output framerate.

showspectrum 🌸

Converts input audio into a video output representing the audio frequency spectrum. This is the one everybody screenshots for thumbnails, ngl 💅✨

The filter accepts the following options:

size, s — specify the video size for the output. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Default value is "640x512".
slide — specify how the spectrum should slide along the window. It accepts the following values:
- replace — the samples start again on the left when they reach the right.
- scroll — the samples scroll from right to left.
- fullframe — frames are only produced when the samples reach the right.
- rscroll — the samples scroll from left to right.
- lreplace — the samples start again on the right when they reach the left.

Default value is "replace".

mode — specify display mode. It accepts the following values:
- combined — all channels are displayed in the same row.
- separate — all channels are displayed in separate rows.

Default value is combined.

color — specify display color mode. It accepts the following values:
- channel — each channel is displayed in a separate color.
- intensity — each channel is displayed using the same color scheme.
- rainbow — each channel is displayed using the rainbow color scheme.
- moreland — each channel is displayed using the moreland color scheme.
- nebulae — each channel is displayed using the nebulae color scheme.
- fire — each channel is displayed using the fire color scheme.
- fiery — each channel is displayed using the fiery color scheme.
- fruit — each channel is displayed using the fruit color scheme.
- cool — each channel is displayed using the cool color scheme.
- magma — each channel is displayed using the magma color scheme.
- green — each channel is displayed using the green color scheme.
- viridis — each channel is displayed using the viridis color scheme.
- plasma — each channel is displayed using the plasma color scheme.
- cividis — each channel is displayed using the cividis color scheme.
- terrain — each channel is displayed using the terrain color scheme.

Default value is channel.

scale — specify scale used for calculating intensity color values. It accepts the following values:
- lin — linear
- sqrt — square root, default
- cbrt — cubic root
- log — logarithmic
- 4thrt — 4th root
- 5thrt — 5th root

Default value is sqrt.

fscale — specify frequency scale. It accepts the following values:
- lin — linear
- log — logarithmic

Default value is lin.

saturation — set saturation modifier for displayed colors. Negative values provide alternative color scheme. 0 is no saturation at all. Saturation must be in [-10.0, 10.0] range. Default value is 1.
win_func — set window function. It accepts the following values:

Default value is "hann".

orientation — set orientation of time vs frequency axis. Can be "vertical" or "horizontal". Default is "vertical".
overlap — set ratio of overlap window. Default value is 0. When value is 1 overlap is set to recommended size for specific window function currently used.
gain — set scale gain for calculating intensity color values. Default value is 1.
data — set which data to display. Can be "magnitude", default, or "phase", or unwrapped phase: "uphase".
rotation — set color rotation, must be in [-1.0, 1.0] range. Default value is 0.
start — set start frequency from which to display spectrogram. Default is 0.
stop — set stop frequency to which to display spectrogram. Default is 0.
fps — set upper frame rate limit. Default is "auto", unlimited 🏃‍♀️💨
legend — draw time and frequency axes and legends. Default is disabled.
drange — set dynamic range used to calculate intensity color values. Default is 120 dBFS. Allowed range is from 10 to 200.
limit — set upper limit of input audio samples volume in dBFS. Default is 0 dBFS. Allowed range is from -100 to 100.
opacity — set opacity strength when using pixel format output with alpha component.

The usage is very similar to the showwaves filter; see the examples in that section 👀

Examples ✨

Large window with logarithmic color scaling:

showspectrum=s=1280x480:scale=log

Complete example for a colored and sliding spectrum per channel using ffplay:

ffplay -f lavfi 'amovie=input.mp3, asplit [a][out1];
             [a] showspectrum=mode=separate:color=intensity:slide=1:scale=cbrt [out0]'

showspectrumpic 🎀

Converts input audio into a single video frame, representing the audio frequency spectrum — one still image of the whole vibe check, no scrolling 🖼️✨

The filter accepts the following options:

size, s — the output video size, obvi 💖 For the syntax, go peek at the "Video size" section in the ffmpeg-utils manual. Default value is 4096x2048 — she's large, she's in charge.
mode — sets the display mode. It accepts these values:
- combined — all the channels are displayed in the same row
- separate — all the channels are displayed in separate rows Default value is combined.
color — sets the display color mode. It accepts:
- channel — each channel is displayed in a separate color
- intensity — each channel is displayed using the same color scheme
- rainbow — each channel is displayed using the rainbow color scheme
- moreland — each channel is displayed using the moreland color scheme
- nebulae — each channel is displayed using the nebulae color scheme
- fire — each channel is displayed using the fire color scheme 🔥
- fiery — each channel is displayed using the fiery color scheme
- fruit — each channel is displayed using the fruit color scheme 🍓
- cool — each channel is displayed using the cool color scheme
- magma — each channel is displayed using the magma color scheme
- green — each channel is displayed using the green color scheme
- viridis — each channel is displayed using the viridis color scheme
- plasma — each channel is displayed using the plasma color scheme
- cividis — each channel is displayed using the cividis color scheme
- terrain — each channel is displayed using the terrain color scheme Default value is intensity.
scale — sets the scale used for calculating intensity color values. It accepts:
- lin — linear
- sqrt — square root, default
- cbrt — cubic root
- log — logarithmic
- 4thrt — 4th root
- 5thrt — 5th root Default value is log. (yes bestie, the page says both things. I don't make the rules, I just deliver them 💅)
fscale — sets the frequency scale. It accepts:
- lin — linear
- log — logarithmic Default value is lin.
saturation — sets the saturation modifier for the displayed colors. Negative values give you an alternative color scheme, and 0 is no saturation at all. Gotta stay in the [-10.0, 10.0] range, no cap. Default value is 1.
win_func — sets the window function. It accepts:

rect
bartlett
hann
hanning
hamming
blackman
welch
flattop
bharris
bnuttall
bhann
sine
nuttall
lanczos
gauss
tukey
dolph
cauchy
parzen
poisson
bohman
kaiser

Default value is hann.

orientation — sets the orientation of the time vs frequency axis. Can be vertical or horizontal. Default is vertical.
gain — sets the scale gain for calculating intensity color values. Default value is 1.
legend — draws time and frequency axes and legends. Default is enabled. ✨
rotation — sets color rotation, must be in the [-1.0, 1.0] range. Default value is 0.
start — sets the start frequency from which to display the spectrogram. Default is 0.
stop — sets the stop frequency to which to display the spectrogram. Default is 0.
drange — sets the dynamic range used to calculate intensity color values. Default is 120 dBFS. Allowed range is from 10 to 200.
limit — sets the upper limit of input audio samples volume in dBFS. Default is 0 dBFS. Allowed range is from -100 to 100.
opacity — sets the opacity strength when you're using a pixel format output with an alpha component.

Examples

Extract an audio spectrogram of a whole audio track into a 1024x1024 picture using ffmpeg:

ffmpeg -i audio.flac -lavfi showspectrumpic=s=1024x1024 spectrogram.png

🔊 showvolume

Converts input audio volume into a video output. Literally turning vibes into visuals.

The filter accepts the following options:

rate, r — sets the video rate.
b — sets the border width, allowed range is [0, 5]. Default is 1.
w — sets the channel width, allowed range is [80, 8192]. Default is 400.
h — sets the channel height, allowed range is [1, 900]. Default is 20.
f — sets the fade, allowed range is [0, 1]. Default is 0.95.
c — sets the volume color expression. The expression can use these variables:
- VOLUME — current max volume of the channel in dB
- PEAK — current peak
- CHANNEL — current channel number, starting from 0
t — if set, displays channel names. Default is enabled.
v — if set, displays volume values. Default is enabled.
o — sets the orientation, can be horizontal: h or vertical: v, default is h.
s — sets the step size, allowed range is [0, 5]. Default is 0, which means the step is disabled.
p — sets the background opacity, allowed range is [0, 1]. Default is 0.
m — sets the metering mode, can be peak: p or rms: r, default is p.
ds — sets the display scale, can be linear: lin or log: log, default is lin.
dm — in second. If set to > 0., displays a line for the max level in the previous seconds. default is disabled: 0.
dmc — the color of the max line. Use this when the dm option is set to > 0. default is: orange 🍊

🌊 showwaves

Converts input audio to a video output, representing the sample waves. It's giving "sound, but make it visible" 🦋

The filter accepts the following options:

size, s — the video size for the output. For the syntax, check the "Video size" section in the ffmpeg-utils manual. Default value is 600x240.
mode — sets the display mode. Available values:
- point — draws a point for each sample
- line — draws a vertical line for each sample
- p2p — draws a point for each sample and a line between them
- cline — draws a centered vertical line for each sample Default value is point.
n — sets the number of samples which are printed on the same column. A larger value will decrease the frame rate. Must be a positive integer. Heads up bestie: this option can be set only if the value for rate is not explicitly specified.
rate, r — sets the (approximate) output frame rate. This is done by setting the n option. Default value is 25.
split_channels — sets whether channels should be drawn separately or overlap. Default value is 0.
colors — sets colors separated by '|' which are going to be used for drawing each channel. 🌈
scale — sets the amplitude scale. Available values:
- lin — linear
- log — logarithmic
- sqrt — square root
- cbrt — cubic root Default is linear.
draw — sets the draw mode. This is mostly useful to set for high n. Available values:
- scale — scales pixel values for each drawn sample
- full — draws every sample directly Default value is scale.

Examples

Output the input file audio and the corresponding video representation at the same time:

amovie=a.mp3,asplit[out0],showwaves[out1]

Create a synthetic signal and show it with showwaves, forcing a frame rate of 30 frames per second:

aevalsrc=sin(1*2*PI*t)*sin(880*2*PI*t):cos(2*PI*200*t),asplit[out0],showwaves=r=30[out1]

🖼 showwavespic

Converts input audio into a single video frame, representing the sample waves. One frame, one iconic lil snapshot 💖

The filter accepts the following options:

size, s — the video size for the output. For the syntax, check the "Video size" section in the ffmpeg-utils manual. Default value is 600x240.
split_channels — sets whether channels should be drawn separately or overlap. Default value is 0.
colors — sets colors separated by '|' which are going to be used for drawing each channel.
scale — sets the amplitude scale. Available values:
- lin — linear
- log — logarithmic
- sqrt — square root
- cbrt — cubic root Default is linear.
draw — sets the draw mode. Available values:
- scale — scales pixel values for each drawn sample
- full — draws every sample directly Default value is scale.
filter — sets the filter mode. Available values:
- average — uses average sample values for each drawn sample
- peak — uses peak sample values for each drawn sample Default value is average.

Examples

Extract a channel split representation of the waveform of a whole audio track into a 1024x800 picture using ffmpeg:

ffmpeg -i audio.flac -lavfi showwavespic=split_channels=1:s=1024x800 waveform.png

🗑 sidedata, asidedata

Deletes frame side data, or selects frames based on it. It's the "block and/or pick" filter of the friend group.

