FFPLAY-ALL(1)

✨ NAME ✨

ffplay — the FFplay media player, bestie 💖 (yes that's the whole name section, she's lowkey humble like that)

🎀 SYNOPSIS 🎀

ffplay [options] [input_url]

🦋 DESCRIPTION 🦋

Okay so picture this: FFplay is a very simple and portable media player that runs on the FFmpeg libraries plus the SDL library. That's the whole vibe. ✨

And like… its main character job? It's mostly used as a testbed for all the various FFmpeg APIs. So it's less "download this to watch your movies" and more "the devs use it to poke at stuff." Ngl it still slays at playing media tho 💅

🌸 OPTIONS 🌸

So every numerical option here — unless it says otherwise — accepts a string that represents a number as input, and that string can be followed by one of the SI unit prefixes. Examples given: 'K', 'M', or 'G'. Cute, right? 🎀

Now if you append 'i' to that SI unit prefix, the whole prefix gets interpreted as a unit prefix for binary multiples, which are based on powers of 1024 instead of powers of 1000. Math girlie behavior fr fr. And if you append 'B' to the SI unit prefix, it multiplies the value by 8. So this is how you get suffixes like 'KB', 'MiB', 'G' and 'B' as number suffixes. Obsessed with the flexibility honestly 🫶

Options that don't take arguments are boolean options, and they set the corresponding value to true. Want it false? Just prefix the option name with "no". So using -nofoo will set the boolean option named "foo" to false. Iconic shortcut, no cap 💀

Options that do take arguments support a special syntax where the argument you gave on the command line is interpreted as a path to the file the actual argument value gets loaded from. To use this feature, add a forward slash / immediately before the option name (after the leading dash). Like so:

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

That will load a filtergraph description from the file named filter.script. She's reading from a file, your honor 📖✨

💜 Stream specifiers 💜

Some options are applied per-stream — think bitrate or codec. Stream specifiers are how you precisely specify which stream(s) a given option belongs to. Vibe check on your streams, basically 🌈

A stream specifier is a string that's generally appended to the option name and separated from it by a colon. Example: -codec:a:1 ac3 contains the "a:1" stream specifier, which matches the second audio stream. So it would select the ac3 codec for the second audio stream. Slay 🎯

A stream specifier can match several streams too, so the option gets applied to all of them. Example: the stream specifier in -b:a 128k matches all audio streams.

An empty stream specifier matches all streams. For example, -codec copy or -codec: copy would copy all the streams without reencoding. Copy queen 👑

Possible forms of stream specifiers are:

stream_index — Matches the stream with this index. Example: -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 (peep 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 that case it's based on the ordering of the streams in the group or program. Order matters, bestie 🧚
stream_type[:additional_stream_specifier] — stream_type is one of the following: 'v' or 'V' for video, 'a' for audio, 's' for subtitle, 'd' for data, and 't' for attachments. 'v' matches all video streams; 'V' only matches video streams which are not attached pictures, video thumbnails or cover arts. If additional_stream_specifier is used, then it matches streams which both have this type and match the additional_stream_specifier. Otherwise, it matches all streams of the specified type. 💅
g:group_specifier[:additional_stream_specifier] — Matches streams which are in the group with the specifier group_specifier. If additional_stream_specifier is used, then it matches streams which both are part of the group and match the additional_stream_specifier. And 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 which are in the program with the id program_id. If additional_stream_specifier is used, then it matches streams which both are part of the program and match the additional_stream_specifier. 🍓
#stream_id or i:stream_id — Match the stream by stream id (like a PID in an MPEG-TS container, iykyk).
m:key[:value] — Matches streams with the metadata tag key having the specified value. If value isn't 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 '+'. Go off 💖
u — Matches streams with usable configuration: the codec must be defined and the essential information such as video dimension or audio sample rate must be present.

Okay small important note: in ffmpeg, matching by metadata will only work properly for input files. Just so you know, bestie 😭

🌟 Generic options 🌟

These options are shared amongst the ff* tools. Family sharing, we love to see it 🫶

-L, -license — Show license. 📜
-h, -?, -help, --help [arg] — Show help. An optional parameter may be specified to print help about a specific item. If no argument is specified, only basic (non advanced) tool options are shown.

Possible values of arg are:

long — Print advanced tool options in addition to the basic tool options.
full — Print complete list of options, including shared and private options for encoders, decoders, demuxers, muxers, filters, etc. The whole buffet 🍓
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.

Little note: the term 'codec' is used throughout this documentation as a shortcut for what is more correctly called a media bitstream format. Precision, bestie 💅

-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. 🌈 (my personal fave, obviously)
-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. And the returned list cannot be assumed to be always complete. Ngl that's just how it is 😭

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 or -v [flags+]loglevel — Set logging level and flags used by the library. Let's control the chatter 💬

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 loglevel. And when setting both flags and loglevel, a '+' separator is expected between the last flags value and before loglevel. 🎀

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. Heads up: 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. She's chatty.
debug, 48 — Show everything, including debugging information.
trace, 56 — (that's it, that's the entire description, she said what she said)

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

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

okay so 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. 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

Errors in parsing the environment variable are not fatal, and will not appear in the report.

-hide_banner — Suppress printing the banner.

All FFmpeg tools will normally show a copyright notice, build options and library versions. This option can be used to suppress printing this information.

-cpuflags flags (global) — Allows setting and clearing cpu flags. This option is intended for testing. Do not use it unless you know what you're doing.

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

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) — Override detection of CPU count. This option is intended for testing. Do not use it unless you know what you're doing.

ffmpeg -cpucount 2

-max_alloc bytes — Set the maximum size limit for allocating a block on the heap by ffmpeg's family of malloc functions. Exercise extreme caution when using this option. Don't use if you do not understand the full consequence of doing so. Default is INT_MAX.

✨ AVOptions ✨

These options are provided directly by the libavformat, libavdevice and libavcodec libraries. To see the list of available AVOptions, use the -help option. They are separated into two categories:

generic — These options can be set for any container, codec or device. Generic options are listed under AVFormatContext options for containers/devices and under AVCodecContext options for codecs.
private — These options are specific to the given container, device or codec. Private options are listed under their corresponding containers/devices/codecs.

For example, to write an ID3v2.3 header instead of a default ID3v2.4 to an MP3 file, use the id3v2_version private option of the MP3 muxer:

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

All codec AVOptions are per-stream, and thus a stream specifier should be attached 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 the above example, a multichannel audio stream is mapped twice for output. The first instance is encoded with codec ac3 and bitrate 640k. The second instance is downmixed to 2 channels and encoded with codec aac. A bitrate of 128k is specified for it using absolute index of the output stream.

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

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

🎀 Main options 🎀

-x width — Force displayed width.
-y height — Force displayed height.
-fs — Start in fullscreen mode.
-an — Disable audio.
-vn — Disable video.
-sn — Disable subtitles.
-ss pos — Seek to pos. Note that in most formats it is not possible to seek exactly, so ffplay will seek to the nearest seek point to pos.

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

-t duration — Play duration seconds of audio/video.

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

-bytes — Seek by bytes.
-seek_interval — Set custom interval, in seconds, for seeking using left/right keys. Default is 10 seconds.
-nodisp — Disable graphical display.
-noborder — Borderless window.
-alwaysontop — Window always on top. Available on: X11 with SDL >= 2.0.5, Windows SDL >= 2.0.6.
-volume — Set the startup volume. 0 means silence, 100 means no volume reduction or amplification. Negative values are treated as 0, values above 100 are treated as 100.
-f fmt — Force format.
-window_title title — Set window title (default is the input filename).
-left title — Set the x position for the left of the window (default is a centered window).
-top title — Set the y position for the top of the window (default is a centered window).
-loop number — Loops movie playback number times. 0 means forever.
-showmode mode — Set the show mode to use. Available values for mode are:
- 0, video — show video
- 1, waves — show audio waves
- 2, rdft — show audio frequency band using RDFT ((Inverse) Real Discrete Fourier Transform)

Default value is video; if video is not present or cannot be played, rdft is automatically selected.

You can interactively cycle through the available show modes by pressing the key w.

-vf filtergraph — Create the filtergraph specified by filtergraph and use it to filter the video stream.

filtergraph is a description of the filtergraph to apply to the stream, and must have a single video input and a single video output. 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.

You can specify this parameter multiple times and cycle through the specified filtergraphs along with the show modes by pressing the key w.

-af filtergraph — filtergraph is a description of the filtergraph to apply to the input audio. Use the option -filters to show all the available filters (including sources and sinks).
-i input_url — Read input_url.

💅 Advanced options 💅

-stats — Print several playback statistics, in particular show the stream duration, the codec parameters, the current position in the stream and the audio/video synchronisation drift. It is shown by default, unless the log level is lower than info. Its display can be forced by manually specifying this option. To disable it, you need to specify -nostats.
-fast — Non-spec-compliant optimizations.
-genpts — Generate pts.
-sync type — Set the master clock to audio (type=audio), video (type=video) or external (type=ext). Default is audio. The master clock is used to control audio-video synchronization. Most media players use audio as master clock, but in some cases (streaming or high quality broadcast) it is necessary to change that. This option is mainly used for debugging purposes.
-ast audio_stream_specifier — Select the desired audio stream using the given stream specifier. The stream specifiers are described in the Stream specifiers chapter. If this option is not specified, the "best" audio stream is selected in the program of the already selected video stream.
-vst video_stream_specifier — Select the desired video stream using the given stream specifier. The stream specifiers are described in the Stream specifiers chapter. If this option is not specified, the "best" video stream is selected.
-sst subtitle_stream_specifier — Select the desired subtitle stream using the given stream specifier. The stream specifiers are described in the Stream specifiers chapter. If this option is not specified, the "best" subtitle stream is selected in the program of the already selected video or audio stream.
-autoexit — Exit when video is done playing.
-exitonkeydown — Exit if any key is pressed.
-exitonmousedown — Exit if any mouse button is pressed.
-codec:media_specifier codec_name — Force a specific decoder implementation for the stream identified by media_specifier, which can assume the values a (audio), v (video), and s subtitle.
-acodec codec_name — Force a specific audio decoder.
-vcodec codec_name — Force a specific video decoder.
-scodec codec_name — Force a specific subtitle decoder.
-autorotate — Automatically rotate the video according to file metadata. Enabled by default, use -noautorotate to disable it.
-framedrop — bestie, this one just drops video frames when the video's out of sync ✨ It's on by default if the master clock isn't set to video. Wanna force frame dropping for all master clock sources? Use this option. Wanna turn it off? -noframedrop. Slay.
-infbuf — do not limit the input buffer size — read as much data as possible from the input, ASAP 💅 This is enabled by default for realtime streams, where data might get dropped if it's not read in time. Use this option to get infinite buffers for all inputs, or -noinfbuf to disable it. iykyk.
-filter_threads nb_threads — tells you how many threads get used to process a filter pipeline 🧵 Every pipeline spawns a thread pool with this many threads available for parallel processing. Default is 0, which literally means the thread count is decided by the number of available CPUs. main character scheduling fr.
-enable_vulkan — use the vulkan renderer instead of the SDL builtin renderer 🦋 Depends on libplacebo. no cap.
-vulkan_params — vulkan configuration, given as a list of key=value pairs separated by ":" 🎀
-hwaccel — use HW accelerated decoding 🔥 Turning this on will enable the vulkan renderer automatically. it's giving efficiency.
-video_bg pattern — sets the video background pattern used for transparent videos 🖼️ For solid color patterns you can use the color name or color code as specified in the Color section in the ffmpeg-utils(1) manual. You can also use the special "tiles" keyword for a checker board style pattern — and that's also the default. The "none" pattern does rendering without alpha blending, effectively ignoring the alpha channel. periodt.

💖 While playing 💖

okay so this is the fun part — the keyboard is literally your bestie now, here's every binding 👑

q, ESC — quit. bye bestie 👋
f — toggle full screen 🖥️
p, SPC — pause ⏸️
m — toggle mute 🔇
9, 0 and /, * — decrease and increase volume respectively 🔊 (yes, both pairs do the same thing, iconic)
a — cycle audio channel in the current program 🎧
v — cycle video channel 📺
t — cycle subtitle channel in the current program 💬
c — cycle program 🔄
w — cycle video filters or show modes 🌈
s — step to the next frame ⏭️ Pauses if the stream isn't already paused, steps to the next video frame, and pauses. so satisfying ngl.
left/right — seek backward/forward by 10 seconds, or a custom interval if -seek_interval is set ⏪⏩
down/up — seek backward/forward 1 minute ⏱️
page down/page up — seek to previous/next chapter, or backward/forward 10 minutes if there's no chapters 📖
right mouse click — seek to percentage in file corresponding to fraction of width 🖱️ (yes it maps your click position to the file, lowkey genius)
left mouse double-click — toggle full screen 🖱️🖱️

📐 SYNTAX 📐

This section documents the syntax and formats employed by the FFmpeg libraries and tools. buckle up, it's grammar class but make it cute 💅

🧚 Quoting and escaping 🧚

FFmpeg adopts the following quoting and escaping mechanism, unless it's explicitly specified otherwise. The rules, bestie:

' and \ are special characters (used for quoting and escaping respectively) 🎀 On top of those, 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. annoying but true 😭
Leading and trailing whitespaces, unless escaped or quoted, are removed from the parsed string ✂️

Heads up: you may need to add a second level of escaping when using the command line or a script, and that depends on the syntax of the adopted shell language. lowkey the villain of the whole system.

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's interpreted as UTC. If the year-month-day part isn't specified, it takes the current year-month-day 🦋

⏳ Time duration ⏳

There are two accepted syntaxes for expressing time duration, bestie:

[-][<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 the 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 a 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 widthx*height*, 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 💫

The following abbreviations are recognized:

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 🎀

Careful bestie — a ratio with an infinite (1/0) or negative value is considered valid, so you should check on the returned value if you wanna exclude those values. 🙈

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 a 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 isn't specified, then 0xff is assumed 💅

The string random will result in a random color 🎲

The following names of colors are recognized:

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
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 🎧

Okay bestie, real talk — a channel layout is just the map of where all the audio channels are placed in space in a multi-channel stream. Like, which speaker goes where in your room. FFmpeg has its own special syntax for saying all that, and it's honestly kind of a vibe once you get it. 🦋

Individual channels

Every channel has an id, and here's the whole guest list:

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 😌✨

You don't have to spell everything out yourself — these identifiers are the pre-made fits. Just drop the name and you're done:

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

Not me pointing out that 6.1 shows up twice with two different compositions — that's literally what the source says, so I'm keeping both. Not a typo on my end, bestie, I promise. 💀

Custom layouts 🛠️✨

A custom channel layout is just a sequence of terms, glued together with +. 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.)

Standard layouts, the short way

You can also specify a standard channel layout with any of these:

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 — vibe check failed on those, sorry. 😭
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.

⚠️ History lesson, and it's giving "back in my day": before libavutil version 53 the trailing character c to specify a number of channels was optional, but now it is required. Meanwhile a channel layout mask can also be specified as a decimal number (if and only if not followed by c or C). So yes, the c is mandatory now. Periodt.

Also, go check out the function av_channel_layout_from_string defined in libavutil/channel_layout.h. 👑

✨ EXPRESSION EVALUATION ✨

Okay so when FFmpeg needs to do math, it uses its own little internal formula evaluator, implemented through the libavutil/eval.h interface. Yes bestie, it has a whole calculator built in 💅

An expression may contain unary operators, binary operators, constants, and functions. It's giving math class but make it slay.

Two expressions expr1 and expr2 can be combined to form another expression "expr1;expr2". expr1 and expr2 are evaluated in turn, and the new expression evaluates to the value of expr2. So like, the second one is the main character 👑

The following binary operators are available: "+", "-", "*", "/", "^".

The following unary operators are available: "+", "-".

Some internal variables can be used to store and load intermediary results. They can be accessed using the "ld" and "st" functions with an index argument varying from 0 to 9 to specify which internal variable to access. Little numbered lockers, basically 🎀

🧮 The Functions 🧮

The following functions are available:

`abs(x)`

Compute absolute value of x. No more negative vibes ✨

`acos(x)`

Compute arccosine of x.

`asin(x)`

Compute arcsine of x.

`atan(x)`

Compute arctangent of x.

`atan2(y, x)`

Compute 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. A true vibe check 🦋

`bitand(x, y)` and `bitor(x, y)`

Compute bitwise and/or operation on x and y.

The results of the evaluation of x and y are converted to integers before executing 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). 😭 Precision loss is not the moment, fr fr.

`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. Keeping it in range, periodt 💅

`cos(x)`

Compute cosine of x.

`cosh(x)`

Compute hyperbolic cosine of x.

`eq(x, y)`

Return 1 if x and y are equivalent, 0 otherwise.

`exp(x)`

Compute 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 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. 💀 So like, don't do that.

`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)`

This function is 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. Pythagoras could never 🎀

`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 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)` and `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. Little peek at what's going on inside 👀

`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. ⚠️ Not continuous? Not our problem, but also it is your problem.

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 sign of x.

`sin(x)`

Compute sine of x.

`sinh(x)`

Compute hyperbolic sine of x.

`sqrt(expr)`

Compute the square root of expr. This is equivalent to "(expr)^.5".

`squish(x)`

Compute expression "1/(1 + exp(4*x))". Love that name honestly, iconic 🌸

`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. So no passing notes between expressions, bestie, they don't talk 😭

`tan(x)`

Compute tangent of x.

`tanh(x)`

Compute hyperbolic tangent of x.

`taylor(expr, x)` and `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. Clock check ⏰

`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 returns the value of the last expr evaluation, or NAN if cond was always false.

🌈 The Constants 🌈

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

💖 Truthy Business 💖

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. Go off, build your own toolkit 🫶

🧚 Unit Prefixes 🧚

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. Everyone's got their own little settings, love that for them 💅

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. Read the label before you use it, bestie 🎀

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 follow:

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. 😬 So like, don't crank it down too hard.
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 loop filter 💫
qscale — use fixed qscale, periodt.
pass1 — use the internal 2pass ratecontrol in first pass mode 🎬
pass2 — use the internal 2pass ratecontrol in second pass mode 🎬
gray — only decode/encode grayscale 🖤🤍
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. Error AVERROR_INPUT_CHANGED is returned when a frame is dropped. so real 💅
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. it's giving reproducibility icon 🧪
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. lowkey chaotic but okay 🫠

time_base rational number — set codec time base.

It is the fundamental unit of time (in seconds) in terms of which frame timestamps are represented. For fixed-fps content, timebase should be 1 / frame_rate and timestamp increments should be identically 1. the way this is literally the backbone of timestamp math 🦋

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.) iykyk, check your encoder's docs 📖
frame_size integer (encoding,audio) — set audio frame size.

Each submitted frame except the last must contain exactly frame_size samples per channel. May be 0 when the codec has CODEC_CAP_VARIABLE_FRAME_SIZE set, in that case the frame size is not restricted. It is set by some decoders to indicate constant frame size. ✨

frame_number integer — set the frame number.
delay integer — (that's it, that's the option 💅)
qcomp float (encoding,video) — set video quantizer scale compression (VBR). It is 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 that B-frames are disabled. If a value of -1 is used, it will choose 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 — (just vibes and a tag 🏷️)
bug flags (decoding,video) — workaround for not auto-detected encoder bugs 🐛

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. ⚠️ bestie pls heed this one

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. 😭 fr fr it's for analysis not aesthetics
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 — (yep, that's the whole thing 🫶)
block_align integer — (no extra deets, it is what it is)
rc_override_count integer — (just a count, bestie)
maxrate integer (encoding,audio,video) — set max bitrate tolerance (in bits/s). Requires bufsize to be set. don't forget the bufsize or it won't slay 🔥
minrate integer (encoding,audio,video) — set min bitrate tolerance (in bits/s). Most useful in setting up a CBR encode. It is 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 an IDCT compatible with the simple one
arm
altivec
sh4
simplearm
simplearmv5te
simplearmv6
simpleneon
xvid
faani — floating point AAN IDCT

slice_count integer — (that's all, just a 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 — (no description, it's literally just the number of 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. same thing, different name bestie 🎀
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:

sad — sum of absolute differences, fast (this is the default btw 💅)
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 — like, it's giving "why would you" energy)
bit — number of bits needed for the block
rd — rate distortion optimal, slow 🐌
zero — literally just 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) — sets the interlaced dct compare function ✨

The possible values are, and yes it's the exact same lil' family of compare functions:

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) — sets the diamond type & size for motion estimation 💎

The ranges go like this, bestie, so pay attention:

(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) — sets the amount of motion predictors from the previous frame 🫶
precmp integer (encoding,video) — sets the pre motion estimation compare function ✨

The possible values (yep, the same iconic 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

pre_dia_size integer (encoding,video) — sets the diamond type & size for the motion estimation pre-pass 💎
subq integer (encoding,video) — sets the sub pel motion estimation quality ⭐
me_range integer (encoding,video) — sets the limit of motion vectors range (1023 for DivX player, iykyk)
global_quality integer (encoding,audio,video)
slice_flags integer
mbd integer (encoding,video) — sets the 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) — sets the number of bits which should be loaded into the rc buffer before decoding even starts 🫶
flags2 flags (decoding/encoding,audio,video,subtitles)

Possible values, bestie:

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. Also peep 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 the 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. Also see 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) — sets 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. Slay.

dc integer (encoding,video) — sets intra_dc_precision
nssew integer (encoding,video) — sets the nsse weight ⚖️
skip_top integer (decoding,video) — sets the number of macroblock rows at the top which are skipped ⬆️
skip_bottom integer (decoding,video) — sets the number of macroblock rows at the bottom which are skipped ⬇️
profile integer (encoding,audio,video) — sets the encoder codec profile. Default value is unknown. Encoder specific profiles are documented in the relevant encoder documentation 📚
level integer (encoding,audio,video) — sets 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) — sets the min macroblock lagrange factor (VBR)
mblmax integer (encoding,video) — sets the max macroblock lagrange factor (VBR)

skip_loop_filter integer (decoding,video)
skip_idct integer (decoding,video)
skip_frame integer (decoding,video)

Okay so these three besties make the 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, and 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) — sets the minimum interval between IDR-frames
refs integer (encoding,video) — sets reference frames to consider for motion compensation
trellis integer (encoding,audio,video) — sets 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) — bestie, if you're using this as an input parameter it's literally just a lil hint to the decoder telling it which color_range the input has. The possible values, slay:
- 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 samples sit, no cap. The possible values:
- left
- center
- topleft
- top
- bottomleft
- bottom
alpha_mode integer (decoding/encoding,video) — the transparency vibe. Possible values:
- premultiplied
- straight
log_level_offset integer — sets the log level offset. That's it, that's the whole gig ✨
slices integer (encoding,video) — the number of slices, used in parallelized encoding. More slices = more parallelism, iykyk 🧚
thread_type flags (decoding/encoding,video) — picks which multithreading methods to use. Heads up though: using frame will increase the decoding delay by one frame per thread, so clients that can't provide future frames should not use it 🙈

Possible values:

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

Default value is slice+frame. Periodt 💅

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".
pkt_timebase rational number — that's the whole description, bestie, it's giving mysterious main character energy 🎀
sub_charenc encoding (decoding,subtitles) — sets the input subtitles character encoding.
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). It works just like the gray flag in the flags option, except it skips chroma information instead of alpha. Default is 0.
codec_whitelist list (input) — a ","-separated list of allowed decoders. By default all are allowed 💖
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

max_pixels integer (decoding/encoding,video) — maximum number of pixels per image. Use this to avoid out-of-memory failures from huge images. Protect your RAM, queen 👑
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 isn't met, the cropping will be partially applied to maintain alignment. Default is 1 (enabled).

Note, bestie: the required alignment depends on whether AV_CODEC_FLAG_UNALIGNED is set and on the CPU. AV_CODEC_FLAG_UNALIGNED cannot be changed from the command line. Also, hardware decoders will not apply left/top cropping. 💀

🎬 DECODERS 🎬

Decoders are these configured elements in FFmpeg that let you decode multimedia streams. Iconic behaviour, honestly.

When you configure your FFmpeg build, all the supported native decoders are enabled by default. Decoders that need an external library have to be enabled manually via the corresponding --enable-lib option. You can list every available decoder with the configure option --list-decoders 🦋

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

And the -decoders option of the ff* tools will display the list of enabled decoders. Slay 🌸

📹 VIDEO DECODERS 📹

A description of some of the currently available video decoders follows. Buckle up.

`av1` ✨

AOMedia Video 1 (AV1) decoder.

Options

operating_point — selects 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.

The decoder supports MV-HEVC multiview streams with at most two views. Which views get output is selected by supplying a list of view IDs to the decoder (the view_ids option). You can set this option either statically before decoder init, or from the get_format() callback — which is super useful for when the view count or IDs change dynamically during decoding.

Only the base layer is decoded by default.

Note: if you're using the ffmpeg CLI tool, you should be using view specifiers as documented in its manual, rather than the options documented here. Okay? Okay 💅

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. All of them. The whole squad 🫶
view_ids_available (MV-HEVC) — the caller may read this option 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 then the value is informational only and may be incorrect (e.g. when the stream contains multiple distinct VPS NALUs). So like… read it at your own risk, bestie 😭

view_pos_available (MV-HEVC) — the caller may read this option 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 the view with ID view_ids_available[i].

Same validity restrictions as for view_ids_available apply to this option. Same rules, different option, no cap.

`rawvideo` 🎥

Raw video decoder.

This decoder decodes rawvideo streams.

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.

libdav1d allows libavcodec to decode the AOMedia Video 1 (AV1) codec. It 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 — sets the max amount of frames the decoder may buffer internally. The default value is 0 (autodetect).
filmgrain — applies 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. Don't say I didn't warn you 💀

oppoint — selects an operating point of a scalable AV1 bitstream (0 - 31). Defaults to the internal default of the library.
alllayers — outputs 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 the davs2 library. Short and sweet, we love that for her.

`libuavs3d` ⭐

AVS3-P2/IEEE1857.10 video decoder.

libuavs3d allows libavcodec to decode AVS3 streams. It 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 — sets the 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, bestie 💖

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

threads (threads) — force the use of a specific number of threads.

QSV Decoders 🚀

The family of Intel QuickSync Video decoders (VC1, MPEG-2, H.264, HEVC, JPEG/MJPEG, VP8, VP9, AV1, VVC). A whole dynasty, honestly.

Common Options

The following options are supported by all qsv decoders:

async_depth — internal parallelization depth. The higher the value, the higher the latency. Choose wisely 🫠
gpu_copy — a GPU-accelerated copy between video and system memory. The 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. Simple. Effective. Iconic.

Options

custom_stride — sets 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, that one 🌟

🎧 AUDIO DECODERS 🎧

okay bestie so this is the part where we meet the decoders that turn your squiggly audio bytes into actual sound 💖 here's the rundown on some of the ones chillin' in the codebase right now.

🌀 ac3

the AC-3 audio decoder, periodt.

this one implements part of ATSC A/52:2010 and ETSI TS 102 366, plus the undocumented RealAudio 3 (a.k.a. dnet). yes, the undocumented one — she read the forbidden texts, no cap 😭

✨ AC-3 Decoder Options ✨

-drc_scale value — the Dynamic Range Scale Factor, aka the factor applied to dynamic range values from the AC-3 stream. it's applied exponentially, fyi. default value is 1. there are 3 notable scale factor ranges and they're giving very different vibes:
- 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 lands somewhere between full range and full compression, iykyk.
- drc_scale > 1 — DRC enabled. applies drc_scale asymmetrically. loud sounds are fully compressed, soft sounds get enhanced ✨

🌸 flac

the FLAC audio decoder 🦋

this decoder aims to implement the complete FLAC specification from Xiph. ambitious queen, we love to see it 👑

🎀 FLAC Decoder options 🎀

-use_buggy_lpc — so the lavc FLAC encoder used to produce buggy streams with high lpc values (like the default value, lol 💀). this option makes it possible to decode such streams correctly by using lavc's old buggy lpc logic for decoding. yes bestie, we're using the bug on purpose to fix the bug. iconic.

🧚 ffwavesynth

internal wave synthesizer 🌈

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 like… don't ask, she's mysterious.

💜 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 ⭐

🍓 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 💅 versatile.

🦋 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 like, off by default, you gotta opt in 🫶

👑 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.

heads up babes: that --enable-nonfree is right there in the configure line — it's what the build needs, fr fr.

💅 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 🫶 good to know, right?

💖 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 ✨ same energy, wideband edition.

🌈 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 💜 and that's on having options.

✨ SUBTITLES DECODERS ✨

Okay bestie, this is the subtitle decoder zone — where the little words at the bottom of your screen get their whole personality. Let's go 💅

🎀 libaribb24

A decoder for ARIB STD-B24 captions. Yes, it's a whole standard, and yes it's iconic.

It implements profiles A and C of the ARIB STD-B24 standard. Slay.

💖 libaribb24 Decoder Options

-aribb24-base-path path — Sets the base path for the libaribb24 library. This is what it uses for reading config files (for custom unicode conversions) and for dumping non-text symbols as images under that location. Unset by default — no cap, it just vibes with no path until you give it one 🫶
-aribb24-skip-ruby-text boolean — Tells the decoder wrapper to skip text blocks that contain half-height ruby text. Enabled by default ✨

🦋 libaribcaption

Yet another ARIB STD-B24 caption decoder, but this one uses the external libaribcaption library. More decoders, more fun, periodt.

It implements profiles A and C of the Japanese ARIB STD-B24 standard, Brazilian ABNT NBR 15606-1, and the Philippines version of ISDB-T. Trilingual legend behavior fr fr.

It requires the presence of the libaribcaption headers and library (<https://github.com/xqq/libaribcaption>) during configuration. You need to explicitly configure the build with "--enable-libaribcaption". And listen — if both libaribb24 and libaribcaption are enabled, the libaribcaption decoder precedes. It's giving main character energy 👑

🌸 libaribcaption Decoder Options

-sub_type subtitle_type — Specifies the format of the decoded subtitles. The options are:
- 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. iykyk 💅

-caption_encoding encoding_scheme — Specifies the encoding scheme of input subtitle text. Your choices:
- 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 the 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 — ARIB STD-B24 specifies that some captions may be displayed at different positions at a time (multi-rectangle subtitle). 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. 🫶

-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 fonts support 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. 💜
-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 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 actual input video size.

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

⭐ libaribcaption decoder usage examples

Display MPEG-TS file with ARIB subtitle by "ffplay" tool:

ffplay -sub_type bitmap MPEG.TS

Display MPEG-TS file with input frame size 1920x1080 by "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 — the math-y glow-up option:
- -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.
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; the same subtitles can also be found in VobSub file pairs and in some Matroska files. Retro but make it fashion 💅

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.

The format for this option is a string containing 16 24-bit 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". So many hex codes, so little time 😭

ifo_palette — Specify the IFO file from which the global palette is obtained. (experimental) — it's giving work in progress, handle with care 🧚
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". 🔥

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 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. Ngl, vintage transmission problems are a mood 💀
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. Cleaning up the mess, love that for us 🎀
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. Live forever or until the next thing happens, honestly same 💅
txt_transparent — Force 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.

🔥 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". Look at all those filters, obsessed 😍

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". Mix and match, it's your build, your rules 👑

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 '='. Periodt ✨

ffmpeg -i INPUT -c:v copy -bsf:v filter1[=opt1=str1:opt2=str2][,filter2] OUTPUT

Okay so below is the ✨ current ✨ roster of bitstream filters, with all their lil parameters (if they have any). Buckle up bestie, some of these have a lot of options 💅

🎀 aac_adtstoasc

Converts an MPEG-2/4 AAC ADTS stream into an MPEG-4 Audio Specific Configuration bitstream. Glow-up, but for headers.

What it actually does: builds an MPEG-4 AudioSpecificConfig out of an MPEG-2/4 ADTS header, then yeets the ADTS header entirely. Bye 👋

And it's required in some situations — for example when you're copying an AAC stream from a raw ADTS AAC container or an MPEG-TS container over to MP4A-LATM, to an FLV file, or to MOV/MP4 files and related formats like 3GP or M4A. Good news tho: for MP4A-LATM and MOV/MP4 and friends it gets auto-inserted, so you don't even have to think about it 🫶

🦋 av1_metadata

Modify metadata embedded in an AV1 stream. Little header surgery, no cap.

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, transfer_characteristics, 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

Removes zero padding at the end of a packet. Just… monch. 🍓

💜 dca_core

Extracts the core from a DCA/DTS stream, dropping extensions such as DTS-HD. It's giving "keep the main character, lose the entourage".

👑 dovi_rpu

Manipulate Dolby Vision metadata in a HEVC/AV1 bitstream, optionally enabling metadata compression. Fancy fancy.

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

Adds extradata to the beginning of the filtered packets — except when those packets already exactly begin with the extradata that's intended to be added. We're not double-dipping, bestie.

freq — The additional argument specifies which packets should be filtered. It accepts the values:
- k, keyframe — add extradata to all key packets
- e, all — add extradata to all packets

If not specified it is assumed k.

For example, the following ffmpeg command forces a global header (thus disabling individual packet headers) in the H.264 packets generated by the libx264 encoder, but 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 which are marked as damaged get replaced by blocks of the specified color. It's giving "we fixed it in post" ✨

color — The color to replace damaged blocks by
sta — A 16 bit mask which specifies which of the 16 possible error status values are to be replaced by colored blocks. 0xFFFE is the default, which replaces all 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

🌈 eac3_core

Extracts the core from an E-AC-3 stream, dropping extra channels. Same energy as dca_core but for E-AC-3, iykyk.

🫶 eia608_to_smpte436m

Convert from a EIA_608 stream to a SMPTE_436M_ANC data stream, wrapping the closed captions in CTA-708 CDP VANC packets. Very niche, very slay.

line_number — Choose which line number the generated VANC packets should go on. You generally want either line 9 (the default) or 11.
wrapping_type — Choose the SMPTE 436M wrapping type, defaults to vanc_frame. It accepts the values:
- vanc_frame — VANC frame (interlaced or segmented progressive frame)
- vanc_field_1
- vanc_field_2
- vanc_progressive_frame
sample_coding — Choose the SMPTE 436M sample coding, defaults to 8bit_luma. It accepts the values:
- 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 — Set the CDP's cdp_frame_rate field. 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

Extract the in-band extradata. Okay so context first, because this one's actually kinda interesting:

Certain codecs allow the long-term headers (e.g. MPEG-2 sequence headers, or H.264/HEVC (VPS/)SPS/PPS) to be transmitted either "in-band" (i.e. as a part of the bitstream containing the coded frames) or "out of band" (e.g. on the container level). That latter form is called "extradata" in FFmpeg terminology. The more you know 💫

This bitstream filter detects the in-band headers and makes them available as extradata.

remove — When this option is enabled, the long-term headers are removed from the bitstream after extraction.

💀 filter_units

Remove units with types in — or not in — a given set from the stream. Bouncer energy: some units get in, some don't.

pass_types — List of unit types or ranges of unit types to pass through while removing all others. This is specified as a '|'-separated list of unit type values or ranges of values with -.
remove_types — Identical to pass_types, except the units in the given set removed and all others passed through.

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 heads up: if the stream contains inline parameter sets then the output may be unusable if they are removed. Consider yourself warned bestie 😭

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

Extract the Rgb or Alpha part of an HAPQA file, without recompression, in order to create an HAPQ or an HAPAlphaOnly file. No quality loss, no drama 💅

texture — Specifies the 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

Modify metadata embedded in an H.264 stream. Yes, another one. Yes, it's also iconic.

📌 Heads up: this list continues past what you see here — more options are coming right up in the next chunk. Don't panic, we're not skipping anything 🫶

aud — Insert or remove AUD NAL units in all access units of the stream.
- pass
- insert
- remove

Default is pass.

sample_aspect_ratio — Set the sample aspect ratio of the stream in the VUI parameters. See H.264 table E-1.
overscan_appropriate_flag — Set whether the stream is suitable for display using overscan or not (see H.264 section E.2.1).
video_format, video_full_range_flag — Set the video format in the stream (see H.264 section E.2.1 and table E-2).
colour_primaries, transfer_characteristics, matrix_coefficients — Set 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 — Set the chroma sample location in the stream (see H.264 section E.2.1 and figure E-1).
tick_rate — Set 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 represents the field rate of the stream (double the frame rate).
fixed_frame_rate_flag — Set whether the stream has fixed framerate — typically this indicates that the framerate is exactly half the tick rate, but the exact meaning is dependent on interlacing and the picture structure (see H.264 section E.2.1 and table E-6).

`zero_new_constraint_set_flags` 🎀

okay so this one zeroes out constraint_set4_flag and constraint_set5_flag inside the SPS. bestie, those bits were reserved in an older version of the H.264 spec, which means some hardware decoders straight up require them to be zero 💀. the slay part? zeroing them still leaves you with a perfectly valid bitstream, so no cap, nothing breaks. ✨

`crop_left`, `crop_right`, `crop_top`, `crop_bottom` 🦋

these four set the frame cropping offsets in the SPS. if the stream is already cropped, these values will replace the current ones — like a glow-up for your crop box 💅.

heads up though: these fields are set in pixels, and some sizes may just not be representable if the chroma is subsampled or the stream is interlaced (see H.264 section 7.4.2.1.1). so like, do a vibe check on your numbers first 🙈.

`sei_user_data` ⭐

this one inserts a string as SEI unregistered user data. the argument has to be of the form UUID+string, where the UUID is hex digits possibly separated by hyphens, and the string can literally be anything you want 🫶.

for example, 086f3693-b7b3-4f2c-9653-21492feee5b8+hello will insert the string hello associated with the given UUID. iconic.

`delete_filler` 💀

deletes both filler NAL units and filler SEI messages. that's it. that's the whole thing. bye filler 👋

`display_orientation` 🌸

insert, extract or remove Display orientation SEI messages. see H.264 section D.1.27 and D.2.27 for the syntax and semantics (iykyk).

the modes are:

pass
insert
remove
extract

default is pass.

insert mode works in conjunction with the rotate and flip options, which is lowkey cute. any pre-existing Display orientation messages will be removed in insert or remove mode. and extract mode? it attaches the display matrix to the packet as side data ✨.

`rotate` 🌀

set rotation in display orientation SEI (anticlockwise angle in degrees, bestie — anticlockwise, don't forget). range is -360 to +360. default is NaN.

`flip` 🪞

set flip in display orientation SEI. your options:

horizontal
vertical

default is unset.

`level` 👑

set the level in the SPS. refer to H.264 section A.3 and tables A-1 to A-5, fr fr.

the argument must be one of these three flavors:

the name of a level (for example, 4.2)
a level_idc value (for example, 42)
the special name auto, which tells the filter to attempt to guess the level from the input stream properties 🔥

`h264_mp4toannexb` 🧚

convert 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 your thing refuses to play, this might be why 🙈.

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. so like, you don't even have to ask 💅

`h264_redundant_pps` 🎀

this applies a specific fixup to some Blu-ray BDMV H264 streams which contain redundant PPSs. those PPSs modify irrelevant parameters of the stream, confusing other transformations which require the correct extradata (rude, honestly).

the encoder used on these impacted streams adds extra PPSs throughout the stream, varying the initial QP and whether weighted prediction was enabled. this causes issues after copying the stream into a global header container, as the starting PPS is not suitable for the rest of the stream 😭. one side effect, for example, is seeking will return garbled output until a new PPS appears. literally the worst.

this BSF removes the extra PPSs and rewrites the slice headers such that the stream uses a single leading PPS in the global header, which resolves the issue. queen behavior 👑✨

`hevc_metadata` 🌈

modify metadata embedded in an HEVC stream. let's go through the options one by one, bestie:

`aud`

insert or remove AUD NAL units in all access units of the stream. pick your fighter:

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). three tasty little options, one colour glow-up 💖

`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 🕐.

ngl, note that it is likely to be overridden by container parameters when the stream is in a container. so don't get too attached, bestie.

`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 — so it's a package deal, no solo acts 🙈.

`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). math is math, unfortunately 💀

`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)
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) — same energy as the H.264 one, different codec 💅.

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. auto-magic, no cap ✨

`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. so like, check your version first, bestie 🙈

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.

so here's the lore: 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."

so 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. it's giving magical little JPEG makeover 🧚✨

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. short, sweet, slays.

`mov2textsub` 🎀

extract a representable text file from MOV subtitles, stripping the metadata header from each subtitle packet.

see also the text2movsub filter — they're like a matching set, bestie 🫶

`mpeg2_metadata` 🌸

modify metadata embedded in an MPEG-2 stream. options time:

`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). so like, choose wisely 😭

`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). close enough is literally the whole vibe here 🙈

`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). slay, colour trio 💖

`mpeg4_unpack_bframes` 🔥

unpack DivX-style packed B-frames.

okay so DivX-style packed B-frames are not valid MPEG-4 — they 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. not them being a whole mess 😭

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 — like a lil chaos agent, but make it scientific 🧪✨

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`

okay so this one takes an expression that gets evaluated per-packet, and whatever it spits out decides whether that packet gets yeeted. Positive value → packet dropped, periodt. Negative value → it's a variable chance of dropping, roughly inverse in proportion to how big the value is (bigger negative = less likely). Default is 0, which means nothing gets dropped. Peep the variable list below, bestie.

💅 `dropamount`

takes a non-negative integer, and gives you a variable chance of the packet getting dropped — again, roughly inverse in proportion to the value. Default 0 = no drops, no drama. This one's kept around for backwards compatibility and it's literally equivalent to setting drop to a negative value of the same magnitude, i.e. "dropamount=4" is the same as "drop=-4". Also: it gets ignored if drop is also specified. so don't do both and expect magic, okay?

Both amount and drop accept expressions containing these 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. literally does nothing. iconic behaviour honestly.

🎀 `pcm_rechunk`

Repacketize PCM audio to a fixed number of samples per packet, or a fixed packet rate per second. Think of it as the asetnsamples audio filter's cousin — except it works on audio packets instead of audio frames.

nb_out_samples, n — sets 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 pads the last audio packet with silence so it contains the same number of samples as the previous ones (or roughly the same number — see frame_rate). Default value is 1.
frame_rate, r — 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 isn't divisible by the frame rate, then the number of samples won't be constant — it'll vary slightly so each packet starts as close to the frame boundary as possible. Using this option takes precedence over nb_out_samples.

You can generate that well known 1602-1601-1602-1601-1602 pattern of 48kHz audio for NTSC frame rate using the frame_rate option. slay:

ffmpeg -f lavfi -i sine=r=48000:d=1 -c pcm_s16le -bsf pcm_rechunk=r=30000/1001 -f framecrc -

🌸 `pgs_frame_merge`

Merges a sequence of PGS Subtitle segments ending with an "end of display set" segment into a single packet. This is required by some containers that support PGS subtitles (muxer "matroska"). so like, not optional, bestie.

💜 `prores_metadata`

Modifies color property metadata embedded in a prores stream. cute.

color_primaries — sets 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 — sets 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 — sets 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`

Removes extradata from packets. It accepts the following parameter:

freq — sets 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`

Sets PTS and DTS in packets. 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 get 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, just so you know 💀):

ffmpeg -i INPUT -c:a copy -bsf:a setts=pts=DTS out.mkv

🫶 `showinfo`

Logs basic packet information. Mainly useful for testing, debugging, and development. your bestie for "wait, what is this packet actually".

🧚 `smpte436m_to_eia608`

Converts 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. bye to everything else ✨

🍓 `text2movsub`

Converts text subtitles to MOV subtitles (as used by the "mov_text" codec) with metadata headers. See also the mov2textsub filter — the other direction, iykyk.

🔥 `trace_headers`

Logs 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. check your build bestie.

👑 `truehd_core`

Extracts the core from a TrueHD stream, dropping ATMOS data. so long, atmos.

💖 `vp9_metadata`

Modifies metadata embedded in a VP9 stream.

color_space — sets 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. Available values:
- unknown
- bt601
- bt709
- smpte170
- smpte240
- bt2020
- rgb
color_range — sets the color range value in the frame header. Note that any value imposed by the color space will take precedence over this value. Available values:
- tv
- pc

🦋 `vp9_superframe`

Merges 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. it's giving reparative surgery 💅

💀 `vp9_superframe_split`

Splits VP9 superframes into single frames. the reverse vibe.

🌟 `vp9_raw_reorder`

Given a VP9 stream with correct timestamps but possibly out of order, inserts additional show-existing-frame packets to correct the ordering. it just wants things in the right order, honestly same.

💅 FORMAT OPTIONS 💅

Okay bestie, so the libavformat library is out here giving you a whole bunch of generic global options — and the iconic part? You can set them on all the muxers and demuxers. Every single one. ✨ And on top of that, each muxer or demuxer might have its own little private options that are specific to that one component, like a signature vibe.

You can set options by specifying -option value in the FFmpeg tools, or by setting the value explicitly in the AVFormatContext options, or by using the libavutil/opt.h API if you're doing the programmatic thing. 🫶

The list of supported options follows, and yes, we're doing all of them:

avioflags flags (input/output) — Possible values are giving:
- direct — reduce buffering. Lowkey less buffering, highkey less latency.
probesize integer (input) — sets the probing size in bytes, i.e. the size of the data to analyze to get stream information. A higher value lets you detect more information in case it's dispersed into the stream, but it will increase latency — trade-off city. Must be an integer not lesser than 32. It's 5000000 by default. 💖
max_probe_packets integer (input) — sets the maximum number of buffered packets when probing a codec. Default is 2500 packets.
packetsize integer (output) — sets packet size. That's it, that's the whole option.
fflags flags — sets format flags. Some are implemented for a limited number of formats, so don't come for me if your fave format doesn't do all of them.

Possible values for input files:

discardcorrupt — discard corrupted packets. Bye, trash. 💀
fastseek — enables fast, but inaccurate seeks for some formats. Speed over precision, we love a trade-off.
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. Fr fr, remember that.
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. Not me needing a whole extra flag, but okay.
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. She's the default for a reason. 👑
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. Important detail, don't skip it.

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. A higher value will enable detecting more accurate information, but will increase latency (again with the latency, I know). It defaults to 5,000,000 microseconds = 5 seconds.
cryptokey hexadecimal string (input) — set decryption key. Secretive. Mysterious. Main character energy. 🕵️♀️
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 though: f_err_detect is deprecated and should be used only via the ffmpeg tool. Don't be out here using the deprecated one, bestie. 🙈

Possible values:

crccheck — verify embedded CRCs.
bitstream — detect bitstream specification deviations.
buffer — detect improper bitstream length.
explode — abort decoding on minor error detection. Dramatic, honestly iconic.
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. The duration is expressed in microseconds, and defaults to 10000000 (10 seconds). Ngl, that's the whole point so let me explain it properly:

To ensure all the streams are interleaved correctly, libavformat will wait 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 which 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, regardless of 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 are 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. So like, everything moves together, no chaos. 🎀

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 that the underlying protocol will decide; 1 enables it, and has the effect of reducing the latency; 0 disables it and may increase IO throughput in some cases. Three whole moods. ✨
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 that the corresponding streams are delayed by the time duration specified in offset. Default value is 0 (meaning that no offset is applied).

format_whitelist list (input) — , separated list of allowed demuxers. By default all are allowed. A little VIP list for your demuxers. 🎟️
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.
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 is aimed at users interested in better durations probing for itself, or indirectly because 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. 😮‍💨 Okay that was a lot but every word of it matters.
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. No notes.
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. 🚨 Bestie I need you to actually read that one, it's not a drill.

🦋 Format stream specifiers 🦋

Format stream specifiers allow selection of one or more streams that match specific properties. It's giving "choose your fighter." 💅

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. So if you want the full lore, that's where it lives.

🌸 DEMUXERS 🌸

Demuxers are configured elements in FFmpeg that can read the multimedia streams from a particular type of file. Think of them as the little readers that crack open a file type and hand you the streams inside — each one knowing its own format's whole personality. ✨

When you configure your FFmpeg build, literally all the supported demuxers are enabled by default — slay, no setup needed 💅. Wanna see the whole roster? Hit up the configure option --list-demuxers and it'll spill the entire list.

You can also turn the whole squad off with --disable-demuxers, then selectively let one back in with --enable-demuxer=DEMUXER, or kick a single one out with --disable-demuxer=DEMUXER. Pick your fighters, bestie 🎮.

Oh and the -demuxers option of the ff* tools will show you the list of enabled demuxers. Want a combined glow-up list of enabled demuxers and muxers? That's -formats. Both, together, iconic.

And now — the description of some of the currently available demuxers. Buckle up 🦋.

🌸 aa 🌸

Audible Format 2, 3, and 4 demuxer.

This one's whole job is demuxing Audible Format 2, 3, and 4 (.aa) files. That's it, that's the vibe.

🎧 aac 🎧

Raw Audio Data Transport Stream AAC demuxer.

This demuxer is used to demux an ADTS input containing a single AAC stream, along with any ID3v1/2 or APE tags hanging out in there. Tags included, bestie 🏷️.

🦋 apng 🦋

Animated Portable Network Graphics demuxer.

This one demuxes APNG files. Here's the tea: all headers — except the PNG signature — up to (but not including) the first fcTL chunk get transmitted as extradata. Then frames are split as being all the chunks between two fcTL ones, or between the last fcTL and IEND chunks. Frame boundaries, sorted ✨.

-ignore_loop bool — Ignore the loop variable in the file if it's set. Default is enabled.
-max_fps int — Maximum framerate in frames per second. Default of 0 imposes no limit, go wild 🏃‍♀️.
-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 demuxer handles ASF files and MMS network streams, both at once, multitasking queen 👑.

-no_resync_search bool — Do not try to resynchronize by looking for a certain optional start code. Like, just don't go looking, okay? 🙈

🎀 concat 🎀

Virtual concatenation script demuxer.

Okay this one's kind of main character energy: it reads a list of files and other directives from a text file, and demuxes them one after the other, as if all their packets had been muxed together. Stitching the whole timeline into one, periodt ✂️.

The timestamps in the files are adjusted so that the first file starts at 0 and each next file starts where the previous one finishes. Lowkey important caveat: this is done globally and may cause gaps if all streams do not have exactly the same length. Slight mismatch → little timeline gaps, ngl 😭.

All files must have the same streams (same codecs, same time base, etc.). No mixing and matching, bestie — they gotta match fr fr.

The duration of each file is used to adjust the timestamps of the next file: if the 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 can be used to override the duration stored in each file. Fix it manually if the vibes are off 💅.

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. Exactly. As. Is. No sneaky BOM 🚨.

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. Roughly-in-point vibes 😬.

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. Overlap is real, bestie.

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. >= Out point? Not on the menu 🙅‍♀️.

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 be able to decode the earliest stream until Out point. Messy but honest 💀.

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. Old and busted 🫠.

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. The fresh new version, we stan 🫶.
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. Order said no, the id said yes 💅.
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. Chaos mode, your call 🙈.

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. Resolution glow-up? Gotta convert, girl ✨.

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.

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. Mirroring the metadata, cute 💖.

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). A whole dictionary of keys, spoiled 💅.

🧚 dvdvideo 🧚

DVD-Video demuxer, powered by libdvdnav and libdvdread.

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). Best effort menu ripping, we respect the hustle 🫡.

Block devices (DVD drives), ISO files, and directory structures are accepted. Activate with -f dvdvideo in front of one of these inputs.

Big ol' warning: 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. Read that again, bestie 😭.

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. Build it right or it's not gonna work ✨.

You will need to provide either the desired "title number" or exact PGC/PG coordinates. Lots of open-source DVD players and tools can help you dig those up. If you don't specify anything, the demuxer just defaults to title 1, which honestly works for a whole lot of discs. BUT — because the format is so flexible, it's recommended you check manually. There are so many discs out there authored strangely or with straight-up invalid headers. 🙈

If your input is an actual real-life DVD drive, heads up: some drives will silently fail on reading bad sectors from the disc and just hand you back random bits instead, which is effectively corrupt data. No cap. This is especially prominent on aging or rotting discs. You'd need a second pass and integrity checks to even detect the corruption. And btw — this is not an FFmpeg issue.

🎬 Background 🎬

DVD-Video is not a directly accessible, linear container format in the traditional sense. 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, in easier terms. 💅

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 not enough to produce the right sequence in many cases. It's giving ✨main character energy✨ from the navigation layer.

There are many other DVD structures (a long subject) that will not be discussed here. NAV packets, in particular, are handled by this demuxer to build accurate timing but 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 to signal 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. 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. 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. 🕹️

This format is used by various Electronic Arts games. iykyk.

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.

🎞️ 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. 🫶

🦋 flv, live_flv, kux

Adobe Flash Video Format demuxer.

This demuxer is used to demux FLV files and RTMP network streams. In case of live network streams, if you force format, you may use live_flv option instead of flv to survive timestamp discontinuities. KUX is a 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 little looping icon that it is. 💖

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 hundredth 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. 🌀

🌈 hls

HLS demuxer — Apple HTTP Live Streaming demuxer.

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". 👑

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 reads from a list of image files, and the list is specified by a pattern. The syntax and the meaning of the pattern is specified by the pattern_type option. 🖼️

The pattern may contain a suffix which is used to automatically determine the format of the images contained in the files.

The size, the pixel format, and the format of each image must be the same for all the files in the sequence — like, all of them, no mixing, bestie. 🙅‍♀️

This demuxer 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 — selects the pattern type used to interpret the filename you handed it. It accepts one of these values:
- none — disables pattern matching, so the video will only contain the specified image. You should use this option if you don't want to create sequences from multiple images and your filenames may contain special pattern characters. 💅
- sequence — selects a sequence pattern type, used 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. 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. 🌀
Default value is sequence.

pixel_format — sets the pixel format of the images to read. If not specified, the pixel format is guessed from the first image file in the sequence.
start_number — sets the index of the file matched by the image file pattern to start to read from. Default value is 0.
start_number_range — sets the index interval range to check when looking for the first image file in the sequence, starting from start_number. Default value is 5.
ts_from_file — if set to 1, will set frame timestamp to modification time of 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, will set 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 is guessed from the first image file in the sequence.
export_path_metadata — if set to 1, will add two extra fields to the metadata found in input, making them also available for other filters (see drawtext filter for examples). Default value is 0. The extra fields are described below:
- 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 for creating 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

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 collection of video game music file emulators. Nostalgia unlocked fr fr 🕹️

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, so pick your fave wisely. Track indexes start at 0. Default is to pick the first track. Number of tracks is exported as 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 amount of reverb. Range 0-100. Default is 0.
reverb_delay — sets 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 cutoff i.e. 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 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 — string which is evaluated using the eval API to assign colors to the generated video stream. Variables which can be used are "x", "y", "w", "h", "t", "speed", "tempo", "order", "pattern" and "row".
video_stream — generate video stream. Can be 1 (on) or 0 (off). Default is 0.
video_stream_w — sets video frame width in 'chars' where one char indicates 8 pixels. Range is 20-512. Default is 30.
video_stream_h — sets video frame height in 'chars' where one char indicates 8 pixels. Range is 20-512. Default is 30.
video_stream_ptxt — print metadata on 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 this 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.
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 a "EIA_608" stream, ignoring all other VANC data 🙈. You can change 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. 💀

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.

🎬 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). The whole iconic family 👑

Registered extensions: mov, mp4, m4a, 3gp, 3g2, mj2, psp, m4b, ism, ismv, isma, f4v

Options

This demuxer accepts the following options:

enable_drefs — enable loading of external tracks, disabled by default. Enabling this can theoretically leak information in some use cases. ⚠️ Like, be aware.
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. No cap, don't just flip this on for random files.
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.
ignore_chapters — okay so this one just tells it to not parse chapters at all, and yes that includes those lil GoPro 'HiLight' tags/moments 🎬. Heads up tho: chapters only get parsed when the input is seekable in the first place. Default is false.
use_mfra_for — for seekable fragmented input, this sets the fragment's starting timestamp from the media fragment random access box, if it's there. The options are:
- auto — auto-detect whether to set mfra timestamps as PTS or DTS (this is the default) ✨
- dts — set mfra timestamps as DTS
- pts — set mfra timestamps as PTS
- 0 — don't use the mfra box to set timestamps at all
use_tfdt — for fragmented input, this sets the fragment's starting timestamp to baseMediaDecodeTime from the tfdt box. Default is enabled, which means it'll prefer using the tfdt box to set DTS 💅. Disable it and it'll use earliest_presentation_time from the sidx box instead. Either way, the timestamp from the mfra box gets used if it's available and use_mfra_for is set to pts or dts.
export_all — exports unrecognized boxes inside the udta box as metadata entries. The first four characters of the box type become the key 🔑. Default is false.
export_xmp — exports the entire contents of the XMP_ box and uuid box as a string with the key "xmp". Small catch: if export_all is set and this option isn't, the contents of the XMP_ box still get exported, but with the key "XMP_" instead. 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 — the fixed key used for handling Audible AAX/AAX+ files. It's been pre-set already, so you really shouldn't need 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 — a 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 — so sometimes a trak's stts box has these very high sample deltas that are actually intended, but usually they're written in error, or they're used to stash a negative value for dts correction when treated as signed 32-bit integers 🫠. This option lets you set an upper limit, and beyond that the delta gets clamped to 1. Values greater than the limit, if negative when cast to int32, are used to adjust onward dts.

The 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 still accommodating 99.9% of the uint32 range.

interleaved_read — interleaves packets from multiple tracks right at the demuxer level. For badly interleaved files, this prevents playback issues caused by large gaps between packets in different tracks, since MOV/MP4 don't have packet placement requirements 💀. But — and it's a big but — this can cause excessive seeking on very badly interleaved files, because of all the 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 you can decrypt them by specifying a 4 byte activation secret. Like so:

ffmpeg -activation_bytes 1CEB00DA -i test.aax -vn -c:a copy output.mp4

🌊 mpegts

MPEG-2 transport stream demuxer. This demuxer accepts the following options:

resync_size — sets the size limit for looking up a new synchronization. Default value is 65536.
skip_unknown_pmt — skips PMTs for programs not defined in the PAT. Default value is 0.
fix_teletext_pts — overrides 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, so set this option to 0 if you want your teletext packet PTS and DTS values left untouched 👀.
ts_packetsize — output option carrying the raw packet size in bytes. It shows the detected raw packet size — you literally cannot set this yourself, bestie.
scan_all_pmts — scan and combine all PMTs. The value is an integer from -1 to 1 (-1 means automatic setting, 1 means enabled, 0 means disabled). Default value is -1.
merge_pmt_versions — reuses existing streams when a PMT's version gets updated and elementary streams move to different PIDs. Default value is 0.
max_packet_size — sets the maximum size, in bytes, of a packet emitted by the demuxer. Payloads above this size get split across multiple packets. Range is 184 to INT_MAX/2. Default is 204800 bytes.

🖼️ mpjpeg

MJPEG encapsulated in multi-part MIME demuxer. This one lets you read MJPEG where each frame is represented as a part of a multipart/x-mixed-replace stream ✨.

strict_mime_boundary — the default implementation applies a relaxed standard to multi-part MIME boundary detection, so that we don't regress with the numerous existing endpoints out there not generating a proper MIME MJPEG stream 😭. Turning this option on by setting it to 1 results in a stricter check of the boundary value.

🎥 rawvideo

Raw video demuxer. This demuxer lets you read raw video data — and because there's no header specifying the assumed video parameters, you have to specify them yourself in order to decode the data correctly. No cap.

This demuxer accepts the following options:

framerate — sets the input video frame rate. Default value is 25.
pixel_format — sets the input video pixel format. Default value is "yuv420p".
stride — sets the frame line size in bytes, only required if there's extra padding. For multiplane formats, stride is a list of line size for each plane.
video_size — sets 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

And to read a rawvideo with ffplay, assuming a pixel format of "yuv420p", a video size of "1080x1920", with 8 bytes of padding in each line of the Y plane, and 4 bytes of padding with the 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 info on the format, see .

This demuxer implements the specification as of March 2024, which has been stable and unchanged since April 2014. Iconic.

✨ 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 gets taken from the current time of day at the time the script is read, and the script layout gets frozen according to that reference. Which means: 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 get shifted accordingly. Facts.

📺 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.

This demuxer accepts the following option:

start_time — sets the start time of the TED talk, in milliseconds. The default is 15000 (15s). It's 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.

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 options:

max_size — see the same option for the wav demuxer.

🌈 wav

RIFF Wave Audio demuxer. This demuxer accepts the following options:

max_size — specifies the maximum packet size in bytes for the demuxed packets. By default this is set to 0, which means a sensible value gets chosen based on the input format 💖.

🎀 METADATA 🎀

Okay bestie, so FFmpeg can literally dump metadata from media files into a simple UTF-8-encoded INI-like text file, and then load it right back with the metadata muxer/demuxer. It’s giving “notes app but for media tags.” 💅

The file format goes like this:

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 the 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 the 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 , with a trailing space, and value is bar, with a leading space).

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. The workflow is lowkey iconic, ngl.

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. Cute, right? 🫶

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 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. It’s basically the backstage pass system for I/O. 🎟️

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.

You can disable all the protocols using the configure option --disable-protocols, and selectively enable a protocol using the option --enable-protocol=PROTOCOL, or you can disable a particular protocol using the option --disable-protocol=PROTOCOL.

The option -protocols of the ff* tools will display the list of supported protocols.

All protocols accept the following options:

rw_timeout — Maximum time to wait for (network) read/write operations to complete, in microseconds.

A description of the currently available protocols follows. Buckle up, bestie. 🦋

🐇 amqp

Advanced Message Queueing Protocol (AMQP) version 0-9-1 is a broker based publish-subscribe communication protocol.

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. It fills data in a background thread, to decouple I/O operation from demux thread. Main character energy for your I/O. 👑

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. It caches the input stream to temporary file. It brings seeking capability to live streams. No cap, it’s literally giving rewind to the live vibe. ⏪

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.

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. Shells stay dramatic, we love that for them. 💀

📜 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. It’s the “everything is a URI if you believe hard enough” protocol. 🌈

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 corresponds to a regular file. fd protocol doesn't support passing a file descriptor via URL for security.

This protocol accepts the following options:

blocksize — Sets the max block size for I/O ops, in bytes. Default is "INT_MAX", which means the requested block size is literally not limited. Setting it reasonably low improves how fast the user termination request gets handled, which is super valuable if data transmission is slow. 🐢
fd — Set file descriptor.

🌸 file

File access protocol, bestie. Read from or write to a file. 💖

A file URL can look like this:

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.

🌈 gophers

Gophers protocol. The Gopher protocol with TLS encapsulation. 🔒

👑 http

HTTP (Hyper Text Transfer Protocol). This protocol accepts the following options:

seekable — Control seekability of connection. If set to 1 the resource is supposed to be seekable, if set to 0 it is assumed not to be seekable, if set to -1 it will try to autodetect if it is seekable. Default value is -1. 🔍
chunked_post — If set to 1 use chunked Transfer-Encoding for posts, default is 1. 📦
http_proxy — set HTTP proxy to tunnel through e.g. http://example.com:1234. 🛰️
headers — Set custom HTTP headers, can override built in default headers. 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 not specified the protocol will use a string describing the libavformat build. ("Lavf/<version>") 🕵️
referer — Set the Referer header. Include 'Referer: URL' header in HTTP request. 🔗
multiple_requests — Use persistent connections if set to 1, default is 0. ♻️
request_size — Limit the size of requests made. This is useful for some pathological servers that throttle unbounded range requests, as well as when expecting to seek frequently. Disabled (set to 0) by default.

Note that if enabling 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 allows FFmpeg to reuse a single HTTP connection wherever possible. 🫶

initial_request_size — Limit 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. Lasts until the demuxer makes a read request larger than this size (without a seek in between), after which the implementation will continue using requests as usual. Disabled (set to 0) by default.

Note that if enabling 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 — Export the MIME type. 🏷️
http_version — Exports the HTTP response version number. Usually "1.0" or "1.1". 🔢
cookies — Set 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. It should be polled in regular intervals by applications interested in mid-stream metadata updates. ⏱️
metadata — Set an exported dictionary containing Icecast metadata from the bitstream, if present. Only useful with the C API. 🧊
auth_type — Set HTTP authentication type. No option for Digest, since this 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 the HTTP basic authentication.

⚠️ Security tea: 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. 🚨

send_expect_100 — Send an Expect: 100-continue header for POST. If set to 1 it will send, if set to 0 it won't, if set to -1 it will try to send if it is 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 do not 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 — set this to 1 and boom, you've got an experimental HTTP server 💅. Use it as an output option to send data, or as an input option to read data from a client via HTTP POST. Set it to 2 and you get an experimental multi-client HTTP server — but heads up bestie, that one is not yet implemented in ffmpeg.c so you literally must not use it 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 that a client requested, when the experimental HTTP server is the one in play.
reply_code — the HTTP code that gets sent back to the client, again only when the experimental HTTP server is running the show.
short_seek_size — sets the threshold, in bytes, for when readahead should be preferred over doing a seek + a brand new HTTP request. Genuinely useful, like, for making sure you stay on the same connection while reading big video packets with lil baby audio packets sandwiched in between 🫶.

🍪 HTTP Cookies 🍪

Some HTTP requests are gonna get straight-up denied unless you pass cookie values along with the request 😭. The cookies option is how you hand those over. Bare minimum, every cookie has to specify a value plus a path and a domain. Any HTTP request matching both the domain and the path will automatically get the cookie value dropped into the HTTP Cookie header field. You can separate multiple cookies with a newline. Slay.

The required syntax to play a stream while specifying a cookie is:

ffplay -cookies "nlqptid=nltid=tsn; path=/; domain=somedomain.com;" http://somedomain.com/somestream.m3u8

🧊 Icecast 🧊

Icecast protocol — for streaming to Icecast servers.

This protocol accepts the following options:

ice_genre — sets the stream genre.
ice_name — sets the stream name.
ice_description — sets the stream description.
ice_url — sets the stream's website URL.
ice_public — sets whether the stream should be public or not. Default is 0 (aka not public, private vibes only).
user_agent — overrides the User-Agent header. If you don't specify one, a string of the form "Lavf/<version>" gets used.
password — sets the Icecast mountpoint password.
content_type — sets the stream content type. You must set this if it's anything other than audio/mpeg.
legacy_icecast — enables support for Icecast versions < 2.4.0, which don't support the HTTP PUT method and use the SOURCE method instead.
tls — establishes a TLS (HTTPS) connection to Icecast.

icecast://[<username>[:<password>]@]<server>:<port>/<mountpoint>

🪐 ipfs 🪐

InterPlanetary File System (IPFS) protocol support, bestie. You can access files stored on the IPFS network through what are called gateways — these are http(s) endpoints. This protocol wraps the IPFS native protocols (ipfs:// and ipns://) so they get sent to one of those gateways. Users can (and honestly should) host their own node, which means this protocol will use your local gateway to reach files on the IPFS network 🌈.

This protocol accepts the following options:

gateway — defines which gateway to use. When it's not set, the protocol first tries to locate the local gateway by checking $IPFS_GATEWAY, then $IPFS_PATH, then "$HOME/.ipfs/", in that exact order 👑

You can use this protocol in 2 ways. With IPFS:

ffplay ipfs://<hash>

Or with the IPNS protocol (IPNS is the mutable version of IPFS):

ffplay ipns://<hash>

📺 mmst 📺

MMS (Microsoft Media Server) protocol over TCP. That's it, that's the whole vibe.

📺 mmsh 📺

MMS (Microsoft Media Server) protocol over HTTP.

The required syntax is:

mmsh://<server>[:<port>][/<app>][/<playpath>]

🔐 md5 🔐

MD5 output protocol.

It computes the MD5 hash of the data being written, and on close it writes that hash to the designated output — or to stdout if you didn't specify one. Super handy for testing muxers without actually writing a real file ✨.

Some examples follow.

# 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) need the output protocol to be seekable, so they will fail with the MD5 output protocol. Not a drill 💀

🚰 pipe 🚰

UNIX pipe access protocol. Read and write from UNIX pipes, iconic behaviour.

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). If number isn't specified at all, the stdout file descriptor is used by default for writing, and stdin for reading 🫶.

For example, to read from stdin with ffmpeg:

cat test.wav | ffmpeg -i pipe:0
# ...this is the same as...
cat test.wav | ffmpeg -i pipe:

For writing 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 means the requested block size doesn't get limited at all. Setting this reasonably low improves how fast your termination requests get reacted to — genuinely valuable when data transmission is slow 🐌.
fd — sets the file descriptor.

⚠️ And again, bestie: some formats (typically MOV) require the output protocol to be seekable, so they'll fail with the pipe output protocol. Same energy as the MD5 situation 😭

📡 prompeg 📡

Pro-MPEG Code of Practice #3 Release 2 FEC protocol.

The Pro-MPEG CoP#3 FEC is a 2D parity-check forward error correction mechanism for MPEG-2 Transport Streams sent over RTP. Fancy, right? 💅

This protocol must be used in conjunction with the "rtp_mpegts" muxer and the "rtp" protocol. No solo missions.

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. Reliable, like your bestie who always texts back.

The accepted options are:

rist_profile — supported values:
- simple
- main — this one is the default.
- advanced
buffer_size — sets the internal RIST buffer size in milliseconds, used for retransmission of data. Default value is 0, which means "use 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 (versus the librist default of 1024).
overrun_nonfatal=1|0 — survive in case of a librist fifo buffer overrun. Default value is 0.
pkt_size — sets the maximum packet size for sending data. 1316 by default.
log_level — sets the 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 an override of the encryption secret; by default it's unset.
encryption — set the encryption type; by default it's 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. Classic, iconic, still here.

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 it's 1935).
app — 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 also override the value parsed from the URI through the "rtmp_app" option 👑
playpath — the path or name of the resource to play, with reference to the application specified in app; it may be prefixed by "mp4:". You can also override the value parsed from the URI through the "rtmp_playpath" option.
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 "AVOption"s):

rtmp_app — name of the application to connect to 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 😤
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>).)
rtmp_flush_interval — how many packets get flushed in the same request (this one's RTMPT only bestie 💅). Default is 10.
rtmp_live — says "this media is a live stream, ok?". When it's live, no resuming and no seeking — like, it is what it is, you can't scrub it 😭. 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 isn't found, it plays the recorded stream instead. The other possible values are "live" and "recorded".
rtmp_pageurl — URL of the web page where the media was embedded. By default, no value will be sent.
rtmp_playpath — the stream identifier to play or to publish. This option overrides the parameter specified in the URI.
rtmp_subscribe — name of the live stream to subscribe to. By default no value will be sent. It's 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 the player swf file; it computes the hash/size automatically ✨.
rtmp_tcurl — URL of the target stream. Defaults to proto://host[:port]/app.
tcp_nodelay=1|0 — sets 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 that reduces the efficiency / effect of TCP_NODELAY. So like… it's giving "the setting is there but the vibe is mid" 💀.

tcp_keepalive=1|0 — enables 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.

So 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

And 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.

The 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. It's giving cryptography class but make it ✨streaming✨.

rtmps 🔒

Real-Time Messaging Protocol over a secure SSL connection.

The Real-Time Messaging Protocol (RTMPS) is used for streaming multimedia content across an encrypted connection. Secure lil guy 🎀.

rtmpt 🧚

Real-Time Messaging Protocol tunneled through HTTP.

The Real-Time Messaging Protocol tunneled through HTTP (RTMPT) is used for streaming multimedia content within HTTP requests to traverse firewalls. Firewall said "no" and RTMPT said "bet" 💅.

rtmpte 👑

Encrypted Real-Time Messaging Protocol tunneled through HTTP.

The Encrypted Real-Time Messaging Protocol tunneled through HTTP (RTMPTE) is used for streaming multimedia content within HTTP requests to traverse firewalls.

rtmpts ⭐

Real-Time Messaging Protocol tunneled through HTTPS.

The Real-Time Messaging Protocol tunneled through HTTPS (RTMPTS) is used for streaming multimedia content within HTTPS requests to traverse firewalls.

libsmbclient 🦋

libsmbclient permits one to manipulate CIFS/SMB network resources.

Following syntax is required, no cap:

smb://[[domain:]user[:password@]]server[/share[/path[/file]]]

This protocol accepts the following options.

timeout — sets the timeout in milliseconds of socket I/O operations used by the underlying low level operation. By default it's set to -1, which means the timeout is not specified.
truncate — truncate existing files on write, if set to 1. A value of 0 prevents truncating. Default value is 1.
workgroup — sets the workgroup used for making connections. By default the workgroup is not specified.

For more information see: <http://www.samba.org/>.

libssh 🌸

Secure File Transfer Protocol via libssh.

Read from or write to remote resources using SFTP protocol.

Following syntax is required:

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 set to -1, which means the timeout is not specified.
truncate — truncate existing files on write, if set to 1. A value of 0 prevents truncating. Default value is 1.
private_key — specifies 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.

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 👀.

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).

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.

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 "AVOption"s).

The list of supported options follows, bestie ✨:

ttl=n — set the TTL (Time-To-Live) value (for multicast only).
rtcpport=n — set the remote RTCP port to n.
localport, local_rtpport, localrtpport=n — set the local RTP port to n.

Using the localport option name is deprecated and should not be used 🙈.

local_rtcpport, localrtcpport=n — set the local RTCP port to n.
pkt_size=n — set max packet size (in bytes) to n.
buffer_size=size — set the maximum UDP socket buffer size in bytes.
connect=0|1 — do a connect() on the UDP socket (if set to 1) or not (if set to 0).
sources=ip[,ip] — list allowed source IP addresses.
block=ip[,ip] — list disallowed (blocked) source IP addresses.
write_to_source=0|1 — send 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 — set timeout (in microseconds) of socket I/O operations to n.

Important notes, don't skip these 🔥:

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, it is 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 a RTSP url is:

rtsp://<hostname>[:<port>]/<path>

Options can be set on the ffmpeg/ffplay command line, or set in code via "AVOption"s or in avformat_open_input.

Muxer 💅

The following options are supported.

rtsp_transport — set 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 — set 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 — set minimum local UDP port. Default value is 5000.
max_port — set maximum local UDP port. Default value is 65000.
buffer_size — set the maximum socket buffer size in bytes.
pkt_size — set 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 — set 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 the lower transport protocol, which is lowkey so useful for slipping past proxies and is widely used for security consideration, bestie 🔐

You can specify multiple lower transport protocols, and in that case they get tried one at a time (so if the setup of one fails, the next one is tried, no big deal). For the muxer, only the tcp and udp options are supported, iykyk.

🦋 `rtsp_flags`

Set RTSP flags. The following values are accepted:

filter_src — accept packets only from the negotiated peer address and port. Like, boundaries, we stan 💅
listen — act as a server, listening for an incoming connection. main character energy fr
prefer_tcp — try TCP for RTP transport first, if TCP is available as RTSP RTP transport
satip_raw — export the raw MPEG-TS stream instead of demuxing it. The flag will simply write out the raw stream, with the original PAT/PMT/PIDs intact. no touchy, just vibes ✨

Default value is none.

🎀 `allowed_media_types`

Set the media types to accept from the server. The following flags are accepted:

video
audio
data
subtitle

By default it accepts all media types. generous queen 🫶

💜 `min_port`

Set the minimum local UDP port. Default value is 5000.

💜 `max_port`

Set the maximum local UDP port. Default value is 65000.

🌸 `listen_timeout`

Set the 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. literally forever, bestie 😭

🧚 `reorder_queue_size`

Set the number of packets to buffer for handling of reordered packets.

⭐ `timeout`

Set the socket TCP I/O timeout in microseconds (yes, microseconds, do not come for me).

🦋 `user_agent`

Override the User-Agent header. If not specified, it defaults to the libavformat identifier string.

🍓 `buffer_size`

Set 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 — rude, honestly). 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. iconic behaviour tbh 👑

✨ 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 — 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. it's giving "hey besties, stream's over here" 📢

🌸 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 — specify the port to send the announcements on, defaults to 9875 if not specified.
ttl=ttl — specify the time to live value for the announcements and RTP packets, defaults to 255.
same_port=0|1 — if set to 1, send all RTP streams on the same port pair. If zero (the default), all streams are sent on unique ports, with each stream on a port 2 numbers higher than the previous. VLC/Live555 requires this to be set to 1, to be able to receive the stream. 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 given to the demuxer is:

sap://[<address>][:<port>]

address is the multicast address to listen for announcements on; if omitted, the default 224.2.127.254 (sap.mcast.net) is used. port is the port that is listened on, 9875 if omitted.

The demuxers listens for announcements on the given address and port. Once an announcement is received, it tries to receive that particular stream. obsessed with the follow-through ✨

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.

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 "AVOption"s). yes, both ways, we're flexible 💅

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. go off, unlimited 🫶

⭐ `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 "AVOption"s).

The list of supported options follows.

connect_timeout=milliseconds — connection timeout; 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). ngl this one's messy, but we move 😭
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, therefore 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). 🔐 once it's gone it's gone bestie
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 — this is the lil gap between "hey bestie here's ur new encryption key ✨" and the moment the switchover actually happens. Lowkey it also covers the next interval — the one between switchover happening and the old key getting decommissioned. Default is -1, and -1 means auto (0x1000 in the srt library). Range is integers from 0 to INT_MAX. 🎀
snddropdelay=microseconds — the sender's extra delay before it starts dropping packets, added on top of the default drop delay time interval. Special value -1: don't drop packets on the sender at all. Not one, not ever. 🙈
payload_size=bytes — sets the maximum declared size of a packet transferred during a single call to the sending function in Live mode. Use 0 if this value isn't used (which is the default in file mode). Default is -1 (automatic), which typically means MPEG-TS; if you're gonna use SRT to send any different kind of payload — like, say, 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. Same vibes, different name. 🦋
peerlatency=microseconds — the latency value (as described down in rcvlatency) that the sender side sets as a minimum value for the receiver. It's giving "here's the floor, bestie." ⭐
pbkeylen=bytes — sender encryption key length, in bytes. Can only be set to 0, 16, 24 and 32. Enables sender encryption if not 0. Not required on the receiver (set it to 0 there), key size is obtained from the sender in the 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. Think of it as buffer time big enough to cover the time spent 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 option's value and the value of peerlatency set by the peer side. Before version 1.3.0 this option was only available as latency. 😭
recv_buffer_size=bytes — set the UDP receive buffer size, expressed in bytes. 🫶
send_buffer_size=bytes — set the UDP send buffer size, expressed in bytes. 🫶
timeout=microseconds — set raise error timeouts for read, write and connect operations. Heads up: the SRT library has its own internal timeouts that can be controlled separately, and the value you set here is only a cap on those. 👑
tlpktdrop=1|0 — Too-late Packet Drop. When enabled on the receiver, it skips missing packets that haven't 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 the 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 the send buffer size, expressed in bytes. 🌸
rcvbuf=bytes — set the receive buffer size, expressed in bytes.

Bestie listen to me: the 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 the 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 this whole mechanism is turned off, and the loss report is always sent immediately upon experiencing a "gap" in sequences. 🧚
minversion — the minimum SRT version that is required from the peer. A connection to a peer that does not satisfy the minimum version requirement will be rejected. Periodt.

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 prior to 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 doesn't enforce any special interpretation of the contents of this string. Also 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. Crisis averted. 🎀
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. No mismatched vibes allowed. 💅
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 to use 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 a 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. Lowkey 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. Very secure, very demure. ✨

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 — no cap, if it can't seek, it can't do the thing. 😭

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 |. It's giving "why choose one when you can have both." 💅

tee:file://path/to/local/this.avi|file://path/to/local/that.avi

⭐ tcp ⭐

Transmission Control Protocol. The reliable one. The one who texts back. 🫶

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 AVOptions).

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 in bytes. 💖
send_buffer_size=bytes — set send buffer size, expressed in 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. Ngl that's just how it is bestie. 💀

tcp_mss=bytes — set maximum segment size for outgoing TCP packets, expressed in bytes. 📦

okay so real talk — here's how you actually spin up a listening TCP connection with ffmpeg and then go vibe-check it with ffplay:

ffmpeg -i <input> -f <format> tcp://<hostname>:<port>?listen
ffplay tcp://<hostname>:<port>

🔐 tls

Transport Layer Security (TLS) / Secure Sockets Layer (SSL). yes both names, it's giving legacy-acronym energy but we move 💅

The required syntax for a TLS/SSL url is literally:

tls://<hostname>:<port>[?<options>]

options is 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. clean URLs ate, no cap.

Options can also be specified via command line options (or in code via AVOptions).

The list of supported options follows, bestie:

ca_file, cafile=filename — a file holding certificate authority (CA) root certificates that you're treating as trusted. If the linked TLS library ships with a default, you might not need to specify this 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, it tries to verify the peer we're talking to. Heads up tho: if you're using OpenSSL, this currently only makes sure that the peer certificate is signed by one of the root certificates in the CA database — it does not validate that the certificate actually matches the host name we're trying to connect to. (With other backends, the host name is validated as well, so slay for them.)

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. main character energy fr.
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 is 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).

DTLS shares most options with TLS, but operates over UDP instead of TCP. same vibes, different transport 🦋

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 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.
cert_pem=string — certificate PEM string ✨
key_pem=string — private key PEM string ✨
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.
mtu=size — set the Maximum Transmission Unit (MTU) for DTLS packets.
use_srtp=1|0 — enable the use_srtp DTLS extension. This is used in WebRTC applications to establish SRTP encryption keys through the DTLS handshake. Default is disabled.
external_sock=1|0 — use an external socket instead of creating a new one. This option only makes sense to pass when interacting with the code via API — enabling this from the CLI will cause immediate failure 💀. 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 classic. the icon. unreliable in the most lovable way.

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 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).

In case threading is enabled on the system, a circular buffer is used to store the incoming data, which allows one to reduce loss of data due to UDP socket buffer overruns. The fifo_size and overrun_nonfatal options are related to this buffer. iykyk 🧚

The list of supported options follows:

buffer_size=size — set the UDP maximum socket buffer size in bytes. This is used to set either the receive or send buffer size, depending on what the socket is used for. Default is 32 KB for output, 384 KB for input. See also fifo_size.
bitrate=bitrate — if set to nonzero, the output will have the specified constant bitrate if the input has enough packets to sustain it.
burst_bits=bits — when using bitrate this specifies the maximum number of bits in packet bursts.
localport=port — override 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 — set the size in bytes of UDP packets.
reuse=1|0 — explicitly allow or disallow reusing UDP sockets.
ttl=ttl — set the time to live value (for multicast only).
dscp=dscp — set the 6-bit DSCP field for outgoing packets.
connect=1|0 — initialize the UDP socket with connect(). In this case, the destination address can't be changed with ff_udp_set_remote_url later. If the destination address isn't known at the start, this option can be specified in ff_udp_set_remote_url, too. This allows finding out the source address for the packets with getsockname, and makes writes return with AVERROR(ECONNREFUSED) if "destination unreachable" is received. For receiving, this gives the benefit of only receiving packets from the specified peer address/port. lowkey a glow-up ngl.
sources=address[,address] — only receive packets sent from the specified addresses. In case of multicast, also subscribe to multicast traffic coming from these addresses only.
block=address[,address] — ignore packets sent from the specified addresses. In case of multicast, also exclude the source addresses in the multicast subscription. not me blocking an IP 🚫
fifo_size=units — set the UDP receiving circular buffer size, expressed as a number of packets with size of 188 bytes. If not specified defaults to 7*4096.
overrun_nonfatal=1|0 — survive in case of UDP receiving circular buffer overrun. Default value is 0.
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.

broadcast=1|0 — explicitly allow or disallow UDP broadcasting.

Note that broadcasting may not work properly on networks having a 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, using 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. yes it's just a file path, and yes that's iconic of it.

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 — create the Unix socket in listening mode.
type — select 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". Ignored for SOCK_STREAM. Default is 0.

🌪️ zmq

ZeroMQ asynchronous messaging using the libzmq library.

This library supports unicast streaming to multiple clients without relying on an external server. no middleman, we love to see it 🫶

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. so chill about it ✨

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 — bestie this one forces the maximum packet size for sending/receiving data, and the default is 131,072 bytes 💅 On the server side it sets the max size of packets sent via ZeroMQ, and on the clients it sets an internal buffer size for receiving packets. And listen — pkt_size on the clients should literally be equal to or greater than pkt_size on the server, ok? Otherwise the received message may get truncated and that causes decoding errors 😭 no cap.

🌸 DEVICE OPTIONS 🌸

So the libavdevice library gives you the exact same interface as libavformat, periodt. Like, an input device is considered like a demuxer, and an output device like a muxer, and the interface plus the generic device options are the same ones provided by libavformat (go peek at the ffmpeg-formats manual, iykyk).

On top of that, each input or output device may support its own so-called private options which are specific to that component 🎀

You can set options 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. Multiple lanes, all valid ✨

🎀 INPUT DEVICES 🎀

Input devices are configured elements in FFmpeg that enable accessing 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 --enable-indev=INDEV, or you can disable a particular input device using --disable-indev=INDEV. Mix and match, she contains multitudes 💅

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

The ALSA (Advanced Linux Sound Architecture) input device 🐧

To enable this input device during configuration you need libasound installed on your system.

This device allows capturing from an ALSA device, and 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 — set the sample rate in Hz. Default is 48000.
channels — set the number of channels. Default is 2.

android_camera

The 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 allows capturing from all cameras on an Android device, which 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, literally ignored 💀

Generally the back facing camera has index 0 while the front facing camera has index 1 — selfie mode activated 🫶

Options

video_size — set 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 is not available or by default.
framerate — set the video framerate. Falls back to the first available configuration reported by Android if the requested framerate is not available or by default (-1).
camera_index — set the index of the camera to use. Default is 0.
input_queue_size — set the maximum number of frames to buffer. Default is 5.

avfoundation

The AVFoundation input device 🍎

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:

-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 👀

All available devices can be enumerated by using -list_devices true, listing all device names and corresponding indices.

There are two device name aliases:

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:

-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, ngl. 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.

Heads up bestie, this part picks up right in the middle of the DeckLink tea ☕ — no new headings needed, we're still in there.

Note: DeckLink SDK 11.0 straight up replaced the duplex property with a profile property. So for the DeckLink Duo 2 and DeckLink Quad 2, a profile is shared between any 2 sub-devices that use the same connectors. For the DeckLink 8K Pro, a profile is shared between all 4 sub-devices. Meaning the DeckLink 8K Pro supports four profiles, periodt. 💅

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`

The timecode type to include in the frame and video stream metadata. Gotta be one of none, rp188vitc, rp188vitc2, rp188ltc, rp188hfr, rp188any, vitc, vitc2, or serial. Defaults to none (aka not included, she's not showing up to the party).

So 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. Lowkey note that this is slightly different from the ordering used by the DeckLink API, which goes HFR, VITC1, LTC, VITC2. Don't mix them up bestie 🙈

`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 get drawn in the event of a signal loss. Defaults to true.

⚠️ This option is deprecated, bestie — please use the signal_loss_action option instead. She's retired but still hanging around 💀

`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 the maximum input buffer size in bytes. If the buffering reaches this value, incoming frames will be dropped, no cap. 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 the configured value. An 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.

Ngl 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:

ffmpeg -f decklink -list_formats 1 -i 'Intensity Pro'

Capture video clip at 1080i50:

ffmpeg -format_code Hi50 -f decklink -i 'Intensity Pro' -c:a copy -c:v copy output.avi

Capture video clip at 1080i50 10 bit:

ffmpeg -raw_format yuv422p10 -format_code Hi50 -f decklink -i 'UltraStudio Mini Recorder' -c:a copy -c:v copy output.avi

Capture video clip at 1080i50 with 16 audio channels:

ffmpeg -channels 16 -format_code Hi50 -f decklink -i 'UltraStudio Mini Recorder' -c:a copy -c:v copy output.avi

🎀 dshow 🎀

Windows DirectShow input device. Yes, Windows besties, this one's for you 🫶

DirectShow support is enabled when FFmpeg is built with the mingw-w64 project. Currently only audio and video devices are supported.

Multiple devices may be opened as separate inputs, but they may also be opened on the same input, which should improve synchronism between them. Iconic behaviour honestly.

The input name should be in the format:

<TYPE>=<NAME>[:<TYPE>=<NAME>]

where TYPE can be either audio or video, and NAME is the device's name or alternative name.

Options

If no options are specified, the device's defaults are used. If the device does not support the requested options, it will fail to open. Like, it just won't, sorry 💀

`video_size`

Set the video size in the captured video.

`framerate`

Set the frame rate in the captured video.

`sample_rate`

Set the sample rate (in Hz) of the captured audio.

`sample_size`

Set the sample size (in bits) of the captured audio.

`channels`

Set the number of channels in the captured audio.

`list_devices`

If set to true, print a list of devices and exit.

`list_options`

If set to true, print a list of the selected device's options and exit.

`video_device_number`

Set the video device number for devices with the same name (starts at 0, defaults to 0).

`audio_device_number`

Set the audio device number for devices with the same name (starts at 0, defaults to 0).

`pixel_format`

Select the pixel format to be used by DirectShow. This may only be set when the video codec is not set or set to rawvideo.

`audio_buffer_size`

Set the audio device buffer size in milliseconds (which can directly impact latency, depending on the device). Defaults to using the audio device's default buffer size (typically some multiple of 500ms). Setting this value too low can degrade performance. See also <http://msdn.microsoft.com/en-us/library/windows/desktop/dd377582(v=vs.85).aspx>

`video_pin_name`

Select the video capture pin to use by name or alternative name.

`audio_pin_name`

Select the audio capture pin to use by name or alternative name.

`crossbar_video_input_pin_number`

Select video input pin number for the crossbar device. This will be 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 occurs. She's got main character energy and she remembers 🔥

`crossbar_audio_input_pin_number`

Select audio input pin number for the crossbar device. This will be routed to the crossbar device's Audio Decoder output pin.

Note that changing this value can affect future invocations (sets a new default) until a system reboot occurs. Same deal, bestie.

`show_video_device_dialog`

If set to true, before capture starts, popup a display dialog to the end user, allowing them to change video filter properties and configurations manually.

Note that for crossbar devices, adjusting values in this dialog may be needed at times 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. 🙈

`show_audio_device_dialog`

If set to true, before capture starts, popup a display dialog to the end user, allowing them to change audio filter properties and configurations manually.

`show_video_crossbar_connection_dialog`

If set to true, before capture starts, popup a display dialog to the end user, allowing them to manually modify crossbar pin routings, when it opens a video device.

`show_audio_crossbar_connection_dialog`

If set to true, before capture starts, popup a display dialog to the end user, allowing them to manually modify crossbar pin routings, when it opens an audio device.

`show_analog_tv_tuner_dialog`

If set to true, before capture starts, popup a display dialog to the end user, allowing them to manually modify TV channels and frequencies.

`show_analog_tv_tuner_audio_dialog`

If set to true, before capture starts, popup a display dialog to the end user, allowing them to manually modify TV audio (like mono vs. stereo, Language A, B or C).

`audio_device_load`

Load an audio capture filter device from file instead of searching it by name. It may load additional parameters too, if the filter supports the serialization of its properties. To use this, an audio capture source has to be specified, but it can be anything — even a fake one. Fake it till you make it 💅

`audio_device_save`

Save 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 exists it will be overwritten. Rude but that's the rule.

`video_device_load`

Load a video capture filter device from file instead of searching it by name. It may load additional parameters too, if the filter supports the serialization of its properties. To use this, a video capture source has to be specified, but it can be anything — even a fake one.

`video_device_save`

Save 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 will be derived from the wallclock instead of the timestamp provided by the capture device. This lets you work around devices that provide unreliable timestamps. Very slay of you to fix their mess 🌸

✨ Examples ✨

Print the list of DirectShow supported devices and exit:

$ ffmpeg -list_devices true -f dshow -i dummy

Open video device Camera:

$ ffmpeg -f dshow -i video="Camera"

Open second video device with name Camera:

(and we're picking that one up in the next part, hold that thought 🦋)

$ ffmpeg -f dshow -video_device_number 1 -i video="Camera" — picks the second video device that literally goes by the name Camera, because the index here is 1 (it's zero-based, bestie 💅).
Open video device Camera and audio device Microphone:

$ ffmpeg -f dshow -i video="Camera":audio="Microphone"

Print the list of supported options in selected device and exit — like a little menu peek, then it dips out:

$ ffmpeg -list_options true -f dshow -i video="Camera"

Specify pin names to capture by name or alternative name, specify alternative device name — yes this line is as long as it looks, do not fear it:

$ 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, allow user to adjust video capture properties at startup — crossbars are giving main-character energy ngl:

$ 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 🌈

The Linux framebuffer input device. Say hi 👋

So the Linux framebuffer is this graphic hardware–independent abstraction layer for showing graphics on a computer monitor — typically on the console. You reach it through a file device node, and that node is usually /dev/fb0 💖.