This filter accepts the following options:

mode — sets the mode of operation of the filter. Can be one of the following:
- select — selects every frame with side data of type
- delete — deletes side data of type. If type is not set, deletes all side data in the frame
type — sets the side data type used with all modes. Must be set for select mode. Possible values are:

PANSCAN
A53_CC
STEREO3D
MATRIXENCODING
DOWNMIX_INFO
REPLAYGAIN
DISPLAYMATRIX
AFD
MOTION_VECTORS
SKIP_SAMPLES
AUDIO_SERVICE_TYPE
MASTERING_DISPLAY_METADATA
GOP_TIMECODE
SPHERICAL
CONTENT_LIGHT_LEVEL
ICC_PROFILE
S12M_TIMECODE
DYNAMIC_HDR_PLUS
REGIONS_OF_INTEREST
VIDEO_ENC_PARAMS
SEI_UNREGISTERED
FILM_GRAIN_PARAMS
DETECTION_BOUNDING_BOXES
DETECTION_BBOXES
DOVI_RPU_BUFFER
DOVI_METADATA
DYNAMIC_HDR_VIVID
AMBIENT_VIEWING_ENVIRONMENT
VIDEO_HINT

🧪 spectrumsynth

Synthesizes audio from 2 input video spectrums — the first input stream represents magnitude across time and the second represents phase across time. The filter will transform from the frequency domain as displayed in the videos back to the time domain as presented in the audio output. It's literally a glow-up in reverse, and it ate 🔥👑

This filter is basically the reverse-engineering queen — she's built first and foremost to undo processed showspectrum filter output, but she can also synthesize sound from other spectrograms too 🎀. Ngl though, in that case the results are going to be poor if the phase data isn't available, because then the phase data has to be recreated, and usually it's just recreated from random noise 😭. For best results use gray only output (that's "channel" color mode in the showspectrum filter) and "log" scale for magnitude video and "lin" scale for phase video, bestie. To produce phase, for the 2nd video, use the "data" option. Input videos should generally use "fullframe" slide mode because that saves resources needed for decoding video — lowkey efficient 💅.

The filter accepts the following options:

sample_rate — specify sample rate of output audio, since the sample rate of the audio the spectrum was generated from may differ ✨
channels — set the number of channels represented in the input video spectrums
scale — set the scale that was used when generating the magnitude input spectrum. Can be "lin" or "log". Default is "log" 🦋
slide — set the slide that was used when generating the input spectrums. Can be "replace", "scroll", "fullframe" or "rscroll". Default is "fullframe"
win_func — set the window function used for resynthesis
overlap — set the window overlap, in range "[0, 1]". Default is 1, which means the optimal overlap for the selected window function will be picked 👑
orientation — set the orientation of the input videos. Can be "vertical" or "horizontal". Default is "vertical"

✨ Examples ✨

First create magnitude and phase videos from audio — assuming the audio is stereo with 44100 sample rate — then resynthesize those videos back to audio with spectrumsynth 💖:

ffmpeg -i input.flac -lavfi showspectrum=mode=separate:scale=log:overlap=0.875:color=channel:slide=fullframe:data=magnitude -an -c:v rawvideo magnitude.nut
ffmpeg -i input.flac -lavfi showspectrum=mode=separate:scale=lin:overlap=0.875:color=channel:slide=fullframe:data=phase -an -c:v rawvideo phase.nut
ffmpeg -i magnitude.nut -i phase.nut -lavfi spectrumsynth=channels=2:sample_rate=44100:win_func=hann:overlap=0.875:slide=fullframe output.flac

🦋 split, asplit 🦋

Split the input into several identical outputs — she's giving "cloning machine, but make it cute" ✨.

"asplit" works with audio input, and "split" with video.

The filter accepts a single parameter which specifies the number of outputs. If unspecified, it defaults to 2 💅.

✨ Examples ✨

Create two separate outputs from the same input:

[in] split [out0][out1]

To create 3 or more outputs, you need to specify the number of outputs, like in:

[in] asplit=3 [out0][out1][out2]

Create two separate outputs from the same input, one cropped and one padded 🎀:

[in] split [splitout1][splitout2];
[splitout1] crop=100:100:0:0    [cropout];
[splitout2] pad=200:200:100:100 [padout];

Create 5 copies of the input audio with ffmpeg:

ffmpeg -i INPUT -filter_complex asplit=5 OUTPUT

💜 zmq, azmq 💜

Receive commands sent through a libzmq client, and forward them to filters in the filtergraph 🔥.

"zmq" and "azmq" work as pass-through filters. "zmq" must be inserted between two video filters, and "azmq" between two audio filters. Both are capable of sending messages to any filter type — versatile icons, honestly 👑.

To enable these filters you need to install the libzmq library and headers and configure FFmpeg with --enable-libzmq.

For more information about libzmq see: http://www.zeromq.org/

The "zmq" and "azmq" filters work as a libzmq server, which receives messages sent through a network interface defined by the bind_address (or the abbreviation "b") option. Default value of this option is tcp://localhost:5555. You may want to alter this value to your needs, but do not forget to escape any ':' signs (see filtergraph escaping) 🙈.

The received message must be in the form:

<TARGET> <COMMAND> [<ARG>]

TARGET specifies the target of the command, usually the name of the filter class or a specific filter instance name. The default filter instance name uses the pattern Parsed_<filter_name>_<index>, but you can override this by using the filter_name@id syntax (see Filtergraph syntax) ✨.

COMMAND specifies the name of the command for the target filter.

ARG is optional and specifies the optional argument list for the given COMMAND.

Upon reception, the message is processed and the corresponding command is injected into the filtergraph. Depending on the result, the filter will send a reply to the client, adopting the format:

<ERROR_CODE> <ERROR_REASON>
<MESSAGE>

MESSAGE is optional.

✨ Examples ✨

Look at tools/zmqsend for an example of a zmq client which can be used to send commands processed by these filters 🫶.

Consider the following filtergraph generated by ffplay. In this example the last overlay filter has an instance name. All other filters will have default instance names.

ffplay -dumpgraph 1 -f lavfi "
color=s=100x100:c=red  [l];
color=s=100x100:c=blue [r];
nullsrc=s=200x100, zmq [bg];
[bg][l]   overlay     [bg+l];
[bg+l][r] overlay@my=x=100 "

To change the color of the left side of the video, the following command can be used:

echo Parsed_color_0 c yellow | tools/zmqsend

To change the right side:

echo Parsed_color_1 c pink | tools/zmqsend

To change the position of the right side:

echo overlay@my x 150 | tools/zmqsend

🌈 MULTIMEDIA SOURCES 🌈

okay so below is the rundown of the multimedia sources we have rn, bestie 💖

🎀 amovie

literally the same as the movie source, except it picks an audio stream by default instead of video. that's it, that's the whole thing. she's movie's audio-first sister 💅

🦋 avsynctest

generates an Audio/Video Sync Test. yes bestie, a whole test stream, built right in ✨

the generated stream periodically shows a flash video frame and emits a beep in the audio. super useful for inspecting A/V sync issues — like, if the beep and the flash don't line up, you know something's off 😭

it accepts the following options:

size, s — sets the output video size. default value is "hd720" 🎀
framerate, fr — sets the output video frame rate. default is 30.
samplerate, sr — sets the output audio sample rate. default value is 44100 ✨
amplitude, a — sets the output audio beep amplitude. default value is 0.7.
period, p — sets the output audio beep period in seconds. default value is 3.
delay, dl — sets the output video flash delay in number of frames. default value is 0.
cycle, c — enables cycling of video delays, by default is disabled.
duration, d — sets the stream output duration. by default duration is unlimited 👑
fg, bg, ag — set foreground/background/additional color 🎨

Commands

this source supports the some above options as commands. (yes, that's the actual wording — some of the options up there double as commands, iykyk 💅)

🧚 movie

reads audio and/or video stream(s) out of a movie container. main character energy fr fr.

it accepts the following parameters:

filename — the name of the resource to read (not necessarily a file; it can also be a device or a stream accessed through some protocol). so like, it's flexible, she's versatile 🌸
format_name, f — specifies the format assumed for the movie to read, and can be either the name of a container or an input device. if not specified, the format is guessed from movie_name or by probing 🔍
seek_point, sp — specifies the seek point in seconds. the frames will be output starting from this seek point. the parameter is evaluated with av_strtod, so the numerical value may be suffixed by an IS postfix. the default value is "0" ✨
streams, s — specifies the streams to read. several streams can be specified, separated by "+". the source will then have as many outputs, in the same order. the syntax is explained in the "Stream specifiers" section in the ffmpeg manual. two special names, "dv" and "da" specify respectively the default (best suited) video and audio stream. default is "dv", or "da" if the filter is called as "amovie" 🎀
stream_index, si — specifies the index of the video stream to read. if the value is -1, the most suitable video stream will be automatically selected. the default value is "-1". deprecated (so like, don't be doing this anymore bestie 🙈). if the filter is called "amovie", it will select audio instead of video.
loop — specifies how many times to read the stream in sequence. if the value is 0, the stream will be looped infinitely. default value is "1" 🔁

heads up bestie: when the movie is looped the source timestamps are not changed, so it will generate non monotonically increasing timestamps. that's a real caveat, not a vibe, ok? 💀

discontinuity — specifies the time difference between frames above which the point is considered a timestamp discontinuity which is removed by adjusting the later timestamps.
dec_threads — specifies the number of threads for decoding 🏃‍♀️
format_opts — specify format options for the opened file. format options can be specified as a list of key=value pairs separated by ':'. the following example shows how to add protocol_whitelist and protocol_blacklist options:

ffplay -f lavfi
"movie=filename='1.sdp':format_opts='protocol_whitelist=file,rtp,udp\:protocol_blacklist=http'"

it allows overlaying a second video on top of the main input of a filtergraph, as shown in this graph:

input -----------> deltapts0 --> overlay --> output
                                    ^
                                    |
movie --> scale--> deltapts1 -------+

Examples

skip 3.2 seconds from the start of the AVI file in.avi, and overlay it on top of the input labelled "in":

movie=in.avi:seek_point=3.2, scale=180:-1, setpts=PTS-STARTPTS [over];
[in] setpts=PTS-STARTPTS [main];
[main][over] overlay=16:16 [out]

read from a video4linux2 device, and overlay it on top of the input labelled "in":

movie=/dev/video0:f=video4linux2, scale=180:-1, setpts=PTS-STARTPTS [over];
[in] setpts=PTS-STARTPTS [main];
[main][over] overlay=16:16 [out]

read the first video stream and the audio stream with id 0x81 from dvd.vob; the video is connected to the pad named "video" and the audio is connected to the pad named "audio":

movie=dvd.vob:s=v:0+#0x81 [video] [audio]

Commands

both movie and amovie support the following commands 💅

seek — perform seek using av_seek_frame. the syntax is: seek stream_index|timestamp|flags
- stream_index: if stream_index is -1, a default stream is selected, and timestamp is automatically converted from AV_TIME_BASE units to the stream specific time_base.
- timestamp: timestamp in AVStream.time_base units or, if no stream is specified, in AV_TIME_BASE units.
- flags: flags which select direction and seeking mode.
get_duration — get movie duration in AV_TIME_BASE units ⏱️

👑 EXTERNAL LIBRARIES 👑

FFmpeg can be hooked up with a number of external libraries to add support for more formats. none of them are used by default — their use has to be explicitly requested by passing the appropriate flags to ./configure. so like, opt-in only, she's not gonna assume 🫶

🌟 Alliance for Open Media (AOM)

FFmpeg can make use of the AOM library for AV1 decoding and encoding.

go to <http://aomedia.org/> and follow the instructions for installing the library. then pass --enable-libaom to configure to enable it ✨

🔥 AMD AMF/VCE

FFmpeg can use the AMD Advanced Media Framework library for accelerated H.264 and HEVC (only windows) encoding on hardware with Video Coding Engine (VCE).

to enable support you must obtain the AMF framework header files (version 1.4.9+) from <https://github.com/GPUOpen-LibrariesAndSDKs/AMF.git>.

create an AMF/ directory in the system include path. copy the contents of AMF/amf/public/include/ into that directory. then configure FFmpeg with --enable-amf.

initialization of amf encoder occurs in this order: 1) trying to initialize through dx11 (only windows) 2) trying to initialize through dx9 (only windows) 3) trying to initialize through vulkan 🌀

to use h.264 (AMD VCE) encoder on linux amdgru-pro version 19.20+ and amf-amdgpu-pro package (amdgru-pro contains, but does not install automatically) are required.

this driver can be installed using amdgpu-pro-install script in official amd driver archive.