Want the deep lore? Go read the file Documentation/fb/framebuffer.txt that ships inside the Linux source tree. Also peep <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 absolutely snag a single screenshot image with this one:

ffmpeg -f fbdev -framerate 1 -i /dev/fb0 -frames:v 1 screenshot.jpeg

Options

framerate — sets the frame rate. Default is 25. slay.

🖥️ gdigrab 🖥️

The Win32 GDI–based screen capture device — Windows girlies rise 🙈✨

This device 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>

The first one captures the entire desktop, or a fixed region of the desktop. The second and third instead capture the contents of a single window, no matter where that window is sitting on your screen. Iconic, honestly.

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 — tells it whether to draw the mouse pointer. Use the value 0 to not draw the pointer. Default value is 1. (Yes, cursor can stay, she's part of the aesthetic 🖱️)
framerate — sets the grabbing frame rate. Default value is ntsc, which corresponds to a frame rate of 30000/1001. periodt.
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 being grabbed if you're only grabbing a portion of the screen 💖

Heads up bestie: show_region is incompatible with grabbing the contents of a single window.

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, this sets the distance from the left edge of the screen or desktop.

Note note note: the offset calculation is from the top left corner of the primary monitor on Windows. If you have 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. negative girlie 💅

offset_y — when capturing a region with video_size, this sets the distance from the top edge of the screen or desktop.

Same deal: offset calculation starts at the top left corner of the primary monitor on Windows. Monitor sitting above your primary? Negative offset_y value, go go go ⬆️

🔥 iec61883 🔥

FireWire DV/HDV input device using libiec61883. Retro tech but make it cute 🎀

To enable this input device you need libiec61883, libraw1394 and libavc1394 installed on your system. Use the configure option --enable-libiec61883 to compile with the device enabled.

The iec61883 capture device supports capturing from a video device connected via IEEE1394 (FireWire), using libiec61883 and the new Linux FireWire stack (juju). This is the default DV/HDV input method in Linux Kernel 2.6.37 and later, since the old FireWire stack was removed. rip old stack 💀

Specify the FireWire port to be used as input file, or auto to choose the first port connected.

Options

dvtype — overrides autodetection of DV/HDV. This should only be used if auto detection doesn't work, or if usage of a different device type should be prohibited.

🚨 Treating a DV device as HDV (or the other way around) will not work and results in undefined behavior. The values auto, dv and hdv are supported.

dvbuffer — sets the maximum size of buffer for incoming data, in frames. For DV this is an exact value. For HDV it is not frame exact, since HDV doesn't have a fixed frame size. iykyk.
dvguid — selects the capture device by specifying its GUID. Capturing will only be performed from the specified device and fails if no device with the given GUID is found. This is useful to select the input if multiple devices are connected at the same time. Peep /sys/bus/firewire/devices to find out 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. Audio besties, this one's for you 🫶

To enable this input device during configuration you need libjack installed on your system.

A JACK input device creates one or more JACK writable clients, one for each audio channel, with name client_name:input_N, where client_name is the name provided by the application, and N is a number which identifies the channel. Each writable client will send the acquired data to the FFmpeg input device.

Once you've created one or more JACK readable clients, you need to 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 — for example with qjackctl.

To list the JACK clients and their properties you can invoke the command 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 information read: <http://jackaudio.org/>

Options

channels — sets the number of channels. Default is 2. stereo supremacy 🍓

💜 kmsgrab 💜

The KMS video input device. Certification: advanced. Vibes: immaculate.

Captures the KMS scanout framebuffer associated with a specified CRTC or plane as a DRM object that can be passed 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, that's literally what the page says 😭

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 for 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 if you're running Linux 5.7 or later, but it will need to be provided explicitly when needed in earlier versions. See the libdrm documentation for possible values 📚
crtc_id — KMS CRTC ID to define the capture source. The first active plane on the given CRTC will be used.
plane_id — KMS plane ID to define the capture source. Defaults to the first active plane found if neither crtc_id nor plane_id are specified.
framerate — framerate to capture at.

Important: this is not synchronised to any page flipping or framebuffer changes — it 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

To capture only part of a plane the output can be cropped — 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. She's the filtergraph girlie and she's everything ✨

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 which is mapped to the generated output. The filtergraph is specified through the option graph.

Options

`graph`

Okay bestie, this one's the big one 💖 — it specifies the filtergraph to use as input.

Every video open output has got to be labelled with a unique string of the form "outN", where N is a number starting from 0 that corresponds to the mapped input stream the device generates. The first unlabelled output gets auto-assigned the "out0" label (she's the main character, she doesn't need to introduce herself 💅), but all the others? Yeah, those need to be specified explicitly. No shortcuts.

You can tack the suffix "+subcc" onto the output label to create an extra stream with the closed captions packets attached to that output — this is experimental, and for now it's only for EIA-608 / CEA-708. The subcc streams get created after all the normal streams, in the order of the corresponding stream. Example time, because honestly this is so much clearer with numbers: if there's "out19+subcc", "out7+subcc" and up to "out42", then stream #43 is subcc for stream #7 and stream #44 is subcc for stream #19. See? Math. Iconic. 🧚

If you don't specify it at all, it defaults to the filename specified for the input device.

`graph_file`

Set the filename of the filtergraph to be read and sent to the other filters 🎀. The syntax of the filtergraph is literally the same as the one specified by the graph option. Same vibes, different delivery method.

`dumpgraph`

Dump graph to stderr. That's it, that's the whole thing 🫶.

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 during configuration you need libcdio installed on your system. It requires the configure option --enable-libcdio.

This device allows playing and grabbing from an Audio-CD. So like, ripping. But make it ✨open source✨.

For example, to copy with ffmpeg the entire Audio-CD in /dev/sr0, you may run the command:

ffmpeg -f libcdio -i /dev/sr0 cd.wav

Options

speed — Set 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. Heads up 🫶: on many CD-ROM drives, specifying a value that's too large will just result in using the fastest speed. She'll do what she wants.

paranoia_mode — Set paranoia recovery mode flags. It accepts one of the following values:
- disable
- verify
- overlap
- neverskip
- full

Default value is disable.

For more information 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 — Set the frame rate. Default is "ntsc", corresponding to a frame rate of 30000/1001. Very specific of her, honestly.
pixel_format — Select the pixel format. Default is "uyvy422".
video_size — Set 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 during configuration, you need OpenAL headers and libraries installed on your system, and you need to configure FFmpeg with --enable-openal.

OpenAL headers and libraries should be provided as part of your OpenAL implementation, or as an additional download (an SDK). Depending on your installation you may need to specify additional flags via the --extra-cflags and --extra-ldflags for allowing the build system to locate the OpenAL headers and libraries. Annoying? Lowkey yes. Necessary? Also yes 😭.

An incomplete list of OpenAL implementations follows:

Creative — The official Windows implementation, providing hardware acceleration with supported devices and software fallback. See <http://openal.org/>.
OpenAL Soft — Portable, open source (LGPL) software implementation. Includes 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 allows one to capture from an audio input device handled through OpenAL.

You need to specify the name of the device to capture in the provided filename. If the empty string is provided, the device will automatically select the default device. You can get the list of the supported devices by using the option list_devices.

Options

channels — Set the number of channels in the captured audio. Only the values 1 (monaural) and 2 (stereo) are currently supported. Defaults to 2.
sample_size — Set the sample size (in bits) of the captured audio. Only the values 8 and 16 are currently supported. Defaults to 16.
sample_rate — Set the sample rate (in Hz) of the captured audio. Defaults to 44.1k.
list_devices — If set to true, print a list of devices and exit. 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, on purpose 🙈):

$ 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. She's trying her best 💀

oss

Open Sound System input device 🔊.

The filename to provide 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 — Set the sample rate in Hz. Default is 48000.
channels — Set the number of channels. Default is 2.

pulse

PulseAudio input device 🦋.

To enable this output device you need to configure FFmpeg with --enable-libpulse.

The filename to 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 — Connect to a specific PulseAudio server, specified by an IP address. Default server is used when not provided.
name — Specify the application name PulseAudio will use when showing active clients; by default it is the "LIBAVFORMAT_IDENT" string.
stream_name — Specify the stream name PulseAudio will use when showing active streams; by default it is "record".
sample_rate — Specify the samplerate in Hz; by default 48kHz is used.
channels — Specify the channels in use; by default 2 (stereo) is set.
frame_size — This option does nothing and is deprecated. Rip to her, she's a ghost 👻
fragment_size — Specify 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 — Set 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 to 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 — Set the sample rate in Hz. Default is 48000.
channels — Set the number of channels. Default is 2.

video4linux2, v4l2

Video4Linux2 input video device 📹.

"v4l2" can be used as alias for "video4linux2". Same girl, shorter name 💅.

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 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 widthx*height* 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. Clock lore fr fr ⭐

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

Okay, back in the v4l2 option buffet — grab a plate, bestie, because these knobs are lowkey essential:

standard — Sets the standard. It has to be the name of a standard that's actually supported. Wanna see the roster? Use the list_standards option. ✨
channel — Sets the input channel number. Default is -1, which means it uses the previously selected channel. No cap.
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 lets you 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 the 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. 🫶

✨ Options ✨

video_size — Set the video frame size.
framerate — Set 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. This device allows one to capture a region of an X11 display. periodt.

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 — Specify 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 — Specify whether to draw the mouse pointer. A value of 0 specifies not to draw the pointer. Default value is 1. 🙈
follow_mouse — Make the grabbed area follow the mouse. The argument can be centered or a number of pixels PIXELS.

When it is 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 — Set the grabbing frame rate. Default value is ntsc, corresponding to a frame rate of 30000/1001. 🌈
show_region — Show 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 is easy to know what is being grabbed if only a portion of the screen is grabbed. ⭐
region_border — Set 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 — Grab 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 — Set 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. 🍓

💖 RESAMPLER OPTIONS 💖

Okay bestie, so the audio resampler has a whole little universe of named options and we're gonna go through every single one, no skips, no cap ✨

You can set these a bunch of ways, and literally all of them count:

by specifying -option value in the FFmpeg tools
option=value for the aresample filter
by setting the value explicitly in the SwrContext options
or using the libavutil/opt.h API for programmatic use (iykyk 🧚)

uchl, used_chlayout — sets the used input channel layout. Default is unset. Heads up: this one is only used for special remapping, so it's giving niche-but-iconic 💅
isr, in_sample_rate — sets the input sample rate. Default value is 0.
osr, out_sample_rate — sets the output sample rate. Default value is 0.
isf, in_sample_fmt — specifies the input sample format. It defaults to none.
osf, out_sample_fmt — specifies the output sample format. Defaults to none.
tsf, internal_sample_fmt — sets the internal sample format. Default value is none. When you don't set it explicitly, it just picks one for you automatically — main character energy, it handles itself 🎀
ichl, in_chlayout and ochl, out_chlayout — set the input/output channel layout. For the exact syntax you gotta check the Channel Layout section in the ffmpeg-utils(1) manual, okay? Don't guess, go read it 📖
clev, center_mix_level — sets the center mix level. It's a value in deciBel, and it must be in the interval [-32,32].
slev, surround_mix_level — sets the surround mix level. Also in deciBel, also has to be in the interval [-32,32].
lfe_mix_level — sets the LFE mix into non-LFE level. It kicks in when you've got an LFE input but no LFE output. Value in deciBel, and yeah, interval [-32,32].
rmvol, rematrix_volume — sets rematrix volume. Default value is 1.0.
rematrix_maxval — sets the maximum output value for rematrixing. This is your lil tug-of-war: preventing clipping vs. preventing volume reduction. A value of 1.0 prevents clipping. Choose wisely bestie 🌸
flags, swr_flags — sets flags used by the converter. Default value is 0.

It supports this individual flag:

res — force resampling. Like the name says, it forces resampling to be used even when the input and output sample rates match.

dither_scale — sets the dither scale. Default value is 1.
dither_method — sets the dither method. Default value is 0.

Supported values, going down the list:

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; filter options precision and cheby are not applicable in this case.
soxr — select the SoX Resampler (where available); compensation, and 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 resampling phase shift, default value is 10, and it must be in the interval [0,30].
linear_interp — use linear interpolation when enabled (and it's enabled by default). Disable it if you'd rather preserve speed instead of quality when exact_rational fails. Trade-offs, bestie 💀
exact_rational — for swr only. When enabled, it tries to use an exact phase_count based on the input and output sample rate. BUT — if that ends up larger than 1 << phase_shift, then the phase_count becomes 1 << phase_shift as a fallback. Default is enabled.
cutoff — sets cutoff frequency ratio (swr: 6dB point; soxr: 0dB point). Must be a float value between 0 and 1. Default value is 0.97 with swr, and 0.91 with soxr — and that soxr one, with a sample-rate of 44100, preserves the entire audio band to 20kHz. Obsessed with her attention to detail 👑
precision — for soxr only, the precision in bits to which the resampled signal will be calculated. The default value of 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'. Glow-up unlocked ✨
cheby — for soxr only, selects passband rolloff none (Chebyshev) & higher-precision approximation for 'irrational' ratios. Default value is 0.
async — for swr only. It's a simple 1-parameter audio sync to timestamps, doing it via stretching, squeezing, filling and trimming. Setting this to 1 enables filling and trimming, and larger values represent the maximum amount in samples that the data may be stretched or squeezed for each second. Default value is 0, so no compensation gets applied to make the samples match the audio timestamps.
first_pts — for swr only, assumes the first pts should be this value. The time unit is 1 / sample rate. This lets you do padding/trimming at the start of the stream. By default, no assumption is made about the first frame's expected pts, so no padding or trimming is done. For example, 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 due to encoder delay. Very "let me fix this for you" of it 🫶
min_comp — for swr only, sets the minimum difference between timestamps and audio data (in seconds) needed to trigger stretching/squeezing/filling or trimming of the data so it 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) to trigger adding/dropping samples so it matches the timestamps. This option is basically a threshold that picks between hard (trim/fill) and soft (squeeze/stretch) compensation. Note: 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 is stretched/squeezed to make it 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 is stretched/squeezed to make it match the timestamps. Must be a non-negative double float value, default value is 0.
matrix_encoding — selects matrixed stereo encoding.

It accepts the following 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. Just so you know, this only affects resampling operations.

It accepts the following values:

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. Frugal queen 💅

✨ SCALER OPTIONS ✨

okay so the video scaler has its own whole lil set of named options and we're gonna go through every single one, no skipping 💅

You can set these by doing -option value in the FFmpeg tools, with a couple of API-only exceptions flagged down below (iykyk). And if you're doing the programmatic thing, you can set them explicitly in the SwsContext options or through the libavutil/opt.h API. 🫶

`sws_flags`

Sets the scaler flags. This is ALSO how you pick the scaling algorithm, bestie — and you should only select a single algorithm. Default value is bicubic.

It accepts the following values:

fast_bilinear — picks the fast bilinear scaling algorithm.
bilinear — picks the bilinear scaling algorithm.
bicubic — picks the bicubic scaling algorithm.
experimental — picks the experimental scaling algorithm 👀
neighbor — picks the nearest neighbor rescaling algorithm.
area — picks the averaging area rescaling algorithm.
bicublin — bicubic for the luma component, bilinear for the chroma components. two vibes at once ✨
gauss — picks the Gaussian rescaling algorithm.
sinc — picks the sinc rescaling algorithm.
lanczos — picks the Lanczos rescaling algorithm. The default width (alpha) is 3 and can be changed by setting param0.
spline — picks 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. May subtly affect the output or even produce wrong results. For testing only, fr fr 🙈

`srcw` (API only)

Sets the source width.

`srch` (API only)

Sets the source height.

`dstw` (API only)

Sets the destination width.

`dsth` (API only)

Sets the destination height.

`src_format` (API only)

Sets the source pixel format (must be expressed as an integer).

`dst_format` (API only)

Sets the destination pixel format (must be expressed as an integer).

`src_range` (boolean)

If the value is set to 1, it means the source is full range. Default value is 0, which means the source is limited range.

`dst_range` (boolean)

If the value is set to 1, full range gets enabled for the destination. Default value is 0, which enables limited range.

`gamma` (boolean)

If the value is set to 1, gamma correct scaling gets enabled. Default value is 0.

`param0`, `param1`

Set scaling algorithm parameters. The values you specify are specific to some scaling algorithms and just ignored by others, nbd. The specified values are floating point number values.

`sws_dither`

Sets 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 that a_dither, if you're into that 💅)

`alphablend`

Sets 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 🦋

So filtering in FFmpeg is enabled through the libavfilter library. That's the whole vibe check. 👑

In libavfilter, a filter can have multiple inputs AND multiple outputs — main character energy, honestly. To illustrate the sorts of things that are possible, we consider the following filtergraph:

                        [main]
        input --> split ---------------------> overlay --> output
                    |                             ^
                    |[tmp]                  [flip]|
                    +-----> crop --> vflip -------+

This filtergraph splits the input stream in 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 the following command to achieve this:

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 will be that the top half of the video is mirrored onto the bottom half of the output video. literally so iconic 😭

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, 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 overlay on its lower half the output generated by the crop,vflip filterchain.

Some filters take in input a list of parameters: they are specified after the filter name and an equal sign, and are separated from each other by a colon. 🎀

There exist so-called source filters that do not have an audio/video input, and sink filters that will not have audio/video output. slay, they're special like that.

🌈 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. fancy 💅

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.

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. ngl this part trips people up, so like, double check it bestie 🫶

🌸 FILTERGRAPH DESCRIPTION 🌸

okay bestie so a filtergraph is literally a directed graph of connected filters — yes, an actual graph, we're doing graph theory in our media player, it's giving CS major era 💅

It can contain cycles, and there can be multiple links between a pair of filters. And like… 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. Pads are the little handshake points, iykyk 🤝

Each filter in a filtergraph is an instance of a filter class registered in the application, which defines the features and the number of input and output pads of the filter.

A filter with no input pads is called a "source", and a filter with no output pads is called a "sink". Sources give, sinks take, main character energy on both ends ✨

🎀 Filtergraph syntax 🎀

A filtergraph has a textual representation, which is 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.

So here's the tea on how it's built up, layer by layer 🧚

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. It may have one of two forms:

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. 👀

A special syntax implemented in the ffmpeg CLI tool allows loading option values from files. This is done be 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. it's literally matchmaking for pads 💖

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.

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 slay 🫶

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 orphan pads allowed, we don't do that here 😤

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. aesthetic formatting is valid formatting bestie ✨

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 — iconic, legendary, no notes:

<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. 😭 yes it's that deep

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. 💀 real

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. way easier, trust 💅

⭐ 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. on/off era, we love a boolean moment 🔥

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. live editing, we love to see it ✨

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, the list is right there bestie 👑

👑 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. 🎀

So like, 'T' = "this one's live, you can poke it while it's running." periodt 💅

✨ OPTIONS FOR FILTERS WITH SEVERAL INPUTS ✨

Okay bestie, some filters with several inputs support a common set of options. Important: these options can only be set by name, not with the short notation. No shortcuts here, iykyk 💅

eof_action — the action to take when EOF is encountered on the secondary input. It accepts one of these values:
- repeat — repeat the last frame (the default).
- endall — end both streams.
- pass — pass the main input through.
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 these 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.

Below is a description of the currently available audio filters. Let’s gooo 🌈

🌀 `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. Iconic behavior fr 💅

The filter accepts the following options. ✨

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.

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.

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 all of the above options as commands. Slay 💅

🌸 `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. Lowkey the simplest icon here 🫶

🦋 `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.

The filter accepts the following options:

split — set split frequencies. Those must be positive and increasing.
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 with enhanced functionality — and it’s genuinely iconic, no cap. A bit crusher is used to audibly reduce the number of bits an audio signal is sampled with. The important part: this doesn’t change the bit depth at all, it just produces the effect. Material reduced in bit depth sounds more harsh and “digital.” This filter can even round to continuous values instead of discrete bit depths. It also has a D/C offset, which results in different crushing of the lower half and the upper half of the signal. And an Anti-Aliasing setting can produce “softer” crushing sounds. 💅

Another feature of this filter is the logarithmic mode. This setting switches from linear distances between bits to logarithmic ones. The result is a much more “natural” sounding crusher which doesn’t gate low signals, for example. The human ear has a logarithmic perception, so this kind of crushing is much more pleasant. Logarithmic crushing is also able to get anti-aliased. ✨

The filter accepts the following options:

level_in — Set level in.
level_out — Set level out.
bits — Set bit reduction.
mix — Set mixing amount.
mode — Can be linear: "lin" or logarithmic: "log".
dc — Set DC.
aa — Set anti-aliasing.
samples — Set sample reduction.
lfo — Enable LFO. By default disabled.
lforange — Set LFO range.
lforate — Set LFO rate.

💅 Commands 💅

This filter supports the all above options as commands.

🎀 acue 🎀

Delay audio filtering until a given wallclock timestamp. See the cue filter. That’s the whole vibe — very “wait for it” behavior. 🫶

🦋 adeclick 🦋

Remove impulsive noise from input audio. Samples detected as impulsive noise are replaced by interpolated samples using autoregressive modelling.

window, w — Set window size, in milliseconds. Allowed range is from 10 to 100. Default value is 55 milliseconds. This sets size of window which will be processed at once.
overlap, o — Set window overlap, in percentage of window size. Allowed range is from 50 to 95. Default value is 75 percent. Setting this to a very high value increases impulsive noise removal but makes whole process much slower.
arorder, a — Set autoregression order, in percentage of window size. Allowed range is from 0 to 25. Default value is 2 percent. This option also controls quality of interpolated samples using neighbour good samples.
threshold, t — Set threshold value. Allowed range is from 1 to 100. Default value is 2. This controls the strength of impulsive noise which is going to be removed. The lower value, the more samples will be detected as impulsive noise.
burst, b — Set burst fusion, in percentage of window size. Allowed range is 0 to 10. Default value is 2. If any two samples detected as noise are spaced less than this value then any sample between those two samples will be also detected as noise.
method, m — Set overlap method. It accepts the following values:
- add, a — Select overlap-add method. Even not interpolated samples are slightly changed with this method.
- save, s — Select overlap-save method. Not interpolated samples remain unchanged.
- Default value is "a".

🌸 adeclip 🌸

Remove clipped samples from input audio. Samples detected as clipped are replaced by interpolated samples using autoregressive modelling.

window, w — Set window size, in milliseconds. Allowed range is from 10 to 100. Default value is 55 milliseconds. This sets size of window which will be processed at once.
overlap, o — Set window overlap, in percentage of window size. Allowed range is from 50 to 95. Default value is 75 percent.
arorder, a — Set autoregression order, in percentage of window size. Allowed range is from 0 to 25. Default value is 8 percent. This option also controls quality of interpolated samples using neighbour good samples.
threshold, t — Set threshold value. Allowed range is from 1 to 100. Default value is 10. Higher values make clip detection less aggressive.
hsize, n — Set size of histogram used to detect clips. Allowed range is from 100 to 9999. Default value is 1000. Higher values make clip detection less aggressive.
method, m — Set overlap method. It accepts the following values:
- add, a — Select overlap-add method. Even not interpolated samples are slightly changed with this method.
- save, s — Select overlap-save method. Not interpolated samples remain unchanged.
- Default value is "a".

💜 adecorrelate 💜

Apply decorrelation to input audio stream.

The filter accepts the following options:

stages — Set decorrelation stages of filtering. Allowed range is from 1 to 16. Default value is 6.
seed — Set random seed used for setting delay in samples across channels.

⭐ adelay ⭐

Delay one or more audio channels. Samples in delayed channel are filled with silence.

The filter accepts the following option:

delays — Set list of delays in milliseconds for each channel separated by '|'. Unused delays will be silently ignored. If number of given delays is smaller than number of channels all remaining channels will not be delayed. If you want to delay exact number of samples, append 'S' to number. If you want instead to delay in seconds, append 's' to number.
all — Use last set delay for all remaining channels. By default is disabled. This option if enabled changes how option "delays" is interpreted.

✨ Examples ✨

Delay 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 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 same number of samples:

adelay=delays=64S:all=1

🧚 adenorm 🧚

Remedy denormals in audio by adding extremely low-level noise. This filter shall be placed before any filter that can produce denormals.

A description of the accepted parameters follows.

level — Set level of added noise in dB. Default is -351. Allowed range is from -451 to -90.
type — Set type of added noise.
- dc — Add DC signal.
- ac — Add AC signal.
- square — Add square signal.
- pulse — Add pulse signal.
- Default is "dc".

💫 Commands 💫

This filter supports the all above options as commands.

🌈 aderivative, aintegral 🌈

Compute derivative/integral of audio stream. Applying both filters one after another produces original audio. It’s giving “undo button” energy, honestly. 😭

🔥 adrc 🔥

Apply spectral dynamic range controller filter to input audio stream.

A description of the accepted options follows.

transfer — Set the transfer expression.
- 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 from 1 to 1000 milliseconds.
release — Set the release in milliseconds. Default is 100 milliseconds. Allowed range is from 5 to 2000 milliseconds.
channels — Set which channels to filter, by default "all" channels in audio stream are filtered.

👑 Commands 👑

This filter supports the all above options as commands.

🍓 Examples 🍓

Apply spectral compression to all frequencies with threshold of -50 dB and 1:6 ratio:

adrc=transfer='if(gt(p,-50),-50+(p-(-50))/6,p)':attack=50:release=100

Similar to above but with 1:2 ratio and filtering only front center channel:

adrc=transfer='if(gt(p,-50),-50+(p-(-50))/2,p)':attack=50:release=100:channels=FC

Apply spectral noise gate to all frequencies with threshold of -85 dB and with short attack time and short release time:

adrc=transfer='if(lte(p,-85),p-800,p)':attack=1:release=5

Apply spectral expansion to all frequencies with threshold of -10 dB and 1:2 ratio:

adrc=transfer='if(lt(p,-10),-10+(p-(-10))*2,p)':attack=50:release=100

Apply limiter to max -60 dB to all frequencies, with attack of 2 ms and release of 10 ms:

adrc=transfer='min(p,-60)':attack=2:release=10

🦄 adynamicequalizer 🦄

Apply dynamic equalization to input audio stream.

A description of the accepted options follows.

threshold — Set the detection threshold used to trigger equalization. Threshold detection is using detection filter. Default value is 0. Allowed range is from 0 to 100.
dfrequency — Set the detection frequency in Hz used for detection filter used to trigger equalization. Default value is 1000 Hz. Allowed range is between 2 and 1000000 Hz.
dqfactor — Set the detection resonance factor for detection filter used to trigger equalization. Default value is 1. Allowed range is from 0.001 to 1000.
tfrequency — Set the target frequency of equalization filter. Default value is 1000 Hz. Allowed range is between 2 and 1000000 Hz.
tqfactor — Set the target resonance factor for target equalization filter. Default value is 1. Allowed range is from 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 from 1 to 200.
mode — Set the mode of filter operation, can be one of the following:
- listen — Output only isolated detection signal.

Okay bestie, back into the option list — these are the detection/threshold modes for this filter:

cutbelow — Cut frequencies below detection threshold.
cutabove — Cut frequencies above detection threshold.
boostbelow — Boost frequencies below detection threshold.
boostabove — Boost frequencies above detection threshold.

The default mode is cutbelow.

dftype — Set the type of detection filter. It can be one of the following:
- bandpass
- lowpass
- highpass
- peak

The default type is bandpass.

tftype — Set the type of target filter. It can be one of the following:
- bell
- lowshelf
- highshelf

The default type is bell.

auto — Automatically gather threshold from the detection filter. By default this is disabled. This option is useful for detecting the threshold in a certain time frame of the input audio stream — in that case the option value is changed at runtime. Available values are:
- disabled — Disable using automatically gathered threshold value.
- off — Stop picking threshold value.
- on — Start picking threshold value.
- adaptive — Adaptively pick threshold value, by calculating sliding window entropy.
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.

💬 Commands

This filter supports all the above options as commands. Slay, runtime tweaking unlocked. ✨

🫧 adynamicsmooth

Apply dynamic smoothing to the input audio stream. Here's what you can tweak:

sensitivity — Set the amount of sensitivity to frequency fluctations. Default is 2. Allowed range is from 0 to 1e+06.
basefreq — Set a base frequency for 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.

Echoes are reflected sound, and they can occur naturally amongst mountains (and sometimes large buildings) when you're talking or shouting. Digital echo effects emulate this behaviour and are often used 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. Iconic behaviour, honestly.

Here are the accepted parameters:

in_gain — Set input gain of the reflected signal. Default is 0.6.
out_gain — Set output gain of the reflected signal. Default is 0.3.
delays — Set the list of time intervals in milliseconds between the original signal and reflections, separated by |. Allowed range for each delay is (0 - 90000.0]. Default is 1000.
decays — Set the 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 the 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

Audio emphasis filter creates or restores material directly taken from LPs or emphased CDs with different filter curves. E.g. to store music on vinyl, the signal has to be altered by a filter first to even out the disadvantages of this recording medium. Once the material is played back, the inverse filter has to be applied to restore the distortion of the frequency response. Literally giving restoration main character energy. 💅

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 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 are evaluated and used to modify a corresponding audio signal.

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.
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

Note: 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. This is done by creating harmonic distortions of the signal which are 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 like an equalizer would do, to create a more "crisp" or "brilliant" sound. It's giving sparkle but make it DSP. 🔥

The filter accepts the following options:

level_in — Set input level prior processing of signal. Allowed range is from 0 to 64. Default value is 1.
level_out — Set output level after processing of signal. Allowed range is from 0 to 64. Default value is 1.
amount — Set the amount of harmonics added to original signal. Allowed range is from 0 to 64. Default is 1.
drive — Set the amount of newly created harmonics. Allowed range is from 0.1 to 10. Default is 8.5.
blend — Set the octave of newly created harmonics. Allowed range is from -10 to 10. Default 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 value is lower than 10000 Hz no limit is applied.
listen — Mute the original signal and output only added harmonics. By default is disabled.

💬 Commands

This filter supports all the above options as commands. 💜

🎬 afade

Apply fade-in/out effect to input audio.

Here's what the accepted parameters do:

type, t — Specify the effect type. Can be 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; see 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 curve for 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 first 15 seconds of audio:

afade=t=in:ss=0:d=15

Fade out last 25 seconds of a 900 seconds audio:

afade=t=out:st=875:d=25

🎧 afftdn

Denoise audio samples with FFT.

okay so here's the full rundown of every parameter this thing eats, no skips, no cap 💅

noise_reduction, nr — sets the noise reduction in dB. allowed range is 0.01 to 97. default is 12 dB. ✨
noise_floor, nf — sets the noise floor in dB. allowed range is -80 to -20. default is -50 dB.
noise_type, nt — sets the noise type, bestie. it accepts these 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 '|'. iykyk 🫶
residual_floor, rf — sets the residual floor in dB. allowed range is -80 to -20. default is -38 dB.
track_noise, tn — enables noise floor tracking. by default it's disabled. with this on, the noise floor gets adjusted automatically. literally hands-free 🦋
track_residual, tr — enables residual tracking. by default it's disabled.
output_mode, om — sets the output mode. it accepts:
- 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, which is like how fast it adapts gain adjustments per each frequency bin. value 0 enables instant adaptation, while higher values react much slower. allowed range is 0 to 1. default is 0.5. she's flexible 💅
floor_offset, fo — sets the noise floor offset factor. this one adjusts the offset applied to the measured noise floor. heads up: it's only effective when noise floor tracking is enabled 🔥. allowed range is -2.0 to 2.0. default is 1.0.
noise_link, nl — sets the noise link used for multichannel audio. accepts:
- 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, used for how much to spread bands across frequency bins. allowed range is 0.2 to 5. default is 1.25.
sample_noise, sn — toggles capturing and measurement of the noise profile from input audio. it accepts:
- 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 gains applied to each frequency bin. super useful for reducing random music noise artefacts ✨. higher values increase the smoothing of gains. allowed range is 0 to 50. default is 0.

🕹️ Commands

this filter supports some of the above mentioned options as commands. periodt 💅

🌈 Examples

reduce white noise by 10dB, and use a previously measured noise floor of -40dB:

afftdn=nr=10:nf=-40

reduce white noise by 10dB, also set the 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 a 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`

applies arbitrary expressions to samples in frequency domain. she's doing math in the frequency domain, main character energy fr fr 👑

✨ `real`

sets 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`

sets 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 real part of frequency bin at location (bin,channel)
imag(b, ch) — return the value of imaginary part of frequency bin at location (bin,channel)

`win_size`

sets the window size. allowed range is 16 to 131072. default is 4096.

`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 is "hann". so many windows, and honestly? each one ate 💅

`overlap`

sets the window overlap. if it's set to 1, the recommended overlap for the selected window function will be picked. default is 0.75.

🌸 Examples

leave almost only low frequencies in audio:

afftfilt="'real=re * (1-clip((b/nb)*b,0,1))':imag='im * (1-clip((b/nb)*b,0,1))'"

apply robotize effect:

afftfilt="real='hypot(re,im)*sin(0)':imag='hypot(re,im)*cos(0)':win_size=512:overlap=0.75"

apply 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`

applies an arbitrary Finite Impulse Response filter. icon behavior honestly.

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. she contains multitudes 💜

heads up on how it works: this filter uses the streams higher than the first one as FIR coefficients. if the non-first stream holds a single channel, it will 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. no cap.

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. 💀 literally a ghost.
irnorm — set norm to be applied to IR coefficients before filtering. allowed range is -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 will be 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 1 to 65536. lower values decrease latency at the cost of higher CPU usage. trade-offs, always 😭
maxp — set maximal partition size used for convolution. default is 8192. allowed range is 8 to 65536. lower values may increase CPU usage.
nbirs — set number of input impulse responses streams which will be switchable at runtime. allowed range is 1 to 32. default is 1.
ir — set 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. slay 🔥
precision — set which precision to use when processing samples. it accepts:
- 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". first one loads and prepares all IRs on initialization, second one 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 setting 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`

sets output format constraints for the input audio. the framework will negotiate the most appropriate format to minimize conversions. she negotiates. she compromises. she minimizes conversions 🫶

it accepts the following parameters:

sample_fmts, f — a |-separated list of the sample formats you're asking for ✨
sample_rates, r — same deal, but for sample rates, also |-separated
channel_layouts, cl — and again for channel layouts, |-separated bestie

Peep the Channel Layout section in the ffmpeg-utils(1) manual for the exact syntax you need here 💅

If you just… don't pass a parameter at all? Then every value is fair game. No restrictions, no vibe check, all allowed 🌈

Wanna force the output to be either unsigned 8-bit or signed 16-bit stereo? Say less:

aformat=sample_fmts=u8|s16:channel_layouts=stereo

afreqshift 🌀

Applies a frequency shift to the input audio samples. Iconic little pitch-scoot, honestly.

The filter accepts these options:

shift — the frequency shift itself. Allowed range is -INT_MAX to INT_MAX. Default? 0.0.
level — output gain slapped onto the final output. Allowed range is 0.0 to 1.0. Default is 1.0.
order — the filter order used for filtering. Allowed range is 1 to 16. Default is 8.

Commands

This filter supports all of the above options as commands 👑

afwtdn 🌊

Reduces broadband noise from your input samples using Wavelets. It's giving noise-cancelling headphones but make it math.

Here's the description of the accepted options:

sigma — sets the noise sigma, allowed range is 0 to 1. Default value is 0. This one controls the strength of the denoising applied to input samples. Most useful way to set this is via decibels, like -45dB.
levels — the number of wavelet levels of decomposition. Allowed range is 1 to 12. Default is 10. Set this too low and the denoising performance gets very poor, like genuinely embarrassing.
wavet — the wavelet type used for decomposition of the input frame. They're sorted by number of coefficients, lowest to highest. More coefficients means worse filtering speed, but overall better quality. Available wavelets are:
- sym2
- sym4
- rbior68
- deb10
- sym10
- coif5
- bl3
percent — sets the percent of full denoising. Allowed range is 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. Heads up though: if that first frame's samples contain non-noise, performance will be very poor 😭
adaptive — if enabled, input frames get analyzed for the presence of noise. If noise is detected with high possibility, then that input frame profile gets used for processing the following frames, until a new noise frame is detected.
samples — sets the size of a single frame in number of samples. Allowed range is 512 to 65536. Default frame size is 8192 samples.
softness — sets the softness applied inside the thresholding function. Allowed range is 0 to 10. Default softness is 1.

Commands

This filter supports all of the above options as commands 🫶

agate 🚪

A gate is mainly used to reduce the lower parts of a signal. This kind of signal processing reduces that annoying disturbing noise between the useful signals. Very "bouncer at the club, only the loud stuff gets in" energy 💅

Gating works 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. That's what attack and release are for.

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, bestie — they get a free pass ✨

level_in — sets the input level before filtering. Default is 1. Allowed range is 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 get amplified, expanding the dynamic range in the upward direction. Otherwise, in the case of "downward", lower parts of the signal get reduced.
range — sets the level of gain reduction when the signal is below the threshold. Default is 0.06125. Allowed range is 0 to 1. Setting this to 0 disables reduction entirely, 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 0 to 1.
ratio — sets a ratio by which the signal is reduced. Default is 2. Allowed range is 1 to 9000.
attack — the amount of milliseconds the signal has to rise above the threshold before gain reduction stops. Default is 20 milliseconds. Allowed range is 0.01 to 9000.
release — the amount of milliseconds the signal has to fall below the threshold before the reduction is increased again. Default is 250 milliseconds. Allowed range is 0.01 to 9000.
makeup — sets the amount of amplification of the signal after processing. Default is 1. Allowed range is 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 1 to 8.
detection — choose whether the exact signal should be taken for detection or an RMS-like one. Default is "rms". Can be "peak" or "rms".
link — choose whether 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 all of the above options as commands 💖

aiir 🎛️

Applies an arbitrary Infinite Impulse Response filter. Yes, arbitrary. Go off, king.

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 the coefficients format:
- ll — lattice-ladder function
- sf — analog transfer function
- tf — digital transfer function
- zp — Z-plane zeros/poles, cartesian (this is the default ✨)
- pr — Z-plane zeros/poles, polar radians
- pd — Z-plane zeros/poles, polar degrees
- sp — S-plane zeros/poles
process, r — set the type of processing:
- d — direct processing
- s — serial processing
- p — parallel processing
precision, e — set the filtering precision:
- dbl — double-precision floating-point (default)
- flt — single-precision floating-point
- i32 — 32-bit integers
- i16 — 16-bit integers
normalize, n — normalize the filter coefficients. By default this is enabled. Enabling it normalizes 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's disabled.
channel — set which IR channel to display the frequency response for. By default the first channel is displayed. This option is used only when response is enabled.
size — set the 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 the coefficients doesn't matter. Coefficients in "zp" format are complex numbers with i as the imaginary unit 🦋

You can provide different coefficients and gains for every channel — in that case, use '|' to separate the coefficients or gains. The last provided coefficients get used for all the remaining channels.

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 one uses lookahead technology to keep your signal from distorting 👑 That means there's a small delay after the signal is processed, so keep in mind that the delay it produces is the attack time you set. Not a bug, it's a feature, iykyk.

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 the 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 flushes the valid audio data in the lookahead buffer when the stream hits EOF.

Depending on the setting you picked, it's recommended to upsample the input 2x or 4x times with aresample before applying this filter 💀

allpass 🎀

Applies 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. Sneaky like that 🌸

The filter accepts the following options:

frequency, f — set frequency in 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 — 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.
channels, c — tells the filter which channels to actually filter. Default is all the ones you've got 💖
normalize, n — normalizes the biquad coefficients. Off by default, bestie. Turn it on and it normalizes the magnitude response at DC to 0dB ✨
order, o — sets the filter order. Can be 1 or 2. Default is 2 🎀
transform, a — sets the transform type of the IIR filter. The options are giving range:
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — sets the precision of the filtering. Pick your fighter:
- 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

Commands 💅

This filter supports the following commands, and yes they're all lowkey the same vibe:

frequency, f — change allpass frequency. The syntax for the command is: "frequency" ✨
width_type, t — change allpass width_type. Syntax is: "width_type"
width, w — change allpass width. Syntax is: "width"
mix, m — change allpass mix. Syntax is: "mix"

🌀 `aloop`

Loops your audio samples. It's giving replay button, fr fr 🔁

The filter accepts the following options:

loop — sets the number of loops. Setting this value to -1 will result in infinite loops. Default is 0 💀
size — sets the maximal number of samples. Default is 0
start — sets the first sample of the loop. Default is 0
time — sets the time of loop start, in seconds. Only used if the option named start is set to -1 ⭐

🫶 `amerge`

Merges two or more audio streams into a single multi-channel stream. Glow-up for your channels ✨

All inputs must have the same sample rate, and format. No mixing and matching, bestie.

If the inputs don't have the same duration, the output will stop with the shortest one. Shortest king wins, I don't make the rules 😭

The filter accepts the following options:

inputs — sets the number of inputs. Default is 2
layout_mode — this one controls how the output channel layout is determined, and whether the audio channels get reordered during the merge. Three moods:
- legacy — this is the mode the filter behaved historically in, so it's 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're 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.

For example, 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. Yeah. It's giving chaos 🙈

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. That's it, that's the whole vibe.

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. The group project that actually works 🫶

Heads up bestie: 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. Not me inserting a whole filter for you 💅

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. (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 the weight of each input audio stream as a sequence of numbers separated by a space. If fewer weights are specified compared to the 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's 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, but make it a command ✨

✖️ `amultiply`

Multiplies the first audio stream with the second audio stream and stores the result in the output audio stream. The multiplication is done by multiplying each sample from the first stream with the sample at the same position from the second stream. Math but make it audio 🔥

With this element-wise multiplication you can create amplitude fades and amplitude modulations. Iconic behaviour, honestly 👑

🌈 `anequalizer`

A 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 '|'. Let's break it down:
- chn — set the channel number to which equalization will be applied. If the input doesn't have that channel, the entry is ignored 🙈
- f — set the central frequency for the band. If the input doesn't have that frequency, the entry is ignored
- w — set the band width in Hertz
- g — set the band gain in dB
- t — set the filter type for the band. Optional, and can be:
  - 0 — Butterworth, this is default 👑
  - 1 — Chebyshev type 1
  - 2 — Chebyshev type 2
curves — with this option activated, the frequency response of anequalizer is displayed in the video stream. Visuals, we love to see it ✨
size — sets the video stream size. Only useful if the curves option is activated
mgain — sets the max gain that will be displayed. Only useful if the 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 — sets the frequency scale used to draw the frequency response in the video output. Can be linear or logarithmic. Default is logarithmic 🎀
colors — sets the color for each channel curve which is going to be displayed in the video stream. This is a list of color names separated by space or by '|'. Unrecognised or missing colors will be replaced by white color 🤍

Examples ✨

Lower gain by 10 of central frequency 200Hz and width 100 Hz for the first 2 channels using a 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. The syntax for the commands is: "fN|f=freq|w=width|g=gain"

fN is the existing filter number, starting from 0 — if no such filter is available, an error is returned. freq sets the new frequency parameter. width sets the new width parameter in Hertz. gain sets the new gain parameter in dB.

A full filter invocation with asendcmd may look like this:

asendcmd=c='4.0 anequalizer change 0|f=200|w=50|g=1',anequalizer=...

🔇 `anlmdn`

Reduces broadband noise in audio samples using the Non-Local Means algorithm. Your audio's skincare routine, no cap 🧖‍♀️

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. Detectives, but for waveforms 🔍

The filter accepts the following options:

strength, s — sets the denoising strength. Allowed range is from 0.00001 to 10000. Default value is 0.00001 ⭐
patch, p — sets the patch radius duration. Allowed range is from 1 to 100 milliseconds. Default value is 2 milliseconds
research, r — sets the research radius duration. Allowed range is from 2 to 300 milliseconds. Default value is 6 milliseconds
output, o — sets 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 — sets the smooth factor. Default value is 11. Allowed range is from 1 to 1000

Commands 💅

This filter supports the all above options as commands. Yes, the all above options — we love the grammar, it ate 💀

🧠 `anlmf`, `anlms`

Applies the Normalized Least-Mean-(Squares|Fourth) algorithm to the first audio stream using the second audio stream. Two filters, one algorithm family, very sibling energy 👯‍♀️

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). It's literally learning in real time, obsessed 😭

A description of the accepted options follows.

order — sets the filter order
mu — sets the filter mu
eps — sets the filter's eps. that's it, that's the whole vibe ✨
leakage — sets the filter leakage. lowkey a lil leaky on purpose, iykyk 💅
out_mode — picks which signal actually slides out the other side. the options are:
- i — pass the 1st input. straightforward icon behavior 👑
- d — pass the 2nd input.
- o — pass the difference between the desired signal, the 2nd input, and the error signal estimate.
- n — pass the difference between the input, the 1st input, and the error signal estimate.
- e — pass the error signal's estimated samples.

default value is o, bestie 🫶

precision — tells it how much precision to flex when processing samples:
- auto — auto-pick the internal sample format depending on the other filters. let it cook 🔥
- float — always use single-floating point precision sample format.
- double — always use double-floating point precision sample format. max effort, no notes 💀

Examples

one of the many usages of this filter is noise reduction — the input audio gets filtered with the same samples delayed by a fixed amount. here's one such example for stereo audio, and yes the line is long, just copy it exactly bestie:

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, except the option order. everything else? command-able, periodt 💅

`anull`

passes the audio source through to the output completely unchanged. it's giving "i'm just here for moral support" 🫶 zero drama, zero processing, iconic.

`apad`

pads the end of an audio stream with silence. yes bestie, it's the "we need more runtime" filter 🌸

this can be used together with ffmpeg -shortest to extend audio streams to the same length as the video stream. very demure, very matching-lengths 💖

a description of the accepted options follows:

packet_size — set the silence packet size. default value is 4096.
pad_len — set the number of samples of silence to add to the end. after the value is reached, the stream is terminated. this option is mutually exclusive with whole_len — pick one, you cannot have both 🙈
whole_len — set the minimum total number of samples in the output audio stream. 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_len.
pad_dur — specify the duration of samples of silence to add. see the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. only used if set to a non-negative value.
whole_dur — specify the minimum total duration in the output audio stream. see the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. only used 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. no cap, it will just keep going 🌈

note that for ffmpeg 4.4 and earlier, a zero pad_dur or whole_dur also caused the filter to add silence indefinitely. different era, same chaos 💀

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, and pad 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 series of peaks and troughs in the frequency spectrum. the position of the peaks and troughs are modulated so that they vary over time, creating a sweeping effect. it's giving ocean wave but make it audio 🌊

a description of the accepted parameters follows:

in_gain — set input gain. default is 0.4.
out_gain — set output gain. default is 0.74.
delay — set delay in milliseconds. default is 3.0.
decay — set decay. default is 0.4.
speed — set modulation speed in Hz. default is 0.5.
type — set modulation type. default is triangular. it accepts the following values:
- triangular, t
- sinusoidal, s

`aphaseshift`

apply phase shift to input audio samples. shifting, but make it phase ✨

the filter accepts the following options:

shift — specify phase shift. allowed range is from -1.0 to 1.0. default value is 0.0.
level — set output gain applied to final output. allowed range is from 0.0 to 1.0. default value is 1.0.
order — set filter order used for filtering. allowed range is from 1 to 16. default value is 8.

Commands

this filter supports all the above options as commands. every single one, we love a flexible queen 💅

`apsnr`

measure Audio Peak Signal-to-Noise Ratio. it's a vibe check for your audio, fr fr 📊

this filter takes two audio streams for input, and outputs the first audio stream. results are in dB per channel at end of either input. so polite, it reports back at the end 🫶

`apsyclip`

apply a Psychoacoustic clipper to the input audio stream. crunchy but make it science 🔥

the filter accepts the following options:

level_in — set input gain. by default it is 1. range is [0.015625 - 64].
level_out — set output gain. by default it is 1. range is [0.015625 - 64].
clip — set the clipping start value. default value is 0dBFS or 1.
diff — output only difference samples, useful to hear introduced distortions. by default is disabled. little peek behind the curtain 🙈
adaptive — set strength of adaptive distortion applied. default value is 0.5. allowed range is from 0 to 1.
iterations — set number of iterations of the psychoacoustic clipper. allowed range is from 1 to 20. default value is 10.
level — auto level the output signal. default is disabled. this normalizes audio back to 0dBFS if enabled.

Commands

this filter supports all the above options as commands. iconic behavior 👑

`apulsator`

audio pulsator is something between an autopanner and a tremolo. but it can produce funny stereo effects as well 🎀

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 that both LFO shapes match each other — the left and right channel are altered equally, which is a conventional tremolo. an offset of 50% means that the shape of the right channel is exactly shifted in phase (or moved backwards about half of the frequency), and pulsator acts as an autopanner. at 1 both curves match again. every setting in between moves the phase shift gaplessly between all 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. physics but make it a bop 💫

the filter accepts the following options:

level_in — set input gain. by default it is 1. range is [0.015625 - 64].
level_out — set output gain. by default it is 1. range is [0.015625 - 64].
mode — set the waveform shape the LFO will use. can be one of: sine, triangle, square, sawup or sawdown. default is sine.
amount — set modulation. defines how much of the original signal is affected by the LFO.
offset_l — set left channel offset. default is 0. allowed range is [0 - 1].
offset_r — set right channel offset. default is 0.5. allowed range is [0 - 1].
width — set pulse width. default is 1. allowed range is [0 - 2].
timing — set possible timing mode. can be one of: bpm, ms or hz. default is hz.
bpm — set bpm. default is 120. allowed range is [30 - 300]. only used if timing is set to bpm.
ms — set ms. default is 500. allowed range is [10 - 2000]. only used if timing is set to ms.
hz — set frequency in Hz. default is 2. allowed range is [0.01 - 100]. only used if timing is set to hz.

`aresample`

resample 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. it's the "i'll figure it out" filter 🧚

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. choose your fighter 🌈

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`

reverse an audio clip 🔄 yes, it does exactly what it says, no plot twist.

warning: this filter requires memory to buffer the entire clip, so trimming is suggested. it's a hungry girl, feed it less audio 😭

Examples

take the first 5 seconds of a clip, and reverse it:

atrim=end=5,areverse

`arls`

apply the Recursive Least Squares algorithm to the first audio stream using the second audio stream. math girlies rise 👑

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 (the 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 the filter order.
lambda — set the forgetting factor. it forgives, but it doesn't forget... until lambda says so 💅
delta — set the coefficient to initialize the 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 the difference between the desired signal, the 2nd input, and the error signal estimate.
- n — pass the difference between the input, the 1st input, and the error signal estimate.
- e — pass the error signal's estimated samples.

default value is o 🫶

precision — set which precision to use when processing samples:
- auto — auto-pick the internal sample format depending on the other filters.
- float — always use single-floating point precision sample format.
- double — always use double-floating point precision sample format. go big or go home 💀

`arnndn`

reduce noise from speech using Recurrent Neural Networks 🧠✨ your voice, but with the background chaos politely escorted out.

this filter accepts the following options:

model, m — set the train model file to load. this option is always required. no model, no filter, that's the rule bestie 🌸
mix — sets how much of the filtered samples get mixed back into the final output, bestie. Allowed range is from -1 to 1, and the default is 1. Negative values are ✨special✨ — they set how much of the filtered noise you keep in the final output instead. Set this 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

Measures the Audio Signal-to-Distortion Ratio, no cap. 📏

This filter takes two audio streams as input, and outputs the first audio stream. Results are in dB per channel, delivered at the end of either input. 🎀

asetnsamples

Sets the number of samples in each output audio frame. 🧚

The last output packet may contain a different number of samples, because the filter flushes out all the remaining samples when the input audio signals its end. So like, don't panic if the last one looks different. 😭

The filter accepts the following options:

nb_out_samples, n — sets 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 pads 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

Sets the sample rate without altering the PCM data. Heads up: this results in a change of speed and pitch. 🎵 (iykyk — chipmunk mode incoming.)

The filter accepts the following options:

sample_rate, r — sets the output sample rate. Default is 44100 Hz.

ashowinfo

Shows one line containing various info for each input audio frame. The input audio is not modified — it's purely here to spill the tea. ☕

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

Measures the Audio Scaled-Invariant Signal-to-Distortion Ratio. Fancy name, iconic job. 👑

This filter takes two audio streams as input, and outputs the first audio stream. Results are in dB per channel at the end of either input.

asoftclip

Applies 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 glow-up instead of a brick wall 💀.

This filter accepts the following options:

type — sets the type of soft-clipping.

It accepts the following values:

hard
tanh
atan
cubic
exp
alg
quintic
sin
erf

threshold — sets the threshold from where to start clipping. Default value is 0dB or 1.
output — sets the gain applied to output. Default value is 0dB or 1.
param — sets an additional parameter which controls the sigmoid function.
oversample — sets the oversampling factor.

Commands

This filter supports the all above options as commands. 💖

aspectralstats

Displays frequency-domain statistical information about the audio channels. Stats are calculated and stored as metadata for each audio channel and for each audio frame. 🦋

It accepts the following option:

win_size — sets the window length in samples. Default value is 2048. Allowed range is from 32 to 65536.
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. Allowed range is from 0 to 1. Default value is 0.5.
measure — selects 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 🎤 — literally the filter that listens so you don't have to.

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 — sets the sampling rate of the input audio. Default is 16000. This needs to match the speech models, otherwise you'll get poor results. (Vibe check your models first, bestie. 🙈)
hmm — sets the dictionary containing acoustic model files.
dict — sets the pronunciation dictionary.
lm — sets the language model file.
lmctl — sets the language model set.
lmname — sets which language model to use.
logfn — sets the output for log messages.

The filter exports recognized speech as the frame metadata "lavfi.asr.text". ✨

astats

Displays time-domain statistical info about the audio channels. Stats are calculated and displayed for each audio channel, and where applicable an overall figure is also given. 📊

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 — sets 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 — sets the number of frames over which cumulative stats are calculated before being reset. Default is disabled.
measure_perchannel — selects 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 — selects 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 follows:

none — no measures.
all — all measures.
Bit_depth — overall bit depth of audio, i.e. the 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 the whole audio; entropy of a 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 the 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 the 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

Boosts subwoofer frequencies. 🔊 Turn the bass all the way up, respectfully.

The filter accepts the following options:

dry — sets the dry gain, i.e. how much of the original signal is kept. Allowed range is from 0 to 1. Default value is 1.0.
wet — sets the wet gain, i.e. how much of the filtered signal is kept. Allowed range is from 0 to 1. Default value is 1.0.
boost — sets the max boost factor. Allowed range is from 1 to 12. Default value is 2.
decay — sets the delay line decay gain value. Allowed range is from 0 to 1. Default value is 0.0.
feedback — sets the delay line feedback gain value. Allowed range is from 0 to 1. Default value is 0.9.
cutoff — sets the cutoff frequency in Hertz. Allowed range is 50 to 900. Default value is 100.
slope — sets how steep the cutoff slope is 🎢 allowed range is 0.0001 to 1, and the default is 0.5.
delay — sets the delay, bestie. allowed range is 1 to 100, default 20.
channels — picks which channels get processed. default is literally all available ones 💜

✨ Commands ✨

this filter supports all of the above options as commands, periodt 💅

🦋 asubcut

cuts the subwoofer frequencies. like, bye bass 👋

the whole point of this one is that it lets you set a custom, steeper roll off than highpass gives you, so it can attenuate the frequency content in the stop-band way more aggressively. it's giving extra-strength 💪

here's what it accepts:

cutoff — cutoff frequency in Hertz. allowed range is 2 to 200. default 20.
order — the filter order. available values are 3 to 20. default 10.
level — input gain level. allowed range is 0 to 1. default 1.

Commands

this filter supports all the above options as commands 💖

🌈 asupercut

cut super frequencies ✂️✨

options:

cutoff — cutoff frequency in Hertz. allowed range is 20000 to 192000. default 20000.
order — filter order. available values are 3 to 20. default 10.
level — input gain level. allowed range is 0 to 1. default 1.

Commands

yep, all the above options work as commands 🫶

🎀 asuperpass

applies a high order Butterworth band-pass filter. fancy name, iconic behaviour 👑

options:

centerf — center frequency in Hertz. allowed range is 2 to 999999. default 1000.
order — filter order. available values are 4 to 20. default 4.
qfactor — the Q-factor. allowed range is 0.01 to 100. default 1.
level — input gain level. allowed range is 0 to 2. default 1.

Commands

all of the above = commands, no cap 💅

⭐ asuperstop

applies a high order Butterworth band-stop filter. the anti-band-pass, if you will 🙈

options:

centerf — center frequency in Hertz. allowed range is 2 to 999999. default 1000.
order — filter order. available values are 4 to 20. default 4.
qfactor — Q-factor. allowed range is 0.01 to 100. default 1.
level — input gain level. allowed range is 0 to 2. default 1.

Commands

all above options, usable as commands 💜

🧚 atempo

adjusts the audio tempo. the speed-up/slow-down filter, and it does it without touching the pitch, which is lowkey the coolest thing ever ✨

it accepts exactly one parameter: the audio tempo. if you don't specify it, the filter assumes the nominal 1.0 tempo. tempo has to be in the [0.5, 100.0] range.

⚠️ heads up bestie: tempo greater than 2 will skip some samples rather than blending them in. if that's a problem for any reason, you can always daisy-chain several atempo instances to get the product tempo you want 🫶

Examples

slow the audio down to 80% tempo:

atempo=0.8

speed the audio up to 300% tempo:

atempo=3

speed it up to 300% by daisy-chaining two atempo instances:

atempo=sqrt(3),atempo=sqrt(3)

Commands

this filter supports these commands:

tempo — changes the filter tempo scale factor. syntax for the command is: "tempo"

🌸 atilt

applies a spectral tilt filter to the audio stream. it can apply any spectral roll-off slope over any frequency band you specify 🎢

options:

freq — central frequency of the tilt, in Hz. default is 10000 Hz.
slope — slope direction of the tilt. default is 0. allowed range is -1 to 1.
width — width of the tilt. default is 1000. allowed range is 100 to 10000.
order — order of the tilt filter.
level — input volume level. allowed range is 0 to 4. default is 1.

Commands

all of the above options double as commands 😌💅

🍓 atrim

trims the input so the output contains one continuous subpart of it. it's the scissors filter ✂️

it takes these parameters:

start — timestamp (in seconds) of the start of the section to keep. like, the audio sample with the timestamp start becomes the first sample in the output.
end — the time of the first audio sample that gets dropped. i.e. the sample immediately before the one with the timestamp end is the last sample in the output.
start_pts — same as start, except this one sets the start timestamp in samples instead of seconds.
end_pts — same as end, except this 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 — peep the Time duration section in the ffmpeg-utils(1) manual if you need the deets 📖

💀 okay so this part matters: the first two sets of 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 you different results when the timestamps are wrong, inexact, or don't start at zero. also, this filter doesn't modify the timestamps at all. if you want the output timestamps to start at zero, put the asetpts filter after atrim.

if you set multiple start or end options, this filter tries to be greedy and keeps all samples matching at least one of the constraints. want only the part that matches all the constraints at once? chain multiple atrim filters, bestie 🫶

the defaults are set so that all the input is kept. so you can totally set, like, 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

calculates the normalized windowed cross-correlation between two input audio streams. very math, very main character energy 🧮✨

the resulting samples are always between -1 and 1 inclusive. if the result is 1 it means the two input samples are highly correlated in that selected segment. result 0 means they're not correlated at all. and if the result is -1 it means the two input samples are out of phase, meaning they cancel each other out 💀

options:

size — sets the size of the segment over which cross-correlation is calculated. default is 256. allowed range is 2 to 131072.
algo — sets the algorithm for cross-correlation. can be "slow", "fast" or "best". default is "best". the fast algorithm assumes mean values over any given segment are always zero, and so it needs way fewer calculations to do its thing. that's generally not true, but it is valid for typical audio streams 🎧

Examples

calculate the correlation between channels in a stereo audio stream:

ffmpeg -i stereo.wav -af channelsplit,axcorrelate=size=1024:algo=fast correlation.wav

🎧 bandpass

applies 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 rolls off at 6dB per octave (that's 20dB per decade, iykyk) 📉

options:

frequency, f — sets the filter's central frequency. default is 3000.
csg — constant skirt gain if set to 1. defaults to 0.
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 — specifies the band-width of the filter, in width_type units.
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 — normalize the biquad coefficients. disabled by default. 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 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. if you set this value 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 🤢

🙈 also note: the filter delay will be exactly this many samples when this is set to a non-zero value.

Commands

this filter supports these commands:

frequency, f — change the bandpass frequency. syntax for the command is: "frequency"
width_type, t — change the bandpass width_type. syntax for the command is: "width_type"
width, w — change the bandpass width. syntax for the command is: "width"
mix, m — change the bandpass mix. syntax for the command is: "mix"

💜 bandreject

applies a two-pole Butterworth band-reject filter with central frequency frequency and (3dB-point) band-width width. it rolls off at 6dB per octave (20dB per decade) 🌀

options:

frequency, f — sets the filter's central frequency. default is 3000.
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 — specifies the band-width of the filter, in width_type units.
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 — normalize the biquad coefficients. disabled by default. 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 — picks the sample format automatically, based on whatever the surround filters are doing ✨
s16 — always signed 16-bit, periodt.
s32 — always signed 32-bit, no cap.
f32 — always float 32-bit, she's ✨precise✨.
f64 — always float 64-bit, maximum delulu precision 💅
block_size, b — sets the block size used for reverse IIR processing 🎀 if you set it high enough (higher than the impulse response length truncated when it reaches near-zero values) the filtering becomes linear phase — iconic. but if it's not big enough it will literally just produce nasty artifacts 😭

heads up bestie: the filter delay will be exactly this many samples when you set it to a non-zero value. fr fr.

Commands

This filter supports the following commands:

frequency, f — change bandreject frequency. Syntax for the command is: "frequency"
width_type, t — change bandreject width_type. Syntax for the command is: "width_type"
width, w — change bandreject width. Syntax for the command is: "width"
mix, m — change bandreject mix. Syntax for the command is: "mix"

bass, lowshelf

Boost or cut the bass (lower) 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), iykyk.

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, that's the vibe check you do NOT want to fail.
frequency, f — set the filter's central frequency, so you can extend or reduce the frequency range to be boosted or cut. Default value is 100 Hz. 🎀
width_type, t — set the method to specify band-width of 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 — set 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 biquad coefficients; by default it's disabled. Enabling it will normalize the 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 the 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 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 bass frequency. Syntax for the command is: "frequency"
width_type, t — change bass width_type. Syntax for the command is: "width_type"
width, w — change bass width. Syntax for the command is: "width"
gain, g — change bass gain. Syntax for the command is: "gain"
mix, m — change bass mix. Syntax for the command is: "mix"

biquad

Apply a biquad IIR filter with the given coefficients 🧮 where b0, b1, b2 and a0, a1, a2 are the numerator and denominator coefficients respectively. and channels, c specify which channels to filter, by default all available are filtered.

Commands

This filter supports the following commands:

a0
a1
a2
b0
b1
b2 — change biquad parameter. Syntax for the command is: "value"
mix, m — how much to use the 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 the impulse response length truncated when it reaches near zero values) filtering will become linear phase; otherwise, if not big enough, it will just produce nasty artifacts 💀

note that the filter delay will be exactly this many samples when set to a non-zero value.

bs2b

Bauer stereo to binaural transformation, which improves headphone listening of stereo audio records 🎧🦋

To enable compilation of this filter you need to configure FFmpeg with --enable-libbs2b.

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

Remap input channels to new locations 🗺️✨

It accepts the following parameters:

map — map 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 is not given then it is 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. 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

Split each channel from an input audio stream into a separate output stream 🎀 like, one stream per channel, she's organized.

It accepts the following parameters:

channel_layout — the channel layout of the input stream. The 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. The default is "all".

choosing channels not present in the channel layout in 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

Add a chorus effect to the audio 🌸

Can make a single vocal sound like a chorus, but can also be applied to instrumentation. She's versatile, okurrr.

Chorus resembles an echo effect with a short delay, but whereas with echo the delay is constant, with chorus it is varied using sinusoidal or triangular modulation. The modulation depth defines the range the modulated delay is played before or after the delay. Hence the delayed sound will sound slower or faster, that is the delayed sound tuned around the original one, like in a chorus where some vocals are slightly off key 🎶💜

It accepts the following parameters:

in_gain — set input gain. Default is 0.4.
out_gain — set output gain. Default is 0.4.
delays — set delays. A typical delay is around 40ms to 60ms.
decays — set decays.
speeds — set speeds.
depths — set 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

Compress or expand the audio's dynamic range 🎚️✨

It accepts the following parameters:

attacks
decays — a list of times in seconds for each channel over which the instantaneous level of the input signal is averaged to determine its volume. attacks refers to increase of volume and decays refers to decrease of volume. For most situations, the attack time (response to the audio getting louder) should be shorter than the decay time, because the human ear is more sensitive to sudden loud audio than sudden soft audio. A typical value for attack is 0.3 seconds and a typical value for decay is 0.8 seconds. If the specified number of attacks & decays is lower than the number of channels, the last set attack/decay will be used for all remaining channels 🫶
points — the list of points for the transfer function, given in dB relative to the maximum possible signal amplitude, bestie. Each key points list has to be defined with this syntax: "x0/y0|x1/y1|x2/y2|...." or "x0/y0 x1/y1 x2/y2 ...."

The input values have to be in strictly increasing order, but lowkey the transfer function does not have to be monotonically rising. 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. It defaults to 0.01. 🎀
gain — sets the additional gain in dB to be applied at all points on the transfer function, so you can adjust the overall gain super easily. It defaults to 0.
volume — sets an initial volume, in dB, to be assumed for each channel when filtering starts. This lets you supply a nominal level up front so that, for example, a very large gain is not applied to initial signal levels before the companding has begun to operate. A typical value for audio that's initially quiet is -90 dB. It 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. Specifying a delay approximately equal to the attack/decay times lets the filter effectively operate in predictive rather than reactive mode. It 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 (making 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 to compensate for differing positions of microphones or speakers. Iconic, honestly.

So like, say you recorded guitar with two microphones placed in different locations. Because the front of the sound wave has a fixed speed in normal conditions, the phasing of the microphones can vary and depends on their location and interposition. The best sound mix can be achieved when these microphones are in phase (synchronized). Note that a distance of ~30 cm between microphones makes one microphone capture the signal in antiphase to the other microphone — and that makes the final mix sound moody 😭. This filter helps solve phasing problems by adding different delays to each microphone track and making them synchronized. Slay.

The best result can be reached when you take one track as base and synchronize other tracks one by one with it. Remember that synchronization/delay tolerance depends on sample rate, too — higher sample rates will give more tolerance. No cap.

The filter accepts the following parameters:

mm — set millimeters distance. This is compensation distance for fine tuning. Default is 0.
cm — set cm distance. This is compensation distance for tightening distance setup. Default is 0.
m — set meters distance. This is compensation distance for hard distance setup. Default is 0.
dry — set dry amount. Amount of unprocessed (dry) signal. Default is 0.
wet — set wet amount. 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 all the above options as commands. 💅

🦋 crossfeed 🦋

Apply the headphone crossfeed filter.

Crossfeed is the process of blending the left and right channels of a stereo audio recording. It's mainly used to reduce extreme stereo separation of low frequencies.

The intent is to produce a more speaker-like sound to the listener. Periodt.

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 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 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 will 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 all the 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 to 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 all the above options as commands.

💜 dcshift 💜

Apply a DC shift to the audio.

This can be useful to remove a DC offset (caused perhaps by a hardware problem in the recording chain) from the 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. Good to know tbh.

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 🍓

Apply 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 amount of ducking on the treble part of 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 🌈

Enhance 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 originally available in both stereo channels. This filter outputs front left and front right channels the same as available in the stereo input. Main character energy fr fr.

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 all the above options as commands 🫶.

⭐ drmeter ⭐

Measure audio dynamic range.

DR values of 14 and higher are found in very dynamic material. DR of 8 to 13 is found in transition material. And anything less than 8 has very poor dynamics and is very compressed.

The filter accepts the following options:

length — set the window length in seconds used to split the audio into segments of equal length. Default is 3 seconds.

💅 dynaudnorm 💅

Dynamic Audio Normalizer.

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). However, in contrast to more "simple" normalization algorithms, the Dynamic Audio Normalizer dynamically re-adjusts the gain factor to the input audio. This lets you apply extra gain to the "quiet" sections of the audio while avoiding distorting 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. Obsessed, honestly 🔥.

framelen, f — sets the frame length, in milliseconds. Valid range is 10 to 8000 ms, default is 500 ms. Okay so real talk: the Dynamic Audio Normalizer chews through your audio in little chunks called frames, and it has to do this because a peak magnitude means literally nothing for one single sample value 💀 — you need a whole contiguous run of samples to figure out what the peak actually is. A "standard" normalizer would just grab the peak of the entire file, but this one is extra and finds the peak individually for every frame. Frame length is given in ms, and the default 500 ms has been found to give good results with most files. Oh and heads up: the exact frame length in samples gets determined automatically from the sampling rate of each 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 in the whole Dynamic Audio Normalizer, no cap. The window size is specified in frames, centered on the current frame, and for simplicity's sake it has to be odd. So the default 31 means: the current frame, plus the 15 frames before it, plus the 15 frames after it 🫶 Bigger window → stronger smoothing → less gain variation → slower gain adaptation. Smaller window → weaker smoothing → more gain variation → faster gain adaptation. In other words bestie, crank this up and it starts acting like a "traditional" normalization filter; dial it down and it starts acting like a dynamic range compressor. Pick your fighter 💅
peak, p — sets the target peak value, i.e. the highest permissible magnitude level for the normalized audio. The filter tries to get as close to that target as it can while 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 ⚠️
maxgain, m — sets the maximum gain factor. Range 1.0 to 100.0, default 10.0. For each input frame the normalizer works out the maximum possible local gain factor — the biggest gain that won'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, specifically so it doesn't go feral with huge gain on "silent" or almost-silent frames 🙈 Default max is 10.0, and for most inputs that's plenty; usually it is not recommended to increase it. That said, if your input has an extremely low overall volume level, you might need to allow even higher gain factors. But note: the normalizer doesn't just slap a "hard" threshold on it (like cutting off values above the threshold) — instead a "sigmoid" threshold function gets applied, so gain factors smoothly approach the threshold value but never exceed it. Iconic behaviour honestly 👑
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 get amplified as much as possible without exceeding the max signal level, i.e. 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 therefore considered a better approximation of "perceived loudness" than just staring at the signal's peak magnitude 👀 So by adjusting all frames to a constant RMS value, you get a uniform "perceived loudness" — slay. If a target RMS value is specified, a frame's local gain factor is defined as the factor that would result in exactly that RMS value. But note: the maximum local gain factor is still restricted by the frame's highest magnitude sample, in order to prevent clipping 🎀
coupling, n — enable channels coupling. Enabled by default. Default behaviour: amplify all channels by the same amount, meaning the same gain factor applies to every channel, so the maximum possible gain factor is determined by the "loudest" channel. But some recordings have uneven volume across channels — one channel might be "quieter" than the other(s) — and in that case you can use this option to disable the channel coupling 💜 Then the gain factor is determined independently for each channel, depending only on that individual channel's highest magnitude sample. Which lets you harmonize the volume of the different channels. Chef's kiss 🍓
correctdc, c — enable DC bias correction. Disabled by default. So, an audio signal (in the time domain) is a sequence of sample values, and in the Dynamic Audio Normalizer those sample values are represented in the -1.0 to 1.0 range, regardless of the original input format. Normally the audio signal (the "waveform") should be centered around the zero point — meaning if you calculate the mean of all samples in a file, or in a single frame, the result should be 0.0 or at least very close to it. If there's a significant deviation of the mean from 0.0, in either the positive or negative direction, that's called a DC bias or DC offset. Since a DC bias is clearly undesirable, the normalizer offers optional DC bias correction ✨ With it enabled, the Dynamic Audio Normalizer determines the mean value (the "DC correction" offset) of each input frame and subtracts that value from all of the frame's sample values, which puts those samples back centered around 0.0. And to avoid "gaps" at the frame boundaries, the DC correction offset values get interpolated smoothly between neighbouring frames. Smooth like butter 🧈
altboundary, b — enable alternative boundary mode. Disabled by default. The normalizer looks at a certain neighbourhood around each frame — the frames before it and the frames after it. But for the "boundary" frames at the very beginning and the very end of the audio file, not all neighbouring frames are available 😭 Like, for the first few frames of the file the preceding frames aren't known, and for the last few frames the subsequent frames aren't known. So the question arises: what gain factors do we assume for the missing frames in that "boundary" region? The Dynamic Audio Normalizer implements two modes for this. The default boundary mode assumes a gain factor of exactly 1.0 for the missing frames, which results in a smooth "fade in" and "fade out" at the beginning and end of the input respectively 🌸
compress, s — sets the compress factor. Range 0.0 to 30.0, default 0.0. By default the Dynamic Audio Normalizer does not apply "traditional" compression — signal peaks don't get pruned, so the full dynamic range is retained within each local neighbourhood. But sometimes you want to combine the normalization algorithm with a more "traditional" compression, and that's exactly what this optional compression (thresholding) function is for. If (and only if) the compression feature is enabled, all input frames get processed by a soft knee thresholding function prior to the actual normalization process. Put simply, the thresholding function prunes all samples whose magnitude exceeds a certain threshold value. But — and you knew this was coming — the normalizer doesn't just apply a fixed threshold value; the threshold gets adjusted for each individual frame 🦋 In general, smaller parameters give stronger compression, and vice versa. Values below 3.0 are not recommended, because audible distortion may appear 💀
threshold, t — sets the target threshold value. This specifies the lowest permissible magnitude level for the audio input that will be normalized. If an input frame's volume is above this value, the frame gets normalized; otherwise the frame may not be normalized at all. Default is 0, which means every input frame gets normalized. This option is mostly useful when you don't want digital noise to get 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 you less conservative gain adjustments, kind of like when the framelen option is set to a smaller value — and if the framelen option value is compensated for non-zero overlap, then gain adjustments will be smoother across time compared to the zero overlap case 🌈
curve, v — specify the peak mapping curve expression, which gets 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. 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 — current frame peak value 👑

💬 Commands 💬

This filter supports the all above options as commands. yes bestie, all of them, go wild 🫶

🎧 earwax 🎧

Makes audio easier to listen to on headphones. that's it, that's the whole vibe 💖

What it actually does: it adds cues to 44.1kHz stereo (i.e. audio CD format) audio so that when you listen on headphones, the stereo image gets moved from inside your head (which is the standard headphone experience, lowkey cursed) to outside and in front of the listener (which is the standard speaker experience). it's giving "I'm not trapped in your skull anymore" ✨

Ported from SoX. iconic lineage fr fr 🌈

🎚️ 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, whilst — and this is the whole point, unlike bandpass and bandreject filters — that at all other frequencies is unchanged. no cap, everything else stays exactly where it was 💅

So you can build complex equalisation curves by giving this filter several times, each with a different central frequency. stack 'em like bracelets 🎀

The filter accepts the following options:

frequency, f — set the filter's central frequency in Hz.
width_type, t — set method to specify band-width of filter. the options are:
- h — Hz
- q — Q-Factor
- o — octave
- s — slope
- k — kHz
width, w — specify the band-width of a filter in width_type units.
gain, g — set the required gain or attenuation in dB. > ⚠️ Beware of clipping when using a positive gain. like genuinely, watch it, positive gain will clip on you and it will not be cute 😭
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 channels are filtered.
normalize, n — normalize biquad coefficients, by default it's disabled. Enabling it will normalize magnitude response at DC to 0dB.
transform, a — set transform type of IIR filter. your picks are:
- 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.

🌸 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. main character energy for your stereo field ✨

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 sound will be unchanged, with -1.0 left and right channels will be swapped. yes, swapped. chaos mode, we love her 💅
c — enable clipping. by default is enabled. (so if you want no clipping you gotta turn it off yourself, iykyk)

💬 Commands 💬

This filter supports the all above options as commands. 🫶

🔥 firequalizer 🔥

Apply FIR Equalization using arbitrary frequency response. this one is the overachiever of the EQ filters fr fr 👑

The filter accepts the following option:

gain — set 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 command. Default is gain_interpolate(f).

gain_entry — set gain entry for 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 command.

delay — set filter delay in seconds. Higher value means more accurate. Default is 0.01.
accuracy — set filter accuracy in Hz. Lower value means more accurate. Default is 5.
wfunc — set window function. Acceptable values are:
- rectangular — rectangular window, useful when 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 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 — set 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 — set file for dumping, suitable for gnuplot. 📈
dumpscale — set scale for dumpfile. Acceptable values are same with 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. swooshy swoosh, you know the one 🦋

The filter accepts the following options:

delay — set base delay in milliseconds. Range from 0 to 30. Default value is 0.
depth — set added sweep delay in milliseconds. Range from 0 to 10. Default value is 2.
regen — set percentage regeneration (delayed signal feedback). Range from -95 to 95. Default value is 0.
width — set percentage of delayed signal mixed with original. Range from 0 to 100. Default value is 71.
speed — set sweeps per second (Hz). Range from 0.1 to 10. Default value is 0.5.
shape — set swept wave shape, can be triangular or sinusoidal. Default value is sinusoidal.
phase — set swept wave percentage-shift for multi channel. Range from 0 to 100. Default value is 25.
interp — set delay-line interpolation, linear or quadratic. Default is linear.

🫶 haas 🫶

Apply Haas effect to audio.

💡 Note that this makes most sense to apply on mono signals. With this filter applied to mono signals it give some directionality and stretches its stereo image. so like... mono in, suddenly-it's-everywhere out ✨

The filter accepts the following options:

level_in — set input level. By default is 1, or 0dB.
level_out — set output level. By default is 1, or 0dB.
side_gain — set gain applied to side part of signal. By default is 1.
middle_source — set 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 is disabled.
left_delay — set left channel delay. By default is 2.05 milliseconds.
left_balance — set left channel balance. By default is -1.
left_gain — set left channel gain. By default is 1.
left_phase — change left phase. By default is disabled.
right_delay — set right channel delay. By defaults is 2.12 milliseconds.
right_balance — set right channel balance. By default is 1.
right_gain — set right channel gain. By default is 1.
right_phase — change right phase. By default is enabled.

💿 hdcd 💿

Decodes High Definition Compatible Digital (HDCD) data. A 16-bit PCM stream with embedded HDCD codes is expanded into a 20-bit PCM stream. glow-up, but for bits 💖

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 the default encoding for wav is 16-bit, so the resulting 20-bit stream will be truncated back to 16-bit. like, all that work and then wav just chops it 😭 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 — processes the stereo channels together as one unit 💞 and if target_gain doesn't match between the channels, it's treated as invalid and the last valid target_gain gets reused instead. consistency is the vibe.
cdt_ms — sets the code detect timer period, in ms ⏱️
force_pe — always extends peaks above -3dBFS, even if PE isn't signaled. no invitation needed, the peaks are getting extended anyway 💅
analyze_mode — replaces the audio with a solid tone and adjusts the amplitude so it signals 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`

okay so this one applies head-related transfer functions (HRTFs) to create virtual loudspeakers all around the listener, for binaural listening through headphones 🎶🫶 the HRIRs come in through additional streams — for each channel you need one stereo input stream.

it accepts these options:

map — sets the mapping of input streams for convolution. the argument is a '|'-separated list of channel names, in the order they're given as additional stream inputs for the filter. this also specifies the number of input streams, and the number of input streams must be not less than the number of channels in the first stream plus one 🗺️
gain — sets the gain applied to the audio. the value is in dB. default is 0.
type — sets the 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 — sets a custom gain for LFE channels. the value is in dB. default is 0.
size — sets the size of the frame, in number of samples, that will be processed at once. default value is 1024. allowed range is from 1024 to 96000.
hrir — sets the format of the hrir stream. default value is stereo, the alternative value is multich. if it's 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 it's set to multich, the number of additional streams should be exactly one — and the number of input channels of the additional stream should be equal to 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 a 7.1 downmix, where 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 a 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`

applies a high-pass filter with a 3dB point frequency. the filter can be single-pole, or double-pole (that's the default). it rolls off at 6dB per pole per octave (which is 20dB per pole per decade) 📉

the filter accepts these options:

frequency, f — sets the frequency in Hz. default is 3000.
poles, p — sets the number of poles. default is 2.
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 — specifies the band-width of the filter, in width_type units. applies only to the double-pole filter. the default is 0.707q, and that gives you a Butterworth response 🧈
mix, m — how much of the filtered signal ends up 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 the 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 — 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. 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 will 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 these commands, so you can change stuff live:

frequency, f — change the highpass frequency. syntax for the command is "frequency"
width_type, t — change the highpass width_type. syntax for the command is "width_type"
width, w — change the highpass width. syntax for the command is "width"
mix, m — change the highpass mix. syntax for the command is "mix"

🔗 `join`

joins multiple input streams into one multi-channel stream. it literally just brings everybody together 🫶

it accepts these parameters:

inputs — the number of input streams. defaults to 2.
channel_layout — the desired output channel layout. defaults to stereo.
map — maps channels from inputs to the output. the argument is a '|'-separated list of mappings, each one 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 (like FL for front left) or its index in the specified input stream. out_channel is the name of the output channel.

if you don't specify the mappings explicitly, the filter will attempt to guess them — it does that by first trying to find an unused matching input channel, and if that fails it just picks the first unused input channel 👍

✨ Examples

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`

loads a LADSPA (Linux Audio Developer's Simple Plugin API) plugin — yes that's the whole name, they really committed to the acronym 😭

to enable compilation of this filter you need to configure FFmpeg with --enable-ladspa 🛠️

file, f — specifies the name of the LADSPA plugin library to load. if the environment variable LADSPA_PATH is defined, the LADSPA plugin is searched for in each one of the directories specified by the colon-separated list in LADSPA_PATH; otherwise it looks 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 isn't set, the filter will list all available plugins within the specified library 👑
controls, c — sets the '|'-separated list of controls, which are zero or more floating point values that determine the behavior of the loaded plugin (things like delay, threshold or gain). controls need to be defined using this 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 this syntax: value0|value1|value2|..., where valuei is the value set on the i-th control. and if controls is set to "help", all available controls and their valid ranges get printed 📋
sample_rate, s — specifies the sample rate, defaults to 44100. only used if the plugin has zero inputs.
nb_samples, n — sets the number of samples per channel per each output frame, default is 1024. only used if the plugin has zero inputs.
duration, d — sets 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, since the generated audio is always cut at the end of a complete frame. if not specified, or if the expressed duration is negative, the audio is supposed to be generated forever ♾️ only used if the plugin has zero inputs.
latency, l — enables latency compensation. by default it's disabled. only used if the plugin has inputs.

💖 Examples

list all available plugins within the amp (LADSPA example plugin) library:

ladspa=file=amp

list all available controls and their valid ranges for the "vcf_notch" plugin from the "VCF" library:

ladspa=f=vcf:p=vcf_notch:c=help

simulate low quality audio equipment using the "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 the plugin "C* Click - Metronome" from the "C* Audio Plugin Suite" (CAPS) library:

ladspa=file=caps:Click:c=c1=20'

apply the "C* Eq10X2 - Stereo 10-band equaliser" effect:

ladspa=caps:Eq10X2:c=c0=-48|c9=-24|c3=12|c4=2

increase volume by 20dB using the fast lookahead limiter from Steve Harris's "SWH Plugins" collection:

ladspa=fast_lookahead_limiter_1913:fastLookaheadLimiter:20|0|2

attenuate low frequencies using the Multiband EQ from Steve Harris's "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:
🦋 and the last one of the previous batch, ladspa=caps:Narrower, still living its best life:

ladspa=caps:Narrower

💜 Another white noise moment, this time pulling from the "C* Audio Plugin Suite" (CAPS) library:

ladspa=caps:White:.2

🌸 Some fractal noise, also courtesy of "C* Audio Plugin Suite" (CAPS):

ladspa=caps:Fractal:c=c1=1

🎀 Dynamic volume normalization using the "VLevel" plugin:

ladspa=vlevel-ladspa:vlevel_mono

💅 Commands

Bestie, this filter takes commands too — here's the tea:

cN — tweak the N-th control value. And if the value you hand over isn't valid? It just gets ignored, no cap, and the previous one stays exactly where it was. 🙈

🎧 `loudnorm`

EBU R128 loudness normalization, and yes she does both dynamic and linear normalization modes. She's got support for single pass (livestreams, files) and double pass (files) — versatile icon. This algorithm can target IL, LRA, and maximum true peak. Heads up: in dynamic mode, to accurately detect true peaks, the audio stream gets upsampled to 192 kHz, so use the -ar option or the aresample filter to explicitly set an output sample rate. ✨

The filter accepts the following options:

I, i — set the integrated loudness target. Range is -70.0 - -5.0, default is -24.0.
LRA, lra — set the loudness range target. Range is 1.0 - 50.0, default is 7.0.
TP, tp — set the maximum true peak. Range is -9.0 - +0.0, default is -2.0.
measured_I, measured_i — measured IL of the input file. Range is -99.0 - +0.0.
measured_LRA, measured_lra — measured LRA of the input file. Range is 0.0 - 99.0.
measured_TP, measured_tp — measured true peak of the input file. Range is -99.0 - +99.0.
measured_thresh — measured threshold of the input file. Range is -99.0 - +0.0.
offset — set the offset gain. The gain gets applied before the true-peak limiter. Range is -99.0 - +99.0, default is +0.0.
linear — normalize by linearly scaling the source audio. For this, 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 that exceeds the target TP. If any of those 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 meant for playback on a stereo system, its EBU R128 measurement will be perceptually incorrect, which is giving broken. If set to true, this option compensates for that effect. Multi-channel input files are not affected by this option at all. Options are true or false, default is false.
print_format — set the print format for stats. Options are summary, json, or none. Default value is none.
stats_file — write stats to the specified file. The 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 (double-pole is the default, obvi). It rolls off at 6dB per pole per octave (20dB per pole per decade). Slay.

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 the method to specify band-width of the 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 the double-pole filter. Default is 0.707q, which gives a Butterworth response. 👑
mix, m — how much of the filtered signal to use in the output. Default is 1, range is between 0 and 1.
channels, c — specify which channels to filter; by default all available channels get filtered.
normalize, n — normalize biquad coefficients, disabled by default. Enabling it will normalize magnitude response at DC to 0dB.
transform, a — set transform type of the IIR filter:
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — set 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 — set the block size used for reverse IIR processing. If you set this high enough (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

Lowpass only the LFE channel — and if LFE isn't present, it does absolutely nothing:

lowpass=c=LFE

💅 Commands

This filter supports the following 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 an 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 (think delay, threshold, or gain). If controls is set to help, all available controls and their valid ranges get printed. 🫶
sample_rate, s — specify the sample rate, defaults to 44100. Only used if the plugin has zero inputs.
nb_samples, n — set the number of samples per channel per each output frame, default is 1024. Only used if the plugin has 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, since the generated audio is always cut at the end of a complete frame. If it's not specified, or the expressed duration is negative, the audio is supposed to be generated forever. Only used if the plugin has zero inputs.

🌈 Examples

Apply the bass enhancer plugin from Calf:

lv2=p=http\\\\://calf.sourceforge.net/plugins/BassEnhancer:c=amount=2

Apply the vinyl plugin from Calf:

lv2=p=http\\\\://calf.sourceforge.net/plugins/Vinyl:c=drone=0.2|aging=0.5

Apply the 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 the plugin as commands. iykyk.

🎚️ `mcompand`

Multiband compress or expand the audio's dynamic range. The input audio gets divided into bands using 4th order Linkwitz-Riley IIRs — this is akin to the crossover of a loudspeaker, and it results in a flat frequency response when there's no compander action happening. ✨

It accepts the following parameters:

args — the option syntax is: attack,decay,[attack,decay..] soft-knee points crossover_frequency [delay [initial_volume [gain]]] | attack,decay ... — for an explanation of each item, refer to the compand filter documentation.

🎛️ `pan`

Mix channels with specific gain levels. The filter accepts the output channel layout followed by a set of channel definitions. She's also designed to efficiently remap the channels of an audio stream — main character energy, honestly.

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 leaves 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 — pick a lane, bestie. 💅

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. Obsessed with a smart little optimization.

💖 Mixing examples

Say you wanna 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-channel 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

Ngl, ffmpeg integrates a default down-mix (and up-mix) system that should be preferred (see the -ac option) unless you have very specific needs. So don't reinvent the wheel for no reason. 🫶

🌸 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. Ate. 🔥

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 the 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 into 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 — literally zero notes, she just observes. At the 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. That's it, that's the whole vibe — she's an internal icon, not a public one. 💅

🧑‍🎤 `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 🎵 like, how fast the whole thing grooves.
pitch — sets the pitch scale factor 🎤 up or down, your call bestie.
transients — picks the transients detector. The possible values are literally:
- crisp
- mixed
- smooth
detector — picks the detector. Possible values, periodt:
- compound
- percussive
- soft
phase — sets phase. Possible values are:
- laminar
- independent
window — sets the processing window size. Possible values:
- standard
- short
- long
smoothing — sets smoothing. Possible values:
- off
- on
formant — turns formant preservation on when you're shift-pitching 💅. Possible values are:
- shifted
- preserved
pitchq — sets pitch quality. Possible values:
- quality
- speed
- consistency
channels — sets channels. Possible values:
- apart
- together

🎛️ Commands

This filter supports the following commands:

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 filter is literally a normal compressor but with main-character energy: it can compress the detected signal using a second input signal. It needs two input streams and hands back one output stream. The first input stream gets processed depending on the second stream's signal, and then that filtered signal can go hang out with other filters in later stages of processing. Peep pan and amerge if you wanna see the family.

The filter accepts the following options:

level_in — sets the 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. Like, 1:2 means that if the level raised 4dB above the threshold, it'll only be 2dB above after the reduction. Default is 2. Range is between 1 and 20.
attack — the 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 — the 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 you enter 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 the 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 all the above options as commands. no cap 💅

Examples

Full ffmpeg example taking 2 audio inputs, 1st input to be compressed depending on the signal of 2nd input, and later the compressed signal to be merged with the 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's got 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 the 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 the 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 the 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 — the 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 — the 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 the amount of amplification of the signal after processing. Default is 1. Allowed range is from 1 to 64.
knee — curves the sharp knee around the threshold so you enter 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 the sidechain gain. Default is 1. Range is from 0.015625 to 64.

Commands

This filter supports all the above options as commands. yup, all of them 🫶

silencedetect 🤫

Detects silence in an audio stream. Lowkey iconic for finding dead air.

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 the 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 the 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 — process each channel separately, instead of combined. By default it's 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. It's giving cleanup era 🧹✨

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 noises 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 the 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 the mode of detection of silence end at the 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 it trigger the end of silence trimming at the start of the audio stream — limited usage, ngl.
stop_periods — sets the count for trimming silence from the end of the 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.

okay so continuing the silenceremove options—these are the ones that control how the tail end and silence detection behave, and honestly they’re kind of the main characters for trimming, fr fr. 💅

stop_threshold — this is literally the same as start_threshold but for trimming silence from the end of the audio. You can give it in dB (if "dB" is appended to the value) or as an amplitude ratio. Default value is 0. 🎀
stop_silence — sets the max duration of silence at the end that will be kept after trimming. Default is 0, which is equal to trimming all samples detected as silence. no cap, that’s the full chop. ✨
stop_mode — sets the mode for detecting the start of silence 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 will trigger start of silence trimming after start of audio stream—limited usage, bestie. With all, only if every sample from every channel is detected as silence will trigger start of silence trimming after start of audio stream. It’s giving strict versus chaotic. 😭
detection — sets how silence is detected. The options are:
- avg — mean of absolute values of samples in moving window.
- rms — root squared mean of absolute values of samples in moving window.
- peak — maximum of absolute values of samples in moving window.
- median — median of absolute values of samples in moving window.
- ptp — absolute of max peak to min peak difference of samples in moving window.
- dev — standard deviation of values of samples in moving window.