🦋 AviSynth

FFmpeg can read AviSynth scripts as input. to enable support, pass --enable-avisynth to configure after installing the headers provided by <https://github.com/AviSynth/AviSynthPlus>. AviSynth+ can be configured to install only the headers by either passing -DHEADERS_ONLY:bool=on to the normal CMake-based build system, or by using the supplied GNUmakefile 💖

for Windows, supported AviSynth variants are <http://avisynth.nl> for 32-bit builds and <http://avisynth.nl/index.php/AviSynth+> for 32-bit and 64-bit builds.

for Linux, macOS, and BSD, the only supported AviSynth variant is <https://github.com/AviSynth/AviSynthPlus>, starting with version 3.5. 🐧

In 2016, AviSynth+ added support for building with GCC. However, due to the eccentricities of Windows' calling conventions, 32-bit GCC builds of AviSynth+ are not compatible with typical 32-bit builds of FFmpeg. 😭 By default, FFmpeg assumes compatibility with 32-bit MSVC builds of AviSynth+ since that is the most widely-used and entrenched build configuration. Users can override this and enable support for 32-bit GCC builds of AviSynth+ by passing -DAVSC_WIN32_GCC32 to --extra-cflags when configuring FFmpeg. 64-bit builds of FFmpeg are not affected, and can use either MSVC or GCC builds of AviSynth+ without any special flags. 🎀 AviSynth(+) is loaded dynamically. Distributors can build FFmpeg with --enable-avisynth, and the binaries will work regardless of the end user having AviSynth installed. If/when an end user would like to use AviSynth scripts, then they can install AviSynth(+) and FFmpeg will be able to find and use it to open scripts. slay 💅

🍓 Chromaprint

FFmpeg can make use of the Chromaprint library for generating audio fingerprints. pass --enable-chromaprint to configure to enable it. see <https://acoustid.org/chromaprint>.

🌸 codec2

FFmpeg can make use of the codec2 library for codec2 decoding and encoding. there is currently no native decoder, so libcodec2 must be used for decoding. important detail, don't forget it 🫶

go to <http://freedv.org/>, download Codec 2 source archive. build and install using CMake. Debian users can install the libcodec2-dev package instead. once libcodec2 is installed you can pass --enable-libcodec2 to configure to enable it.

the easiest way to use codec2 is with .c2 files, since they contain the mode information required for decoding. to encode such a file, use a .c2 file extension and give the libcodec2 encoder the -mode option: ffmpeg -i input.wav -mode 700C output.c2. playback is as simple as ffplay output.c2. for a list of supported modes, run ffmpeg -h encoder=libcodec2. raw codec2 files are also supported. to make sense of them the mode in use needs to be specified as a format option: ffmpeg -f codec2raw -mode 1300 -i input.raw output.wav. ✨

💜 dav1d

FFmpeg can make use of the dav1d library for AV1 video decoding.

go to <https://code.videolan.org/videolan/dav1d> and follow the instructions for installing the library. then pass --enable-libdav1d to configure to enable it.

🌈 davs2

FFmpeg can make use of the davs2 library for AVS2-P2/IEEE1857.4 video decoding.

go to <https://github.com/pkuvcl/davs2> and follow the instructions for installing the library. then pass --enable-libdavs2 to configure to enable it.

libdavs2 is under the GNU Public License Version 2 or later (see <http://www.gnu.org/licenses/old-licenses/gpl-2.0.html> for details), you must upgrade FFmpeg's license to GPL in order to use it. licensing tea, bestie — read it before you commit ☕💀

⭐ uavs3d

FFmpeg can make use of the uavs3d library for AVS3-P2/IEEE1857.10 video decoding.

go to <https://github.com/uavs3/uavs3d> and follow the instructions for installing the library. then pass --enable-libuavs3d to configure to enable it. 🎀

Game Music Emu 🎮🎵

okay so FFmpeg can make use of the Game Music Emu library to read audio from supported video game music file formats — yes, the chiptune thing, we stan. Pass --enable-libgme to configure to enable it. All the deets live here: <https://bitbucket.org/mpyne/game-music-emu/overview>.

Intel QuickSync Video ⚡

FFmpeg can use Intel QuickSync Video (QSV) for accelerated decoding and encoding of multiple codecs. To use QSV, FFmpeg must be linked against the libmfx dispatcher, which loads the actual decoding libraries (she's the middle manager of this whole operation, truly).

The dispatcher is open source and can be downloaded from <https://github.com/lu-zero/mfx_dispatch.git>. FFmpeg needs to be configured with the --enable-libmfx option, and pkg-config needs to be able to locate the dispatcher's .pc files. (pkg-config is so picky but we love her.)

Kvazaar 💅

FFmpeg can make use of the Kvazaar library for HEVC encoding.

Go to <https://github.com/ultravideo/kvazaar> and follow the instructions for installing the library. Then pass --enable-libkvazaar to configure to enable it.

LAME 🎧

FFmpeg can make use of the LAME library for MP3 encoding.

Go to <http://lame.sourceforge.net/> and follow the instructions for installing the library. Then pass --enable-libmp3lame to configure to enable it.

LCEVCdec ✨

FFmpeg can make use of the liblcevc_dec library for LCEVC enhancement layer decoding on supported bitstreams.

Go to <https://github.com/v-novaltd/LCEVCdec> and follow the instructions for installing the library. Then pass --enable-liblcevc-dec to configure to enable it.

fyi bestie, LCEVCdec is under the BSD-3-Clause-Clear License.

libilbc 🎙️

iLBC is a narrowband speech codec that has been made freely available by Google as part of the WebRTC project. libilbc is a packaging friendly copy of the iLBC codec. FFmpeg can make use of the libilbc library for iLBC decoding and encoding.

Go to <https://github.com/TimothyGu/libilbc> and follow the instructions for installing the library. Then pass --enable-libilbc to configure to enable it.

libjxl 🖼️

JPEG XL is an image format intended to fully replace legacy JPEG for an extended period of life. See <https://jpegxl.info/> for more information, and see <https://github.com/libjxl/libjxl> for the library source. You can pass --enable-libjxl to configure in order enable the libjxl wrapper. (yes, "in order enable" — that's how it's written, we're not fixing it, we're just quoting her 💀)

libvpx 🎬

FFmpeg can make use of the libvpx library for VP8/VP9 decoding and encoding.

Go to <http://www.webmproject.org/> and follow the instructions for installing the library. Then pass --enable-libvpx to configure to enable it.

ModPlug 🎼

FFmpeg can make use of this library, originating in Modplug-XMMS, to read from MOD-like music files. See <https://github.com/Konstanty/libmodplug>. Pass --enable-libmodplug to configure to enable it. Tracker kids, this one's for you 🫶

OpenCORE, VisualOn, and Fraunhofer libraries 🏭

Spun off Google Android sources, OpenCore, VisualOn and Fraunhofer libraries provide encoders for a number of audio codecs.

⚠️ license check time, bestie, read carefully: OpenCORE and VisualOn libraries are under the Apache License 2.0 (see <http://www.apache.org/licenses/LICENSE-2.0> for details), which is incompatible to the LGPL version 2.1 and GPL version 2. You have to upgrade FFmpeg's license to LGPL version 3 (or if you have enabled GPL components, GPL version 3) by passing --enable-version3 to configure in order to use it.

⚠️ and also: the license of the Fraunhofer AAC library is incompatible with the GPL. Therefore, for GPL builds, you have to pass --enable-nonfree to configure in order to use it. To the best of our knowledge, it is compatible with the LGPL.

OpenCORE AMR

FFmpeg can make use of the OpenCORE libraries for AMR-NB decoding/encoding and AMR-WB decoding.

Go to <http://sourceforge.net/projects/opencore-amr/> and follow the instructions for installing the libraries. Then pass --enable-libopencore-amrnb and/or --enable-libopencore-amrwb to configure to enable them.

VisualOn AMR-WB encoder library

FFmpeg can make use of the VisualOn AMR-WBenc library for AMR-WB encoding.

Go to <http://sourceforge.net/projects/opencore-amr/> and follow the instructions for installing the library. Then pass --enable-libvo-amrwbenc to configure to enable it.

Fraunhofer AAC library

FFmpeg can make use of the Fraunhofer AAC library for AAC decoding & encoding.

Go to <http://sourceforge.net/projects/opencore-amr/> and follow the instructions for installing the library. Then pass --enable-libfdk-aac to configure to enable it.

Fraunhofer MPEG-H 3D Audio decoder library

FFmpeg can make use of the Fraunhofer MPEG-H decoder library for MPEG-H 3DA decoding.

Go to <https://github.com/Fraunhofer-IIS/mpeghdec> and follow the instructions for installing the library. Then pass --enable-libmpeghdec --enable-nonfree to configure to enable it.

LC3 library

FFmpeg can make use of the Google LC3 library for LC3 decoding & encoding.

Go to <https://github.com/google/liblc3/> and follow the instructions for installing the library. Then pass --enable-liblc3 to configure to enable it.

OpenH264 🌐

FFmpeg can make use of the OpenH264 library for H.264 decoding and encoding.

Go to <http://www.openh264.org/> and follow the instructions for installing the library. Then pass --enable-libopenh264 to configure to enable it.

ngl for decoding this library is much more limited than the built-in decoder in libavcodec; currently, this library lacks support for decoding B-frames and some other main/high profile features. (It currently only supports constrained baseline profile and CABAC.) Using it is mostly useful for testing and for taking advantage of Cisco's patent portfolio license (<http://www.openh264.org/BINARY_LICENSE.txt>).

OpenJPEG 🏞️

FFmpeg can use the OpenJPEG libraries for decoding/encoding J2K videos. Go to <http://www.openjpeg.org/> to get the libraries and follow the installation instructions. To enable using OpenJPEG in FFmpeg, pass --enable-libopenjpeg to ./configure.

rav1e 🦀

FFmpeg can make use of rav1e (Rust AV1 Encoder) via its C bindings to encode videos. Go to <https://github.com/xiph/rav1e/> and follow the instructions to build the C library. To enable using rav1e in FFmpeg, pass --enable-librav1e to ./configure. Rust girlies rise 🫶

SVT-AV1 🚀

FFmpeg can make use of the Scalable Video Technology for AV1 library for AV1 encoding.

Go to <https://gitlab.com/AOMediaCodec/SVT-AV1/> and follow the instructions for installing the library. Then pass --enable-libsvtav1 to configure to enable it.

SVT-JPEG-XS 📸

FFmpeg can make use of the Scalable Video Technology for JPEG-XS library for JPEG-XS decoding and encoding.