Default value is "rms". 🦋

window — sets duration in number of seconds used to calculate size of window in number of samples for detecting silence. Using 0 will effectively disable any windowing and use only single sample per channel for silence detection. In that case it may be needed to also set start_silence and/or stop_silence to nonzero values with also start_duration and/or stop_duration to nonzero values. Default value is 0.02. Allowed range is from 0 to 10. Lowkey crucial, highkey underrated. 🌸
timestamp — sets processing mode of every audio frame output timestamp. The modes are:
- write — full timestamps rewrite, keep only the start time for the first output frame.
- copy — non-dropped frames are left with same timestamp as input audio frame.

Default value is "write". 💖

🌸 Examples 🌸

The following example shows how this filter can be used to start a recording that does not contain the delay at the start which usually occurs between pressing the record button and the start of the performance:

silenceremove=start_periods=1:start_duration=5:start_threshold=0.02

Trim all silence encountered from beginning to end where there is more than 1 second of silence in 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 is more than 0 samples of digital silence in audio and digital silence is detected in all channels at same positions in 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 is more than 1 second of silence per silence period in audio:

silenceremove=stop_periods=-2:stop_duration=1:stop_threshold=-90dB

Similar as above, but keep 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

Similar as above, but keep maximum of 1.5 seconds of silence from start of 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 above options as commands. yes bestie, runtime tweaking is on the menu. 🫶

🎧 sofalizer 🎧

SOFAlizer uses head-related transfer functions (HRTFs) to create virtual loudspeakers around the user for binaural listening via headphones (audio formats up to 9 channels supported). The HRTFs are stored in SOFA files (see http://www.sofacoustics.org/ for a database). SOFAlizer is developed at the Acoustics Research Institute (ARI) of the Austrian Academy of Sciences. iconic academic binaural behavior, honestly. 👑

To enable compilation of this filter you need to configure FFmpeg with --enable-libmysofa. slay, dependency check. ✨

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 loudspeakers and the listener with near-field HRTFs. Default is 1.
type — set processing type. Can be time or freq. time is processing audio in time domain which is slow. freq is processing audio in frequency domain which is fast. Default is freq. chose speed, periodt. 🔥
speakers — set custom positions of virtual loudspeakers. Syntax for this option is: <CH> <AZIM> <ELEV>[|<CH> <AZIM> <ELEV>|...]. Each virtual loudspeaker is described with short channel name following with azimuth and elevation in degrees. Each virtual loudspeaker description is separated by '|'. For example to override front left and front right channel positions use: 'speakers=FL 45 15|FR 345 15'. Descriptions with unrecognised channel names are ignored. iykyk, channel placement is a vibe. 🦋
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 SOFA file. By default is enabled.
interpolate — should nearest IRs be interpolated with neighbor IRs if exact position does not match. By default is disabled.
minphase — minphase all IRs upon loading of 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 ClubFritz6 sofa file:

sofalizer=sofa=/path/to/ClubFritz6.sofa:type=freq:radius=1

Using ClubFritz12 sofa file and bigger radius with small rotation:

sofalizer=sofa=/path/to/ClubFritz12.sofa:type=freq:radius=2:rotation=5

Similar as above but with custom speaker positions for front left, front right, back left and back right and also with 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. This filter expands or compresses each half-cycle of audio samples (local set of samples all above or all below zero and between two nearest zero crossings) depending on threshold value, so audio reaches target peak value under conditions controlled by below options. it’s giving audio glow-up, but make it math. 💜

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 maximum local half-cycle of samples expansion. The maximum expansion would be such that local peak value reaches target peak value but never to surpass it and that ratio between new and previous peak value does not surpass this option value. no overshoot, bestie. ✨
compression, c — set the maximum compression factor. Allowed range is from 1.0 to 50.0. Default value is 2.0. This option controls maximum local half-cycle of samples compression. This option is used only if threshold option is set to value greater than 0.0, then in such cases when local peak is lower or same as value set by threshold all samples belonging to that peak's half-cycle will be compressed by 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 will be compressed and which will be expanded. Any half-cycle samples with their local peak value below or same as this option value will be compressed by current compression factor, otherwise, if greater than threshold value they will be expanded with expansion factor so that it could reach peak target value but never surpass it. the sorting hat of half-cycles, lowkey. 🎀
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 expansion factor is raised per each new half-cycle until it reaches expansion value. Setting this options too high may lead to distortions. be careful, gorgeous. 😭
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 compression factor is raised per each new half-cycle until it reaches 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 interpretation of threshold option. When enabled any half-cycle of samples with their local peak value below or same as threshold option will be expanded otherwise it will be compressed.
link, l — link channels when calculating gain applied to each filtered channel sample, by default is disabled. When disabled each filtered channel 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 the above options as commands. runtime control, we love to see it. 🫶

🌈 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 one is literally a whole toolbox for stereo signals — it converts M/S stereo recordings to L/R while letting you control the parameters, and it can also spread the stereo image of a master track. It's giving "Swiss army knife but make it audio." 🎀

It accepts these options:

level_in — input level before filtering, for both channels. Default is 1. Allowed range is 0.015625 to 64.
level_out — output level after filtering, for both channels. Default is 1. Allowed range is 0.015625 to 64.
balance_in — input balance between both channels. Default is 0. Allowed range is -1 to 1.
balance_out — output balance between both channels. Default is 0. Allowed range is -1 to 1.
softclip — enable softclipping. Gives you analog distortion instead of that harsh digital 0dB clipping. Disabled by default.
mutel — mute the left channel. Disabled by default.
muter — mute the right channel. Disabled by default.
phasel — flip the phase of the left channel. Disabled by default.
phaser — flip the phase of the right channel. Disabled by default.
mode — set the stereo mode. Available values:
- lr>lr — Left/Right to Left/Right, this is the default.
- 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 — level of the side signal. Default is 1. Allowed range is 0.015625 to 64.
sbal — balance of the side signal. Default is 0. Allowed range is -1 to 1.
mlev — level of the middle signal. Default is 1. Allowed range is 0.015625 to 64.
mpan — middle signal pan. Default is 0. Allowed range is -1 to 1.
base — stereo base between mono and inversed channels. Default is 0. Allowed range is -1 to 1.
delay — delay in milliseconds, i.e. how much to delay left from right channel and vice versa. Default is 0. Allowed range is -20 to 20.
sclevel — S/C level. Default is 1. Allowed range is 1 to 100.
phase — the stereo phase in degrees. Default is 0. Allowed range is 0 to 360.
bmode_in, bmode_out — set the balance mode for the balance_in/balance_out options. Can be one of:
- balance — classic balance mode. Attenuates one channel at a time. Gain is raised up to 1.
- amplitude — similar to the classic mode above, but 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. Iconic. 👑

Examples

Apply a karaoke-like effect:

stereotools=mlev=0.015625

Convert an M/S signal to L/R:

"stereotools=mode=ms>lr"

stereowiden 🦋

This filter enhances the stereo effect by suppressing the signal common to both channels, and by delaying the left signal into the right and vice versa — thereby widening the stereo effect. It just makes everything feel bigger, no cap. ✨

It accepts these 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. Gives a delay effect of the left signal in the right output and vice versa, which is what gives you that widening effect. Default is 0.3.
crossfeed — cross feed of left into right with inverted phase. Helps suppress the mono. If the value is 1 it will cancel all the signal common to both channels. Default is 0.3.
drymix — level of input signal of the original channel. Default is 0.8.

Commands

This filter supports all the above options except delay as commands. (Yes, delay is the one that doesn't get to come to the party. 😭)

superequalizer 🎚️

Apply an 18 band equalizer. Eighteen. Bands. Obsessed.

The filter accepts these options:

1b — set 65Hz band gain.
2b — set 92Hz band gain.
3b — set 131Hz band gain.
4b — set 185Hz band gain.
5b — set 262Hz band gain.
6b — set 370Hz band gain.
7b — set 523Hz band gain.
8b — set 740Hz band gain.
9b — set 1047Hz band gain.
10b — set 1480Hz band gain.
11b — set 2093Hz band gain.
12b — set 2960Hz band gain.
13b — set 4186Hz band gain.
14b — set 5920Hz band gain.
15b — set 8372Hz band gain.
16b — set 11840Hz band gain.
17b — set 16744Hz band gain.
18b — set 20000Hz band gain.

surround 🌈

Apply the audio surround upmix filter. This one lets you produce multichannel output from an audio stream — glow-up for your two-channel audio, fr fr. 🔥

The filter accepts these options:

chl_out — set 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 — set 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 — set the input volume level. By default this is 1.
level_out — set 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 — set the LFE low cut off frequency. By default this is 128 Hz.
lfe_high — set the LFE high cut off frequency. By default this is 256 Hz.
lfe_mode — set 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 — set the temporal smoothness strength, used to gradually change factors when transforming stereo sound in time. Allowed range is 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 — set the angle of the stereo surround transform. Allowed range is 0 to 360. Default is 90.
focus — set the focus of the stereo surround transform. Allowed range is -1 to 1. Default is 0.
fc_in — set the front center input volume. By default this is 1.
fc_out — set the front center output volume. By default this is 1.
fl_in — set the front left input volume. By default this is 1.
fl_out — set the front left output volume. By default this is 1.
fr_in — set the front right input volume. By default this is 1.
fr_out — set the front right output volume. By default this is 1.
sl_in — set the side left input volume. By default this is 1.
sl_out — set the side left output volume. By default this is 1.
sr_in — set the side right input volume. By default this is 1.
sr_out — set the side right output volume. By default this is 1.
bl_in — set the back left input volume. By default this is 1.
bl_out — set the back left output volume. By default this is 1.
br_in — set the back right input volume. By default this is 1.
br_out — set the back right output volume. By default this is 1.
bc_in — set the back center input volume. By default this is 1.
bc_out — set the back center output volume. By default this is 1.
lfe_in — set the LFE input volume. By default this is 1.
lfe_out — set the LFE output volume. By default this is 1.
allx — set spread usage of the stereo image across the X axis, for all channels. Allowed range is -1 to 15. By default this value is negative -1, and thus unused.
ally — set spread usage of the stereo image across the Y axis, for all channels. Allowed range is -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 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 0.06 to 15. By default this value is 0.5.
win_size — set the window size. Allowed range is 1024 to 65536. Default size 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 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). It's giving "the tone knob on your dad's stereo, but make it precision." 💜

The filter accepts these 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.
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 channels are filtered.
normalize, n — normalize biquad coefficients; by default this 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 — how precise the filtering math gets to be 💅 pick your sample-format vibe:
- auto — let it pick the sample format automatically, based on the 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 you set this 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 just produces nasty artifacts. 😭

Heads up bestie: when this is set to a non-zero value, the filter delay will be exactly this many samples. No cap.

📜 Commands

This filter supports some options as commands. 🫶

🎚️ `treble`, `highshelf`

Boost or cut 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. It's also known as shelving equalisation (EQ). Iconic behavior honestly 💖

The filter accepts the following options:

gain, g — gives the gain at whichever is the lower of ~22 kHz and the Nyquist frequency. Its useful range is about -20 (big cut) to +20 (big boost). ⚠️ Beware of clipping when you use a positive gain, fr fr.
frequency, f — sets the filter's central frequency, so you can extend or reduce the frequency range being 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 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 — 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 makes nasty artifacts. 😭

Note that the filter delay will be exactly this many samples when it's set to a non-zero value.

📜 Commands

This filter supports the following commands:

frequency, f — change the treble frequency. Syntax for the command is: "frequency"
width_type, t — change the treble width_type. Syntax for the command is: "width_type"
width, w — change the treble width. Syntax for the command is: "width"
gain, g — change the treble gain. Syntax for the command is: "gain"
mix, m — change the treble mix. Syntax for the command is: "mix"

🌊 `tremolo`

Sinusoidal amplitude modulation. Wiggle wiggle 🦋

The filter accepts the following 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. Same energy, different flavor ✨

The filter accepts the following 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. Bass go brrr 💜

This filter accepts stereo input and produces stereo with LFE (2.1) channels as 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, just as they were in the stereo input — no notes taken, no notes added 🫶

The filter accepts the following 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. The main character of audio filters, honestly 👑

It accepts the following parameters:

volume — sets the audio volume expression.

Output values are clipped to the maximum value.

The output audio volume is given by the 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 the input sample format to U8, S16, and S32.
float — 32-bit floating-point; this limits the input sample format to FLT. (default)
double — 64-bit floating-point; this limits the input sample format to DBL.

replaygain — chooses the behaviour when it encounters ReplayGain side data in input frames:
- drop — remove ReplayGain side data, ignoring its contents (the default).
- ignore — ignore ReplayGain side data, but leave it in the frame.
- track — prefer the track gain, if present.
- album — prefer 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 the following values:

once — only evaluate the expression once during filter initialization, or when the volume command is sent
frame — evaluate the expression for each incoming frame

Default value is once.

The volume expression can contain the following 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

Okay important: when eval is set to once, only the sample_rate and tb variables are available — all the other variables will evaluate to NAN. 💀

📜 Commands

This filter supports the following 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. Detective mode activated 🕵️‍♀️

The filter has no parameters. It supports only 16-bit signed integer samples, so the input will be converted when needed. Statistics about the volume get printed in the log when the end of the input stream is reached.

In particular it will show 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. Maths, but make it audio 💅

🗣️ `whisper`

Runs automatic speech recognition using OpenAI's Whisper model. Speech-to-text but make it ffmpeg ✨

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.

The filter has the following options:

model — the file path of the downloaded whisper.cpp model (mandatory). ⚠️
language — the language to use for transcription ('auto' for auto-detect). Default value: "auto"
queue — the max size that gets queued into the filter before whisper actually processes the audio. Small value = the stream gets processed more often, but the transcription quality drops and the CPU bill goes up. Big value (like 10-20s) = more accurate results using less CPU (same as using the whisper-cli tool), but the transcription latency goes up, so it's not useful for real-time streams. Lowkey consider pairing this with the vad_model option and a large queue value. Default value: "3"
use_gpu — whether 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 gets sent to the specified file or URL (use one of the FFmpeg AVIO protocols); otherwise the output is just logged as info messages. The output also gets set in the "lavfi.whisper.text" frame metadata. Heads up: if the destination is a file and it already exists, it will be overwritten 💀 Default value: none.
format — the destination format string; it can be "text" (only the transcribed text goes 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 get split so they don't exceed this length. Super useful for generating subtitles with shorter lines. Default value: "0"
vad_model — path to the VAD model file. If set, the filter loads an additional voice activity detection module (https://github.com/snakers4/silero-vad) which is used to fragment the audio queue; use this option by setting a valid path obtained from the whisper.cpp repository (e.g. "../whisper.cpp/models/ggml-silero-v5.1.2.bin") and bump the queue parameter up 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 ✨

Bestie, run a transcription and generate an srt file 🎀

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 yeet 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 🎤 (this one's giving main character energy ngl)

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 so below is the rundown of all the audio sources that currently exist. Slay.

`abuffer`

Buffers audio frames and makes them available to the filter chain. 💅

This source is mainly intended for programmatic use, specifically through the interface defined in libavfilter/buffersrc.h.

It accepts these parameters:

time_base — the timebase that gets used for timestamps of submitted frames. It has to 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 (no fighting, besties 🫶).

Examples

abuffer=sample_rate=44100:sample_fmt=s16p:channel_layout=stereo

will instruct the source to accept planar 16bit signed stereo at 44100Hz. And since the sample format named "s16p" corresponds to the number 6 and the "stereo" channel layout corresponds to the value 0x3, this is literally equivalent to:

abuffer=sample_rate=44100:sample_fmt=6:channel_layout=0x3

`aevalsrc`

Generates an audio signal specified by an expression. ✨

This source accepts one or more expressions as input (one for each channel), which get evaluated and used to generate a corresponding audio signal.

This source accepts these 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 gets applied to the remaining output channels.
channel_layout, c — set the channel layout. The number of channels in the specified layout must equal 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, since the generated audio is always cut at the end of a complete frame.

If it's 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 is 1024.
sample_rate, s — specify the sample rate, default is 44100.

Each expression in exprs can contain these 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, and 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 (the vibe of a broken TV, iconic)

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`

Generates fractional delay FIR coefficients.

The resulting stream can be used with the afir filter to filter the audio signal.

The filter accepts these 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 the 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`

Generates FIR equalizer coefficients. 🎚️

The resulting stream can be used with the afir filter to filter the audio signal.

The filter accepts these options:

preset, p — set the 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 the custom equalizer gains get 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 represents a frequency in Hz. Default is "25 40 63 100 160 250 400 630 1000 1600 2500 4000 6300 10000 16000 24000".
taps, t — set the number of filter coefficients in the output audio stream. Default value is 4096.
sample_rate, r — set the sample rate of the output audio stream, default is 44100.
nb_samples, n — set the number of samples per each frame in the output audio stream. Default is 1024.
interp, i — set the interpolation method for the FIR equalizer coefficients. Can be "linear" or "cubic".
phase, h — set the phase type of the FIR filter. Can be "linear" or "min" (minimum-phase). Default is a minimum-phase filter.

`afirsrc`

Generates FIR coefficients using the frequency sampling method. 🦋

The resulting stream can be used with the afir filter to filter the audio signal.

The filter accepts these options:

taps, t — set the number of filter coefficients in the output audio stream. Default value is 1025.
frequency, f — set the frequency points from where magnitude and phase get set. This must be in non-decreasing order, and the first element must be 0 while the last element must be 1. Elements are separated by white spaces.
magnitude, m — set the magnitude value for every frequency point set by frequency. The number of values must be the same as the number of frequency points. Values are separated by white spaces.
phase, p — set the phase value for every frequency point set by frequency. The number of values must be the same as the number of frequency points. Values are separated by white spaces.
sample_rate, r — set the sample rate, default is 44100.
nb_samples, n — set the number of samples per each frame. Default is 1024.
win_func, w — set the window function. Default is blackman.

🎀 `anullsrc` 🎀

The null audio source — it literally just hands back unprocessed audio frames. Nothing extra, no seasoning. It's mostly useful as a template, and for analysis/debugging tools, or as the source for filters that ignore their input data anyway (like the sox synth filter, iykyk).

This source accepts the following options:

channel_layout, cl — sets the channel layout, and it can be either an integer or a string that represents a channel layout. Default for channel_layout is "stereo". The mapping between strings and channel layout values lives in the channel_layout_map definition in libavutil/channel_layout.c — go peek if you're curious 💖
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. The accepted syntax is in the Time duration section of the ffmpeg-utils(1) manual. If it's not specified, or if the expressed duration is negative, the audio is assumed to be generated forever. infinite vibe fr.

Examples

Set the sample rate to 48000 Hz and the channel layout to AV_CH_LAYOUT_MONO.

anullsrc=r=48000:cl=4

Same operation, but 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` 🗣️

Synthesizes a voice utterance using the libflite library. She speaks 💅

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, 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 used for the speech synthesis. Default value is "kal". Also see the list_voices option.

Examples

Read from the 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` 🌈

Generates a noise audio signal. Static queen 🎀

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. If you don't specify this option, you get noise with an infinite length. endless static arc 😭
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` 🦋

Generates 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 girlie hours ✨

The filter accepts the following options:

sample_rate, s — sets the sample rate, default is 44100.
taps, t — sets the length of the FIR filter, default is 22051.
nb_samples, n — sets the number of samples per each frame.
win_func, w — sets the window function to be used when generating FIR coefficients.

💜 `sinc` 💜

Generates sinc kaiser-windowed low-pass, high-pass, band-pass, or band-reject FIR coefficients.

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 the sample rate, default is 44100.
nb_samples, n — sets the number of samples per each frame. Default is 1024.
hp — sets the high-pass frequency. Default is 0.
lp — sets the low-pass frequency. Default is 0. If the high-pass frequency is lower than the low-pass frequency and the low-pass frequency is higher than 0, then the filter creates band-pass filter coefficients — otherwise it makes band-reject filter coefficients. choose your fighter 🫶
phase — sets the filter phase response. Default is 50. Allowed range is from 0 to 100.
beta — sets the Kaiser window beta.
att — sets the stop-band attenuation. Default is 120dB, allowed range is from 40 to 180 dB.
round — enables rounding, by default it's disabled.
hptaps — sets the number of taps for the high-pass filter.
lptaps — sets the number of taps for the low-pass filter.

🎵 `sine` 🎵

Generates an audio signal made of a sine wave with amplitude 1/8.

The audio signal is bit-exact. no cap, literally bit-exact 🔥

The filter accepts the following options:

frequency, f — sets the carrier frequency. Default is 440 Hz.
beep_factor, b — enables 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 an 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 the "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.

🫶 `abuffersink` 🫶

Buffers audio frames, and makes 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. genuinely nothing, it's giving "I'm just here" 😭 It's mainly useful as a template and for use in analysis / debugging tools.

🦋 👑 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.

Below is a description of the currently available video filters.

🎀 `addroi` 🎀

Marks 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. You can mark multiple regions by applying the filter multiple times. stack 'em 🫶

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.
clear — if set to true, removes 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` ✨

Extracts the alpha component from the input as a grayscale video. This is especially useful with the alphamerge filter. dynamic duo energy 💖

💅 `alphamerge` 💅

Adds or replaces the alpha component of the primary input with the grayscale value of a second input. This is intended for use with alphaextract, so you can transmit or store frame sequences that have alpha in a format that doesn't support an alpha channel. iconic workaround, honestly 👑

For example, to reconstruct full frames from a normal YUV-encoded video and a separate video created with alphaextract, you might use:

movie=in_alpha.mkv [alpha]; [in][alpha] alphamerge [out]

🌈 amplify

Amplify the differences between the current pixel and the pixels sitting in the same spot on neighbouring frames. Basically glow-up the delta, bestie ✨

This filter accepts the following options:

radius — sets the frame radius. Default is 2. Allowed range is from 1 to 63. So like, a radius of 3 tells the filter to average 7 frames. Math is math 💅
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 than 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 that correspond to the option of the same name:

factor
threshold
tolerance
low
high
planes

🎀 ass

Same as the subtitles filter, except it doesn't need libavcodec and libavformat to work. On the flip side (😭), it's limited to ASS (Advanced Substation Alpha) subtitle files. Trade-offs, bestie.

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. 👑
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. Say bye to the grain, not the vibes ✨

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.

So like, 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 personalities, both iconic 💅

s — sets the number of frames the filter will use for averaging. Default is 9. Must be an odd number in range [5, 129].
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 the frames if they are equal to or below the thresholds. Fr fr, parallel said "I'm done here" 💀

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. 👀

💬 Commands

This filter supports the same commands as the options, except the option "s". The command accepts the same syntax of the corresponding option.

🌸 avgblur

Applies an average blur filter. Soft-focus main character energy 🫶

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 the 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.

🧚 backgroundkey

Turns a static background into transparency. Green screen who? Never heard of her ✨

The filter accepts the following option:

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 of the above options as commands.

⭐ bbox

Computes the bounding box for the non-black pixels in the input frame luma plane.

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. Very "here are the receipts" of it all 📝

The filter accepts the following option:

min_val — sets the minimal luma value. Default is 16.

💬 Commands

This filter supports all of the above options as commands.

💜 bilateral

Applies a bilateral filter — spatial smoothing while preserving edges. Glow and structure, we love a multitasker 💅

The filter accepts the following options:

sigmaS — sets the sigma of the gaussian function used to calculate the spatial weight. Allowed range is 0 to 512. Default is 0.1.
sigmaR — sets the sigma of the gaussian function used to calculate the range weight. Allowed range is 0 to 1. Default is 0.1.
planes — sets the planes to filter. Default is the first only.

💬 Commands

This filter supports all of the above options as commands.

🍓 bitplanenoise

Shows and measures bit plane noise.

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. Lowkey useful for detecting chapter transitions, commercials, or straight-up invalid recordings 👀

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 the key "lavfi.black_start", and to the first frame after the black segment ends with the key "lavfi.black_end". The value is the frame's timestamp. This metadata is added regardless of the minimum duration specified. No cap.

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.

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. Can be useful to detect chapter transitions or commercials. 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. Don't skip this step or you'll be staring at nothing, bestie 😭

This filter exports frame metadata "lavfi.blackframe.pblack". The value represents the percentage of pixels in the picture that are below the threshold value.

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. Collab era ✨

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.

A description of the accepted options follows.

c0_mode
c1_mode
c2_mode
c3_mode
all_mode

Set 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:

okay so like, the blend modes?? there are SO many and they're all iconic in their own way 💅 here's the full roster, no crumbs left behind:

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

🌸 the opacity girlies

c0_opacity, c1_opacity, c2_opacity, c3_opacity, all_opacity — sets the blend opacity for one specific pixel component, or for all of them when you use all_opacity. heads up bestie: this only does anything when you're using pixel-component blend modes. iykyk

✨ the expression girlies (yes there are two flavors)

c0_expr, c1_expr, c2_expr, c3_expr, all_expr — sets the blend expression for a specific pixel component, or for all of them with all_expr. important: if you set these, the related mode options get ignored. like, completely. 🙈

So what can you actually put in these expressions? These variables, bestie:

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 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 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 — value of the pixel component at the current location for the first video frame (the top layer).
BOTTOM, B — 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, obviously 💁‍♀️

🎀 Examples (eat these up)

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 the uncover-left effect:

blend=all_expr='if(gte(N*SW+X,W),A,B)'

apply the uncover-down effect:

blend=all_expr='if(gte(Y-N*SH,0),A,B)'

apply the 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 layer on one side and the bottom layer on the other:

blend=all_expr='if(gt(X,Y*(W/H)),A,B)'

display the differences between the current frame and the previous one:

tblend=all_mode=grainextract

Commands

this filter supports the same commands as options 🫶

🌈 blockdetect

figures out how blocky your frames are — and like, without touching them at all. the input frames come out exactly as they went in, it's just reporting 👑

based on Remco Muijs and Ihor Kirenko: "A no-reference blocking artifact measure for adaptive video processing." 2005 13th European signal processing conference.

options, let's go:

period_min, period_max — set the minimum and maximum values for determining pixel grids (periods). the default values are [3,24].
planes — set which planes to filter. default is first only.

Examples

determine blockiness for the first plane and search for periods within [8,32]:

blockdetect=period_min=8:period_max=32:planes=1

🦋 blurdetect

same vibe, different metric: it determines how blurry your frames are without altering them. based on Marziliano, Pina, et al. "A no-reference perceptual blur metric." it also allows for a block-based abbreviation.

options time:

low, high — set the low and high threshold values used by the Canny thresholding algorithm.

the high threshold picks out the "strong" edge pixels, which then get connected through 8-connectivity with the "weak" edge pixels that the low threshold selected.

low and high have to be in the range [0,1], and low should be lesser or equal to high (don't be chaotic about it 🙈).

the default value for low is 20/255, and the default value for high is 50/255.

radius — defines 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 planes to filter. default is first only.

Examples

determine blur for 80% of the most significant 32x32 blocks:

blurdetect=block_width=32:block_height=32:block_pct=80

💜 bm3d

denoises frames using the Block-Matching 3D algorithm. yes it's as fancy as it sounds ✨

the options:

sigma — set the denoising strength. default value is 1. allowed range is from 0 to 999.9. ngl the denoising algorithm is very sensitive to sigma, so adjust it according to the source. fr fr, tune it
block — set the local patch size. this sets 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 is slower (tragic but real 😭).
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 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

applies a boxblur algorithm to the input video. soft focus moment 💖

it accepts these parameters:

luma_radius, lr
luma_power, lp
chroma_radius, cr
chroma_power, cp
alpha_radius, ar
alpha_power, ap

and here's the full description of the accepted options:

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. respect the limit bestie

the 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 these 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.

the 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. (zero effort, zero blur, we love a lazy queen 👑)

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

deinterlaces the input video. also, bwdif stands for "Bob Weaver Deinterlacing Filter" — yes, a whole acronym for a name, main character energy honestly.

it's motion adaptive deinterlacing based on yadif with the use of w3fdif and cubic interpolation algorithms. it accepts these 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. 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.

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". 💖

🎀 `ccrepack`

Okay so this one repacks CEA-708 closed captioning side data — it's basically the glow-up artist for your captions. ✨

It fixes a whole bunch of issues that people see 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).

No cap, that's it, that's the whole filter. 🫶

🎀 `cas`

Apply the Contrast Adaptive Sharpen filter to your video stream. Crispy edges, bestie. 🔥

The filter accepts the following options:

strength — set the sharpening strength. Default value is 0. (Yeah, zero means no sharpening, so like… turn it up, babe 💅)
planes — set which planes to filter. Default is to filter all planes except the alpha plane.

Commands

This filter supports the same commands as options. 🎮

🎀 `chromahold`

Removes all color information for all colors except for one certain one. It's giving "keep only the main character, everyone else go gray." 👑

The filter accepts the following options:

color — the color which will not be replaced with neutral chroma. The chosen one. ✨
similarity — similarity percentage with the above color. 0.01 matches only the exact key color, while 1.0 matches everything. Wild range, honestly.
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 yuv enabled anymore. This can be used to pass exact YUV values as hexadecimal numbers. 🧚

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.

🎀 `chromakey`

YUV colorspace color/chroma keying. This is THE greenscreen filter, iykyk. 💚

The filter accepts the following options:

color — the color which will be replaced with transparency. Bye bye 👋
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 pixel's color is to the key color. Soft edges, slay. 🦋
yuv — signals that the color passed is already in YUV instead of RGB.

Literal colors like "green" or "red" don't make sense with yuv enabled anymore. This can be used to pass exact YUV values as hexadecimal numbers. 🌸

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.

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. Like a little skincare routine but for your video's color channels. 🧖‍♀️

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 the same commands as options. The command accepts the same syntax of the corresponding option.

🎀 `chromashift`

Shift chroma pixels horizontally and/or vertically. Little scoot for the color channels. 💫

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 a CIE color diagram with pixels overlaid onto it. Science girlie aesthetic, honestly iconic. 🔬🌈

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 the "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. It's the "show me what the codec is actually thinking" filter. 🧠✨

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 the 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. Color grading era, we love to see it. 🎨💅

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. Positive = this way, negative = the opposite way, got it? 🫶

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. Fancy. ✨

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. ⚠️ (so like, weights are literally the on switch, don't forget them)
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. Selective queen. 👑

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`

Bestie, this one remixes your color channels like it's making a playlist 💿 — it adjusts video input frames by re-mixing the color channels.

Here's the vibe: it modifies one color channel by adding up the values of the other channels of the same pixels. So if the value you're tweaking is red, your output value comes out as:

<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 — adjust how much the input red, green, blue and alpha channels contribute to the output green channel. Default is 1 for gg, and 0 for gr, gb and ga.
br, bg, bb, ba — adjust how much the input red, green, blue and alpha channels contribute to the output blue channel. Default is 1 for bb, and 0 for br, bg and ba.
ar, ag, ab, aa — adjust how much the input red, green, blue and alpha channels contribute to the output alpha channel. Default is 1 for aa, and 0 for ar, ag and ab.

Allowed ranges for all of these options are [-2.0, 2.0] 🫶

pc — set the preserve color mode. Accepted values:
- 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 like, disabled by default, no cap.

✨ Examples

Convert source to grayscale:

colorchannelmixer=.3:.4:.3:0:.3:.4:.3:0:.3:.4:.3

Simulate sepia tones (so aesthetic 🎞️):

colorchannelmixer=.393:.769:.189:0:.349:.686:.168:0:.272:.534:.131

🕹️ Commands

This filter supports the all above options as commands.

🔍 `colordetect`

Okay so this one literally analyzes the video frames to figure out the effective value range and alpha mode. It's giving detective energy 🕵️‍♀️

The filter accepts the following options:

mode — the set of properties to detect. Unavailable properties (like alpha mode for an input image with no alpha channel) get ignored automatically, no drama.

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 that 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. It's like putting a gel over the lens, lowkey iconic.

The filter accepts the following options:

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 is set to 1.0. Allowed range is from 0.0 to 1.0.

🕹️ Commands

This filter supports the all above options as commands.

🟩 `colorkey`

RGB colorspace color keying, aka the greenscreen filter of your dreams 💚 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.

The filter accepts the following options:

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.

🎯 `colorhold`

Remove all color information for all RGB colors except for a certain one. It's the "everyone wears gray except the main character" filter, fr fr 👑

The filter accepts the following options:

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. Think of it as the sliders panel in your editing app 🎚️

The filter accepts the following options:

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), and change the balance of bright and dark tones.

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 allow manual selection of a constrained output level range.

preserve — set preserve color mode. Accepted values:
- 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 the all above options as commands.

🗺️ `colormap`

Apply custom color maps to video stream. This one's a whole group project — it needs three input video streams, okay?

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.

The filter accepts the following options:

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 the 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. Simple as that, bestie.

The filter accepts the following options:

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 though — input video needs to have an even size or this whole thing falls apart 😭

The filter accepts the following options:

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

…and here's the rest of the trc squad, because she has range 💅:

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 — okay so this one sets your output color primaries, bestie. Accepted values, no cap:

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 your output color range ✨ accepted values:

tv — TV (restricted) range
mpeg — MPEG (restricted) range
pc — PC (full) range
jpeg — JPEG (full) range

format — sets your output color format 🎀 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 that just… skips gamma/primary correction entirely 🙈 it uses significantly less CPU, but it's mathematically incorrect, so like, know what you're signing up for. If you want output that matches what the colormatrix filter produces, go fast=1.

dither — picks your dithering mode:

none — no dithering at all
fsb — Floyd-Steinberg dithering

wpadapt — whitepoint adaptation mode:

bradford — Bradford whitepoint adaptation
vonkries — von Kries whitepoint adaptation
identity — identity whitepoint adaptation (i.e. no whitepoint adaptation)

clipgamut — controls how out-of-gamut colors get clipped when they show up as a result of 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 the input properties at once 🫶 same accepted values as all.

ispace — overrides input colorspace. Same accepted values as space.

iprimaries — overrides input color primaries. Same accepted values as primaries.

itrc — overrides input transfer characteristics. Same accepted values as trc.

irange — overrides input color range. Same accepted values as range.

So the way this filter works, bestie: it converts the transfer characteristics, color space and color primaries over to whatever values you specified. The output value — if you didn't specify it — gets set to a default based on the all property. And if that property isn't specified either? The filter logs an error, periodt. The output color range and format default to the same value as the input's 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 those are missing, the filter logs an error and no conversion takes 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, so you can simulate variations in ambient color temperature. Iconic little mood-lighting filter honestly.

Accepts these 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 same commands as options. Slay.

🌀 convolution

Applies a convolution of 3x3, 5x5, 7x7, or horizontal/vertical up to 49 elements. It's giving… math but make it fashion.

Accepts these options:

0m, 1m, 2m, 3m — set 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 — set multiplier for calculated value for each plane. If unset or 0, it will be 1/sum of all matrix elements.
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. Default is 0.0.
0mode, 1mode, 2mode, 3mode — set matrix mode for each plane. Can be square, row or column. Default is square.

Commands

This filter supports the all 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. Academic. Obsessed.

Accepts these options:

planes — set which planes to process.
impulse — set which impulse video frames will be processed; can be first or all. Default is all.

Heads up bestie: the convolve filter also supports the framesync options.

🎀 copy

Copies the input video source unchanged to the output. It's mainly useful for testing purposes — like a little "are you even running" vibe check ✨

🍓 coreimage

Video filtering on GPU using Apple's CoreImage API on OSX. Very main character energy.

Hardware acceleration is based on an OpenGL context. Usually, this means it's processed by video hardware. However — software-based OpenGL implementations exist, which means there's genuinely no guarantee of hardware processing. It depends on the respective OSX, ngl.

Apple provides many filters and image generators that come with a large variety of options. The filter has to be referenced by its name along with its options.

The coreimage filter accepts the following options:

list_filters — list all available filters and generators along with all their respective options, plus the possible minimum and maximum values and the default values. Like this:

list_filters=true

filter — specify all filters by their respective name and options. Use list_filters to figure out all valid filter names and options. Numerical options are given as a float value and get 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. And there's a special option name default available to use default options for a filter.

It's required to specify either default or at least one of the filter options. All omitted options get 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 — specify a rectangle where the output of the filter chain gets copied into the input image. It's given by a list of space separated float values:

output_rect=x\ y\ width\ height

If you don't give it, 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, iykyk.

output_rect=25\ 25\ 100\ 100

Several filters can be chained for successive processing without GPU-HOST transfers, which allows for 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 — like they exist but they're not serving the fantasy yet 🙈

Some filters generate output images with additional padding, depending on the respective filter kernel. The padding gets automatically removed so 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, which results in partial or complete coverage of the output image. Blend, slay, repeat 💅

You can use the coreimagesrc video source to generate input images that get fed directly into the filter chain. By using it, providing input images by another video source or an input video isn't required at all.

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 a 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. This filter takes two input videos — yes, two, it's a whole duet situation 🫶

Both input videos must have the same resolution and pixel format for this filter to work correctly — no cap, it just won't vibe otherwise 🙈. It also assumes both inputs have the same number of frames, which get compared one by one.

The obtained per component, average, min and max correlation is printed through the logging system (so keep an eye on your logs, bestie 👀).

The filter stores the calculated correlation of each frame in frame metadata ✨.

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 corr -f null -

🫶 cover_rect

Cover a rectangular object. Iconic behaviour, honestly 💅.

It accepts the following options:

cover — Filepath of the optional cover image, needs to be in yuv420. Yes, yuv420, it's picky like that.
mode — Set covering mode. It accepts the following values:
- cover — cover it by the supplied image 🖼️
- blur — cover it by interpolating the surrounding pixels 🌫️

Default value is blur.

🌸 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

✂️ crop

Crop the input video to given dimensions. Literally the makeover filter 🎀.

It accepts the following parameters:

w, out_w — The width of the output video. It defaults to iw. This expression is 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. It defaults to ih. This expression is evaluated only once during the filter configuration, 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. It defaults to (in_w-out_w)/2. This expression is evaluated per-frame.
y — The vertical position, in the input video, of the top edge of the output video. It defaults to (in_h-out_h)/2. This expression is evaluated per-frame.
keep_aspect — If set to 1 will force the output display aspect ratio to be the same of the input, by changing the output sample aspect ratio. It defaults to 0.
exact — Enable exact cropping. If enabled, subsampled videos will be cropped at exact width/height/x/y as specified and will not be rounded to nearest smaller value. It defaults to 0.

The out_w, out_h, x, y parameters are expressions containing the following constants:

x, y — The computed values for x and y. They are 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 is 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, as x and y are 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 are evaluated for each frame. If the evaluated value is not valid, it is 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, we love to see it 💞.

🌈 Examples

Crop area with 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 from the left and right borders, and 20 pixels from 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 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 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 width/height of the output video and the horizontal/vertical position in 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.

🔍 cropdetect

Auto-detect the crop size. It literally does your homework for you 😭✨.

It calculates the necessary cropping parameters and prints the recommended parameters via the logging system. The detected dimensions correspond to the non-black or video area of the input video according to mode.

It accepts the following parameters:

mode — Depending on 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 option limit.
- mvedges — Detect the playing video by the motion vectors inside the video and scanning for edge pixels typically forming the border of a playing video.
limit — Set higher black value threshold, which can be optionally specified from nothing (0) to everything (255 for 8-bit based formats). An intensity value greater to the set value is considered non-black. It defaults to 24. You can also specify a value between 0.0 and 1.0 which will be scaled depending on the bitdepth of the pixel format.
round — The value which the width/height should be divisible by. It 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 — Set the number of initial frames for which evaluation is skipped. Default is 2. Range is 0 to INT_MAX.
reset_count, reset — Set 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 useful when channel logos distort the video area. 0 indicates 'never reset', and returns the largest area encountered during playback.

mv_threshold — Set motion in pixel units as threshold for motion detection. It defaults to 8.
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 5/255, and default value for high is 15/255.

🦋 Examples

Find video area surrounded by black borders:

ffmpeg -i file.mp4 -vf cropdetect,metadata=mode=print -f null -

Find an embedded video area, generate motion vectors beforehand:

ffmpeg -i file.mp4 -vf mestimate,cropdetect=mode=mvedges,metadata=mode=print -f null -

Find an embedded video area, use motion vectors from 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 of the corresponding option. If the specified expression is not valid, it is kept at its current value.

⏰ cue

Delay video filtering until a given wallclock timestamp ⏳. 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 its input.

The filter can be used 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 on data to output almost immediately after the target wallclock timestamp is reached.

Perfect frame accuracy cannot be guaranteed, but the result is good enough for some use cases. Lowkey honest of it 🫶.

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

Apply color adjustments using curves. It's giving Adobe Photoshop meets GIMP, and honestly? Ate 👑.

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 from each other using a smooth curve. The x-axis represents the pixel values from the input frame, and the y-axis 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). This creates a straight line where each original pixel value is "adjusted" to its own value, which means no change to the image. Iconic minimalism, honestly ✨.

So like, this filter lets you redefine those two points and add even more — a brand new curve gets drawn so it passes smoothly through all your new coordinates. The new points have to be strictly increasing along the x-axis, and both their x and y values gotta live in the [0;1] interval. The curve itself is made with either a natural or monotonic cubic spline interpolation, depending on the interp option (default: "natural"). The "natural" spline gives you a smoother curve overall, while the monotonic ("pchip") spline guarantees the transitions between your points stay monotonic. And if the computed curves wander outside the vector spaces? Values get clipped accordingly. 💅

The filter accepts the following options:

preset — pick one of the available color presets. You can totally use this in addition to the r, g, b parameters — and in that case, the later options win over the preset values. The 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 define a second pass mapping, sometimes called a "luminance" or "value" mapping. Works 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 (not including master). Can be used alongside the other per-component key point options — and in that case, any component you didn't set falls back on this all setting.
psfile — specify a Photoshop curves file (".acv") to import the settings from.
plot — save a Gnuplot script of the curves into the file you specify.
interp — specify the kind of interpolation. The available algorithms are:
- natural — natural cubic spline using a piece-wise cubic polynomial that's 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 has to be defined using this syntax: "x0/y0 x1/y1 x2/y2 ...".

Commands

This filter supports the same commands as options. ✨

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 components:

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:

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. 🕵️‍♀️

This filter shows hexadecimal pixel values of part of video — literally the pixels but in numbers, iykyk.

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 a color background picked from the input video, and the text color is picked in such a way so it's always visible.
axis — draw rows and columns numbers on the left and top of the 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 the 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 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. 😌

dctdnoiz

Denoise frames using 2D DCT (frequency domain filtering).

Heads up bestie — this filter is not designed for real time. 🫠

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 processing speed. Also, a larger block size doesn't necessarily mean 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 the 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 the current pixel and the reference pixel is less than the threshold, it'll 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 a square of four pixels around the current pixel.
direction, d — set direction in radians from which four pixels will be compared. If positive, a random direction from 0 to the set direction will be picked. If negative, the exact absolute value will be picked. For example, direction 0, -PI or -2*PI radians will pick only pixels on the same row, and -PI/2 will pick only pixels on the same column.
blur, b — if enabled, the current pixel is compared with the average value of all four surrounding pixels. The default is enabled. If disabled, the 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, the current pixel is changed if and only if all pixel components are banded — e.g. the 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 the exact edge of the 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 threshold for duplicate detection. If a frame's difference metric is less than or equal to this value, it gets declared a duplicate. Default is 1.1. So basically the "are these two frames the same pic, bestie?" line ✨
scthresh — sets the scene change threshold. Default is 15. Lowkey the "is this a new scene or nah" detector 🎬
blockx, blocky — set the size of the x and y-axis blocks used during metric calculations. Bigger blocks = better noise suppression, but also = worse detection of small movements. Must be a power of two. Default is 32. Not optional, she's doing math 💅
ppsrc — marks the main input as a pre-processed input and activates the clean source input stream. This lets the input be pre-processed with various filters to help the metrics calculation while keeping the frame selection lossless. When set to 1, the first stream is the pre-processed input, and the second stream is the clean source from where the kept frames are chosen. Default is 0. Two streams, one glow-up, zero quality loss 🫶
chroma — sets whether or not chroma is considered in the metric calculations. Default is 1. Color girlies stay winning 🌈
mixed — sets whether or not the input only partially contains content to be decimated. Default is "false". If enabled, the video output stream will be in variable frame rate. Heads up, that VFR is real 📉

✨ deconvolve ✨

Applies 2D deconvolution of the video stream in the frequency domain using the second stream as impulse. It's giving "un-blur me with science" 🔬💜

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 noise when doing divisions. Default is 0.0000001. Super useful when width and height are not the same and not a power of 2, or if the stream prior to convolving had noise. Tiny number, huge job 💅

The deconvolve filter also supports the framesync options.

✨ dedot ✨

Reduces cross-luminance (aka dot-crawl) and cross-color (aka rainbows) from video. The most cursed little visual artifacts, gone ✨🌈

It accepts the following options:

m — set mode of operation. Can be a combination of dotcrawl for cross-luminance reduction and/or rainbows for cross-color reduction.
lt — set spatial luma threshold. Lower values increase reduction of cross-luminance.
tl — set tolerance for temporal luma. Higher values increase reduction of cross-luminance.
tc — set tolerance for chroma temporal variation. Higher values increase reduction of cross-color.
ct — set temporal chroma threshold. Lower values increase reduction of cross-color.

✨ deflate ✨

Applies a deflate effect to the video. This filter replaces the pixel by the local (3x3) average, but only takes into account values lower than the pixel. Squish, but make it respectful 🎈

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. Like a "do not touch" sign for that channel 🚫

Commands — this filter supports all the above options as commands. Yes bestie, you can change them at runtime 👑

✨ deflicker ✨

Removes temporal frame luminance variations. The flickering was giving migraine energy, so bye 💀

It accepts the following options:

size, s — set moving-average filter size in frames. Default is 5. Allowed range is 2 - 129.
mode, m — set averaging mode to smooth 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 one only wants metadata. All vibes, no pixels 🙈

✨ deinterlace_d3d12 ✨

Deinterlaces the input video using D3D12 hardware-accelerated video processing. GPU said "I got this" 👑🔥

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 party — that's a whole requirement, not a suggestion.

The filter automatically queries the hardware for the required number of temporal reference frames (if needed) and manages a frame queue to supply them. Self-sufficient icon 💅

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 falls back to bob. This is the default.
- bob — bob deinterlacing. Each field is independently scaled to full frame height. Simple and fast, but 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 drama 🔥

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 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 stutter was NOT it 😭

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. May change the recorded frame rate of the container. Aside from that change, this filter will not affect constant frame rate video. So yeah — VFR warning, bestie, keep that in mind 🫶

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 a mixture of the two.

The default is 4.

✨ delogo ✨

Suppresses a TV station logo by a simple interpolation of the surrounding pixels. Just set a rectangle covering the logo and watch it disappear (and sometimes something even uglier appear — your mileage may vary). Ngl this filter is a little chaotic, but iconic 💀🎀

It accepts the following parameters:

x, y — specify the top left corner coordinates of the logo. They must be specified. Non-negotiable, she needs the corner 📍
w, h — specify the width and height of the logo to clear. They must be specified. Also non-negotiable.
show — when set to 1, a green rectangle is drawn on the screen to simplify finding the right x, y, w, and h parameters. 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. So the green box is basically a little aura showing you where the magic happens 💚

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. Yes, she's an AI filter, and yes, she ate 🌧️🔥

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>. Go be a scientist, bestie 🧚

The filter accepts the following options:

filter_type — specify which filter to use. This option accepts the following values:
- derain — Derain filter. To conduct derain filter, you need to use a derain model.
- dehaze — Dehaze filter. To conduct 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 path to the model file specifying network architecture and its parameters. Note that different backends use different file formats. TensorFlow can load files for only its format. Kind of a "read the room" situation for file formats 🎀

To get full functionality (such as async execution), please use the dnn_processing filter.

✨ 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. Your shaky walking-vlog era is over, we're stabilizing her 🦋

The filter accepts the following options:

x, y, w, h — specify a rectangular area where to limit the search for motion vectors. If desired, 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 be confused for camera motion by the motion vector search. Fr, if your bestie is dancing in frame, she's not camera shake — don't let the vectors get confused 💅

If any or all of x, y, w and h are set to -1, then the full frame is used. This lets later options be set without specifying the bounding box for the motion vector search.

Default: search the whole frame.

rx, ry — Specify the maximum extent of movement in x and y directions in the range 0-64 pixels. Default 16.
edge — Specify how to generate pixels to fill blanks at the edge of the frame. 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 — Specify the blocksize to use for motion search. Range 4-128 pixels, default 8.
contrast — Specify the contrast threshold for blocks. Only blocks with more than the specified contrast (difference between darkest and lightest pixels) will be considered. Range 1-255, default 125.
search — Specify the search strategy. Available values are:
- exhaustive, 0 — Set exhaustive search.
- less, 1 — Set less exhaustive search. Default value is exhaustive.
filename — If set, a detailed log of the motion search is written to the specified file.

💅 despill 💅

Remove unwanted contamination of foreground colors, caused by reflected color of greenscreen or bluescreen.

This filter accepts the following options, bestie:

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 all the above options as commands.

🎀 detelecine 🎀

Apply an exact inverse of the telecine operation. It requires a predefined pattern specified using the pattern option, which must be the same as that passed to the telecine filter.

This filter accepts the following options:

first_field — Available values are:
- 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. This is to be used if the stream is cut. The default value is 0.

🌸 dilation 🌸

Apply dilation effect to the video.

This filter replaces the pixel by the local(3x3) maximum.

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.
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 all the above options as commands.

🦋 displace 🦋

Displace pixels as indicated by second and third input stream.

It takes three input streams and outputs one stream, the first input is the source, and second and third input 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 displacement map streams terminates, last frame from that displacement map will be used.

Note that once generated, displacements maps can be reused over and over again.

A description of the accepted options follows.

edge — Set 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 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 🧚

Do classification with deep neural networks based on bounding boxes.

The filter accepts the following options:

dnn_backend — Specify which DNN backend to use for model loading and execution. This option accepts only openvino now, tensorflow backends will be added.
model — Set path to 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 label file specifying the mapping between label id and name. Each label name is written in one line, trailing 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 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 🍓

Do object detection with deep neural networks.

The filter accepts the following options:

dnn_backend — Specify which DNN backend to use for model loading and execution. This option accepts only openvino now, tensorflow backends will be added.
model — Set path to 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 label file specifying the mapping between label id and name. Each label name is written in one line, trailing spaces and empty lines are skipped. The first line is the name of label id 0 (usually it is 'background'), and the second line is the name of label id 1, etc. The label id is considered as 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 💜

Do 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.

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 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 bestie, this one draws a colored box right on top of the input image. It's giving literal highlighter energy. ✨

It accepts the following parameters:

x, y — the expressions that specify the top left corner coordinates of the box. Defaults to 0.
width, w and height, h — the expressions specifying 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 — specify the color of the box to write. For the general syntax of this option, check the "Color" section in the ffmpeg-utils manual. If the special value invert is used, the box edge color is the same as the video with inverted luma. 👀
thickness, t — the expression that sets the thickness of the box edge. A value of fill will create a filled box (obsessed actually). Default value is 3.

See below for the list of accepted constants.

replace — applicable if the 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, leaving 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, 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.
in_h, ih and 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 still use box data in detection bboxes of side data. So 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. Mutual reference slay 💅

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 crash, no drama, it just stays exactly where it is. 🫶

📈 `drawgraph`

Draw a graph using input video metadata. Literally turning your metadata into a whole little chart, main character energy fr.

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`

Draw a grid on the input image. Aesthetic grid moment, we love to see it. 🦋

It accepts the following parameters:

x, y — the expressions which specify the coordinates of some point of grid intersection (meant to configure offset). Both default to 0.
width, w and height, h — the expressions which specify the width and height of the grid cell; if 0 they are interpreted as the input width and height, respectively, minus thickness, so 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 which sets the thickness of the grid line. Default value is 1.

See below for the list of accepted constants.

replace — applicable 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, leaving 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, 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.
in_h, ih and 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. Same deal as before, it just vibes where it is. ✨

✍️ `drawtext`

Draw a text string or text from a specified file on top of a video, using the libfreetype library. This is THE text-on-video filter, the icon, the legend. 👑

Heads up, build-flag bestie: 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. Miss these and the filter simply won't be there, no cap. 😭

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 — set the width of all the borders to 10
- boxborderw=10|20 — set 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 — set 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 — set 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.

Fr fr, strftime is deprecated — if you're building something new, clock that and reach for normal expansion instead. 💜

fix_bounds — if true, it checks your text coords and fixes them so nothing gets clipped. no more text hiding behind the edge of the frame 💅
fontcolor — the color used for drawing the fonts. 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. highkey the glow-up ✨
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. like, no fontconfig? then this is your only way in, bestie 🫶
alpha — draws 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. transparency slay 🎀
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). respectful of every script fr fr 🌈
ft_load_flags — the flags used for loading the fonts.

the flags map the corresponding flags supported by libfreetype, and are a combination of these 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

the 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 — sets the width of the box drawn around the text. the default value of boxw is computed automatically to match the text width.
boxh — sets the height of the box drawn around the 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 — sets the initial timecode representation in "hh:mm:ss[:;.]ff" format. it can be used with or without the text parameter. the timecode_rate option must be specified. ⏰
timecode_rate, rate, r — sets the timecode frame rate (timecode only). the value will be rounded to the 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). midnight reset energy 🌙
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 a lane, bestie 😭

text_source — the 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 will 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. like, seriously, only if you know 👑

reload — the textfile will be reloaded at the 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.

for real, atomic updates only — half-written files are giving chaos 💀

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? iconic 💫

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. like, it still works but it's giving legacy 🙈

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 layers of escaping is genuinely not the vibe 😭 use the file, bestie.

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.
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.

Okay so, the first argument? That's the format of the timestamp, and if you don't say anything it just defaults to "flt" — which is seconds as a decimal number with microsecond accuracy. Then "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, and "localtime" stands for the timestamp of the frame formatted as local time zone time. 🌍

The second argument is an offset that gets added to the timestamp. Little nudge, y'know? 💫

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). Bestie, 24h format is the superior format, I don't make the rules. 🕐

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 ✨

This filter 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'

Heads up bestie: if the entire argument can't be parsed or applied as valid values, then the filter will continue with its existing parameters. It does not freak out. It just vibes with what it already had. 🫶

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 🎀

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. 💅

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 behaviour 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 the https://ayosec.github.io/ffmpeg-drawvg/playground/docs/langref.html.

Graphics are rendered using the https://cairographics.org/.

To enable compilation of this filter, you need to configure FFmpeg with --enable-cairo. No cairo, no drawing, them's the rules. 🫶

Parameters

Either "script" or "file" must be set. Like, pick one bestie, you gotta give it something.

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. Literally don't skip this step, the vibes will be off. 😭

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. Serving outline realness. ✏️

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. So pretty. 🦋

📊 entropy 📊

Measure graylevel entropy in histogram of color channels of video frames.

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. Math girl summer fr fr. 🧮

🔍 epx 🔍

Apply the EPX magnification filter which is designed for pixel art. Retro gamers rise up. 👾

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 — sets the saturation expression 💅 must be a float in the range 0.0 to 3.0, and the default is "1".
gamma — sets the gamma expression. float, range 0.1 to 10.0, default "1".
gamma_r — sets the gamma expression for red. float, range 0.1 to 10.0, default "1".
gamma_g — sets the gamma expression for green. float, range 0.1 to 10.0, default "1".
gamma_b — sets the gamma expression for blue. float, range 0.1 to 10.0, default "1".
gamma_weight — sets the gamma weight expression ✨ this one's lowkey the glow-up saver: it lets you dial back how hard a high gamma value hits the bright parts of the image, so they don't get overamplified and just turn into plain white 😭 must be a float in range 0.0 to 1.0. 0.0 turns gamma correction all the way down, 1.0 leaves it at full strength. default is "1".
eval — sets when the expressions for brightness, contrast, saturation and gamma get evaluated. it takes these values:
- init — only evaluate expressions once, during filter initialization or when a command is processed
- frame — evaluate expressions for every 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, fr fr
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 🎀

bestie this filter takes commands too, here's the whole set:

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.

each command accepts the same syntax as its matching option 🫶

and if the expression you hand it isn't valid? it just keeps whatever value it's currently on. no drama, no crash 💀

erosion 🪨

applies an erosion effect to the video — literally replaces each pixel with the local (3x3) minimum. it accepts these options:

threshold0, threshold1, threshold2, threshold3 — limit the maximum change for each plane. default is 65535. if you set 0, that plane stays completely unchanged ✨
coordinates — a flag that says which pixels to refer to. default is 255, i.e. all eight pixels get used.

the flags map onto local 3x3 coordinates like this:

1 2 3
4   5
6 7 8

Commands 💫

this filter supports all of the above options as commands.

estdif 🦋

deinterlaces the input video (the name stands for "Edge Slope Tracing Deinterlacing Filter", iykyk). it's a spatial only filter that uses an edge slope tracing algorithm to interpolate the missing lines. it accepts these parameters:

mode — the interlacing mode to adopt, one of:
- frame — output one frame for each frame
- field — output one frame for each field

default value is "field".

parity — the picture field parity assumed for the input interlaced video, one of:
- tff — assume the top field is first
- bff — assume the bottom field is first
- auto — enable automatic detection of field parity

default value is "auto". if the interlacing is unknown or the decoder doesn't export this info, top field first gets assumed.

deint — which frames to deinterlace, one of:
- all — deinterlace all frames
- interlaced — only deinterlace frames marked as interlaced

default value is "all".

rslope — search radius for edge slope tracing. default 1, allowed range 1 to 15.
redge — search radius for best edge matching. default 2, allowed range 0 to 15.
ecost — edge cost for edge matching. default 2, allowed range 0 to 50.
mcost — middle cost for edge matching. default 1, allowed range 0 to 50.
dcost — distance cost for edge matching. default 1, allowed range 0 to 50.
interp — the interpolation used. default is 4-point interpolation. one of:
- 2p — two-point interpolation
- 4p — four-point interpolation
- 6p — six-point interpolation

Commands 🌸

this filter supports the same commands as options.

exposure ☀️

adjusts the exposure of the video stream. options:

exposure — set the exposure correction in EV. allowed range is -3.0 to 3.0 EV, default value is 0 EV.
black — set the black level correction. allowed range is -1.0 to 1.0, default value is 0.

Commands ✨

this filter supports the same commands as options.

extractplanes 🧚

pulls the color channel components out of the input video stream and into separate grayscale video streams. one option:

planes — set the plane(s) to extract. the available values are:
- y
- u
- v
- a
- r
- g
- b

🚨 picking planes that aren't in the input = error. and that means you literally cannot select "r", "g", "b" planes at the same time as "y", "u", "v" planes. no cap.

Examples 🎀

extract the luma, u and v color channel components from an 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 🌗

applies a fade-in/fade-out effect to the input video. parameters:

type, t — the effect type, either "in" for a fade-in or "out" for a fade-out. default is "in".
start_frame, s — the number of the frame where the fade effect starts. default is 0.
nb_frames, n — how many frames the fade effect lasts. at the end of the fade-in effect the output video has the same intensity as the input video; at the end of the fade-out transition the output video is filled with the selected color. default is 25.
alpha — if set to 1, fade only the alpha channel, if one exists on the input. default value is 0.
start_time, st — the timestamp (in seconds) of the frame where the fade effect starts. if both start_frame and start_time are specified, the fade starts at whichever comes last. default is 0.
duration, d — the number of seconds the fade effect has to last. at the end of the fade-in effect the output video has the same intensity as the input video; at the end of the fade-out transition the output video is 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 — the color of the fade. default is "black".

Examples 🌈

fade in the first 30 frames of video:

fade=in:0:30

that 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 the 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 🔁

applies the feedback video filter. this one's giving main character energy: it passes cropped input frames to the 2nd output, and from there you can filter them with other video filters. once the filter receives a frame from the 2nd input, that frame gets combined on top of the original frame from the 1st input and passed to the 1st output 🫶

the typical usage is filtering only part of a frame. options:

x, y — set the top left crop position.
w, h — set the crop size.

Examples ⭐

blur only the top left rectangular part of a video frame of size 100x100 with the gblur filter:

[in][blurin]feedback=x=0:y=0:w=100:h=100[out][blurout];[blurout]gblur=8[blurin]

draw a black box on the top left part of a video frame of size 100x100 with the 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 a rectangular part of a video frame of size 100x100 with the pixelize filter:

[in][blurin]feedback=x=320:y=240:w=100:h=100[out][blurout];[blurout]pixelize[blurin]

fftdnoiz 🌀

denoises frames using 3D FFT (that's frequency domain filtering, bestie — different realm entirely). options:

sigma — set the noise sigma constant. this sets the denoising strength. default value is 1, allowed range is 0 to 30.

⚠️ using very high sigma with low overlap may give you blocking artifacts. be so serious about that.

amount — set the amount of denoising. by default all detected noise gets reduced. default value is 1, allowed range is 0 to 1.
block — set the size of the block in pixels. default is 32, can be 8 to 256.
overlap — set the block overlap. default is 0.5, allowed range is 0.2 to 0.8.
method — set the denoising method. default is "wiener", can also be "hard".
prev — set the number of previous frames to use for denoising. by default it's set to 0.
next — set the number of next frames to use for denoising. by default it's set to 0.
planes — set which planes get filtered. by default all available planes are filtered except alpha.

fftfilt 🎛️

applies arbitrary expressions to samples in the frequency domain

dc_Y — this is the DC value (aka the gain, bestie, it's literally the gain 💅) for the luma plane of the image. Takes an integer in range 0 to 1000. Default? 0. slay.
dc_U — same energy, but for the 1st chroma plane. Integer in range 0 to 1000, and the default is 0 too.
dc_V — you guessed it 🎀: DC value (gain) of the 2nd chroma plane. Integer 0 to 1000, default 0.
weight_Y — sets the frequency domain weight expression for the luma plane. yes it's an expression, yes you can be fancy ✨
weight_U — sets the frequency domain weight expression for the 1st chroma plane.
weight_V — sets the frequency domain weight expression for the 2nd chroma plane.
eval — controls when those expressions get evaluated. lowkey important. it accepts these values:
- init — only evaluate expressions once, during filter initialization. one and done 🫶
- frame — evaluate expressions for each incoming frame. every single one, no skips.

Default value is init.

The filter accepts the following variables:

X, Y — the coordinates of the current sample 📍
W, H — the width and height of the image
N — the number of the input frame, starting from 0
WS, HS — the size of the FFT array for horizontal and vertical processing

Examples

High-pass 💀✨

fftfilt=dc_Y=128:weight_Y='squish(1-(Y+X)/100)'

Low-pass 🧚

fftfilt=dc_Y=0:weight_Y='squish((Y+X)/100-1)'

Sharpen 🔥

fftfilt=dc_Y=0:weight_Y='1+squish(1-(Y+X)/100)'

Blur 🌸

fftfilt=dc_Y=0:weight_Y='exp(-4 * ((Y+X)/(W+H)))'

field

Okay so this one extracts a single field from an interlaced image, and it does it using stride arithmetic — that's the whole trick, it avoids wasting CPU time. Iconic efficiency. The output frames are marked as non-interlaced.

The filter accepts the following options:

type — specify whether to extract the top field (if the value is 0 or "top") or the bottom field (if the value is 1 or "bottom").

fieldhint

This one creates new frames by copying the top and bottom fields from surrounding frames — and which frames? The numbers come from a hint file 📝

hint — set the file containing hints: absolute/relative frame numbers.

Here's the deal, bestie: there must be one line for each frame in a clip. Each line must contain two numbers separated by the comma, optionally followed by "-" or "+".

Numbers supplied 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. First number tells from which frame to pick up the top field, and the second number tells from which frame to pick up the bottom field.

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 will be marked the same as the input frame.

If optionally followed by "t" the output frame will use only the top field, or in case of "b" it will use only the bottom field.

If the 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 has, it seeks back to the start. loop it 🔁

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 — its whole personality is reconstructing progressive frames from a telecined stream 🎬

Important thing: 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. fr fr, it can't do it alone.

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 these remaining combed frames will be marked as interlaced, and thus can be de-interlaced by a later filter such as yadif before decimation. see? the split is a feature not a bug 💅

In addition to the various configuration options, "fieldmatch" can take an optional second stream, activated through the ppsrc option. If enabled, the frames reconstruction will be based on the fields and frames from this second stream. This allows the first input to be pre-processed in order to help the various algorithms of the filter, while keeping the output lossless (assuming the fields are matched properly). Typically, a field-aware denoiser, or brightness/contrast adjustments can help.

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 from which "fieldmatch" is based on. While the semantics and usage are very close, some behaviour and option names can differ. iykyk 🌈

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 the following options:

order — specify the assumed field order of the input stream. Available values are:
- auto — auto detect parity (use FFmpeg's internal parity value).
- bff — assume bottom field first.
- tff — assume top field first.

Note that it is sometimes recommended not to trust the parity announced by the stream 👀

Default value is auto.

mode — set 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 will 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 will almost always find a good frame if there is one. The other values are all 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 are available in the p/c/n/u/b meaning section.

Available values are:

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 indicate 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 — mark the main input stream as a pre-processed input, and enable the secondary input stream as the clean source to pick the fields from. See the filter introduction for more details. It is similar to the clip2 feature from VFM/TFM.

Default value is 0 (disabled).

field — set the field to match from. It is 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 is used to match from can have a large impact on matching performance. Available values are:
- auto — automatic (same value as order).
- bottom — match from the bottom field.
- top — match from the top field.

Default value is auto.

mchroma — set whether or not chroma is included during the match comparisons. In most cases it is 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 will disable the feature. y0 and y1 default to 0.
scthresh — set 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 what match to use as the final match. Available values are:
- 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 — force "fieldmatch" to calculate the combed metrics for certain matches and print them. This setting is known as micout in TFM/VFM vocabulary. Available values are:
- 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 cthresh is the area combing threshold for combed frame detection — basically it's the vibe-check for how strong or visible the combing has to be before the filter is like "yep, that's combed." 💅

Bigger values = combing has to be super obvious to get flagged. Smaller values = the filter is way more sensitive and will catch even faint combing.

Valid settings go from -1 (literally every pixel gets detected as combed, chaotic) all the way up to 255 (no pixel will ever be detected as combed, king of denial). It's basically a pixel difference value, and a good range is [8, 12].

Default is 9.

chroma

sets whether chroma counts toward the combed-frame decision. Bestie, only turn this off if your source has chroma problems (rainbowing, etc.) that are messing with the detection while chroma is enabled. Honestly chroma=0 is usually more reliable — except for the specific case where the source has chroma-only combing.

Default is 0.

blockx, blocky

these set the x-axis and y-axis size of the window used during combed frame detection. It's about the size of the area where combpel pixels need to be detected as combed before a frame gets declared combed — peek the combpel description for the details.

Possible values: any number that's a power of 2, starting at 4 and going up to 512.

Default is 16.

combpel

the number of combed pixels inside any of the blocky by blockx size blocks on the frame for the frame to count as combed. So like — cthresh controls how visible the combing must be, and this one controls how much combing there has to be in any localized area (a window defined by blockx and blocky) on the frame. 🎀

Minimum value is 0, maximum is blocky x blockx — and at that max, no frames will ever be detected as combed. This setting is known as MI in TFM/VFM vocabulary.

Default is 80.

p/c/n/u/b meaning ✨

p/c/n

let's assume this 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. So 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 that 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

so because of the field matching, we can see some frames get duplicated. To do a complete inverse telecine you need to rely on a decimation filter after this operation — see for instance the decimate filter.

Same operation but 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 you can see what p, c and n mean — 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, because they match from the opposite parity flag. In the examples below we're assuming we're currently matching the 2nd frame (Top:2, bottom:2). Depending on the match, an 'x' gets placed above and below each matched field. 👑

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

transforms the field order of the input video. iconic little order-flipper 🦋

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 happens by shifting the picture content up or down by one line, then filling the remaining line with appropriate picture content. This method is consistent with most broadcast field order converters.

If the input video isn't flagged as being interlaced, or it's already flagged as being of the required output field order, then this filter does not alter the incoming video. no notes, no changes, we move. 💀

It's 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

fills the borders of the input video without changing the video stream dimensions. Sometimes video has garbage at the four edges and you don't want to crop the input just to keep the size a multiple of some number. so this is the moment. ✨

This filter accepts the following options:

left — number of pixels to fill from the left border.
right — number of pixels to fill from the right border.
top — number of pixels to fill from the top border.
bottom — number of pixels to fill from the bottom border.
mode — set the fill mode.

mode accepts the following values:

smear — fill pixels using the outermost pixels
mirror — fill pixels using mirroring (half sample symmetric)
fixed — fill pixels with a constant value
reflect — fill pixels using reflecting (whole sample symmetric)
wrap — fill pixels using wrapping
fade — fade pixels to a constant value
margins — fill pixels at top and bottom with weighted averages of pixels near the borders

Default is smear.

color — set the color for pixels in fixed or fade mode. Default is black.

Commands

this filter supports the same commands as options, and the command accepts the same syntax as the corresponding option.

If the specified expression isn't valid, it's kept at its current value.

find_rect

finds a rectangular object in the input video. detective mode fr 🔥

The object to search for must be specified as a gray8 image, given via the object option.

For each possible match a score is computed, and 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. (yeah. just one. pick a favourite.)

When an object is found, these metadata entries get 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 the object is not detected. Default is disabled.

Examples

Cover a rectangular object with 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's giving paint bucket 🎨

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

converts the input video to one of the specified pixel formats. Libavfilter will try to pick one that's suitable as input to the next filter. team player energy 🫶

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` 🎬

Converts the video to a specified constant frame rate by duplicating or dropping frames as necessary — no cap, it just does the thing.

It accepts the following parameters:

fps — the desired output frame rate. It takes expressions containing these constants:
- 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 — pretend the first PTS should be the value you give, in seconds. This is what lets you pad/trim at the start of a 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 are:
- 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 done when reading the last frame. Possible values are:
- 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, the options can be specified as a flat string: fps[:start_time[:round]].

`framepack` 🥽

Packs 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 will stop when the shorter video ends. Please note that you may conveniently adjust view properties with the scale and fps filters — bestie behaviour, honestly.

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. Glow-up but make it temporal.

Ok so important bestie: this filter is not designed to function correctly with interlaced media. If you wish to change the frame rate of interlaced media then you are required to deinterlace before this filter and re-interlace after this filter. No shortcuts, iykyk.

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. Genuinely the simplest filter here and she ate.

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. Crime scene investigation but for your stream that stopped moving.

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. Like pressing pause on a vibe.

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. Plugins, but make it video.

To enable the compilation of this filter, you need to install the frei0r header and configure FFmpeg with --enable-frei0r. Non-negotiable prerequisite, sorry babes 😭

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. You're welcome 💅

`fspp` ⚡

Apply fast and simple postprocessing. It is a faster version of spp.

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 — highkey iconic optimisation behaviour.

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).

The flicker warning on use_bframe_qp is real, bestie — enabling it can cause flicker because B-Frames often have larger QP. Don't say I didn't warn you 💀

`fsync` 🗺️

Synchronize video frames with an external mapping from a file. Fr fr the map is the moment.

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 era, we love her.

The filter accepts the following options:

sigma — set horizontal sigma, standard deviation of Gaussian blur. Default is 0.5.
steps — how many steps the Gaussian approximation gets. default is 1.
planes — which planes get filtered. literally all of them by default, bestie.
sigmaV — vertical sigma, and if you hand it a negative number it just copies whatever sigma is. default is -1.

Commands

This filter is down with commands the same way it's down with options — the command speaks the exact same syntax as the matching option. 💅

If the expression you typed is not valid, no panic: it just stays at its current value. living its truth.

geq

Apply a generic equation to each pixel. yes really, every single pixel. math girl summer. 🔥

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 actually specified. If one of lum_expr, cb_expr, or cr_expr is in play, the filter automatically selects a YCbCr colorspace. If one of red_expr, green_expr, or blue_expr is in play, it selects an RGB colorspace instead. slay, it just knows. ✨

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 an opaque value. And if none of the chrominance expressions are specified, they evaluate to the luma expression. the way it has a whole backup plan 😭

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. 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), which lets you obtain sums of samples within a rectangle. See the functions without the sum postfix. so basically the same icons but with receipts 📊
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. it's not mad, it's just helpful 🫶

Please note that this filter can use multiple threads, in which case each slice will have its own expression state. If you want to use only a single expression state because your expressions depend on previous state, then you should limit the number of filter threads to 1. bestie read that twice, it's important 💀

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

Fix the banding artifacts that sometimes sneak into nearly flat regions when you truncate to 8-bit color depth. It interpolates the gradients that should go where the bands are, and dithers them. banding? never heard of her. 🦋

It is designed for playback only. Do not use it prior to lossy compression, because compression tends to lose the dither and bring back the bands. so do not come for me when it doesn't work, I warned you 😭

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 it 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 will fall back to the default value):

gradfun=radius=8

graphmonitor

Show various filtergraph stats. it's giving dashboard, it's giving observability 👑

With this filter you can debug a complete filtergraph. Especially issues with links filling up with queued frames. we love a diagnostic icon.

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 the 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 the 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 the output video frame rate will not be higher than this value. a little ceiling, as a treat 💅

grayworld

A color constancy filter that applies color correction based on the grayworld assumption. no cap, it's just out here fixing your colors 🌈

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). do the prep work 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. it's the science-y sibling of the one above 🧚

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 the 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 the default value is 1. Set to 0 for getting the max value instead of calculating the Minkowski distance. icon behavior honestly.
sigma — the standard deviation of Gaussian blur to be applied on the scene. Must be chosen in the range [0,1024.0] and the default value = 1. floor( sigma * break_off_sigma(3) ) can't be equal to 0 if difford is greater than 0. yes that's a real constraint, I don't make the rules 💀

Examples

Grey Edge:

greyedge=difford=1:minknorm=5:sigma=2

Max Edge:

greyedge=difford=1:minknorm=0:sigma=2

guided

Apply guided filter for edge-preserving smoothing, dehazing and so on. smoothing but make it smart ✨

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, a single input is required. If on, two inputs of the same resolution and pixel format are required. The second input serves as the guidance. team work makes the dream work 🫶
planes — set planes to filter. Default is first only.

Commands

This filter supports all of the above options as commands. yes, all of them, go off 👑

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 generating the guidance image, bestie, go read the 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 💅

Okay so haldclut applies a Hald CLUT to a video stream. The first input is the video stream to process, and the second one is the Hald CLUT. That Hald CLUT input can be a simple picture or a whole complete video stream.

The filter accepts these options:

clut — sets which CLUT video frames will be processed from the second input stream. It can be first or all. Default is all.
shortest — forces termination when the shortest input terminates. Default is 0.
repeatlast — keeps applying the last CLUT after the end of the stream. A value of 0 disables 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, iykyk.

This filter also supports the framesync options.

More info about the Hald CLUT lives 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 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 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. That's it, that's the vibe.

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.

The filter accepts these options:

strength — determines 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 — sets the maximum intensity that can generated and scales 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 — sets 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. It displays a color graph for each color component. It shows the distribution of the Y, U, V, A or R, G, B components, depending on the input format, in the current frame. Below each graph a color component scale meter is shown.

The filter accepts these options:

level_height — sets height of level. Default value is 200. Allowed range is [50, 2048].
scale_height — sets height of color scale. Default value is 12. Allowed range is [0, 40].
display_mode — sets display mode. It accepts these 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 — sets mode. Can be either linear, or logarithmic. Default is linear.
components — sets what color components to display. Default is 7.
fgopacity — sets foreground opacity. Default is 0.7.
bgopacity — sets background opacity. Default is 0.5.
colors_mode — sets colors mode. It accepts these 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.

It accepts these 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. No cap, this part matters.

🗺️ 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 — sets 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 one is dangerous. It can break the filter that came right before it in undefined ways if that filter has any extra constraints on its output. Do not touch it unless you fully understand the implications. Like, fr fr. 💀

🖼️ hwupload

Uploads your system-memory frames up onto hardware surfaces. Slay. ✨

The device you're uploading to has to be supplied when the filter is initialised. If you're running through ffmpeg, pick the right device with the -filter_hw_device option, or with the derive_device option. Your input and output devices must be of different types and compatible — the exact meaning of that is system-dependent, but typically it means they've gotta refer to the same underlying hardware context (think: same graphics card, iykyk).

Extra parameters it accepts:

derive_device type — instead of using the device handed over at initialisation, derive a brand-new device of type type from the device that the input frames are already living on. Cute, right? 🎀

💜 hwupload_cuda

Uploads system-memory frames to a CUDA device.

Accepts these optional parameters:

device — the number of the CUDA device to use.

🧚 hqx

Applies a high-quality magnification filter built specifically for pixel art. This one was originally created by Maxim Stepin — icon behavior, honestly. 👑

Accepts this option:

n — set the scaling dimension: 2 for hq2x, 3 for hq3x, and 4 for hq4x. Default is 3.

🌈 hstack

Stacks your input videos horizontally. Side by side, bestie. 🫶

Every stream has to be the same pixel format and the same height.

Ngl, this filter is faster than doing the same thing with the overlay and pad filters to build the same output.

Options:

inputs — set the number of input streams. Default is 2.
shortest — if set to 1, force the output to terminate the moment the shortest input terminates. Default value is 0.

🌸 hsvhold

Turns a certain HSV range into gray values.

Here's the vibe: this filter measures the color difference between the HSV color you set in the options and the ones it actually measures in the video stream. Depending on your options, output colors can be changed to gray — or not. Your call. 💅

Options:

hue — set the hue value used in the color difference calculation. Allowed range is from -360 to 360. Default value is 0.
sat — set the saturation value used in the color difference calculation. Allowed range is from -1 to 1. Default value is 0.
val — set the value used in the color difference calculation. Allowed range is from -1 to 1. Default value is 0.
similarity — set the 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 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 pixel's color is to the key color.

🦋 hsvkey

Turns a certain HSV range into transparency.

Same energy as above — it measures the color difference between the HSV color you set in the options and the ones it measures in the video stream. Depending on your options, output colors can be made transparent by adding an alpha channel. ✨

Options:

hue — set the hue value used in the color difference calculation. Allowed range is from -360 to 360. Default value is 0.
sat — set the saturation value used in the color difference calculation. Allowed range is from -1 to 1. Default value is 0.
val — set the value used in the color difference calculation. Allowed range is from -1 to 1. Default value is 0.
similarity — set the 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 pixel's color is to the key color.

🎀 hue

Modifies the hue and/or the saturation of the input. Main character filter energy, honestly. 🌈

Accepts these parameters:

h — specify the hue angle as a number of degrees. Accepts an expression, defaults to 0.
s — specify the saturation in the [-10,10] range. Accepts an expression and defaults to 1.
H — specify the hue angle as a number of radians. Accepts an expression, defaults to 0.
b — specify the brightness in the [-10,10] range. Accepts an expression and defaults to 0.

h and H are mutually exclusive — you literally cannot specify both at the same time. Pick a lane bestie. 💅

The b, h, H and s option values are expressions containing these 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 second 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 second 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 these commands:

b
s
h
H — modify the hue and/or the saturation and/or the brightness of the input video. The command accepts the same syntax as the corresponding option. If the specified expression isn't valid, it's kept at its current value. No drama. 🫶

👑 huesaturation

Applies hue-saturation-intensity adjustments to the input video stream.

Important: this filter operates in RGB colorspace.

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 gonna be adjusted. You set this by providing one or multiple values, so you 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 the strength of filtering. Allowed range is from 0 to 100. Default value is 1.
rw, gw, bw — set the weight for each RGB component. Allowed range is from 0 to 1. By default it's set to 0.333, 0.334, 0.333. These options get used in the saturation and lightness processing.
lightness — set preserving lightness, disabled by default. Adjusting hues can change lightness from the original RGB triplet; with this option enabled, lightness is kept at the same value. 💖

🍓 hysteresis

Grows the first stream into the second stream by connecting components. Sounds unhinged, but it makes it possible to build more robust edge masks. Slay.

Options:

planes — set which planes get processed as a bitmap; unprocessed planes get copied from the first stream. Default value 0xf — all planes get processed.
threshold — set the threshold used in filtering. If a pixel component value is higher than this value, the filter algorithm for connecting components kicks in. By default the value is 0.

The hysteresis filter also supports the framesync options. 🧚

🔍 iccdetect

Detects the colorspace from an embedded ICC profile (if there is one), and updates the frame's tags to match accordingly.

Options:

force — if true, the frame's existing colorspace tags will always be overridden by the values detected from an ICC profile. Otherwise they'll only be assigned if they contain "unknown". Enabled by default.

🎨 iccgen

Generates ICC profiles and attaches them to frames.

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. Check out the setparams filter for a list of possible values, but note that "unknown" values are not valid 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.

This filter takes two input videos.

Both input videos 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, which get compared one by one.

The obtained per component, average, min and max identity score gets printed through the logging system.

The filter stores the calculated identity scores of each frame in frame metadata. 📊

It also supports the framesync options.

In the example below, the input file main.mpg being processed is compared against the reference file ref.mpg:

ffmpeg -i main.mpg -i ref.mpg -lavfi identity -f null -

🌀 idet

Detects the video interlacing type.

This filter tries to detect whether the input frames are interlaced, progressive, top field first, or bottom field first. It'll also try to detect fields that are repeated between adjacent frames — which is a sign of telecine, if you were wondering.

Single frame detection considers only immediately adjacent frames when it classifies each frame. Multiple frame detection incorporates the classification history of the previous frames. Two different vibes, both valid. 💅

The filter logs these metadata values:

single.current_frame — detected type of the current frame using single-frame detection. One of: tff'' (top field first), bff'' (bottom field first), progressive'', or undetermined''
single.tff — cumulative number of frames detected as top field first using single-frame detection.
multiple.tff — cumulative number of frames detected as top field first using multiple-frame detection.
single.bff — cumulative number of frames detected as bottom field first using single-frame detection.
multiple.current_frame — detected type of the current frame using multiple-frame detection. One of: tff (top field first), bff (bottom field first), progressive, or undetermined.
multiple.bff — cumulative number of frames detected as bottom field first using multiple-frame detection.
single.progressive — cumulative number of frames detected as progressive using single-frame detection.
multiple.progressive — cumulative number of frames detected as progressive using multiple-frame detection.
single.undetermined — cumulative number of frames that could not be classified using single-frame detection.
multiple.undetermined — cumulative number of frames that could not be classified using multiple-frame detection.
repeated.current_frame — which field in the current frame is repeated from the last. One of neither, top, or bottom.
repeated.neither — cumulative number of frames with no repeated field.
repeated.top — cumulative number of frames with the top field repeated from the previous frame's top field.
repeated.bottom — cumulative number of frames with the bottom field repeated from the previous frame's bottom field.

The filter accepts the following options, bestie:

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 is halved (i.e., it contributes only 0.5 to its classification). The default of 0 means that all frames seen are given full weight of 1.0 forever.
analyze_interlaced_flag — when this is not 0 then idet will use the specified number of frames to determine if the interlaced flag is accurate, and it will not count undetermined frames. If the flag is found to be accurate it will be used without any further computations; if it's found to be inaccurate it will be cleared without any further computations. This allows inserting the idet filter as a low computational method 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 will add analysis metadata to each frame, which will then be discarded. At the end, the filter will also print a final report with statistics. obsessed — you get the receipts AND the cleanup.

🌀 il

Deinterleave or interleave fields — literally what it says on the tin.

This filter lets you process interlaced image fields without deinterlacing them. Deinterleaving splits the input frame into 2 fields (so-called half pictures). Odd lines get moved to the top half of the output image, even lines to the bottom half. You can process (filter) them independently and then re-interleave them. iconic separation behavior ✨

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.
deinterleave, d — deinterleave fields, placing one above the other.
interleave, i — interleave fields. Reverse the effect of deinterleaving.

Default value is "none".

luma_swap, ls
chroma_swap, cs
alpha_swap, as

Swap 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 by the local (3x3) average, but only taking into account values higher than the pixel. It's giving selective glow-up, no cap.

It accepts the following options:

threshold0
threshold1
threshold2
threshold3

Limit the maximum change for each plane. Default is 65535. If 0, the plane will remain unchanged.

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.

              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 — this 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 will slightly less reduce twitter and moire, but better retain detail and the subjective sharpness impression.

🌸 kerndeint

Deinterlace input video by applying Donald Graft's adaptive kernel deinterlacing. Works on the interlaced parts of a video to produce progressive frames — surgical, honestly.

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.
map — paint pixels exceeding the threshold value to white if set to 1. Default is 0.
order — set the field 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 will be copied. By default value 0xf, all planes will be processed.
scale — set value which will be multiplied with the filtered result.
delta — set value which will be 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 — the drama is extended, respectfully. This filter accepts the following options:

decay — set 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.

This filter can be used to correct for 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. iykyk

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 will have to take care of ordering, i.e. whether vignetting should be applied before or after lens correction.

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 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 half of the image diagonal and r_src and r_tgt are the distances from the focal point in the source and target images, respectively.

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. no skipping the homework on this one 💅

ok so to get a list of available makes and models, you just leave out one or both of the make and model options — the filter will yeet the full list into the log at level INFO. Column one is the make, column two is the model. 💅 And if you want the lens list, set any values for make and model and leave out the lens_model option — boom, full list of lenses in the log, again at level INFO. Heads up: the ffmpeg tool will literally exit right after the list is printed. Like, it's done, it's leaving. 🫶

The filter accepts these options, bestie:

make — the make of the camera (for example, "Canon"). This option is required.
model — the model of the camera (for example, "Canon EOS 100D"). This option is required.
lens_model — the model of the lens (for example, "Canon EF-S 18-55mm f/3.5-5.6 IS STM"). This option is required.
db_path — the full path to the lens database folder. If you don't set it, the filter will try to load the database from the install path from when the library was built. Default is unset.
mode — the type of correction to apply. Valid values are:
- vignetting — enables fixing lens vignetting.
- geometry — enables fixing lens geometry. This is the default.
- subpixel — enables fixing chromatic aberrations.
- vig_geo — enables fixing lens vignetting and lens geometry.
- vig_subpixel — enables fixing lens vignetting and chromatic aberrations.
- distortion — enables fixing both lens geometry and chromatic aberrations.
- all — enables all possible corrections. the whole menu, no skips. 👑
focal_length — the focal length of the image/video (that's the zoom; expected to be constant for video). For example, an 18--55mm lens has a focal length range of [18--55], so pick a value in that range when you're using that lens. Default 18.
aperture — the aperture of the image/video (again, expected constant for video). ngl, aperture is only used for vignetting correction. Default 3.5.
focus_distance — the focus distance of the image/video (expected constant for video). This one 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 (which is 1000).
scale — the scale factor applied after transformation. After correction the video is no longer necessarily rectangular, 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 is done. Lower values may show you more of the corrected image, higher values may avoid unmapped areas in the output. pick your fighter 💀
target_geometry — the target geometry of the output image/video. Valid options are:
- rectilinear (default)
- fisheye
- panoramic
- equirectangular
- fisheye_orthographic
- fisheye_stereographic
- fisheye_equisolid
- fisheye_thoby
reverse — apply the reverse of image correction (so instead of correcting distortion, apply it). un-correcting, iconic behaviour 😭
interpolation — the type of interpolation used when correcting distortion. Valid options are:
- 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 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

Apply the same as before, but only for the first 5 seconds of video. cute little snack-sized correction ✨

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).

`libplacebo` 🔥

Flexible GPU-accelerated processing filter, built on libplacebo (https://code.videolan.org/videolan/libplacebo).

Options 💅

The options for this filter are split into the following sections:

Output mode 🎬

These options control the overall output mode. By default, libplacebo will try 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.

inputs — set the number of inputs. This can be used, alongside the idx variable, to let you place/blend multiple inputs inside the output frame. Basically it gives you hstack, overlay, etc. energy. 🫶
w, h — set the output video dimension expression. Default values are iw and ih. Allows for the same expressions as the scale filter.
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. Your options:
- 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" (the default), input timestamps get passed through to the output unmodified instead. Otherwise, the input video frames will be interpolated as necessary to rescale the video to your specified target framerate, in a manner as 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 gets performed.
force_original_aspect_ratio, force_divisible_by — 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 will always have a pixel aspect ratio of 1:1. If disabled (the default), any aspect ratio mismatches — including ones from, like, anamorphic video sources — get forwarded to the output pixel aspect ratio.
normalize_sar — like reset_sar, but instead of stretching the video content to fill the new output aspect ratio, the content gets 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 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 a value of 1.0 always crops off parts of the content. Intermediate values are possible, giving you a mix of the two approaches. bestie's blender settings 💀
fit_mode — specify the content fit strategy according to a list of predefined modes. 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 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. The input 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 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 — fully replaces the conversion from the image color space to the output color space. If such a LUT is present, it has the highest priority, and overrides any ICC profiles, as well as options related to tone mapping and output colorimetry (color_primaries, color_trc). Periodt 💅
extra_opts — pass extra libplacebo internal configuration options. These can be specified as a list of key=value pairs separated by :. The following example shows how to configure a custom filter kernel ("EWA LanczosSharp") and 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. This cache is not cleaned up automatically. If the path does not end in a directory separator, the generated filenames will be effectively prefixed by the last path component. All directories must already exist.

-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, leading to no change. 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, when the output format has an alpha channel. See the setparams filter for a list of possible values.
apply_filmgrain — apply film grain (e.g. AV1 or H.274) if present in source frames, and strip it from the output. Enabled by default. ✨
apply_dolbyvision — apply Dolby Vision RPU metadata if present in source frames, and strip it from the output. Enabled by default. Note that Dolby Vision will always output BT.2020+PQ, overriding the usual input frame metadata. These will also be picked as the values of auto for the respective frame output options.

In addition to 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 the following constants:

in_idx, idx — the (0-based) numeric index of the currently active input stream.
crop_w, cw / crop_h, ch — the computed values of crop_w and crop_h.
pos_w, pw / pos_h, ph — the computed values of pos_w and pos_h.
in_t, t — the input frame timestamp, in seconds. NAN if input timestamp is unknown.
out_t, ot — the input frame timestamp, in seconds. NAN if 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. Note that libplacebo will always internally operate on 4:4:4 content, so any sub-sampled chroma formats such as yuv420p will necessarily be upsampled and downsampled as part of the rendering process. That means scaling might be in effect even if the source and destination resolution are the same. 😭 lowkey important

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.
- 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 — 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.
- 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. 👑
- gaussian — Gaussian kernel. Has certain ideal mathematical properties, but subjectively very blurry.
- 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 — 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). The value (between 0.0 and 1.0) configures the strength of the anti-ringing algorithm. May increase aliasing if set too high. 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 frames are not deinterlaced unless a deinterlacing algorithm is chosen.

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 good at counteracting many common sources of banding and blocking. Turning this on is highly recommended whenever quality is desired. 🫶

deband — enable (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. 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.

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), 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 exactly reproducible frame-perfect results are needed, it's recommended to turn this feature off.