Go to <https://github.com/OpenVisualCloud/SVT-JPEG-XS> and follow the instructions for installing the library. Then pass --enable-libsvtjpegxs to configure to enable it.

TwoLAME 💿

FFmpeg can make use of the TwoLAME library for MP2 encoding.

Go to <http://www.twolame.org/> and follow the instructions for installing the library. Then pass --enable-libtwolame to configure to enable it.

VapourSynth 🌫️

FFmpeg can read VapourSynth scripts as input. To enable support, pass --enable-vapoursynth to configure. Vapoursynth is detected via pkg-config. Versions 42 or greater supported. See <http://www.vapoursynth.com/>.

🔒 okay security moment, listen up: due to security concerns, Vapoursynth scripts will not be autodetected so the input format has to be forced. For ff* CLI tools, add -f vapoursynth before the input -i yourscript.vpy.

-f vapoursynth -i yourscript.vpy

x264 🎥

FFmpeg can make use of the x264 library for H.264 encoding.

Go to <http://www.videolan.org/developers/x264.html> and follow the instructions for installing the library. Then pass --enable-libx264 to configure to enable it.

⚠️ licensing tea: x264 is under the GNU Public License Version 2 or later (see <http://www.gnu.org/licenses/old-licenses/gpl-2.0.html> for details), you must upgrade FFmpeg's license to GPL in order to use it.

x265 🎞️

FFmpeg can make use of the x265 library for HEVC encoding.

Go to <http://x265.org/developers.html> and follow the instructions for installing the library. Then pass --enable-libx265 to configure to enable it.

⚠️ same energy: x265 is under the GNU Public License Version 2 or later (see <http://www.gnu.org/licenses/old-licenses/gpl-2.0.html> for details), you must upgrade FFmpeg's license to GPL in order to use it.

xavs 🇨🇳

FFmpeg can make use of the xavs library for AVS encoding.

Go to <http://xavs.sf.net/> and follow the instructions for installing the library. Then pass --enable-libxavs to configure to enable it.

xavs2 🎀

FFmpeg can make use of the xavs2 library for AVS2-P2/IEEE1857.4 video encoding.

Go to <https://github.com/pkuvcl/xavs2> and follow the instructions for installing the library. Then pass --enable-libxavs2 to configure to enable it.

⚠️ heads up: libxavs2 is under the GNU Public License Version 2 or later (see <http://www.gnu.org/licenses/old-licenses/gpl-2.0.html> for details), you must upgrade FFmpeg's license to GPL in order to use it.

eXtra-fast Essential Video Encoder (XEVE) 🏎️

FFmpeg can make use of the XEVE library for EVC video encoding.

Go to <https://github.com/mpeg5/xeve> and follow the instructions for installing the XEVE library. Then pass --enable-libxeve to configure to enable it.

eXtra-fast Essential Video Decoder (XEVD) 🏁

FFmpeg can make use of the XEVD library for EVC video decoding.

Go to <https://github.com/mpeg5/xevd> and follow the instructions for installing the XEVD library. Then pass --enable-libxevd to configure to enable it.

ZVBI 📺

ZVBI is a VBI decoding library which can be used by FFmpeg to decode DVB teletext pages and DVB teletext subtitles.

Go to <http://sourceforge.net/projects/zapping/> and follow the instructions for installing the library. Then pass --enable-libzvbi to configure to enable it.

🌈 SUPPORTED FILE FORMATS 🌈

You can use the -formats and -codecs options to have an exhaustive list. Like, an actually exhaustive list — use those, bestie, don't come for me if your fave isn't in the highlights below 💅

📁 File Formats

FFmpeg supports the following file formats through the libavformat library. The columns in the OG table are Name, Encoding, Decoding, Comments — I'm giving you the same tea, just in list form:

3dostr — decoding only
4xm — decoding only — 4X Technologies format, used in some games.
8088flex TMV — decoding only
AAX — decoding only — Audible Enhanced Audio format, used in audiobooks.
AA — decoding only — Audible Format 2, 3, and 4, used in audiobooks.
ACT Voice — decoding only — contains G.729 audio
Adobe Filmstrip — encoding ✅ and decoding ✅
Audio IFF (AIFF) — encoding ✅ and decoding ✅
American Laser Games MM — decoding only — Multimedia format used in games like Mad Dog McCree.

okay so the actual format list starts here, alphabetical-ish, numbers first 💅 quick legend before we go: demux = ffmpeg can read it, mux = ffmpeg can write it. no cap, that's the whole table.

3GPP AMR — demux ✅ · mux ✅
Amazing Studio Packed Animation File — mux ✅ only
- the multimedia format used in the game Heart Of Darkness
Apple HTTP Live Streaming — mux ✅ only
Artworx Data Format — mux ✅ only
Interplay ACM — mux ✅ only
- audio-only format, used in some Interplay games
ADP — mux ✅ only
- audio format used on the Nintendo Gamecube
AFC — mux ✅ only
- also an audio format used on the Nintendo Gamecube
ADS/SS2 — mux ✅ only
- audio format used on the PS2
APNG — demux ✅ · mux ✅
ASF — demux ✅ · mux ✅
- that's Advanced / Active Streaming Format
AST — demux ✅ · mux ✅
- audio format used on the Nintendo Wii
AVI — demux ✅ · mux ✅
AviSynth — mux ✅ only
AVR — mux ✅ only
- audio format used on Mac
AVS — mux ✅ only
- multimedia format used by the Creature Shock game
Beam Software SIFF — mux ✅ only
- audio + video format used in some games by Beam Software
Bethesda Softworks VID — mux ✅ only
- used in some games from Bethesda Softworks
Binary text — mux ✅ only
Bink — mux ✅ only
- multimedia format used by many games
Bink Audio — mux ✅ only
- audio-only multimedia format used by some games
Bitmap Brothers JV — mux ✅ only
- used in the Z and Z95 games
BRP — mux ✅ only
- Argonaut Games format
Brute Force & Ignorance — mux ✅ only
- used in the game Flash Traffic: City of Angels
BFSTM — mux ✅ only
- audio format used on the Nintendo WiiU (based on BRSTM)
BRSTM — mux ✅ only
- audio format used on the Nintendo Wii
BW64 — mux ✅ only
- Broadcast Wave 64bit
BWF — demux ✅ · mux ✅
codec2 (raw) — demux ✅ · mux ✅
- you must pass the -mode format option or it won't decode correctly 🫶
codec2 (.c2 files) — demux ✅ · mux ✅
- these carry a header with version + mode info, so playback is way simpler
CRI ADX — demux ✅ · mux ✅
- audio-only format used in console video games
CRI AIX — mux ✅ only
CRI HCA — mux ✅ only
- audio-only format used in console video games
Discworld II BMV — mux ✅ only
Interplay C93 — mux ✅ only
- used in the game Cyberia from Interplay
Delphine Software International CIN — mux ✅ only
- multimedia format used by Delphine Software games
Digital Speech Standard (DSS) — mux ✅ only
CD+G — mux ✅ only
- video format used by CD+G karaoke disks 🎤
Phantom Cine — mux ✅ only
Commodore CDXL — mux ✅ only
- Amiga CD video format
Core Audio Format — demux ✅ · mux ✅
- Apple Core Audio Format
CRC testing format — demux ✅ only
Creative Voice — demux ✅ · mux ✅
- created for the Sound Blaster Pro
CRYO APC — mux ✅ only
- audio format used in some games by CRYO Interactive Entertainment
D-Cinema audio — demux ✅ · mux ✅
Deluxe Paint Animation — mux ✅ only
DCSTR — mux ✅ only
DFA — mux ✅ only
- this format is used in the Chronomaster game
DirectDraw Surface — mux ✅ only
DSD Stream File (DSF) — mux ✅ only
DV video — demux ✅ · mux ✅
DXA — mux ✅ only
- used in the non-Windows version of the Feeble Files game, and in different game cutscenes repacked for use with ScummVM
Electronic Arts cdata — mux ✅ only
Electronic Arts Multimedia — mux ✅ only
- used in various EA games; files have extensions like WVE and UV2
Ensoniq Paris Audio File — mux ✅ only
FFM (FFserver live feed) — demux ✅ · mux ✅
Flash (SWF) — demux ✅ · mux ✅
Flash 9 (AVM2) — demux ✅ · mux ✅
- only the embedded audio gets decoded, js
FLI/FLC/FLX animation — mux ✅ only
- .fli/.flc files
Flash Video (FLV) — demux ✅ · mux ✅
- Macromedia Flash video files
framecrc testing format — demux ✅ only
FunCom ISS — mux ✅ only
- audio format used in various games from FunCom like The Longest Journey
G.723.1 — demux ✅ · mux ✅
G.726 — mux ✅ only
- both left- and right-justified
G.729 BIT — demux ✅ · mux ✅
G.729 raw — demux ✅ only
GENH — mux ✅ only
- audio format for various games
GIF Animation — demux ✅ · mux ✅
GXF — demux ✅ · mux ✅
- General eXchange Format SMPTE 360M, used by Thomson Grass Valley playout servers
HNM — mux ✅ only
- only version 4 is supported, used in some games from Cryo Interactive
iCEDraw File — mux ✅ only
ICO — demux ✅ · mux ✅
- Microsoft Windows ICO
id Quake II CIN video — mux ✅ only
id RoQ — demux ✅ · mux ✅
- used in Quake III, Jedi Knight 2 and other computer games
IEC61937 encapsulation — demux ✅ · mux ✅
IFF — mux ✅ only
- Interchange File Format
IFV — mux ✅ only
- a format used by some old CCTV DVRs
iLBC — demux ✅ · mux ✅
Interplay MVE — mux ✅ only
- format used in various Interplay computer games
Iterated Systems ClearVideo — mux ✅ only
- I-frames only
IV8 — mux ✅ only
- a format generated by the IndigoVision 8000 video server
IVF (On2) — demux ✅ · mux ✅
- a format used by libvpx
Internet Video Recording — mux ✅ only
IRCAM — demux ✅ · mux ✅
LAF — mux ✅ only
- Limitless Audio Format
LATM — demux ✅ · mux ✅
LMLM4 — mux ✅ only
- used by Linux Media Labs MPEG-4 PCI boards
LOAS — mux ✅ only
- contains LATM multiplexed AAC audio
LRC — demux ✅ · mux ✅
LVF — mux ✅ only
LXF — mux ✅ only
- VR native stream format, used by Leitch/Harris' video servers
Magic Lantern Video (MLV) — mux ✅ only
Matroska — demux ✅ · mux ✅
Matroska audio — demux ✅ only
FFmpeg metadata — demux ✅ · mux ✅
- metadata in text format
MAXIS XA — mux ✅ only
- used in Sim City 3000; file extension .xa
MCA — mux ✅ only
- used in some games from Capcom; file extension .mca
MD Studio — mux ✅ only
Metal Gear Solid: The Twin Snakes — demux ✅ only
Megalux Frame — mux ✅ only
- used by Megalux Ultimate Paint
MobiClip MODS — mux ✅ only
MobiClip MOFLEX — mux ✅ only
Mobotix .mxg — mux ✅ only
Monkey's Audio — mux ✅ only
Motion Pixels MVI — mux ✅ only
MOV/QuickTime/MP4 — demux ✅ · mux ✅
- 3GP, 3GP2, PSP and iPod variants supported 🍎
MP2 — demux ✅ · mux ✅
MP3 — demux ✅ · mux ✅
MPEG-1 System — demux ✅ · mux ✅
- muxed audio and video, VCD format supported
MPEG-PS (program stream) — demux ✅ · mux ✅
- also known as a VOB file; SVCD and DVD format supported
MPEG-TS (transport stream) — demux ✅ · mux ✅
- also known as DVB Transport Stream
MPEG-4 — demux ✅ · mux ✅
- MPEG-4 is a variant of QuickTime
MSF — mux ✅ only
- audio format used on the PS3
Mirillis FIC video — mux ✅ only
- no cursor rendering
MIDI Sample Dump Standard — mux ✅ only
MIME multipart JPEG — demux ✅ only
MSN TCP webcam — mux ✅ only
- used by MSN Messenger webcam streams
MTV — mux ✅ only
Musepack — mux ✅ only
Musepack SV8 — mux ✅ only
Material eXchange Format (MXF) — demux ✅ · mux ✅
- SMPTE 377M, used by D-Cinema and the broadcast industry
Material eXchange Format (MXF), D-10 Mapping — demux ✅ · mux ✅
- SMPTE 386M, D-10/IMX Mapping
NC camera feed — mux ✅ only
- NC (AVIP NC4600) camera streams
NIST SPeech HEader REsources — mux ✅ only
Computerized Speech Lab NSP — mux ✅ only
NTT TwinVQ (VQF) — mux ✅ only
- Nippon Telegraph and Telephone Corporation TwinVQ
Nullsoft Streaming Video — mux ✅ only
NuppelVideo — mux ✅ only
NUT — demux ✅ · mux ✅
- NUT Open Container Format
Ogg — demux ✅ · mux ✅
Playstation Portable PMP — mux ✅ only
Portable Voice Format — mux ✅ only
RK Audio (RKA) — mux ✅ only
TechnoTrend PVA — mux ✅ only
- used by TechnoTrend DVB PCI boards