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, somewhere between 0.0 and 1000.0. Higher values make peak detection less responsive to changes in the input — like, it stops freaking out every frame. 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 functionality is disabled, no cap 🌸
percentile — which percentile of the frame brightness histogram gets used as the source peak for tone-mapping. Defaults to 99.995, a pretty conservative value (sensible queen behaviour tbh). Setting it to 100.0 disables frame histogram measurement entirely and instead uses the true peak brightness for tone-mapping. iykyk 💅

Tone mapping 🎨

The options in this section control how libplacebo performs tone-mapping and gamut-mapping when it's dealing with mismatches between wide-gamut or HDR content. In general, libplacebo relies on accurate source tagging and mastering display gamut information to produce the best results — so tag your sources properly, bestie, it literally matters 🙈

gamut_mode — how to handle out-of-gamut colors that can occur as a result of colorimetric gamut mapping. The options are:
- clip — do nothing, simply clip out-of-range colors to the RGB volume. Low quality but extremely fast ⚡
- 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 for which a bright, saturated display is desired 👀
- 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). Great for debugging, cursed for watching.
- 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 💀
tonemapping — the tone-mapping algorithm to use. Available values are:
- 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.
- 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 🌈
- 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" 🫶
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.

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. (That's the move. Trust.)

dithering — 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 — think of it as the "gaming mode" of the filter, bestie 💅

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 🔥 (Yes, really. Almost everything stops working. Maximum speed though.)

Commands 👑

This filter supports almost all of the above options as commands. Slay — you can tweak stuff at runtime, no restart needed 💖

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:
Feelin' fancy with frame mixing? Slap on fps=60:frame_mixer=mitchell_clamp:

libplacebo=fps=60:frame_mixer=mitchell_clamp

Wanna convert your input to standard sRGB JPEG? Say less:

libplacebo=format=yuv420p:colorspace=bt470bg:color_primaries=bt709:color_trc=iec61966-2-1:range=pc

Higher quality debanding settings, for when you're extra:

libplacebo=deband=true:deband_iterations=3:deband_radius=8:deband_threshold=6

Run this filter on the CPU on systems with Mesa installed (with the most expensive options turned off, bestie, we're budgeting):

ffmpeg ... -init_hw_device vulkan:llvmpipe ... -vf libplacebo=upscaler=none:downscaler=none:peak_detect=false

Suppress the CPU-based AV1/H.274 film grain application in the decoder so this filter handles it instead. Lowkey important note: this is only a gain if the frames are already on the GPU, or if you're using libplacebo for other purposes — 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 round-trip through RAM (the glow-up we deserve):

ffmpeg -init_hw_device vulkan -hwaccel vaapi -hwaccel_output_format vaapi ... -vf libplacebo

🌈 libvmaf 🌈

Calculate the VMAF (Video Multi-Method Assessment Fusion) score for a reference/distorted pair of input videos. It's giving ✨quality judge✨.

The first input is the distorted video, and the second input is the reference video.

The VMAF score you get is printed through the logging system.

It requires 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, no drama:

ffmpeg -i distorted.mpg -i reference.mpg -lavfi libvmaf=log_path=output.xml -f null -

Example with multiple models (go off, king):

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:

ffmpeg -i distorted.mpg -i reference.mpg -lavfi libvmaf='feature=name=psnr|name=ciede' -f null -

Example with options and different containers:

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 — that's the whole vibe check.

It requires Netflix's vmaf library (libvmaf) as a pre-requisite. Once you've installed the library, you can enable it using: ./configure --enable-nonfree --enable-ffnvcodec --enable-libvmaf.

✨ Examples ✨

Basic usage showing CUVID hardware decoding and CUDA scaling with scale_cuda:

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 limited difference filter using second and optionally third video stream. Very "spot the difference but make it scientific".

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 the above options as commands, except option reference. (Not me excluding reference from the command squad 💅)

🎀 limiter 🎀

Limits the pixel component values to the specified range [min, max]. Keeping everyone in their lane, respectfully.

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 the above options as commands.

🔥 loop 🔥

Loop video frames. It's giving rewind-the-TikTok energy.

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 single first frame infinitely (periodt):

loop=loop=-1:size=1:start=0

Loop single first frame 10 times:

loop=loop=10:size=1:start=0

Loop 10 first frames 5 times:

loop=loop=5:size=10:start=0

🌸 lut1d 🌸

Apply a 1D LUT to an input video. Iconic lil colour transform.

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 the above options as commands.

👑 lut3d 👑

Apply a 3D LUT to an input video. The big sister of lut1d, no cap.

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.

🫶 lumakey 🫶

Turn certain luma values into transparency. Bye-bye, luma. 👋

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.

💀 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.

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.

The lut filter requires either YUV or RGB pixel formats in input, lutrgb requires RGB pixel formats in input, and lutyuv requires YUV. (Read the dress code, bestie 🎀)

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:
crush the colors into their opposites — negating via maxval+minval-val is literally the same energy as just saying negval, bestie, the long way is just for the girlies who like to see the math:

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, and I cannot stress this enough, the exact same thing as:

lutrgb="r=negval:g=negval:b=negval"
lutyuv="y=negval:u=negval:v=negval"

negate just the luma — one line, no notes 💅

lutyuv=y=negval

yeet the chroma, go full grayscale — u and v get pinned to 128 and the color just... leaves the chat. iconic honestly:

lutyuv="u=128:v=128"

luma burn effect — double the value, let it glow 🔥

lutyuv="y=2*val"

no green, no blue, only vibes — red stays, everything else is deleted:

lutrgb="g=0:b=0"

set a flat alpha channel on your input — needs the rgba format first, she's specific like that:

format=rgba,lutrgb=a="maxval-minval/2"

gamma-correct the luma by 0.5 — half the gamma, all of the slay:

lutyuv=y=gammaval(0.5)

throw away the least significant bits of luma — yes the bitmask is 128+64+32, keep it exactly, and yes the single quotes matter:

lutyuv=y='bitand(val, 128+64+32)'

technicolor glam — push u and v out from the midpoint, very 1960s movie musical:

lutyuv=u='(val-maxval/2)*2+maxval/2':v='(val-maxval/2)*2+maxval/2'

`lut2`, `tlut2` 🎀

Okay so lut2 is the two-input girly 💖 — she takes two input streams and gives you back exactly one output stream.

Meanwhile tlut2 (the time lut2, iykyk) doesn't want two streams at all — she takes two consecutive frames from one single stream. Main character behavior.

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 this one corresponds to the alpha component
d — set output bit depth, only available for the lut2 filter. By default it's 0, which means the bit depth is automatically picked from the first input format.

Also: lut2 supports the framesync options too, no cap 🫶

Each of these bad boys specifies the expression used to compute the lookup table for the corresponding pixel component values.

Which component actually goes with which c* option? Depends on the format in the inputs, babes. It's giving "read the room / read the pixel format".

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. periodt.

Commands 💬

This filter supports all the above options as commands except option d. (She's not like the other options.)

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. Three streams, one clamp, very "hold my hand so I don't go out of bounds" energy.

She returns the value of the first stream, squeezed to live between second input stream − undershoot and third input stream + 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; the unprocessed planes get copied straight from the first stream. By default the value is 0xf, which means all planes get processed. all of them. every single one. 💅

Commands

This filter supports all the above options as commands.

`maskedmax` 👑

Merge the second and third input stream into the output stream using absolute differences between the second input stream and the first input stream, plus the absolute difference between the third input stream and the first input stream. The picked value comes from the second input stream if the second absolute difference is greater than the first one — otherwise it comes from the third input stream. She's dramatic but she's fair.

This filter accepts the following options:

planes — set which planes will be processed as bitmap; unprocessed planes get copied from the first stream. By default 0xf, all planes are processed.

Commands

This filter supports all 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. Yes, the third stream is the boss here, we're just blending.

A value of 0 in the third stream pixel component means that pixel component from the first stream is returned unchanged. The maximum value (e.g. 255 for 8-bit videos) means that pixel component from the second stream is returned unchanged. Intermediate values define the amount of merging between both input stream's pixel components. It's a gradient, bestie, not a switch.

This filter accepts the following options:

planes — set which planes will be processed as bitmap; unprocessed planes get copied from the first stream. By default 0xf, all planes are processed.

Commands

This filter supports all the above options as commands.

`maskedmin` 🧚

Merge the second and third input stream into the output stream using 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 comes from the second input stream if the second absolute difference is less than the first one — otherwise it comes from the third input stream. (Same as maskedmax but she flipped the comparison, as one does.)

This filter accepts the following options:

planes — set which planes will be processed as bitmap; unprocessed planes get copied from the first stream. By default 0xf, all planes are processed.

Commands

This filter supports all the above options as commands.

`maskedthreshold` ⭐

Pick pixels by comparing the absolute difference of two video streams against a fixed threshold.

So: if the absolute difference between the pixel component of the first and second video stream is equal to or lower than the threshold you supplied, then the pixel component from the first video stream is picked. Otherwise the pixel component from the second video stream is picked. Simple as.

This filter accepts the following options:

threshold — set the threshold used when picking pixels from the absolute difference of the two input video streams.
planes — set which planes will be processed as bitmap; unprocessed planes get copied from the second stream (yes, second this time, keep up 💀). By default 0xf, all planes are processed.
mode — set the mode of filter operation. Can be abs or diff. Default is abs.

Commands

This filter supports all the above options as commands.

`maskfun` 🍓

Create a mask from the input video.

For example, it's useful to create motion masks after the tblend filter. Very "I made this for you" of her.

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. 👀

Commands

This filter supports all the above options as commands.

`mcdeint` 🌈

Apply motion-compensation deinterlacing.

Heads up though: she needs one field per frame as input, and must therefore be used together with yadif=1/3 or an equivalent. No solo acts, she's a duet girlie.

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 uses 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. Democracy, but for pixels.

This filter accepts the following options:

radius — set horizontal radius size. Default value is 1. Allowed range is 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 integer from 0 to 127. If it is 0, the value will be picked from the horizontal radius option. (She'll share, but only if you make her.)
percentile — set the median percentile. Default value is 0.5. A default value of 0.5 will always pick median values, while 0 will pick minimum values, and 1 maximum values.

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. No crash, no drama, she just stays where she is. 🫶

`mergeplanes` 👑

Merge color channel components from several video streams.

The filter accepts up to 4 input streams, and merges the selected input planes into the output video. It's a group project and she's the one actually doing the work.

This filter accepts the following options:

mapping — set the 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 the plane number of the corresponding input to use (from 0 to 3). The rest of the mapping is 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. It's giving hex code, but make it planes. 💅

format — set output pixel format. Default is yuva444p.
map0s, map1s, map2s, map3s — set the input to output stream mapping for output Nth plane. Default is 0.
map0p, map1p, map2p, map3p — set the input to output plane mapping for output Nth plane. Default is 0.

Examples 🎀

Merge three gray video streams of the same width and height into a single video stream:

[a0][a1][a2]mergeplanes=0x001020:yuv444p

Merge the 1st yuv444p stream and the 2nd gray video stream into a yuva444p video stream:

[a0][a1]mergeplanes=0x00010210:yuva444p

Swap Y and A plane in a yuva444p stream:
Swap U and V plane in yuv420p stream:

format=yuv420p,mergeplanes=0x000201:yuv420p

Cast a rgb24 clip to yuv444p:

format=rgb24,mergeplanes=0x000102:yuv444p

🌀 mestimate

Okay so this one estimates motion vectors using block matching algorithms and then exports them — the vectors get stashed in frame side data so other filters can use them later. Bestie, it's literally giving "I did the math so you don't have to" 💅

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

Same energy as mestimate but make it GPU ✨ This one estimates and exports motion vectors using D3D12 hardware-accelerated motion estimation. It leans on the GPU hardware motion estimation capabilities hanging out in the DirectX 12 Video APIs, which gets you significant performance improvements over the software-based mestimate filter. No cap, that's a glow-up.

Motion vectors are stored in frame side data to be used by other filters.

Heads up though: this filter requires the input to be in d3d12 hardware pixel format. And it uses quarter-pixel precision for motion vector estimation — fancy.

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 Midway Image Equalization effect using two video streams. 🫶

So what even is that? Midway Image Equalization adjusts a pair of images so they have the same histogram, while keeping their dynamics as much as possible — like, preserving the vibes of each image while matching them up. It's useful for e.g. matching exposures from a pair of stereo cameras. Iconic for stereo work, fr fr.

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. (Yes, the first one — remember that 👀)

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 hours, basically ✨

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 friend group, blended into one 💖

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. Grayscale but make it couture 💅

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.

And like — unlike naive implementation and much slower performance in erosion and dilation filters, when speed is critical morpho filter should be used instead. So if you're in a hurry, bestie, pick this one. 🏃‍♀️💨

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. Yes, it literally just throws frames away when they're basically twins 👯‍♀️

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. It's a comparison filter, bestie — it's giving "who wore it better" but with math 🔥

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.

Fr fr, mismatch the resolution or pixel format and things get weird — this filter is picky about its matchups.

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 real quick, bestie: in this example the input file main.mpg being processed is literally being compared with the reference file ref.mpg.

ffmpeg -i main.mpg -i ref.mpg -lavfi msad -f null -

multiply

Multiply the first video stream's pixel values by the second video stream's pixel values. It's giving math but make it cinema ✨

The filter accepts these options:

scale — sets the scale applied to the second video stream. Default is 1. Allowed range is from 0 to 9.
offset — sets the offset applied to the second video stream. Default is 0.5. Allowed range is from -1 to 1.
planes — specifies which planes from the input video stream get processed. By default all planes are processed.

Commands

This filter supports the same commands as options. 💅

negate

Negate/invert the input video. Main character energy, but flipped 😭

It accepts this option:

components — sets which 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 the same commands as options. 🫶

nlmeans

Denoise frames using the Non-Local Means algorithm. No cap, this one is lowkey iconic for cleaning up noisy footage.

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 these 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 an 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 an 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. It's giving smart-girl deinterlacing 🔥

This filter accepts these 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 the filter processes 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 the 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 the 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 the "girl, not that format" filter 💅

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 make it aesthetic 🌸

The filter accepts these options:

all_seed, c0_seed, c1_seed, c2_seed, c3_seed — set the noise seed for a specific pixel component, or for all pixel components in the 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 for all pixel components in the case of 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 in the case of 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 the 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)

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. Obsessed with the flexibility, honestly 🎀

The normalize filter accepts these options:

blackpt, whitept — colors which define the output range. The minimum input value is mapped to blackpt. The maximum input value is mapped to 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 the 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. No notes, just vibes 💜

ocio

OpenColorIO library filter. This filter lets you 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 these 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 it goes from a display/view to the "input colorspace".
filetransform — lets you point at an external file-transform instead of the OCIO config file. It's lowkey useful for applying a single transform without needing a whole OCIO config file. ✨
format — lets you specify the output pix_fmt of the OCIO filter. This has to be a RGB colorspace, so bestie, you really are limited to rgb24, rgba, rgb48, rgba48, gbrp10, gbrp12, gbrpf32le, gbrapf32le — for most encoding we'd recommend rgb48. 💅
context_params — lets you specify additional context parameters for 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 a sRGB display using the "ACES 1.0 - SDR Video" view transform — this assumes the OCIO file is defined with the OCIO environment variable. Note you will need to wrap the argument in quotes to ensure that the spaces are interpreted correctly.

input=ACEScg:display=sRGB - Display:view=ACES 1.0 - SDR Video:format=rgb48

As above, but using the OCIO file studio-config-v1.0.0_aces-v1.3_ocio-v2.1_ns.ocio rather than 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're converting to YCrCb you still will want to set the color matrix for the conversion. This is a 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 tesseract data. Default is to use whatever was set at installation.
language — set language, default is "eng".
whitelist — set character whitelist.
blacklist — set character blacklist.

The filter exports recognized text as the frame metadata lavfi.ocr.text. The filter exports 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 information: http://docs.opencv.org/master/modules/imgproc/doc/filtering.html

Several libopencv filters are supported; see the following subsections. 💖

`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, the number or 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 the following 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 first three components.
g — draw trace grid. By default is enabled.
st — draw some statistics. By default is enabled.
sc — draw scope. By default is enabled.

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 🌸

Inspect full first row of video frame.

oscilloscope=x=0.5:y=0:s=1

Inspect full last row of video frame.

oscilloscope=x=0.5:y=1:s=1

Inspect full 5th line of video frame of height 1080.

oscilloscope=x=0.5:y=5/1080:s=1

Inspect full last column of 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 that 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 the 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 format of 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, 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, hence, 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 in the same zero timestamp, as the example for the movie filter does.

You can chain together more overlays but you should test the efficiency of such 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 of 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:
Insert 2 different transparent PNG logos (second logo chilling in the 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 — and bestie, WxH has to specify the size of the main input to the overlay filter, no cap:

color=color=red@.3:size=WxH [over]; [in][over] overlay [out]

Play an original video and a filtered version side by side (here we're using the deshake filter) with 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 literally the same as:

ffplay input.avi -vf 'split[b], pad=iw*2[src], [b]deshake, [src]overlay=w'

Make a sliding overlay that appears from the left to the right top part of the screen, starting since time 2 — iconic entrance behavior tbh:

overlay=x='if(gte(t,2), -w+(t-2)*20, NAN)':y=0

Compose output by putting two input videos side to side — a whole collab:

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 — sneaky little redaction moment:

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 — layers on layers, we love a stacked fit:

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. Say bye to the grain, bestie ✨

The filter accepts the following options:

depth — Set depth.

Larger depth values will denoise lower frequency components more, but slow down filtering. So like, more glow-up, more waiting, iykyk.

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 "frame it, bestie" 🖼️

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. Mutual support, we love to see it 💖

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. Cute little auto-center moment ✨

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. It's giving curator energy, fr fr 👑

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 like, the leftovers just sit there in black, humble and unseen 🖤)
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. Main character = whatever's moving 💃
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.

🦋 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. Glow-down, but make it art 🎨

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. (An icon of questionable choices, we respect the archival moment 💀)
- 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). Efficiency queen 💅

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.

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. (Filmed it at a whole angle? Not anymore bestie, we're straightening that vibe out 📏✨)

Okay so here's the whole rundown of the params it takes 💅

x0, y0, x1, y1, x2, y2, x3, y3 — these set the coordinate expressions for the top left, top right, bottom left and bottom right corners. 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 "source", then the points you specified get sent to the corners of the destination. If "sense" is "destination", then it's the other way around — the corners of the source get sent to the coordinates you gave.

The expressions can use these variables:

W, H — the width and height of the video frame.
in — input frame count.
on — output frame count.

interpolation — sets the interpolation used for the perspective correction.

It accepts these values:

linear
cubic

Default value is linear.

sense — sets how the coordinate options get interpreted, bestie.

It accepts these values:

0, source — send the point in the source that's 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 — tells it when to evaluate the expressions for the coordinates x0,y0,...x3,y3.

It accepts these values:

init — only evaluate the expressions once, during filter initialization, or when a command gets processed.
frame — evaluate the expressions for every single incoming frame.

Default value is init.

🌀 `phase`

Delays interlaced video by one field time so that the field order actually flips. Slay.

The intended use? Fixing PAL movies that were captured with the opposite field order to the film-to-video transfer. Like, a rescue mission 🫶

Here's what it accepts:

mode — sets the phase mode.

It accepts these values:

t — capture field order top-first, transfer bottom-first. The filter will delay the bottom field.
b — capture field order bottom-first, transfer top-first. The filter will delay the top field.
p — capture and transfer with the same field order. This mode literally only exists so the other options' documentation has something to refer to — but if you actually select it, the filter will faithfully do nothing. Iconic behaviour, honestly 💀
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 this 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 selecting 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 this 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 the all above options as commands.

⚡ `photosensitivity`

Reduces the various flashes in video, so it helps out users with epilepsy. Genuinely a good-vibes 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, ngl. 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. Aesthetic ✨

The filter accepts these options:

width, w, height, h — set the block dimensions that get 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 — don't try to be cute with anything smaller.

The filters accept these options:

x — set scope X position, relative offset on X axis.
y — set scope Y position, relative offset on 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 X axis.
wy — set window Y position, relative offset on Y axis.

Commands 💬

This filter supports 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 you don't set it, the filter will use the QP from the video stream (if available).
mode — set the 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 an alpha premultiply effect to the input video stream, using the first plane of the second stream as alpha.

Both streams must have same dimensions and same pixel format. No exceptions, bestie.

The filter accepts this option:

planes — set which planes will be processed; unprocessed planes will be copied. By default value 0xf, all planes will be processed.
inplace — don't require the 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 the requirements of the downstream filter graph. This is roughly equivalent to either "premultiply:inplace=yes" or "unpremultiply:inplace=yes", or otherwise a noop.

The filter accepts this option:

planes — set which planes will be processed; unprocessed planes will be copied. By default value 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 will be copied. By default value 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 the all above options as commands.

🌈 `pseudocolor`

Alters frame colors in the video with pseudocolors. It's giving glow-up 💜

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 component to use as 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 opacity of output colors. Allowed range is from 0 to 1. Default value is set to 1.

Each of the expression options specifies the 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 the all 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 is 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, which get compared one by one. So like, keep them matched or it's gonna be a mess 😭

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 the processing it's averaged across all frames equally, and the following 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 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 — tells the filter whether to write that max value into the stats log. Default is 0, and it needs stats_version >= 2. If you set it while stats_version < 2, the filter just returns an error. no cap. 😭

This filter also supports the framesync options. 🫶

The file printed if stats_file is selected? It's a whole sequence of key/value pairs, written as key:value, one for each compared couple of frames.

And 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. Will match stats_version. ✨
fields — a comma separated list of the per-frame-pair parameters included in the log.

A description of each shown per-frame-pair parameter, bestie, coming right up:

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

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. iconic 💅

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 — it can handle mixed hard-telecine, 24000/1001 fps progressive, and 30000/1001 fps progressive content. she's versatile 💜

The pullup filter is built to take advantage of future context when it makes its decisions. It's stateless in the sense that it doesn't lock onto a pattern to follow — instead it looks forward to the following fields so it can identify matches and rebuild progressive frames. lowkey genius.

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.
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.
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. 🧚

For best results (without duplicated frames in the output file) it is necessary to change the output frame rate. For example, to inverse telecine NTSC input:

ffmpeg -i input -vf pullup -r 24000/1001 ...

qp

Change video quantization parameters (QP). ✨

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.
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. okay this one is genuinely so fun 🎀

To enable the compilation of this filter, you need to configure FFmpeg with --enable-libqrencode.

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 overlaid. 🙈

This 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.

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.
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.

🦋 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.

These parameters allow the q and Q expressions to refer to each other, so you can for example specify q=3/4*Q.

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.

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. 😭 seriously, four levels.

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.

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.

okay so, the qrencode text-expansion functions keep going like this, bestie:

expr_formatted, ef — Evaluate the expression's value and output as a formatted string. 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.
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 qrencode filter are also available.
rand(min, max) — return a random number included between min and max.

✨ Examples

QR-code glow-up time — commands exactly as the manual gives them:

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. iykyk ✨

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 and 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 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. she’s giving cleanup era ✨

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.

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.

Ok so if you need to, you can fix little splotches by hand 💅 — but bestie, remember: if logo pixels aren't covered, the filter quality gets much worse. 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 those extra pixels will slow things down on a large logo. So like… be chill about it, ok? 🫶

repeatfields

This filter uses the repeat_field flag from the Video ES headers and hard repeats fields based on its value. Literally just reads the flag and goes "aight, doubling it then" ✨

reverse

Reverse a video clip. It's giving rewind 🎬

Warning: this filter requires memory to buffer the entire clip, so trimming is suggested. Like, your RAM is not infinite babe 😭

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. It's a little pixel shove, ngl 🫸

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. Slay 💅

roberts

Apply roberts cross operator to input video stream. Edge detection moment 👑

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. ✨

rotate

Rotate video by an arbitrary angle expressed in radians. Spin that thing 🌀

The filter accepts the following options:

A description of the optional parameters follows. (Very demure, very mindful of the docs.)

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, bestie 🫡

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. Not anywhere else, ok? 🙈

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 vibes with what it had 💀

sab

Apply Shape Adaptive Blur. Soft-focus fairy vibes 🧚

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. Cute, right? It just copies its sibling 🎀

scale

Scale (resize) the input video, using the libswscale library. This is the main-character filter fr fr 👑

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.

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. (Go read it, iykyk 📚)

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 setting, bestie, lets you override the autodetected value and also force a specific value used for the output and encoder. If you don’t specify it, the color space type depends on the pixel format.

Possible values:

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 value used 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 value used for the output and encoder. Possible values:
- auto — choose automatically. This is the default.
- 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 value used for the output and encoder. Possible values:
- auto — choose automatically. This is the default.
- 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 — output video dimensions will automatically be decreased if needed.
- increase — 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 keeping 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 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 — 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 — 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 the following commands:

width, w — set the output video dimension expression. The command accepts the same syntax as the corresponding option. If the specified expression is not valid, it is kept at its current value.
height, h — set the output video dimension expression. The command accepts the same syntax as 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 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 all of the above options as commands, periodt.

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 the scrolling direction.
vertical, v — set the vertical scrolling speed. Default is 0. Allowed range is from -1 to 1. Negative values change the 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.

In addition, this filter logs a message and sets frame metadata when it detects a scene change by threshold.

lavfi.scd.mafd metadata keys are set with mafd for every frame.

lavfi.scd.score metadata keys are set with scene change score for every frame to detect scene change.

lavfi.scd.time metadata keys are set with current filtered frame time which detect scene change with threshold.

The filter accepts the following options:

threshold, t — sets the scene change detection threshold as a percentage of maximum change. Good values live in the [8.0, 14.0] range, and the 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 enable it if you want to get a snapshot of scene change frames only. 🎬

🎨 selectivecolor 🎨

Adjusts cyan, magenta, yellow and black (CMYK) to certain ranges of colors (like "reds", "yellows", "greens", "cyans", ...). The adjustment range is defined by the "purity" of the color — that is, how saturated it already is. It's basically the Adobe Photoshop Selective Color tool, but make it open source. 💅

The filter accepts the following options:

correction_method — selects 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 maximum)
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 maximum)
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 maximum)
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 the amount of cyan, magenta, yellow and black for the pixels of its range.

🌸 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. Literally doubling everything, main character energy. ✨

This filter uses field-dominance information in the frame to decide which of each pair of fields to place 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.

This is done by changing the specified Sample (aka Pixel) Aspect Ratio, according to the following equation:

<DAR> = <HORIZONTAL_RESOLUTION> / <VERTICAL_RESOLUTION> * <SAR>

Keep in mind that the setdar filter does not modify the pixel dimensions of the video frame. Also, the display aspect ratio set by this filter may be changed by later filters in the filterchain, e.g. in case of scaling or if another setdar or a setsar filter is applied. So yeah, it's giving "teamwork makes the dream work" — later filters can absolutely undo your vibe. 😭

The setsar filter sets the Sample (aka Pixel) Aspect Ratio for the filter output video.

Note that as a consequence of the application of this filter, the output display aspect ratio will change according to the equation above.

Keep in mind that the sample aspect ratio set by the setsar filter may be changed by later filters in the filterchain, e.g. if another setsar or a setdar filter is applied.

It accepts the following 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 is not specified, the value 0 is assumed, meaning that the same input value is used.
max — set the maximum integer value to use for expressing numerator and denominator when reducing the expressed aspect ratio to a rational. Default value is 100.

The parameter sar is an expression containing the following 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. 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.

⭐ 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 does not change the input frame, but only sets the corresponding property, which affects how the frame is treated by following filters (e.g. fieldorder or yadif). Sneaky little metadata queen. 👑

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 frame parameters for the output video frame.

The setparams filter marks interlace and color range for the output frames. It does not change the input frame, but only sets the corresponding property, which affects how the frame is treated by filters/encoders. Same deal as setfield — all vibes, no pixels touched. ✨

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 mode. Available values are:
- auto — keep the same alpha mode (default).
- unspecified, unknown
- premultiplied
- straight

💫 shear 💫

Applies shear transform to the input video. It's slanted, it's cute, it's a whole look. 🦋

This filter supports the following options:

shx — shear factor in X-direction. Default value is 0. Allowed range is from -2 to 2.
shy — shear factor in 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 interpolation type. Can be bilinear or nearest. Default is bilinear.

Commands

This filter supports all the above options as commands. 💅

🔍 showinfo 🔍

Shows a line containing various information for each input video frame. The input video is not modified. Purely a "let me just look at you" filter. 👀

This filter supports the following options:

checksum — calculate checksums of each plane. By default enabled.
udu_sei_as_ascii — try to print user data unregistered SEI as ascii character when possible, in hex format otherwise.

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, 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 syntax of this option, check the "Video size" section in the ffmpeg-utils manual.
i — the type of interlaced mode: "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. iykyk 💅
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. Ngl this filter is only relevant for pal8 pixel format frames — like, don't come for it if your frames aren't pal8, it literally can't help you.

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. Yes, all three, it's giving multitasking queen.

It accepts the following parameter:

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. Tiny pixels, big chaos.

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 filtering process to get original input. For example, to reverse forward shuffle you need to use same parameters and exact same seed and to set direction to inverse. Write that seed down babes, it's your only way back. 🫶

shuffleplanes 🌈

Reorder and/or duplicate video planes. Plane-on-plane crime, but make it aesthetic.

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. It's like a vibe check for your analog capture, fr fr.

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 and out — stat and 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. It's literally fingerprinting your video, main character forensics energy.

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. 🫶 Heads up though: 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) and that bit gets swapped out for the input number. And if you don't specify a filename at all? Nothing gets written. That's the default, bestie. 💅
format — picks the output format. Your options are:
- binary — uses the specified binary representation (this is the default, no cap).
- xml — uses the specified xml representation. ✨
th_d — sets the threshold for detecting one word as similar. The value has to be an integer greater than zero. Default is 9000. 💖
th_dc — sets the threshold for detecting all words as similar. Again, integer greater than zero required. Default is 60000.
th_xh — sets the threshold for detecting frames as similar. Integer greater than zero, please and thank you. Default is 116.
th_di — sets the minimum length of a sequence, in frames, to recognize it as a matching sequence. Value must be a non negative integer. Default is 0.
th_it — sets the minimum relation that matching frames to all frames must have. 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, exactly 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. 🎀

Ngl bestie, heads up: this is a legacy implementation that corresponds to a superseded recommendation. For the latest version, refer to ITU-T Rec. P.910 (07/22): https://www.itu.int/rec/T-REC-P.910-202207-I/en 💀

It accepts literally one option:

print_summary — if set to 1, summary statistics get printed to the console. Default 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 touching the outlines. It's giving soft-focus main character energy. 👑

It accepts the following options:

luma_radius, lr — sets the luma radius. The value must be a float in the range [0.1,5.0] specifying the variance of the gaussian filter used to blur the image (bigger = slower, sorry 😭). Default is 1.0.
luma_strength, ls — sets the luma strength. Must be a float in the range [-1.0,1.0] and it configures the blurring: a value in [0.0,1.0] blurs the image, while a value in [-1.0,0.0] sharpens it. Default 1.0.
luma_threshold, lt — sets the luma threshold, used as a coefficient to decide whether a pixel should be blurred or not. Must be an 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 0.
chroma_radius, cr — sets the chroma radius. Float in [0.1,5.0], same deal: variance of the gaussian filter used to blur (slower if larger). Default is luma_radius.
chroma_strength, cs — sets the chroma strength. Float in [-1.0,1.0]; [0.0,1.0] blurs, [-1.0,0.0] sharpens. Default is luma_strength.
chroma_threshold, ct — sets the chroma threshold, same coefficient situation as luma. Integer in [-30,30]; 0 filters everything, [0,30] filters flat areas, [-30,0] filters edges. Default is luma_threshold.
alpha_radius, ar — sets the alpha radius. Float in [0.1,5.0], variance of the gaussian filter (slower if larger). Default is luma_radius.
alpha_strength, as — sets the alpha strength. Float in [-1.0,1.0]; [0.0,1.0] blurs, [-1.0,0.0] sharpens. Default is luma_strength.
alpha_threshold, at — sets the alpha threshold, coefficient-style. Integer in [-30,30]; 0 filters everything, [0,30] flat areas, [-30,0] edges. Default is luma_threshold.

Fr fr: if a chroma or alpha option isn't explicitly set, the corresponding luma value gets used instead. 🫶

💜 sobel 💜

Applies the sobel operator to the input video stream. Edge detection slay. 🔥

The filter accepts the following option:

planes — sets which planes will be processed; unprocessed planes get copied. Default value is 0xf, which 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. 👑

🎀 spp 🎀

Applies a simple postprocessing filter that compresses and decompresses the image at several shifts — or, in the case of quality level 6, all shifts — and averages the results.

The filter accepts the following options:

quality — sets quality. This option defines the number of levels for averaging, and it accepts an integer in the range 0-6. If set to 0, the filter has no effect at all. A value of 6 means higher quality. For each increment of that value the speed drops by a factor of approximately 2. Default value is 3. 😭
qp — forces a constant quantization parameter. If it's not set, the filter uses the QP from the video stream (if available).
mode — sets the thresholding mode. Available modes are:
- hard — sets hard thresholding (default).
- soft — sets 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. Using this option may cause flicker, since the B-Frames have often 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. The original repository is at https://github.com/HighVoltageRocknRoll/sr.git. 🫶

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.
model — sets the path to the model file specifying network architecture and its parameters. Note that different backends use different file formats. TensorFlow and OpenVINO backends 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. Scale factor is necessary for the SRCNN model, because it accepts input upscaled using bicubic upscaling with proper scale factor.

Lowkey important: to get full functionality (such as async execution), please use the dnn_processing filter. 💅

🌈 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 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.
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.
fill — specifies whether the output image outside the region of interest, which does not 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. 🦋

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.

both video inputs have to match on resolution and pixel format or this filter just won't work right — that's the vibe check it needs to pass 😤. It also lowkey assumes both inputs have the same number of frames, and it compares them one by one, bestie, one by one.

The filter stores the calculated SSIM of each frame 💾.

The description of the accepted parameters follows, and yes we're doing all of them:

stats_file, f — if you specify 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, iconic behaviour honestly.

The file printed if stats_file is selected contains a sequence of key/value pairs 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, iykyk ✨

Examples

for example, this whole situation right here:

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, no cap.

another example with both psnr and ssim at same time, because we multitask 💅:

ffmpeg -i main.mpg -i ref.mpg -lavfi  "ssim;[0:v][1:v]psnr" -f null -

and 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 🥽.

The filters accept the following options (yes, "filters", that's what it says, plural energy):

in — set the 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 the 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, slay:

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 👀.

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, chef's kiss:

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 ✍️.

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, that's the ✨ pipeline ✨.

The filter accepts the following options:

filename, f — set the filename of the subtitle file to read 📄. It must be specified, no skipping.
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 (not our fault, bestie 😭), 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 — yes this is as cute as it sounds 🎀:

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 🎨.

Useful for enlarging pixel art images without reducing sharpness — the retro gamer's bestie fr fr 👾.

swaprect 🔄

Swap two rectangular objects in video, because sometimes you just gotta move the rectangle 📦.

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, and yes, they're giving math class but make it fun:

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. 💅

`tblend`

Blends successive video frames together — one on top of the other, chef's kiss. ✨

See blend (like, go read that one, it's the whole story).

`telecine`

Applies the telecine process to your video. Yes, the pulldown thing. Vintage cinema vibes, no cap. 🎀

It takes these options:

first_field — which field gets to go first. Values:
- 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.

Some typical patterns, bestie — screenshot this:

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 the input video across time. Like the histogram filter but with main character energy over time. 💖

Unlike histogram, which only shows the histogram of a single input frame at a certain time, this filter also shows past histograms, for the number of frames defined by the width option.

The computed histogram is a representation of the color component distribution in an image.

Options, let's go:

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. Accepts:
- 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 which color components to display. Default is 7.
bgopacity, b — sets the background opacity. Default is 0.9.
envelope, e — show envelope. Default is disabled.
ecolor, ec — sets the envelope color. Default is "gold".
slide — sets the slide mode. Available values are:
- 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 "replace".

`threshold`

Applies the threshold effect to your video stream. 🦋

This filter needs four video streams to do its thresholding thing. First stream is the stream we're filtering. Second stream holds the threshold values, third stream holds the min values, and last — fourth — stream holds the max values.

It takes this option:

planes — sets which planes will be processed; unprocessed planes will be copied. By default the value is 0xf, meaning all planes will be processed.

So like, for example: if a pixel's component value in the first stream is less than the threshold value of that pixel component from the 2nd (threshold) stream, the third stream value gets picked — otherwise the fourth stream's pixel component value gets picked. Sorting hat behaviour, honestly. 🎀

Using the color source filter, you can perform various types of thresholding:

Commands

This filter supports all the 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`

Picks the most representative frame out of a given sequence of consecutive frames. The it-girl of every batch, lowkey. 👑

Options:

n — sets the frames batch size to analyze; in a set of n frames, the filter will pick one of them, and then handle 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".

Okay so heads up: since the filter keeps track of the whole frames sequence, a bigger n value will result in higher memory usage, so a high value is not recommended. Don't say I didn't warn you 💀

Examples

Extract one picture each 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. Grid moment ✨

The untile filter can do the reverse. Reversible queen.

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. The default value is 0, meaning all the 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 (such as having 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. Like, actually necessary. 🫶

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 the tilt-and-shift effect. 🌈

What happens when you invert time and space? Buckle up.

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, 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 is either very little motion and the background is static, or when there is a lot of motion and a very wide depth of field (e.g. wide panorama, while moving on a train). Train window aesthetic, literally iconic. 🚂

It accepts these parameters:

tilt — tilt video while shifting (default). When unset, 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 is 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. Old-school but make it ✨aesthetic✨.

Frames are counted starting from 1, so the first input frame is considered odd. (Not me counting from one 🙈)

Options:

mode — specifies the mode of the interlacing. This option can also be specified as a value alone. See below for a list of values for this option. Available values are:
- 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

                       Output:
                       11111                           33333
                       22222                           44444
                       11111                           33333
                       22222                           44444
                       11111                           33333
                       22222                           44444
                       11111                           33333
                       22222                           44444

drop_even, 1 — bestie, we're only letting the odd frames through, the even ones get yeeted. Result: same height, half the 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 — the flip side, fr. Only even frames survive, odds are gone, 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 — every frame gets stretched to full height, but the in-between lines get padded with black. Double height, same input frame rate, very much a vibe 💅

                        ------> 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 — taking the upper field from odd frames and the lower field from even frames, weaving them together like a friendship bracelet 🎀 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 — same energy but flipped: lower field from odd frames, 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 the frame rate, height unchanged. New frames get inserted, each one carrying the second temporal field from the previous input frame plus the first temporal field from the next input frame. Heads up: this mode relies on the top_field_first flag. It's the move 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 go into the upper field, even frames go into the lower field, and you get double height at the same frame rate. Chef's kiss 🌸

                        ------> 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

Numeric values are deprecated but hey, they're still accepted for backward compatibility reasons 🙈

Default mode is "merge", iykyk.

🌈 `flags`

Specify flags influencing the filter process, bestie. That's literally it.

Available value for flags is:

low_pass_filter, vlpf — turns on linear vertical low-pass filtering in the filter. Ngl you need vertical low-pass filtering when you're creating an interlaced destination from a progressive source that contains high-frequency vertical detail. The filtering reduces interlace 'twitter' and Moire patterning ✨
complex_filter, cvlpf — turns on complex vertical low-pass filtering. This reduces interlace 'twitter' and Moire patterning slightly less, but it better retains detail and the subjective sharpness impression 💅
bypass_il — skips over frames that are already interlaced, 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. No cap.

🎀 `tmedian`

Picks the median pixels from several successive input video frames. Iconic behaviour, honestly.

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, by which all planes are 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. Sliding that slider, love that for us.

Commands

This filter supports all the above options as commands — excluding 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, like, matching exposures from a sequence of video frames. Genuinely obsessed.

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 the filtering. Setting this option to 0 effectively does nothing, which, okay, sure 🥲
planes — sets which planes to process. Default is 15, which is all available planes 💜

⭐ `tmix`

Mixes successive video frames together. Group project energy but it actually works.

A description of the accepted options follows.

frames — the number of successive frames to mix. If unspecified, it defaults to 3.
weights — specify the weight of each input video frame. Each weight is separated by a space. If the number of weights is smaller than the number of frames, the last specified weight gets reused for all the remaining unset weights 🎯
scale — specify scale. If it's set, it gets multiplied with the sum of each weight multiplied with pixel values, giving you 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 a simple temporal convolution:

tmix=frames=3:weights="-1 3 -1"

Similar to the 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` 🌈

Tone map colors from different dynamic ranges. HDR bestie meets SDR bestie and they get along 💖

This filter expects data in single precision floating point, because it genuinely needs to operate on (and can output!) out-of-range values. So yeah, you'll need another filter like zscale to convert the resulting frame back to a format that's actually usable.

Heads up! The tonemapping algorithms implemented here only work on linear light, so the input data should be linearized beforehand (and possibly correctly tagged). No skipping that step, bestie 😭

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, bestie. 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. (yeah, it does nothing here, we been knew)
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.

This option works only if the input frame has a supported color tag. No tag, no desat, periodt.

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 a little cushion for your stream, cute right?

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 giving gymnastics, honestly.

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. And numerical values are deprecated too, and should be dropped in favor of symbolic constants. We're moving on, bestie 💅

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. We're cropping the timeline, not the pixels, iykyk.

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.

Ngl this next bit matters a lot: 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. Bestie please don't forget the setpts 🙈

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. Slay, flexible queen 👑

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 mixing and matching, bestie 🫶

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. Bit math era fr fr 🧮

inplace — do not require 2nd input for processing, instead use alpha plane from input stream.

`unsharp` ⭐

Sharpen or blur the input video. Pick a lane, honestly, and the filter will do whichever you set.

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.

Negative values blur the input video, positive values sharpen it, and a value of 0 disables the effect entirely. Default value is 0.0. Slay.

alpha_msize_x, ax — sets the alpha matrix horizontal size. Gotta be an odd integer between 3 and 23. Default? 5. ✨
alpha_msize_y, ay — sets the alpha matrix vertical size. Also an odd integer between 3 and 23, bestie. Default is 5.
alpha_amount, aa — sets the alpha effect strength. Floating point number, and the reasonable zone is between -1.5 and 1.5.

Negative values blur the input video, positive ones sharpen it, and 0 disables 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'. Like, every single one. 🫶

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 giving reverse-engineering energy and I'm obsessed. 🦋

The frame rate of the output video is the frame rate of the input video multiplied by the number of tiles. Math! We love her.

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. iykyk.

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 an 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. She's thorough, she's slow, she's iconic. 💅

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.

Heads up bestie: this filter is not available in ffmpeg versions between 5.0 and 6.0. Check your version before you come for her. 😭

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. (So yes, quality 8 is a commitment. 🐌)
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

Convert 360 videos between various formats. Literally a whole world of projections, buckle up. 🌈🌍

The filter accepts the following options:

input
output — set 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. (She said what she said. 👑)

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. (Diagonal is the main character, apparently. ✨)

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. (Mark Zuckerberg could never, honestly.)
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. She's quirky like that. 🎀
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, okurrr)

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 — set interpolation method. Note: more complex interpolation methods require much more memory to run. (Your RAM is gonna feel it, fr fr. 💀)

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 value is line. ✅

w, h — sets the output video resolution. Lowkey the default resolution depends on formats, so don't come at it expecting a fixed number 💫
in_stereo, out_stereo — sets the input/output stereo format, bestie:
- 2d — 2D mono
- sbs — side by side
- tb — top bottom
- Default value is 2d for input and output format. That's the vibe out of the box 🎀
yaw, pitch, roll — sets rotation for the output video. Values in degrees, periodt.
rorder — sets the rotation order for the output video. Pick one item for each position:
- y, Y — yaw
- p, P — pitch
- r, R — roll
- Default value is ypr. Iconic little triple 🌸
h_flip, v_flip, d_flip — flips the output video horizontally (swaps left-right)/vertically (swaps up-down)/in-depth (swaps back-forward). Boolean values, no cap.
ih_flip, iv_flip — sets whether the input video is flipped horizontally/vertically. Boolean values 🦋
in_trans — sets whether the input video is transposed. Boolean value, and by default it's disabled.
out_trans — sets whether the output video needs to be transposed. Boolean value, by default disabled.
h_offset, v_offset — sets the output horizontal/vertical off-axis offset. Default is set to 0. Allowed range is from -1 to 1. Stay in your lane or it won't work 💅
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 — resets the rotation of the output video. Boolean value, by default disabled. Main character energy: start fresh ✨

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 subset of above options as commands. So yes, you can yeet these at it live 🫶

vaguedenoiser

Applies a wavelet based denoiser. It's giving "glow-up for grainy footage" ✨

So here's the whole situation: it transforms each frame from the video input into the wavelet domain, using Cohen-Daubechies-Feauveau 9/7 (yes, that's the whole name, I didn't make it up 💀). Then it applies some filtering to the obtained coefficients. It does an inverse wavelet transform after. And because wavelet properties are Like That, it should give you a nice smoothed result, and reduced noise, without blurring picture features. Obsessed honestly.