and that's not even the end of the list bestie, we're still going 📋✨

Quick key refresher before we keep going, bestie: in these rows a X under a column means that feature is supported for that format. Enc X = encoding works, Dec X = decoding works, and a — is the blank cell (so, not supported). Now, onward 💅

QCP — Enc — · Dec X
raw ADTS (AAC) — Enc X · Dec X
raw AC-3 — Enc X · Dec X
raw AMR-NB — Enc — · Dec X
raw AMR-WB — Enc — · Dec X
raw APAC — Enc — · Dec X
raw APV — Enc X · Dec X
raw aptX — Enc X · Dec X
raw aptX HD — Enc X · Dec X
raw Bonk — Enc — · Dec X
raw Chinese AVS video — Enc X · Dec X
raw DFPWM — Enc X · Dec X
raw Dirac — Enc X · Dec X
raw DNxHD — Enc X · Dec X
raw DTS — Enc X · Dec X
raw DTS-HD — Enc — · Dec X
raw E-AC-3 — Enc X · Dec X
raw EVC — Enc X · Dec X
raw FLAC — Enc X · Dec X
raw G.728 — Enc — · Dec X
raw GSM — Enc — · Dec X
raw H.261 — Enc X · Dec X
raw H.263 — Enc X · Dec X
raw H.264 — Enc X · Dec X
raw HEVC — Enc X · Dec X
raw Ingenient MJPEG — Enc — · Dec X
raw MJPEG — Enc X · Dec X
raw MLP — Enc — · Dec X
raw MPEG — Enc — · Dec X
raw MPEG-1 — Enc — · Dec X
raw MPEG-2 — Enc — · Dec X
raw MPEG-4 — Enc X · Dec X
raw NULL — Enc X · Dec —
raw video — Enc X · Dec X
raw id RoQ — Enc X · Dec —
raw OBU — Enc X · Dec X
raw OSQ — Enc — · Dec X
raw SBC — Enc X · Dec X
raw Shorten — Enc — · Dec X
raw TAK — Enc — · Dec X
raw TrueHD — Enc X · Dec X
raw VC-1 — Enc X · Dec X
raw PCM A-law — Enc X · Dec X
raw PCM mu-law — Enc X · Dec X
raw PCM Archimedes VIDC — Enc X · Dec X
raw PCM signed 8 bit — Enc X · Dec X
raw PCM signed 16 bit big-endian — Enc X · Dec X
raw PCM signed 16 bit little-endian — Enc X · Dec X
raw PCM signed 24 bit big-endian — Enc X · Dec X
raw PCM signed 24 bit little-endian — Enc X · Dec X
raw PCM signed 32 bit big-endian — Enc X · Dec X
raw PCM signed 32 bit little-endian — Enc X · Dec X
raw PCM signed 64 bit big-endian — Enc X · Dec X
raw PCM signed 64 bit little-endian — Enc X · Dec X
raw PCM unsigned 8 bit — Enc X · Dec X
raw PCM unsigned 16 bit big-endian — Enc X · Dec X
raw PCM unsigned 16 bit little-endian — Enc X · Dec X
raw PCM unsigned 24 bit big-endian — Enc X · Dec X
raw PCM unsigned 24 bit little-endian — Enc X · Dec X
raw PCM unsigned 32 bit big-endian — Enc X · Dec X
raw PCM unsigned 32 bit little-endian — Enc X · Dec X
raw PCM 16.8 floating point little-endian — Enc — · Dec X
raw PCM 24.0 floating point little-endian — Enc — · Dec X
raw PCM floating-point 32 bit big-endian — Enc X · Dec X
raw PCM floating-point 32 bit little-endian — Enc X · Dec X
raw PCM floating-point 64 bit big-endian — Enc X · Dec X
raw PCM floating-point 64 bit little-endian — Enc X · Dec X
RDT — Enc — · Dec X
REDCODE R3D — Enc — · Dec X
- File format used by RED Digital cameras, contains JPEG 2000 frames and PCM audio. 🎥
RealMedia — Enc X · Dec X
Redirector — Enc — · Dec X
RedSpark — Enc — · Dec X
Renderware TeXture Dictionary — Enc — · Dec X
Resolume DXV — Enc X · Dec X
- Encoding is only supported for the DXT1 (Normal Quality, No Alpha) texture format. ⚠️
RF64 — Enc — · Dec X
RL2 — Enc — · Dec X
- Audio and video format used in some games by Entertainment Software Partners.
RPL/ARMovie — Enc — · Dec X
Lego Mindstorms RSO — Enc X · Dec X
RSD — Enc — · Dec X
RTMP — Enc X · Dec X
- Output is performed by publishing stream to RTMP server
RTP — Enc X · Dec X
RTSP — Enc X · Dec X
Sample Dump eXchange — Enc — · Dec X
SAP — Enc X · Dec X
SBG — Enc — · Dec X
SDNS — Enc — · Dec X
SDP — Enc — · Dec X
SER — Enc — · Dec X
Digital Pictures SGA — Enc — · Dec X
Sega FILM/CPK — Enc X · Dec X
- Used in many Sega Saturn console games. 🎮
Silicon Graphics Movie — Enc — · Dec X
Sierra SOL — Enc — · Dec X
- .sol files used in Sierra Online games.
Sierra VMD — Enc — · Dec X
- Used in Sierra CD-ROM games.
Smacker — Enc — · Dec X
- Multimedia format used by many games.
SMJPEG — Enc X · Dec X
- Used in certain Loki game ports.
SMPTE 337M encapsulation — Enc — · Dec X
Smush — Enc — · Dec X
- Multimedia format used in some LucasArts games.
Sony OpenMG (OMA) — Enc X · Dec X
- Audio format used in Sony Sonic Stage and Sony Vegas.
Sony PlayStation STR — Enc — · Dec X
Sony Wave64 (W64) — Enc X · Dec X
SoX native format — Enc X · Dec X
SUN AU format — Enc X · Dec X
SUP raw PGS subtitles — Enc X · Dec X
SVAG — Enc — · Dec X
- Audio format used in Konami PS2 games.
TDSC — Enc — · Dec X
Text files — Enc — · Dec X
THP — Enc — · Dec X
- Used on the Nintendo GameCube.
Tiertex Limited SEQ — Enc — · Dec X
- Tiertex .seq files used in the DOS CD-ROM version of the game Flashback.
True Audio — Enc X · Dec X
VAG — Enc — · Dec X
- Audio format used in many Sony PS2 games.
VC-1 test bitstream — Enc X · Dec X
Vidvox Hap — Enc X · Dec X
Vivo — Enc — · Dec X
VPK — Enc — · Dec X
- Audio format used in Sony PS games.
Marble WADY — Enc — · Dec X
WAV — Enc X · Dec X
Waveform Archiver — Enc — · Dec X
WavPack — Enc X · Dec X
WebM — Enc X · Dec X
Windows Televison (WTV) — Enc X · Dec X
Wing Commander III movie — Enc — · Dec X
- Multimedia format used in Origin's Wing Commander III computer game.
Westwood Studios audio — Enc X · Dec X
- Multimedia format used in Westwood Studios games.
Westwood Studios VQA — Enc — · Dec X
- Multimedia format used in Westwood Studios games.
Wideband Single-bit Data (WSD) — Enc — · Dec X
WVE — Enc — · Dec X
Konami XMD — Enc — · Dec X
XMV — Enc — · Dec X
- Microsoft video container used in Xbox games.
XVAG — Enc — · Dec X
- Audio format used on the PS3.
xWMA — Enc — · Dec X
- Microsoft audio container used by XAudio 2.
eXtended BINary text (XBIN) — Enc X · Dec X
YUV4MPEG pipe — Enc X · Dec X
Psygnosis YOP — Enc — · Dec X

And that's the whole container/format list, babes 🌈 — X means that the feature in that column (encoding / decoding) is supported. Periodt.

🎀 Image Formats 🎀

FFmpeg can read and write images for each frame of a video sequence — yes, every single frame, she's versatile 💅. The following image formats are supported:

This table's got three columns, so read it like: Encoding · Decoding · the lil comment. A X means supported in that column, E means support comes through an external library.

.Y.U.V — Enc X · Dec X
- one raw file per component
Alias PIX — Enc X · Dec X
- Alias/Wavefront PIX image format
animated GIF — Enc X · Dec X
APNG — Enc X · Dec X
- Animated Portable Network Graphics
BMP — Enc X · Dec X
- Microsoft BMP image
BRender PIX — Enc — · Dec X
- Argonaut BRender 3D engine image format.
CRI — Enc — · Dec X
- Cintel RAW
DPX — Enc X · Dec X
- Digital Picture Exchange
EXR — Enc — · Dec X
- OpenEXR
FITS — Enc X · Dec X
- Flexible Image Transport System
HDR — Enc X · Dec X
- Radiance HDR RGBE Image format
IMG — Enc — · Dec X
- GEM Raster image
JPEG — Enc X · Dec X
- Progressive JPEG is not supported. ⚠️ note that, bestie
JPEG 2000 — Enc X · Dec X
JPEG-LS — Enc X · Dec X
JPEG-XS — Enc E · Dec E
- Supported through external library libsvtjpegxs
LJPEG — Enc X · Dec —
- Lossless JPEG
Media 100 — Enc — · Dec X
MSP — Enc — · Dec X
- Microsoft Paint image
PAM — Enc X · Dec X
- PAM is a PNM extension with alpha support.
PBM — Enc X · Dec X
- Portable BitMap image
PCD — Enc — · Dec X
- PhotoCD
PCX — Enc X · Dec X
- PC Paintbrush
PFM — Enc X · Dec X
- Portable FloatMap image
PGM — Enc X · Dec X
- Portable GrayMap image
PGMYUV — Enc X · Dec X
- PGM with U and V components in YUV 4:2:0
PGX — Enc — · Dec X
- PGX file decoder
PHM — Enc X · Dec X
- Portable HalfFloatMap image
PIC — Enc — · Dec X
- Pictor/PC Paint
PNG — Enc X · Dec X
- Portable Network Graphics image
PPM — Enc X · Dec X
- Portable PixelMap image
PSD — Enc — · Dec X
- Photoshop
PTX — Enc — · Dec X
- V.Flash PTX format
QOI — Enc X · Dec X
- Quite OK Image format
SGI — Enc X · Dec X
- SGI RGB image format
Sun Rasterfile — Enc X · Dec X
- Sun RAS image format
TIFF — Enc X · Dec X
- YUV, JPEG and some extension is not supported yet.
Truevision Targa — Enc X · Dec X
- Targa (.TGA) image format
VBN — Enc X · Dec X
- Vizrt Binary Image format
WBMP — Enc X · Dec X
- Wireless Application Protocol Bitmap image format
WebP — Enc E · Dec X
- WebP image format, encoding supported through external library libwebp
XBM — Enc X · Dec X
- X BitMap image format
XFace — Enc X · Dec X
- X-Face image format
XPM — Enc — · Dec X
- X PixMap image format
XWD — Enc X · Dec X
- X Window Dump image format

Two legend girlies for this one 🫶: X means that the feature in that column (encoding / decoding) is supported, and E means that support is provided through an external library.

🌈 Video Codecs 🌈

Same deal, three-column energy: Name · Encoding · Decoding · Comments.

4X Movie — Enc — · Dec X
- Used in certain computer games.
8088flex TMV — decode X only, play it but don't make it 🎬
A64 multicolor — encode X; creates video suitable to be played on a commodore 64 (multicolor mode). literally a C64 maker bestie, iconic 👑
Amazing Studio PAF Video — decode X
American Laser Games MM — decode X; used in games like Mad Dog McCree 🎯
Amuse Graphics Movie — decode X
AMV Video — encode X, decode X; used in Chinese MP3 players 🎧
ANSI/ASCII art — decode X
Apple Intermediate Codec — decode X
Apple MJPEG-B — decode X
Apple Pixlet — decode X
Apple ProRes — encode X, decode X; fourcc: apch,apcn,apcs,apco,ap4h,ap4x 🍎
Apple QuickDraw — decode X; fourcc: qdrw
APV — decode X
Argonaut Video — decode X; used in some Argonaut games
Asus v1 — encode X, decode X; fourcc: ASV1
Asus v2 — encode X, decode X; fourcc: ASV2
ATI VCR1 — decode X; fourcc: VCR1
ATI VCR2 — decode X; fourcc: VCR2
Auravision Aura — decode X
Auravision Aura 2 — decode X
Autodesk Animator Flic video — decode X
Autodesk RLE — decode X; fourcc: AASC
AV1 — encode E, decode E; supported through external libraries libaom, libdav1d, librav1e and libsvtav1 ✨
Avid 1:1 10-bit RGB Packer — encode X, decode X; fourcc: AVrp
AVS (Audio Video Standard) video — decode X; video encoding used by the Creature Shock game
AVS2-P2/IEEE1857.4 — encode E, decode E; supported through external libraries libxavs2 and libdavs2
AVS3-P2/IEEE1857.10 — decode E; supported through external library libuavs3d
AYUV — encode X, decode X; Microsoft uncompressed packed 4:4:4:4 💅
Beam Software VB — decode X
Bethesda VID video — decode X; used in some games from Bethesda Softworks
Bink Video — decode X
BitJazz SheerVideo — decode X
Bitmap Brothers JV video — decode X
y41p Brooktree uncompressed 4:1:1 12-bit — encode X, decode X
Brooktree ProSumer Video — decode X; fourcc: BT20
Brute Force & Ignorance — decode X; used in the game Flash Traffic: City of Angels
C93 video — decode X; codec used in Cyberia game
CamStudio — decode X; fourcc: CSCD
CD+G — decode X; video codec for CD+G karaoke disks 🎤
CDXL — decode X; Amiga CD video codec
Chinese AVS video — encode E, decode X; AVS1-P2, JiZhun profile, encoding through external library libxavs
Delphine Software International CIN video — decode X; codec used in Delphine Software International games
Discworld II BMV Video — decode X
CineForm HD — encode X, decode X
Canopus HQ — decode X
Canopus HQA — decode X
Canopus HQX — decode X
Canopus Lossless Codec — decode X
CDToons — decode X; codec used in various Broderbund games
Cinepak — decode X
Cirrus Logic AccuPak — encode X, decode X; fourcc: CLJR
CPiA Video Format — decode X
Creative YUV (CYUV) — decode X
DFA — decode X; codec used in Chronomaster game
Dirac — encode E, decode X; supported through the native vc2 (Dirac Pro) encoder
Deluxe Paint Animation — decode X
DNxHD — encode X, decode X; aka SMPTE VC3
Duck TrueMotion 1.0 — decode X; fourcc: DUCK 🦆
Duck TrueMotion 2.0 — decode X; fourcc: TM20
Duck TrueMotion 2.0 RT — decode X; fourcc: TR20
DV (Digital Video) — encode X, decode X
Dxtory capture format — decode X
Feeble Files/ScummVM DXA — decode X; codec originally used in Feeble Files game
Electronic Arts CMV video — decode X; used in NHL 95 game 🏒
Electronic Arts Madcow video — decode X
Electronic Arts TGV video — decode X
Electronic Arts TGQ video — decode X
Electronic Arts TQI video — decode X
Escape 124 — decode X
Escape 130 — decode X
EVC / MPEG-5 Part 1 — encode E, decode E; encoding and decoding supported through external libraries libxeve and libxevd 🔥
FFmpeg video codec #1 — encode X, decode X; lossless codec (fourcc: FFV1) — the homegrown icon, we stan 💖
Flash Screen Video v1 — encode X, decode X; fourcc: FSV1
Flash Screen Video v2 — encode X, decode X
Flash Video (FLV) — encode X, decode X; Sorenson H.263 used in Flash
FM Screen Capture Codec — decode X
Forward Uncompressed — decode X
Fraps — decode X
Go2Meeting — decode X; fourcc: G2M2, G2M3
Go2Webinar — decode X; fourcc: G2M4
Gremlin Digital Video — decode X
H.261 — encode X, decode X
H.263 / H.263-1996 — encode X, decode X
H.263+ / H.263-1998 / H.263 version 2 — encode X, decode X
H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 — encode E, decode X; encoding supported through external library libx264 and OpenH264 👑
HEVC — encode X, decode X; encoding supported through external library libx265 and libkvazaar
HNM version 4 — decode X
HuffYUV — encode X, decode X
HuffYUV FFmpeg variant — encode X, decode X
IBM Ultimotion — decode X; fourcc: ULTI
id Cinematic video — decode X; used in Quake II
id RoQ video — encode X, decode X; used in Quake III, Jedi Knight 2, other computer games 🎮
IFF ILBM — decode X; IFF interleaved bitmap
IFF ByteRun1 — decode X; IFF run length encoded bitmap
Infinity IMM4 — decode X
Intel H.263 — decode X
Intel Indeo 2 — decode X
Intel Indeo 3 — decode X
Intel Indeo 4 — decode X
Intel Indeo 5 — decode X
Interplay C93 — decode X; used in the game Cyberia from Interplay
Interplay MVE video — decode X; used in Interplay .MVE files
J2K — encode X, decode X
Karl Morton's video codec — decode X; codec used in Worms games 🪱
Kega Game Video (KGV1) — decode X; Kega emulator screen capture codec
Lagarith — decode X
LCEVC / MPEG-5 LCEVC / MPEG-5 Part 2 — decode E; decoding supported through external library liblcevc-dec
LCL (LossLess Codec Library) MSZH — decode X
LCL (LossLess Codec Library) ZLIB — encode E, decode E
LEAD MCMP — decode X
LOCO — decode X
LucasArts SANM/Smush — decode X; used in LucasArts games / SMUSH animations
lossless MJPEG — encode X, decode X
MagicYUV Video — encode X, decode X
Mandsoft Screen Capture Codec — decode X
Microsoft ATC Screen — decode X; also known as Microsoft Screen 3
Microsoft Expression Encoder Screen — decode X; also known as Microsoft Titanium Screen 2
Microsoft RLE — encode X, decode X
Microsoft Screen 1 — decode X; also known as Windows Media Video V7 Screen
Microsoft Screen 2 — decode X; also known as Windows Media Video V9 Screen
Microsoft Video 1 — decode X
Mimic — decode X; used in MSN Messenger Webcam streams 📹
Miro VideoXL — decode X; fourcc: VIXL
MJPEG (Motion JPEG) — encode X, decode X
Mobotix MxPEG video — decode X
Motion Pixels video — decode X
MPEG-1 video — encode X, decode X
MPEG-2 video — encode X, decode X
MPEG-4 part 2 — encode X, decode X; libxvidcore can be used alternatively for encoding
MPEG-4 part 2 Microsoft variant version 1 — decode X
MPEG-4 part 2 Microsoft variant version 2 — encode X, decode X
MPEG-4 part 2 Microsoft variant version 3 — encode X, decode X
Newtek SpeedHQ — encode X, decode X
Nintendo Gamecube THP video — decode X 🎮
NotchLC — decode X
NuppelVideo/RTjpeg — decode X; video encoding used in NuppelVideo files
On2 VP3 — decode X; still experimental ⚠️
On2 VP4 — decode X; fourcc: VP40
On2 VP5 — decode X; fourcc: VP50
On2 VP6 — decode X; fourcc: VP60,VP61,VP62
On2 VP7 — decode X; fourcc: VP70,VP71
VP8 — encode E, decode X; fourcc: VP80, encoding supported through external library libvpx
VP9 — encode E, decode X; encoding supported through external library libvpx
Pinnacle TARGA CineWave YUV16 — decode X; fourcc: Y216
Q-team QPEG — decode X; fourccs: QPEG, Q1.0, Q1.1
QuickTime 8BPS video — decode X
QuickTime Animation (RLE) video — encode X, decode X; fourcc: 'rle '
QuickTime Graphics (SMC) — encode X, decode X; fourcc: 'smc '
QuickTime video (RPZA) — encode X, decode X; fourcc: rpza
R10K AJA Kona 10-bit RGB Codec — encode X, decode X
R210 Quicktime Uncompressed RGB 10-bit — encode X, decode X
Raw Video — encode X, decode X 🧚
RealVideo 1.0 — encode X, decode X
RealVideo 2.0 — encode X, decode X
RealVideo 3.0 — decode X; still far from ideal 😭
RealVideo 4.0 — decode X
RealVideo 6.0 — decode X
Renderware TXD (TeXture Dictionary) — decode X; texture dictionaries used by the Renderware Engine
RivaTuner Video — decode X; fourcc: 'RTV1'
RL2 video — decode X; used in some games by Entertainment Software Partners

...and we're still rolling through the video codec table, bestie — hydrate, stretch, we're nearly at the bottom of this alphabet 💖

ScreenPressor — decode only.
Screenpresso — decode only.
Screen Recorder Gold Codec — decode only.
Sierra VMD video — decode only. Used in Sierra VMD files.
Silicon Graphics Motion Video Compressor 1 (MVC1) — decode only.
Silicon Graphics Motion Video Compressor 2 (MVC2) — decode only.
Silicon Graphics RLE 8-bit video — decode only.
Smacker video — decode only. Video encoding used in Smacker.
SMPTE VC-1 — decode only.
Snow — encode + decode. experimental wavelet codec (fourcc: SNOW)
Sony PlayStation MDEC (Motion DECoder) — decode only.
Sorenson Vector Quantizer 1 — encode + decode. fourcc: SVQ1
Sorenson Vector Quantizer 3 — decode only. fourcc: SVQ3
Sunplus JPEG (SP5X) — decode only. fourcc: SP5X
TechSmith Screen Capture Codec — decode only. fourcc: TSCC
TechSmith Screen Capture Codec 2 — decode only. fourcc: TSC2
Theora — encode E, decode X. encoding supported through external library libtheora
Tiertex Limited SEQ video — decode only. Codec used in DOS CD-ROM FlashBack game.
Ut Video — encode + decode.
v210 QuickTime uncompressed 4:2:2 10-bit — encode + decode.
v308 QuickTime uncompressed 4:4:4 — encode + decode.
v408 QuickTime uncompressed 4:4:4:4 — encode + decode.
v410 QuickTime uncompressed 4:4:4 10-bit — encode + decode.
VBLE Lossless Codec — decode only.
vMix Video — decode only. fourcc: 'VMX1'
VMware Screen Codec / VMware Video — decode only. Codec used in videos captured by VMware.
Westwood Studios VQA (Vector Quantized Animation) video — decode only.
Windows Media Image — decode only.
Windows Media Video 7 — encode + decode.
Windows Media Video 8 — encode + decode.
Windows Media Video 9 — decode only. not completely working 😭
Wing Commander III / Xan — decode only. Used in Wing Commander III .MVE files.
Wing Commander IV / Xan — decode only. Used in Wing Commander IV.
Winnov WNV1 — decode only.
WMV7 — encode + decode.
YAMAHA SMAF — encode + decode.
Psygnosis YOP Video — decode only.
yuv4 — encode + decode. libquicktime uncompressed packed 4:2:0
ZeroCodec Lossless Video — decode only.
ZLIB — encode + decode. part of LCL, encoder experimental ⚠️
Zip Motion Blocks Video — encode + decode. Encoder works only in PAL8.

Legend time, because the letters matter:

"X" means that the feature in that column (encoding / decoding) is supported. "E" means that support is provided through an external library.

🎧 Audio Codecs

Same deal as above, but for your ears. The columns are Name : Encoding @tab Decoding @tab Comments — slay.

8SVX exponential — decode only.
8SVX fibonacci — decode only.
AAC — encode EX, decode X. encoding supported through internal encoder and external library libfdk-aac
AAC+ — encode E, decode IX. encoding supported through external library libfdk-aac
AC-3 — encode IX, decode IX.
ACELP.KELVIN — decode only.
ADPCM 4X Movie — decode only.
ADPCM Yamaha AICA — decode only.
ADPCM AmuseGraphics Movie — decode only.
ADPCM Argonaut Games — encode + decode.
ADPCM CDROM XA — decode only.
ADPCM Creative Technology — decode only. 16 -> 4, 8 -> 4, 8 -> 3, 8 -> 2
ADPCM Electronic Arts — decode only. Used in various EA titles.
ADPCM Electronic Arts Maxis CDROM XS — decode only. Used in Sim City 3000.
ADPCM Electronic Arts R1 — decode only.
ADPCM Electronic Arts R2 — decode only.
ADPCM Electronic Arts R3 — decode only.
ADPCM Electronic Arts XAS — decode only.
ADPCM G.722 — encode + decode.
ADPCM G.726 — encode + decode.
ADPCM IMA Acorn Replay — decode only.
ADPCM IMA AMV — encode + decode. Used in AMV files
ADPCM IMA Cunning Developments — decode only.
ADPCM IMA Electronic Arts EACS — decode only.
ADPCM IMA Electronic Arts SEAD — decode only.
ADPCM IMA Funcom — decode only.
ADPCM IMA High Voltage Software ALP — encode + decode.
ADPCM IMA Mobiclip MOFLEX — decode only.
ADPCM IMA QuickTime — encode + decode.
ADPCM IMA Simon & Schuster Interactive — encode + decode.
ADPCM IMA Ubisoft APM — encode + decode.
ADPCM IMA Loki SDL MJPEG — decode only.
ADPCM IMA WAV — encode + decode.
ADPCM IMA Westwood — decode only.
ADPCM ISS IMA — decode only. Used in FunCom games.
ADPCM IMA Dialogic — decode only.
ADPCM IMA Duck DK3 — decode only. Used in some Sega Saturn console games.
ADPCM IMA Duck DK4 — decode only. Used in some Sega Saturn console games.
ADPCM IMA Radical — decode only.
ADPCM IMA Xbox — decode only.
ADPCM Microsoft — encode + decode.
ADPCM MS IMA — encode + decode.
ADPCM Nintendo Gamecube AFC — decode only.
ADPCM Nintendo Gamecube DTK — decode only.
ADPCM Nintendo THP — decode only.
ADPCM Playstation — decode only.
ADPCM QT IMA — encode + decode.
ADPCM Sanyo — decode only.
ADPCM SEGA CRI ADX — encode + decode. Used in Sega Dreamcast games.
ADPCM Shockwave Flash — encode + decode.
ADPCM Sound Blaster Pro 2-bit — decode only.
ADPCM Sound Blaster Pro 2.6-bit — decode only.
ADPCM Sound Blaster Pro 4-bit — decode only.
ADPCM VIMA — decode only. Used in LucasArts SMUSH animations.
ADPCM Konami XMD — decode only.
ADPCM Westwood Studios IMA — encode + decode. Used in Westwood Studios games like Command and Conquer.
ADPCM Yamaha — encode + decode.
ADPCM Zork — decode only.
AMR-NB — encode E, decode X. encoding supported through external library libopencore-amrnb
AMR-WB — encode E, decode X. encoding supported through external library libvo-amrwbenc
Amazing Studio PAF Audio — decode only.
Apple lossless audio — encode + decode. QuickTime fourcc 'alac'
aptX — encode + decode. Used in Bluetooth A2DP
aptX HD — encode + decode. Used in Bluetooth A2DP
ATRAC1 — decode only.
ATRAC3 — decode only.
ATRAC3+ — decode only.
ATRAC9 — decode only.
Bink Audio — decode only. Used in Bink and Smacker files in many games.
Bonk audio — decode only.
CELT — decode E. decoding supported through external library libcelt
codec2 — encode E, decode E. en/decoding supported through external library libcodec2
CRI HCA — decode only.
Delphine Software International CIN audio — decode only. Codec used in Delphine Software International games.
DFPWM — encode + decode.
Digital Speech Standard - Standard Play mode (DSS SP) — decode only.
Discworld II BMV Audio — decode only.
COOK — decode only. All versions except 5.1 are supported. 🙈
DCA (DTS Coherent Acoustics) — encode + decode. supported extensions: XCh, XXCH, X96, XBR, XLL, LBR (partially)
Dolby E — decode only.
DPCM Cuberoot-Delta-Exact — decode only. Used in few games.
DPCM Gremlin — decode only.
DPCM id RoQ — encode + decode. Used in Quake III, Jedi Knight 2 and other computer games.
DPCM Marble WADY — decode only.
DPCM Interplay — decode only. Used in various Interplay computer games.
DPCM Squareroot-Delta-Exact — decode only. Used in various games.
DPCM Sierra Online — decode only. Used in Sierra Online game audio files.
DPCM Sol — decode only.
DPCM Xan — decode only. Used in Origin's Wing Commander IV AVI files.
DPCM Xilam DERF — decode only.
DSD (Direct Stream Digital), least significant bit first — decode only.
DSD (Direct Stream Digital), most significant bit first — decode only.
DSD (Direct Stream Digital), least significant bit first, planar — decode only.
DSD (Direct Stream Digital), most significant bit first, planar — decode only.
DSP Group TrueSpeech — decode only.
DST (Direct Stream Transfer) — decode only.
DV audio — decode only.
Enhanced AC-3 — encode + decode.
EVRC (Enhanced Variable Rate Codec) — decode only.
FLAC (Free Lossless Audio Codec) — encode X, decode IX.
FTR Voice — decode only.
G.723.1 — encode + decode.
G.728 — decode only.
G.729 — decode only.
GSM — encode E, decode X. encoding supported through external library libgsm
GSM Microsoft variant — encode E, decode X. encoding supported through external library libgsm
IAC (Indeo Audio Coder) — decode only.
iLBC (Internet Low Bitrate Codec) — encode E, decode EX. encoding and decoding supported through external library libilbc
IMC (Intel Music Coder) — decode only.
Interplay ACM — decode only.
LC3 — encode E, decode E. supported through external library liblc3
MACE (Macintosh Audio Compression/Expansion) 6:1 — decode only.
Marian's A-pac audio — decode only.
MI-SC4 (Micronas SC-4 Audio) — decode only.
MLP (Meridian Lossless Packing) — encode + decode. Used in DVD-Audio discs.
Monkey's Audio — decode only.
MP1 (MPEG audio layer 1) — decode IX.
MP2 (MPEG audio layer 2) — encode IX, decode IX. encoding supported also through external library TwoLAME
MP3 (MPEG audio layer 3) — encode E, decode IX. encoding supported through external library LAME, ADU MP3 and MP3onMP4 also supported
MPEG-4 Audio Lossless Coding (ALS) — decode only.
MPEG-H 3D Audio — decode E. decoding supported through external library libmpeghdec

...and we're picking right back up mid-codec-list, no biggie 💅 the audio codec tea continues:

MobiClip FastAudio — decode only
Musepack SV7 — decode only
Musepack SV8 — decode only
Nellymoser Asao — encode ✅ + decode ✅
On2 AVC (Audio for Video Codec) — decode only
Opus — encode is E, decode is ✅ (and bestie, that E means encoding is supported through the external library libopus 🫶)
OSQ (Original Sound Quality) — decode only
PCM A-law — encode ✅ + decode ✅
PCM mu-law — encode ✅ + decode ✅
PCM Archimedes VIDC — encode ✅ + decode ✅
PCM signed 8-bit planar — encode ✅ + decode ✅
PCM signed 16-bit big-endian planar — encode ✅ + decode ✅
PCM signed 16-bit little-endian planar — encode ✅ + decode ✅
PCM signed 24-bit little-endian planar — encode ✅ + decode ✅
PCM signed 32-bit little-endian planar — encode ✅ + decode ✅
PCM 32-bit floating point big-endian — encode ✅ + decode ✅
PCM 32-bit floating point little-endian — encode ✅ + decode ✅
PCM 64-bit floating point big-endian — encode ✅ + decode ✅
PCM 64-bit floating point little-endian — encode ✅ + decode ✅
PCM D-Cinema audio signed 24-bit — encode ✅ + decode ✅
PCM signed 8-bit — encode ✅ + decode ✅
PCM signed 16-bit big-endian — encode ✅ + decode ✅
PCM signed 16-bit little-endian — encode ✅ + decode ✅
PCM signed 24-bit big-endian — encode ✅ + decode ✅
PCM signed 24-bit little-endian — encode ✅ + decode ✅
PCM signed 32-bit big-endian — encode ✅ + decode ✅
PCM signed 32-bit little-endian — encode ✅ + decode ✅
PCM signed 16/20/24-bit big-endian in MPEG-TS — decode only
PCM unsigned 8-bit — encode ✅ + decode ✅
PCM unsigned 16-bit big-endian — encode ✅ + decode ✅
PCM unsigned 16-bit little-endian — encode ✅ + decode ✅
PCM unsigned 24-bit big-endian — encode ✅ + decode ✅
PCM unsigned 24-bit little-endian — encode ✅ + decode ✅
PCM unsigned 32-bit big-endian — encode ✅ + decode ✅
PCM unsigned 32-bit little-endian — encode ✅ + decode ✅
PCM SGA — decode only
QCELP / PureVoice — decode only
QDesign Music Codec 1 — decode only
QDesign Music Codec 2 — decode only 😭 and fr fr, there are still some distortions.
RealAudio 1.0 (14.4K) — encode ✅ + decode ✅ (the Real 14400 bit/s codec, iconic)
RealAudio 2.0 (28.8K) — decode only (Real 28800 bit/s codec)
RealAudio 3.0 (dnet) — encode IX + decode ✅ — the Real low bitrate AC-3 codec ✨
RealAudio Lossless — decode only
RealAudio SIPR / ACELP.NET — decode only
RK Audio (RKA) — decode only
SBC (low-complexity subband codec) — encode ✅ + decode ✅ — used in Bluetooth A2DP 🎧
Shorten — decode only
Sierra VMD audio — decode only — used in Sierra VMD files 🫶
Smacker audio — decode only
SMPTE 302M AES3 audio — encode ✅ + decode ✅
Sonic — encode ✅ + decode ✅ — experimental codec bestie, noted 📝
Sonic lossless — encode ✅ + decode ✅ — also experimental codec
Speex — encode E + decode EX — supported through the external library libspeex 💜
TAK (Tom's lossless Audio Kompressor) — decode only
True Audio (TTA) — encode ✅ + decode ✅
TrueHD — encode ✅ + decode ✅ — used in HD-DVD and Blu-Ray discs 👑
TwinVQ (VQF flavor) — decode only
VIMA — decode only — used in LucasArts SMUSH animations 🎮
ViewQuest VQC — decode only
Vorbis — encode E + decode ✅ — ngl a native but very primitive encoder exists 😬
Voxware MetaSound — decode only
Waveform Archiver — decode only
WavPack — encode ✅ + decode ✅
Westwood Audio (SND1) — decode only
Windows Media Audio 1 — encode ✅ + decode ✅
Windows Media Audio 2 — encode ✅ + decode ✅
Windows Media Audio Lossless — decode only
Windows Media Audio Pro — decode only
Windows Media Audio Voice — decode only
Xbox Media Audio 1 — decode only
Xbox Media Audio 2 — decode only

Quick legend check so nobody's confused 💅

X means that the feature in that column (encoding / decoding) is supported.
E means that support is provided through an external library.
I means that an integer-only version is available, too (which ensures high performance on systems without hardware floating point support — lowkey important on tiny boxes 🧚).

🎀 Subtitle Formats 🎀

Columns are muxing, demuxing, encoding, decoding, in that order bestie:

3GPP Timed Text — encoding ✅ + decoding ✅
AQTitle — demuxing ✅ + decoding ✅
DVB — muxing ✅ + demuxing ✅ + encoding ✅ + decoding ✅ (full sweep, ate)
DVB teletext — demuxing ✅ + decoding E
DVD — muxing ✅ + demuxing ✅ + encoding ✅ + decoding ✅
JACOsub — muxing ✅ + demuxing ✅ + decoding ✅
MicroDVD — muxing ✅ + demuxing ✅ + decoding ✅
MPL2 — demuxing ✅ + decoding ✅
MPsub (MPlayer) — demuxing ✅ + decoding ✅
PGS — decoding only
PJS (Phoenix) — demuxing ✅ + decoding ✅
RealText — demuxing ✅ + decoding ✅
SAMI — demuxing ✅ + decoding ✅
Spruce format (STL) — demuxing ✅ + decoding ✅
SSA/ASS — muxing ✅ + demuxing ✅ + encoding ✅ + decoding ✅
SubRip (SRT) — muxing ✅ + demuxing ✅ + encoding ✅ + decoding ✅
SubViewer v1 — demuxing ✅ + decoding ✅
SubViewer — demuxing ✅ + decoding ✅
TED Talks captions — demuxing ✅ + decoding ✅
TTML — muxing ✅ + encoding ✅ (demuxing and decoding left blank, that's just how it is 💅)
VobSub (IDX+SUB) — demuxing ✅ + decoding ✅
VPlayer — demuxing ✅ + decoding ✅
WebVTT — muxing ✅ + demuxing ✅ + encoding ✅ + decoding ✅
XSUB — encoding ✅ + decoding ✅

Legend time 🌸

X means that the feature is supported.
E means that support is provided through an external library.

Legend 🦋

X means that the protocol is supported.
E means that support is provided through an external library.

✨ Input/Output Devices ✨

ALSA — input ✅ + output ✅
BKTR — input ✅ (no output, she's input-only)
caca — output ✅
DV1394 — input ✅
Lavfi virtual device — input ✅
Linux framebuffer — input ✅ + output ✅
JACK — input ✅
LIBCDIO — input ✅
LIBDC1394 — input ✅
OpenAL — input ✅
OpenGL — output ✅
OSS — input ✅ + output ✅
PulseAudio — input ✅ + output ✅
SDL — output ✅
Video4Linux2 — input ✅ + output ✅
VfW capture — input ✅
X11 grabbing — input ✅
Win32 grabbing — input ✅

Legend 💖

X means that input/output is supported.

⭐ Timecode ⭐

Read then write, both slaying:

AVI — read ✅ + write ✅
DV — read ✅ + write ✅
GXF — read ✅ + write ✅
MOV — read ✅ + write ✅
MPEG1/2 — read ✅ + write ✅
MXF — read ✅ + write ✅

👑 SEE ALSO 👑

Go read these next bestie, they're all canon:

ffmpeg(1), ffplay(1), ffprobe(1), ffmpeg-utils(1), ffmpeg-scaler(1), ffmpeg-resampler(1), ffmpeg-codecs(1), ffmpeg-bitstream-filters(1), ffmpeg-formats(1), ffmpeg-devices(1), ffmpeg-protocols(1), ffmpeg-filters(1)

💜 AUTHORS 💜

The FFmpeg developers. That's the whole credit line, and honestly? Respect 🫶

For details about the authorship, see the Git history of the project (<https://git.ffmpeg.org/ffmpeg>), e.g. by typing the command git log in the FFmpeg source directory, or browsing the online repository at <https://git.ffmpeg.org/ffmpeg>. Go look, it's giving archaeology but make it code 🧚

Maintainers for the specific components are listed in the file MAINTAINERS in the source code tree — check there if you wanna know who to thank for your fave filter ✨

And that's a wrap on FFMPEG-ALL(1) 🎬✨ the whole manual, every last codec, subtitle format, protocol and device, all accounted for. Signing off with the page's own little footer energy — source/version and date, exactly as the page stamps it, no cap 💅 periodt.

📦 where this page comes from

alpine 3.24: ffmpeg-doc 8.1.3-r0 · GPL-2.0-or-later AND LGPL-2.1-or-later
nixos 26.05: ffmpeg 8.1.3
nixos 26.05: ffmpeg-headless 8.1.3
💖 girlified by deepseek-ai/DeepSeek-V4.1-Flash-1017 on 2026-11-06: a language model's rewrite of the original, under its terms

FFMPEG-ALL(1)

💖 FFMPEG-ALL(1) 💖

🌸 NAME 🌸

✨ SYNOPSIS ✨

🎀 DESCRIPTION 🎀

🦋 DETAILED DESCRIPTION 🦋

🌸 Streamcopy 🌸

🔥 Transcoding 🔥

✨ Filtering ✨

🌸 Simple filtergraphs 🌸

💅 Complex filtergraphs 💅

🦋 Loopback decoders 🦋

🌸 STREAM SELECTION 🌸

🎀 Description 🎀

✨ Automatic stream selection ✨

💅 Manual stream selection 💅

🦋 Complex filtergraphs 🦋

🧚 Stream handling 🧚

🌈 Examples 🌈

🍓 Example: automatic stream selection 🍓

💜 Example: automatic subtitles selection 💜

🔥 Example: unlabeled filtergraph outputs 🔥

👑 Example: labeled filtergraph outputs 👑

✨ OPTIONS ✨

🌈 Stream specifiers 🌈

✨ Generic options ✨

🦋 AVOptions 🦋

generic

private

🌈 Main options 🌈

🎀 target presets: one flag, whole aesthetic

✨ Video Options ✨

🌈 Advanced Video options 🌈

💖 time[,time...]

💖 expr:expr

💖 source

scd_metadata 🎬

-apply_cropping[:stream_specifier] source (input,per-stream) ✂️

-copyinkf[:stream_specifier] (output,per-stream) 📋

-init_hw_device type[=name][:device[,key=value...]] 🛠️

cuda 🟩

dxva2 🪟

d3d11va 🪟✨

vaapi 🐧

vdpau 🐧

qsv 💠

opencl 🌈

vulkan 🔥

-init_hw_device type[=name]@source 🧬

-init_hw_device list 📜

-filter_hw_device name 🎛️

-hwaccel[:stream_specifier] hwaccel (input,per-stream) ⚡

🎶 Audio Options 🎶

🚀 Advanced Audio options 🚀

📝 Subtitle options 📝

🧚 Advanced Subtitle options 🧚

👑 Advanced options 👑

🎀 Preset files 🎀

ffpreset files

avpreset files

🌈 vstats file format 🌈

✨ EXAMPLES ✨

📹 Video and Audio grabbing

🖥️ X11 grabbing

🎞️ Video and Audio file format conversion

🧠 SYNTAX 🧠

🔤 Quoting and escaping

📅 Date

⏱️ Time duration

📐 Video size

🎚️ Video rate

🎀 Ratio 🎀

🎨 Color 🎨

🔊 Channel Layout 🔊

✨ EXPRESSION EVALUATION ✨

🎛️ CODEC OPTIONS 🎛️

🎬 DECODERS

🎥 VIDEO DECODERS

av1

hevc

`scd_metadata` 🎬

`-apply_cropping[:stream_specifier] source` (input,per-stream) ✂️

`-copyinkf[:stream_specifier]` (output,per-stream) 📋

`-init_hw_device type[=name][:device[,key=value...]]` 🛠️

`cuda` 🟩

`dxva2` 🪟

`d3d11va` 🪟✨

`vaapi` 🐧

`vdpau` 🐧

`qsv` 💠

`opencl` 🌈

`vulkan` 🔥

`-init_hw_device type[=name]@source` 🧬

`-init_hw_device list` 📜

`-filter_hw_device name` 🎛️

`-hwaccel[:stream_specifier] hwaccel` (input,per-stream) ⚡