This filter accepts the following options:

threshold — the filtering strength. The higher it is, the more filtered the video will be. 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.
- soft — all values under the threshold will be zeroed. All values above will be reduced by the threshold.
- garrote — scales or nullifies coefficients — an intermediary situation between (more) soft and (less) hard thresholding.
- Default is garrote. She's the middle child and she ate 🎀
nsteps — number of times the wavelet will decompose the picture. 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 — threshold used is same for all decompositions.
- bayes — threshold used depends also on each decomposition coefficients.
- Default is universal.

varblur

Applies a variable blur filter by using a 2nd video stream to set the blur radius. Heads up: the 2nd stream must have the same dimensions, no exceptions, it's giving strict bestie behavior 😤

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. 🧚

Commands

This filter supports all the above options as commands. Every single one, slay 💅

vectorscope

Displays 2 color component values in the two dimensional graph (which is called a vectorscope). It's the aesthetic plot moment of the whole page 🌈

This filter accepts the following options:

mode, m — set vectorscope mode. It accepts the following values:
- gray, tint — gray values are displayed on graph, higher brightness means more pixels have same component color value on location in graph. This is the default mode.
- color — gray values are displayed on graph. Surrounding pixels values which are not present in video frame are drawn in gradient of 2 color components which are set by option x and y. The 3rd color component is static.
- color2 — actual color components values present in video frame are displayed on graph.
- color3 — similar as color2 but higher frequency of same values x and y on graph increases value of another color component, which is luminance by default values of x and y.
- color4 — actual colors present in video frame are displayed on graph. If two different colors map to same position on graph then color with higher value of component not present in graph is picked.
- color5 — gray values are displayed on graph. Similar to color but with 3rd color component picked from radial gradient.
x — set which color component will be represented on X-axis. Default is 1.
y — set which color component will be represented on Y-axis. Default is 2.
intensity, i — set intensity, used by modes: gray, color, color3 and color5 for increasing brightness of color component which represents frequency of (X, Y) location in graph.
envelope, e —
- none — no envelope, this is default.
- instant — instant envelope, even the darkest single pixel will be clearly highlighted. Not one pixel left behind 💖
- peak — hold maximum and minimum values presented in graph over time. This way you can still spot out of range values without constantly looking at vectorscope.
- peak+instant — peak and instant envelope combined together. Both. At once. Periodt.
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 white point.
- black — draw graticule for black point.
- name — draw color points short names.
bgopacity, b — set background opacity.
lthreshold, l — set low threshold for color component not represented on X or Y axis. Values lower than this value will be ignored. Default is 0. Note this value is multiplied with actual max possible value one pixel component can have. So for 8-bit input and low threshold value of 0.1 actual threshold is 0.1 * 255 = 25. Math is mathing 🧮
hthreshold, h — set high threshold for color component not represented on X or Y axis. Values higher than this value will be ignored. Default is 1. Note this value is multiplied with actual max possible value one pixel component can have. So for 8-bit input and high threshold value of 0.9 actual threshold is 0.9 * 255 = 230.
colorspace, c — set what kind of colorspace to use when drawing graticule:
- auto
- 601
- 709
- Default is auto.
tint0, t0 and tint1, t1 — set color tint for gray/tint vectorscope mode. By default both options are zero. This means no tint, and output will remain gray. Respect the neutral era 🩶

vidstabdetect

Analyzes video stabilization/deshaking. It performs pass 1 of 2 — see vidstabtransform for pass 2, because this filter is the setup girlie and that one is the payoff 💅

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 flag, no filter — that's the whole deal 🚨

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. She's thorough 🔥
stepsize — set stepsize of the search process. The region around 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. If set to 0, it is disabled. The frames are counted starting from 1. (Yes, from 1, not 0 — don't come for me 😭)
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). Efficiency queen 👑

Examples 🎬

Use default values:

vidstabdetect

Analyze 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:

...and the code block from that last example wraps up with just this one lonely line:

vidstabdetect=show=1

Want to analyze a video with medium shakiness using ffmpeg? Say less:

ffmpeg -i input -vf vidstabdetect=shakiness=5:show=1 dummy.avi

🌟 vidstabtransform 🌟

Video stabilization/deshaking: this is pass 2 of 2, bestie — if you skipped pass 1, go look at vidstabdetect first, iykyk.

So what it does: it reads a file that has transform information for each frame, then applies/compensates those transforms. Team it up with vidstabdetect and boom, you've got yourself a deshaker. See also <http://public.hronopik.de/vid.stab>. And important: you also gotta use the unsharp filter, see below, fr fr.

To enable compilation of this filter you need to configure FFmpeg with --enable-libvidstab. No flag, no filter. 💅

Options

input — Set the path to the file it reads the transforms from. Default value is transforms.trf.
smoothing — Set 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 that 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 it limits the acceleration of the camera (pan/tilt movements). 0 is a special case where a static camera is simulated.
optalgo — Set the camera path optimization algorithm. Accepted values are:
- gauss — gaussian kernel low-pass filter on camera motion (default)
- avg — averaging on transformations
maxshift — Set the maximal number of pixels to translate frames. Default value is -1, meaning no limit.
maxangle — Set the maximal angle in radians (degree*PI/180) to rotate frames. Default value is -1, meaning no limit.
crop — Specify 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 — Set percentage to zoom. A positive value will result in a zoom-in effect, a negative value in a zoom-out effect. Default value is 0 (no zoom).
optzoom — Set 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 — Set 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 — Specify 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. 👑

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. That's it, that's the filter. ✨

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 the end it will output the number of frames detected as having variable delta pts, and the ones with constant delta pts. If there were frames with variable delta, then it will also show min, max and average delta encountered. Receipts on receipts. 🧾

🌈 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

This filter supports all the above options as commands. 🫶

👁️ 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. Aesthetic. 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 a 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 — picks the output color primaries, bestie. The accepted values (and what they actually are) 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
- 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. Iconic little one-liner:

vpp_amf=-2:720:format=yuv420p

Upscale to 4K and change color profile to bt2020. Glow-up time 🦋:

vpp_amf=4096:2160:color_profile=bt2020

🌸 vstack 🌸

Stacks input videos vertically, like, literally one on top of the other.

Every stream has to have the same pixel format and the same width, no cap.

Fun fact: this filter is faster than doing the same output with overlay and pad, so use it, bestie, it's giving efficiency 💅

The filter accepts the following options:

inputs — sets the number of input streams. Default is 2.
shortest — if set to 1, the output is forced to terminate when the shortest input terminates. Default value is 0.

🫶 w3fdif 🫶

Deinterlaces the input video (w3fdif stands for "Weston 3 Field Deinterlacing Filter" — yes, that's a whole mouthful).

It's 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. BBC really said "we'll handle the fields" 💀

This filter 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 w3fdif filter — that's the fix, iykyk.

There are two sets of filter coefficients, the so-called "simple" and "complex". Which set gets used can be set by passing an optional parameter:

filter — sets the interlacing filter coefficients. Accepts one of:
- 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:
- 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:
- 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 — specifies which frames to deinterlace. Accepts one of:
- all — deinterlace all frames,
- interlaced — only deinterlace frames marked as interlaced.

Default value is all.

Commands 🎀

This filter supports the same commands as options. Slay.

🌈 waveform 🌈

Video waveform monitor. It plots color component intensity — by default luma only. Each column of the waveform corresponds to a column of pixels in the source video. It's giving "data but make it art" ✨

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 — sets 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 — sets 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 — sets 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 — displays a separate graph for the color components side by side in "row" mode or one below the other in "column" mode.
- parade — displays a 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 — sets which color components to display. Default is 1, which means only luma or red color component if input is in RGB colorspace. If it's 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.
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 — sets 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 — sets graticule opacity.
flags, fl — sets graticule flags.
- numbers — draw numbers above lines. By default enabled.
- dots — draw dots instead of lines.
scale, s — sets scale used for displaying graticule.
- digital
- millivolts
- ire

Default is digital.

bgopacity, b — sets 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 — sets 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 — sets 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 joins each two sequential fields into a single frame, producing a new double height clip with half the frame rate and half the frame count. Two fields becoming one frame? That's called commitment 💖

The doubleweave works the same as weave but without halving frame rate and frame count.

It accepts the following option:

first_field — sets 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 💜

Applies 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 girlies stay winning 🎨

It accepts the following option:

n — sets the scaling dimension: 2 for "2xBR", 3 for "3xBR" and 4 for "4xBR". Default is 3.

🦋 xcorrelate 🦋

Applies normalized cross-correlation between the first and second input video stream.

Second input video stream dimensions must be lower than first input video stream. That's the rule, bestie, don't fight it.

The filter accepts the following options:

planes — sets which planes to process.
secondary — sets which secondary video frames will be processed from the second input video stream, can be first or all. Default is all.

The xcorrelate filter also supports the framesync options. Fancy.

🔥 xfade 🔥

Applies cross fade from one input video stream to another input video stream. The cross fade is applied for the specified duration. It's giving dramatic transition, main character energy ✨

Both inputs must be constant frame-rate and have the same resolution, pixel format, frame rate and timebase. No exceptions.

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

Default transition effect is fade. 💅

duration — sets the cross fade duration in seconds. Range is 0 to 60 seconds. Default duration is 1 second. ⏱️
offset — sets the cross fade start relative to the first input stream, in seconds. Default offset is 0.
expr — sets the expression for the 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 — currently processed plane.
A — return value of first input at current location and plane.
B — return value of second input at 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 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 second input.

Examples

Cross fade from one input video to another input video, with fade transition and duration of transition of 2 seconds starting at offset of 5 seconds:

ffmpeg -i first.mp4 -i second.mp4 -filter_complex xfade=transition=fade:duration=2:offset=5 output.mp4

xmedian

Pick median pixels from several input videos. 🎯 like, the group chat votes and the middle opinion wins.

The filter accepts the following options:

inputs — set number of inputs. Default is 3. Allowed range is from 3 to 255. If number of inputs is even number, than result will be mean value between two median values.
planes — set which planes to filter. Default value is 15, by which all planes are processed.
percentile — set median percentile. Default value is 0.5. Default value of 0.5 will pick always median values, while 0 will pick minimum values, and 1 maximum values.

Commands

This filter supports all above options as commands, excluding option "inputs".

xpsnr

Obtain 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. 👑

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

Stack video inputs into custom layout. 🧱 like a collage but make it cinema.

All streams must be of same pixel format.

The filter accepts the following options:

inputs — set number of input streams. Default is 2.
layout — specify layout of inputs. This option requires the desired layout configuration to be explicitly set by the user. This sets position of each video input in 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 video input from which to take width or height. Multiple values can be used when separated by '+'. In such case 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.

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 — specify 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 "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, force the output to terminate when the shortest input terminates. Default value is 0.
fill — if set to 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 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 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 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 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

ngl bestie, quick heads up before we keep it moving:

⚠️ Note that if inputs are of different sizes, gaps or overlaps may occur. like, it's giving "we tried" — not a crash, just a lil visual mess. 🙈

`yadif` 🦋

Deinterlace the input video (and yes, yadif means "yet another deinterlacing filter" — the Unix world has a whole thing about naming stuff "yet another," iykyk 💅).

It accepts the following parameters:

mode — the interlacing mode to adopt, she takes one of these values:
- 0, send_frame — output one frame for each frame. one-to-one, no drama ✨
- 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.
The default value is send_frame. periodt.
parity — the picture field parity assumed for the input interlaced video. it 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.
The default value is auto. and lowkey important: 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 these:
- 0, all — deinterlace all frames. every single one, no skip 💀
- 1, interlaced — only deinterlace frames marked as interlaced.
The default value is all.

`yaepblur` 🌸

Apply blur filter while preserving edges (yaepblur means "yet another edge preserving blur filter" — see, the naming thing again 😭). The algorithm is 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 bestie, actual cited paper, we're academic now 🎓.

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. slay.

`zoompan` 🔍

Apply Zoom & Pan effect. main character energy fr.

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 the effect will last for a 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. (not me getting a NAN, embarrassing 🙈)
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. (compile-time flag, so don't come for me if it's not there 💀)

The zscale filter forces the output display aspect ratio to be the same as the input, by changing the output sample aspect ratio. iconic behaviour honestly.

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.

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.
- 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. (so it just, like, stays. no crash, no fuss 🫶)

🧊 CUDA VIDEO FILTERS 🧊

To enable CUDA and/or NPP filters please refer to configuration guidelines for CUDA and for CUDA NPP filters.

Running CUDA filters requires you to initialize a hardware device and to pass that device to all filters in any filter graph. no shortcuts bestie, the whole graph needs the device 💅

-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

Since CUDA filters operate exclusively on GPU memory, frame data must sometimes be uploaded (<code>hwupload</code>) to hardware surfaces associated with the appropriate CUDA device before processing, and downloaded (<code>hwdownload</code>) back to normal memory afterward, if required. Whether <code>hwupload</code> or <code>hwdownload</code> is necessary depends on the specific workflow:

If the input frames are already in GPU memory (e.g., when using -hwaccel cuda or -hwaccel_output_format cuda), explicit use of hwupload is not needed, as the data is already in the appropriate memory space.
If the input frames are in CPU memory (e.g., software-decoded frames or frames processed by CPU-based filters), it is necessary to use hwupload to transfer the data to GPU memory for CUDA processing.
If the output of the CUDA filters needs to be further processed by software-based filters or saved in a format not supported by GPU-based encoders, hwdownload is required to transfer the data back to CPU memory.

Note that hwupload uploads data to a surface with the same layout as the software frame, so it may be necessary to add a format filter immediately before hwupload to ensure the input is in the correct format. Similarly, hwdownload may not support all output formats, so an additional format filter may need to be inserted immediately after hwdownload in the filter graph to ensure compatibility.

🌈 CUDA 🌈

Below is a description of the currently available Nvidia CUDA video filters.

Prerequisites:

Install Nvidia CUDA Toolkit

Note: If FFmpeg detects the Nvidia CUDA Toolkit during configuration, it will enable CUDA filters automatically without requiring any additional flags. If you want to explicitly enable them, use the following options:

Configure FFmpeg with --enable-cuda-nvcc --enable-nonfree.
Configure FFmpeg with --enable-cuda-llvm. Additional requirement: llvm lib must be installed.

💖 bilateral_cuda 💖

CUDA accelerated bilateral filter, an edge preserving filter. This filter is mathematically accurate thanks to the use of GPU acceleration. For best output quality, use one to one chroma subsampling, i.e. yuv444p format.

The filter accepts the following options:

sigmaS — Set sigma of gaussian function to calculate spatial weight, also called sigma space. Allowed range is 0.1 to 512. Default is 0.1.
sigmaR — Set sigma of gaussian function to calculate color range weight, also called sigma color. Allowed range is 0.1 to 512. Default is 0.1.
window_size — Set window size of the bilateral function to determine the number of neighbours to loop on. If the number entered is even, one will be added automatically. 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 that it can work 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 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.

It is by no means feature complete compared to the software colorspace filter, and at the current time only supports color range conversion between jpeg/full and mpeg/limited range.

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. The underlying input pixel formats have to match.

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:

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; 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.
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 to 720p, aspect ratio preserved, and make sure the output is yuv420p.

scale_cuda=-2:720:format=yuv420p

Upscale to 4K with the nearest neighbour algorithm.

scale_cuda=4096:2160:interp_algo=nearest

Do zero conversion, zero scaling — just copy every input frame into a freshly allocated one. This is lowkey a lifesaver when a filter + 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. Bestie, it literally just looks at a batch and goes "this one. this is the one." 👑

It accepts these options:

n — the frames batch size to analyze. Out of a set of n frames the filter picks one, then handles the next batch of n frames, and so on till the end. Default is 100.

Heads up: since the filter keeps track of the whole frames sequence, a bigger n means higher memory usage — so a high value is not recommended. Don't be greedy. 💀

Example

Thumbnails get extracted from every n=150-frame batch, one pick per batch. Then the chosen frames are 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 using the yadif algorithm — except it's implemented in CUDA, so it can chill as part of a GPU accelerated pipeline with nvdec and/or nvenc. Same brain, GPU body. 🫶

It accepts these parameters:

mode — the interlacing mode to adopt. Takes one of these values:
- 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.

The default value is send_frame.

parity — the picture field parity assumed for the input interlaced video. Takes one of these 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 info, top field first will be assumed.

deint — specify which frames to deinterlace. Accepts one of these:
- 0, all — deinterlace all frames.
- 1, interlaced — only deinterlace frames marked as interlaced.

The default value is all.

🌈 CUDA NPP 🌈

Below is a description of the currently available NVIDIA Performance Primitives (libnpp) video filters. Fancy library, fancy filters. 💅

Prerequisites:

<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 the exact same way as it does for the scale filter. 🎀

The following additional options are accepted:

format — the pixel format of the output CUDA frames. If you set it to the string same (which is the default), the input format gets 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 — enables decreasing or increasing the output video width or height when 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 genuinely useful case for this: 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. Say device A allows 1280x720 playback and your video is 1920x800 — use this option (set to decrease) and pass 1280x720 on the command line, and the output becomes 1280x533. Slay. ✨

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 work. Don't skip that step, bestie. 🙈

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 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. 😬

It can be handy if you need a video to fit within or exceed a defined resolution using force_original_aspect_ratio, but you've also got 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 these values:
- init — only evaluate expressions once during filter initialization, or when a command is processed.
- frame — evaluate expressions for each incoming frame.

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.
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 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. The main character is the reference, honestly. 💅

See the scale_npp filter for available options — scale2ref_npp supports the same ones, but it uses the reference video instead of the main input as its basis. scale2ref_npp also supports the following additional constants for the w and h options:

main_w, main_h — the main input video's width and height.
main_a — the 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 this 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) to match 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 perform image sharpening with border control. Crispy edges, no cap.

The following additional options are accepted:

border_type — the type of sampling to be 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, see the transpose video filter — it shares mostly the same options.

It accepts these 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.
passthrough — do not apply the transposition if the input geometry matches the one specified by the given 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).

🧚 OPENCL VIDEO FILTERS 🧚

Below is a description of the currently available OpenCL video filters. GPU girlies, this one's for you. 💜

To enable compilation of these filters you need to configure FFmpeg with --enable-opencl.

Running OpenCL 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. 🫶

-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 avgblur_opencl 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 bestie, OpenCL lore drop 🫶: these filters literally can’t reach frame data sitting in normal memory, so all of it needs to be uploaded (hwupload) to hardware surfaces connected to the right device before it’s used, then downloaded (hwdownload) back to normal memory. Note that hwupload uploads to a surface with the same layout as the software frame, so you may need to add a format filter immediately before to get the input into the right format. And hwdownload does not support all formats on output — you may need to insert an extra format filter immediately after it in the graph to get output in a supported format. 💅

avgblur_opencl 🎀

Apply average blur filter.

The filter accepts these options:

sizeX — horizontal radius size. Range is [1, 1024], default is 1.
planes — which planes to filter. Default is 0xf, which means all planes get processed.
sizeY — vertical radius size. Range is [1, 1024], default is 0. If zero, the sizeX value is 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 these parameters:

luma_radius, lr
luma_power, lp
chroma_radius, cr
chroma_power, cp
alpha_radius, ar
alpha_power, ap

Okay, description time:

luma_radius, lr; chroma_radius, cr; alpha_radius, ar — set an expression for the box radius in pixels used to blur 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 these 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 radii 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 these 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 pixel 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 these 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 these 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:

1 2 3
4 x 5
6 7 8

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 minus 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 these 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:

threshold0, threshold1, threshold2, threshold3 — the cap on how much each plane is allowed to change 👑 Range is [0, 65535] and the default is 65535. And if you set it to 0, that plane literally just stays exactly as it was, untouched, no notes 💅
coordinates — a flag saying which pixels we're allowed to peek at 👀 Range is [0, 255], default 255 — which means all eight neighbors get used. Each flag points at a spot in the local 3x3 neighborhood centered on x:

Example ✨

Apply the dilation filter with threshold0 set to 30, threshold1 set to 40, threshold2 set to 50, and coordinates set to 231 — which sets every output pixel to the local maximum among pixels 1, 2, 3, 6, 7, 8 of the 3x3 region centered on it in the input 🎯 And if the difference between the input pixel and that local max is more than the threshold of the corresponding plane, the output pixel gets set to input pixel + that plane's threshold:

-i INPUT -vf "hwupload, dilation_opencl=30:40:50:coordinates=231, hwdownload" OUTPUT

`nlmeans_opencl` 🧚

Non-local Means denoise filter, but through OpenCL 🫶 This filter accepts the same options as nlmeans.

`overlay_opencl` 🎀

Overlay one video on top of another, bestie. 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 though: this filter requires the same memory layout for all the inputs, so you might need a format conversion first ngl.

The filter accepts the following options:

x — sets the x coordinate of the overlaid video on the main video. Default value is 0.
y — sets 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 the same memory layout for color channels, and the overlay has an 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` 🦋

Adds padding to the input image and drops the original input at the x, y coordinates you give it. It's giving "scrapbook page with a cute margin" energy honestly 💜

It accepts the following options:

width, w and 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 gets used for the output instead. Also: the width expression can reference the value set by the height expression, and vice versa — they're mutual besties like that 🫶 The default value of width and height is 0.
x and y — specify the offsets for placing the input image inside the padded area, relative to the top/left border of the output image. The x expression can reference the value set by the y expression, and vice versa. Default value of x and y is 0. And if x or y evaluate to a negative number, they get changed so the input image is centered on the padded area 🎯
color — 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 of 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 — same as in_w and in_h.
out_w, out_h — the output width and height (i.e. the size of the padded area), as specified by the width and height expressions.
ow, oh — 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, which is the same as (iw / ih) * sar.

`prewitt_opencl` 🔥

Applies the Prewitt operator (https://en.wikipedia.org/wiki/Prewitt_operator) to the input video stream. Edge detection slay 💅

The filter accepts the following options:

planes — set which planes to filter. Default value is 0xf, which means all planes get processed.
scale — set the value that gets multiplied with the filtered result. Range is [0.0, 65535] and default value is 1.0.
delta — set the value that gets added to the 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. This one's the main character fr — you bring the code, it runs it 🧚

source — OpenCL program source file.
kernel — kernel name in the 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 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 every pixel to be generated. So the global ID offset for each work-item is literally the coordinates of a pixel in the destination image. iykyk.

The kernel function needs to take the following arguments:

Destination image, __write_only image2d_t. This image becomes the output, and the kernel should write all of it.
Frame index, unsigned int. This is a counter starting from zero and going up 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 that increases with the index counter. Pixel values get linearly interpolated by the sampler, and the output doesn't need to 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 how much of each input gets 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 the 2nd: Xmap and 3rd: Ymap input video stream. Yes, that's the numbering — 2nd and 3rd 🫶

So the destination pixel at position (X, Y) gets picked from the source (x, y) position where x = Xmap(X, Y) and y = Ymap(X, Y). And if the mapping values are out of range, a zero value for that pixel gets used for the destination pixel 💀

Xmap and Ymap input video streams must be of the same dimensions, and the output video stream will have Xmap/Ymap video stream dimensions. Also, Xmap and Ymap input video streams are 32bit float pixel format, single channel 📌

interp — specify the interpolation used for remapping 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` 💜

Applies the Roberts cross operator (https://en.wikipedia.org/wiki/Roberts_cross) to the input video stream. More edge detection, we love to see it ✨

The filter accepts the following options:

planes — set which planes to filter. Default value is 0xf, which means all planes get processed.
scale — set the value that gets multiplied with the filtered result. Range is [0.0, 65535] and default value is 1.0.
delta — set the value that gets added to the 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`

Okay so this one literally just applies the Sobel operator (you know, the edge-detection icon — <https://en.wikipedia.org/wiki/Sobel_operator>) to your input video stream. Edge vibes only. ✨

It takes exactly one option:

planes — tells it which planes to filter. Default is 0xf, which means all the planes get processed. slay.
scale — a value that gets multiplied with the filtered result. Range is [0.0, 65535] and the default is 1.0.
delta — a value that gets added to the filtered result. Range is [-65535, 65535] and the default is 0.0.

🎀 Example

Apply the sobel operator with scale set to 2 and delta set to 10:

-i INPUT -vf "hwupload, sobel_opencl=scale=2:delta=10, hwdownload" OUTPUT

🌈 `tonemap_opencl`

This one does HDR (PQ/HLG) → SDR conversion with tone-mapping. It's giving "make your HDR look right on a normal screen" realness. 💅

It accepts these parameters:

tonemap — specify which tone-mapping operator to use. Same as the tonemap option in tonemap.
param — tunes the tone mapping algorithm. Same as the param option in tonemap.
desat — applies desaturation for highlights that go past this level of brightness. The higher you set it, the more color info gets preserved. It helps stop super-highlights from looking unnaturally blown out by (smoothly) turning them into white instead. Result: images feel more natural, but you lose info about those out-of-range colors. The default value is 0.5, and ngl the algorithm here is a little different from the cpu version tonemap has right now. Setting it to 0.0 disables this option entirely.
threshold — the tonemapping algorithm params get fine-tuned per scene, and this threshold is what detects whether the scene actually changed or not. If the distance between the current frame's average brightness and the current running average exceeds the threshold value, it re-calculates the scene average and peak brightness. Default is 0.2.
format — specify the output pixel format. Currently supported formats are:
- p010
- nv12
range, r — set the output color range. Possible values are:
- tv/mpeg
- pc/jpeg

Default is same as input.

primaries, p — set the output color primaries. Possible values are:
- bt709
- bt2020

Default is same as input.

transfer, t — set the output transfer characteristics. Possible values are:
- bt709
- bt2020

Default is bt709.

matrix, m — set the output colorspace matrix. Possible value are:
- bt709
- bt2020

Default is same as input.

lowkey the original says "Possible value are" for matrix — that's just how it's written, don't come for me 🫶

🎀 Example

Convert HDR (PQ/HLG) video to bt2020-transfer-characteristic p010 format using the linear operator:

-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. Two moods, one filter, iconic. ✨

It accepts these parameters:

luma_msize_x, lx — set the luma matrix horizontal size. Range is [1, 23] and the default is 5.
luma_msize_y, ly — set the luma matrix vertical size. Range is [1, 23] and the default is 5.
luma_amount, la — set the luma effect strength. Range is [-10, 10] and the default is 1.0. Negative values blur the input video, positive values sharpen it, and a value of zero disables the effect completely.
chroma_msize_x, cx — set the chroma matrix horizontal size. Range is [1, 23] and the default is 5.
chroma_msize_y, cy — set the chroma matrix vertical size. Range is [1, 23] and the default is 5.
chroma_amount, ca — set the chroma effect strength. Range is [-10, 10] and the default is 0.0. Negative values blur the input video, positive values sharpen it, zero disables the effect.

Every single parameter is optional and defaults to the equivalent of the string '5:5:1.0:5:5:0.0'. iykyk. 💜

🎀 Examples

Apply a strong luma sharpen effect:

-i INPUT -vf "hwupload, unsharp_opencl=luma_msize_x=7:luma_msize_y=7:luma_amount=2.5, hwdownload" OUTPUT

Apply a strong blur of both luma and chroma parameters:

-i INPUT -vf "hwupload, unsharp_opencl=7:7:-2:7:7:-2, hwdownload" OUTPUT

⭐ `xfade_opencl`

Cross fade two videos together with a custom transition effect, using OpenCL. Main character transition energy fr fr. 🔥

It accepts these options:

transition — set 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 — OpenCL program source file for the custom transition.
kernel — set the name of the kernel to use for the custom transition from the program source file.
duration — set duration of the video transition.
offset — set the time of the start of the transition relative to the first video.

Okay so here's how the custom thing works 👑: the program source file must contain a kernel function with the given name, and it 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. periodt.

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 may read from them to generate the output, but they can't be written to.
Transition progress, float. This value is always between 0 and 1 inclusive.

🎀 Example programs

Apply the dots curtain transition effect:

__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);
}

💖 and that's the whole custom-kernel situation — copy that shape and you're golden.

💜 VAAPI VIDEO FILTERS 💜

Okay so VAAPI video filters are normally besties with the VAAPI decoder and the VAAPI encoder — like they literally show up as a trio. Here's the whole rundown of the VAAPI video filters, no gatekeeping.

To even compile these filters you gotta configure FFmpeg with --enable-vaapi. That's the vibe check for the build, fr.

And to actually use vaapi filters you need to set up the vaapi device correctly — no shortcuts, bestie. For more info, please read https://trac.ffmpeg.org/wiki/Hardware/VAAPI.

overlay_vaapi

Overlay one video on the top of another. Classic main character + supporting character situation.

It takes two inputs and has one output. The first input is the "main" video on which the second input is overlaid. 💅

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 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.

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, it's backwards from your gut feeling, iykyk.)
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 — that's the whole guest list rn. 🎀

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

Notice the backslash-space escapes in that display value — copy it exactly, don't get cute with it. 🙈

hstack_vaapi

Stack input videos horizontally. Side-by-side slay.

This is the VA-API variant of the hstack filter — each input stream may have different height, and this filter will scale down/up each input stream while keeping the original aspect. So nobody gets stretched, we love to see it. ✨

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, but stacked up.

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 custom layout. This is the "I make the rules" one, highkey iconic.

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 border, honestly. 🎀

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. 💅 (built-in centering, we stan)
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 time, bestie. 🖍️

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

That last one has negative offsets and math in it — literally just paste it as-is and let the expressions cook. 🔥

🍎 VIDEOTOOLBOX VIDEO FILTERS 🍎

Okay bestie, so these are the VideoToolbox video filters, and they're basically the Apple-silicon-era glow-up of video filtering. ✨

The VideoToolbox filter depends on the VideoToolbox framework, and it's auto detected when building ffmpeg for an Apple platform such as macOS. If autodetect is disabled, just add --enable-videotoolbox to configure. Easy. 💅

✨ `scale_vt` ✨

Scales and converts the color parameters using VTPixelTransferSession. It's giving "resize but make it Apple-native." 🦋

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 touch it, it just matches what came in. Iconic lazy-girl behavior, honestly.) 💖
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

Like, the way that command does HDR→SDR and halves the size in one -vf? Chef's kiss. 🔥

🌸 `transpose_vt` 🌸

Transposes rows with columns in the input video, and optionally flips it too. For more in-depth examples, go check out the transpose video filter — it shares mostly the same options. (Sibling energy fr fr. 👯)

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).

Lowkey important: passthrough only skips the transposition when the geometry matches — it's not out here skipping filters for fun. Read the value, bestie. 💀

The way dir defaults to cclock_flip is such a quirky main-character choice, but periodt, that's the spec. 🎀

💅 VULKAN VIDEO FILTERS

Okay bestie, gather round — here's the rundown on all the Vulkan video filters we currently have, and yes they are exactly as extra as they sound. ✨

First things first: to even compile these filters you gotta configure FFmpeg with --enable-vulkan and either --enable-libglslang or --enable-libshaderc. No shortcuts, no cap. 🎀

And running Vulkan filters? It literally requires you to initialize a hardware device and then pass that device to all filters in any filter graph. Like, all of them. Every single one. It's giving "everybody in the group chat needs to be on the same page." 🫶

-init_hw_device vulkan[=name][:device[,key=value...]] — Initialise a new hardware device of type vulkan called name, using the given device parameters and options in key=value. The following options are supported:
- debug — Switches validation layers on if set to 1.
- linear_images — Allocates linear images. Does not apply to decoding.
- disable_multiplane — Disables multiplane images. Does not apply to decoding.
- avoid_host_import — Avoids the use of dynamic host memory imports, in favor of a regular memcpy() into a previously mapped buffer.
-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 choosing the first device and running nlmeans_vulkan filter with default parameters on it.

-init_hw_device vulkan=vk:0 -filter_hw_device vk -i INPUT -vf "hwupload,nlmeans_vulkan,hwdownload" OUTPUT

Okay so here's the tea ☕ — Vulkan filters are not able to access frame data in normal memory. So all frame data needs to be uploaded (hwupload) to hardware surfaces connected to the appropriate device before being used, and then downloaded (hwdownload) back to normal memory. Note that hwupload will upload to a frame 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 is usually necessary to insert an additional format filter immediately following in the graph to get the output in a supported format. 🙈

🌸 avgblur_vulkan

Apply an average blur filter, implemented on the GPU using Vulkan.

The filter accepts the following options:

sizeX — Set horizontal radius size. Range is [1, 32] and default value is 3.
sizeY — Set vertical radius size. Range is [1, 32] and default value is 3.
planes — Set which planes to filter. Default value is 0xf, by which all planes are processed.

🦋 blend_vulkan

Blend two Vulkan frames into each other.

The blend filter takes two input streams and outputs one stream, the first input is the "top" layer and second input is "bottom" layer. By default, the output terminates when the longest input terminates.

A description of the accepted options follows.

c0_mode
c1_mode
c2_mode
c3_mode
all_mode

Set blend mode for specific pixel component or all pixel components in case of all_mode. Default value is normal.

Available values for component modes are:

normal
multiply

🎀 bwdif_vulkan

Deinterlacer using bwdif, the "Bob Weaver Deinterlacing Filter" algorithm, implemented on the GPU using Vulkan.

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.

🌈 chromaber_vulkan

Apply an effect that emulates chromatic aberration. Works best with RGB inputs, but provides a similar effect with YCbCr inputs too. ✨

dist_x — Horizontal displacement multiplier. Each chroma pixel's position will be multiplied by this amount, starting from the center of the image. Default is 0.
dist_y — Similarly, this sets the vertical displacement multiplier. Default is 0.

⭐ color_vulkan

Video source that creates a Vulkan frame of a solid color. Useful for benchmarking, or overlaying. 👑

It accepts the following parameters:

color — The color to use. Either a name, or a hexadecimal value. The 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 allows the autodetected value to be overridden as well as allows forcing a specific value used 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).

🧚 vflip_vulkan

Flips an image vertically. That's it, that's the whole vibe. 🪞

🧚 hflip_vulkan

Flips an image horizontally.

🧚 flip_vulkan

Flips an image along both the vertical and horizontal axis. Double flip, double slay. 💅

💜 gblur_vulkan

Apply Gaussian blur filter on Vulkan frames.

The filter accepts the following options:

sigma — Set horizontal sigma, standard deviation of Gaussian blur. Default is 0.5.
sigmaV — Set vertical sigma, if negative it will be same as sigma. Default is -1.
planes — Set which planes to filter. By default all planes are filtered.
size — Set the kernel size along the horizontal axis. Default is 19.
sizeV — Set the kernel size along the vertical axis. Default is 0, which sets to use the same value as size.

🔥 nlmeans_vulkan

Denoise frames using Non-Local Means algorithm, implemented on the GPU using Vulkan. Supports more pixel formats than nlmeans or nlmeans_opencl, including alpha channel support. Iconic behaviour honestly. 👑

The filter accepts the following options.

s — Set denoising strength for all components. Default is 1.0. Must be in range [0.0, 100.0].
p — Set patch size for all planes. Default is 7. Must be odd number in range [0, 99].
r — Set research size. Default is 15. Must be odd number in range [0, 99].
t — Set parallelism. Default is 8. Must be a number in the range [1, 64]. Larger values will use more VRAM but may not result in greater speed. The optimal value is hardware and input dependent.
s0, s1, s2, s3 — Set denoising strength for a specific component. Default is 1.0, equal to s. Must be in range [0.0, 100.0]. 0.0 disables denoising in that component.
p0, p1, p2, p3 — Set patch size for a specific component. Default is 7, equal to p. Must be odd number in range [0, 99].

🍓 overlay_vulkan

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 all inputs to use the same pixel format. 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.

🎀 transpose_vulkan

Transpose rows with columns in the input video and optionally flip it. For more in depth examples see the transpose video filter, which shares mostly the same options.

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.

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).

💅 QSV VIDEO FILTERS 💅

Okay so here’s the tea on the currently available QSV video filters, bestie. ✨

To enable compilation of these filters you need to configure FFmpeg with --enable-libmfx or --enable-libvpl.

To use QSV filters, you need to setup the QSV device correctly. For more information, please read https://trac.ffmpeg.org/wiki/Hardware/QuickSync.

🎀 hstack_qsv 🎀

Stack input videos horizontally. It’s the QSV variant of the hstack filter, and each input stream may have a different height — this filter will scale down/up each input stream while keeping the original aspect. 💖

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_qsv 🦋

Stack input videos vertically. This is the QSV variant of the vstack filter, and each input stream may have a 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_qsv 🌈

Stack video inputs into a custom layout. 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 permits the user to supply output size for each input stream.

xstack_qsv=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.

🌟 VIDEO SOURCES 🌟

Below is a description of the currently available video sources. Let’s go, bestie. 🫶

💖 buffer 💖

Buffer video 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:

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.
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). 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 this syntax is deprecated:

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. ✨

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.

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.
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.

🌈 ddagrab 🌈

Captures the Windows Desktop via Desktop Duplication API.

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.

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.

okay so we're still in ddagrab land — here are the rest of the options, no cap 💅

framerate — Maximum framerate at which the desktop will be captured. The interval between successive frames will not be smaller than the inverse of the framerate. When dup_frames is true (the default) and the desktop is not being updated often enough, the filter will duplicate a previous frame. Note: there is no background buffering going on, so when the filter is not polled often enough, the actual inter-frame interval may be significantly larger.

Defaults to 30 FPS.

video_size — Specify the size of the captured video.

Defaults to the full size of the screen.

Cropped from the bottom/right if smaller than screen size.

offset_x — Horizontal offset of the captured video.
offset_y — Vertical offset of the captured video.
output_fmt — Desired filter output format. Defaults to 8 Bit BGRA.

It accepts the following values:

auto — Passes all supported output formats to DDA and returns what DDA decides to use.
8bit, bgra — 8 Bit formats always work, and DDA will convert to them if necessary.
10bit, x2bgr10 — Filter initialization will fail if 10 bit format is requested but unavailable.

dup_frames — When this option is set to true (the default), the filter will duplicate frames when the desktop has not been updated in order to maintain approximately constant target framerate. When this option is set to false, the filter will wait for the desktop to be updated (inter-frame intervals may vary significantly in this case).

✨ Examples ✨

Capture 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. 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, this can be achieved by specifying a smaller size and its offsets into the screen:

ddagrab=video_size=800x600:offset_x=100:offset_y=100

🎀 gfxcapture 🎀

Capture windows or monitors using Windows.Graphics.Capture API.

This source provides low overhead capture of application windows or entire monitors. The filter outputs hardware frames in "d3d11" format; use "hwdownload,format=" if system memory frames are required.

The window to be captured can be selected 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 must match all provided expressions to be selected. The first matching window will be picked, in whatever order Windows returns them.

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. It returns frames at whatever rate the compositor provides them, only capped by the 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 — As window_title, but matched against the window class name.
window_exe — 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 decorations/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. (yes, the manpage says left here too — we don't make the rules 😭)
crop_right — Crop this many pixels from the left side of the captured frames. (same energy 💀)
crop_bottom — Crop this many pixels from the left side of the captured frames. (the copy-paste is real 🙈)
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.

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. Default for each color 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. (yes, the original literally lists y0 twice, bestie 💀)
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. (yes, “the some above” — we're keeping it ✨)

🦋 mandelbrot 🦋

Generate a Mandelbrot set fractal, and progressively zoom towards the point specified with start_x and start_y.

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 so this one generates a bunch of test patterns — literally the same ones the MPlayer test filter makes ✨. Main character energy for your encoding experiments.

The size of the generated video is fixed, and it's 512x512 — no, you can't change it, bestie, that's the whole point. This source is useful in particular for testing encoding features 🎀.

It accepts these options:

rate, r — the frame rate of the sourced video, meaning the number of frames generated per second 💅. Gotta 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 value is "25" ✨.
duration, d — sets 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 😭 (no cap, forever-forever).
test, t — sets the number or the name of the test to perform. Supported tests are:
- dc_luma
- dc_chroma
- freq_luma
- freq_chroma
- amp_luma
- amp_chroma
- cbp
- mv
- ring1
- ring2
- all
- max_frames, m — set the maximum number of frames generated for each test, default value is 30 👑

Default value is "all", which will cycle through the list of all tests 🌈.

Some examples, because we love a demo:

mptestsrc=t=dc_luma

will generate a "dc_luma" test pattern 💖.

🧚 frei0r_src

Provides a frei0r source 🌸.

To enable compilation of this filter you need to install the frei0r header and configure FFmpeg with --enable-frei0r 🛠️.

This source accepts the following parameters:

size — the size of the video to generate. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual 📏.
framerate — the framerate of the generated video. It may be a string of the form num/den or a frame rate abbreviation ⏱️.
filter_name — the name of the frei0r source to load. For more information regarding 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 which is overlaid on the overlay filter main input:

frei0r_src=size=200x200:framerate=10:filter_name=partik0l:filter_params=1234 [overlay]; [in][overlay] overlay

🦋 life

Generates a life pattern 👑 — yes, that life.

This source is based on a generalization of John Conway's life game 🧫.

The sourced input represents a life grid, each pixel represents a cell which can be in one of two possible states: alive or dead 💀. Every cell interacts with its eight neighbours, which are the cells that are horizontally, vertically, or diagonally adjacent 🌐.

At each interaction the grid evolves according to the adopted rule, which specifies the number of neighbor alive cells which will make a cell stay alive or born 🐣. The rule option allows one to specify the rule to adopt.

This source accepts the following options:

filename, f — set the file from which to read the initial grid state 📄. In the file, each non-whitespace character is considered an alive cell, and newline is used to delimit the end of each row. If this option is not specified, the initial grid is generated randomly 🎲.
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 random grid. It's a floating point number value ranging from 0 to 1, defaults to 1/PHI 🌗. It is ignored when a file is specified 🙈.
random_seed, seed — set the seed for filling the initial random grid, 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 🎰.
rule — set the life rule 📜.

A rule can be specified 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 to become alive (i.e. to "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 are used to encode the next cell state if it is alive for each number of neighbor alive cells, the low order bits specify the rule for "borning" new cells. Higher order bits encode for an 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, and will keep a cell alive if it has 2 or 3 neighbor alive cells, and will born a new cell if there are three alive cells around a dead cell 👑.

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 is specified, the size is set by default to the same size of 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 is 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 also. Defaults to 1 🧵.
mold — set cell mold speed. If set, a dead cell will go from death_color to mold_color with a step of mold 🍄. mold can have a value from 0 to 255.
life_color — set the color of living (or new born) cells 💚.
death_color — set the color of dead cells. If mold is set, this is the first color used to represent a dead cell 🖤.
mold_color — set 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 🌊. This can be used to generate patterns with continuity in space and time, e.g. to simulate smoke, fluids, or terrain 🔥.

In case more than one octave is specified through the octaves option, Perlin noise is generated as a sum of components, each one with doubled frequency 🎵. In this case the persistence option specify the ratio of the amplitude with respect to the previous component. More octave components enable to specify more high frequency details in the generated noise (e.g. small size variations due to boulders in a generated terrain) 🪨.

Options

size, s — 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. Default value is "320x240" 🎀.
rate, r — specify the frame rate expected for the video stream, expressed as a number of frames per second. Default value is 25 ⏱️.
octaves — specify the total number of components making up the noise, each one with doubled frequency. Default value is 1 🎶.
persistence — set 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. This can be useful to define 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. This can be useful to change the time variation speed. Default value is 1 🕰️.
random_mode — set random mode used to compute 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 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 🛠️.

The QR code is 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, but an empty colored output is returned instead 🙈.

This source accepts the following options:

qrcode_width, q, 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 only once, when initializing the source 💅. See the qrencode Expressions section for details 🔍.

okay so quick heads-up bestie: some of the constants are lowkey missing for the source (like x or t or ¸n), because they only make sense when you’re evaluating the expression for each frame instead of at initialization time. no cap, that’s just the vibe 💅

rate, r — sets the frame rate of the sourced video, aka how many frames get generated per second. it has to be a string in the format frame_rate_num/frame_rate_den, an integer, a floating point number, or a valid video frame rate abbreviation. default is "25" ✨
case_sensitive, cs — tells libqrencode to use case sensitive encoding. this is enabled by default. you can disable it to reduce the QR encoding size 🎀
level, l — sets the QR encoding error correction level. higher correction level = bigger encoding size, but the code is more robust to corruption. lower level is L. it accepts these values:
- L
- M
- Q
- H
expansion — picks how the input text gets expanded. can be either "none" or "normal" (default). see the qrencode Text expansion section for the details 🦋
text
textfile — define the text to be rendered. if neither is specified, no QR is encoded (just an empty colored frame, iconic but empty 💀). 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
foreground_color, fc — set the QR code and background color. default for foreground_color is "black", 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`

okay so this is the source squad, and they all have main character energy in different ways 👑

the "allrgb" source returns frames of size 4096x4096 of all rgb colors. literally all of them.

the "allyuv" source returns frames of size 4096x4096 of all yuv colors. yuv girlies rise 🫶

the "color" source provides an uniformly colored input. simple, clean, slay.

the "colorchart" source provides a colors checker chart 🎨

the "colorspectrum" source provides a color spectrum input 🌈

the "haldclutsrc" source provides an identity Hald CLUT. see also haldclut filter.

the "nullsrc" source returns unprocessed video frames. it’s mainly useful in analysis / debugging tools, or as the source for filters which ignore the input data. lowkey the “no thoughts head empty” source 💀

the "pal75bars" source generates a color bars pattern, based on EBU PAL recommendations with 75% color levels.

the "pal100bars" source generates a color bars pattern, based on EBU PAL recommendations with 100% color levels.

the "rgbtestsrc" source 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. if you don’t, that’s a vibe check fail 😭

the "smptebars" source generates a color bars pattern, based on the SMPTE Engineering Guideline EG 1-1990.

the "smptehdbars" source generates a color bars pattern, based on the SMPTE RP 219-2002.

the "testsrc" source generates a test video pattern, showing a color pattern, a scrolling gradient and a timestamp. this is mainly intended for testing purposes.

the "testsrc2" source is similar to testsrc, but supports more pixel formats instead of just "rgb24". this allows using it as an input for other tests without requiring a format conversion. glow-up fr fr ✨

the "yuvtestsrc" source generates an YUV test pattern. you should see a y, cb and cr stripe from top to bottom.

the sources accept the following parameters:

level — sets 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 — sets 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 — sets the size of the sourced video. for the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. default value is "320x240". this option is not available with the "allrgb", "allyuv", and "haldclutsrc" filters 🚫
rate, r — sets 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 value is "25".
duration, d — sets 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 — sets the sample aspect ratio of the sourced video.
alpha — sets 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 — sets the number of decimals to show in the timestamp, only available in the "testsrc" source. the displayed timestamp value will correspond to the original timestamp value multiplied by the power of 10 of the specified value. default value is 0.
type — sets the type of the color spectrum, only available in the "colorspectrum" source. can be one of the following:
- black
- white
- all
patch_size — sets patch size of single color patch, only available in the "colorchart" source. default is "64x64".
preset — sets colorchecker colors preset, only available in the "colorchart" source. available values are:
- reference
- skintones

default value 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 — sets the color of the created image. accepts the same syntax of the corresponding color option.

🦋 `openclsrc`

generate video using an OpenCL program. nerdy but make it fashion 💅

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 value 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. iconic fractal behavior ✨

__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`

generates a Sierpinski carpet/triangle fractal, and randomly pans around. art but make it math 🎨

this source accepts the following options:

size, s — sets frame size. for the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. default value is "640x480".
rate, r — sets frame rate, expressed as number of frames per second. default value is "25".
seed — sets seed which is used for random panning.
jump — sets max jump for single pan destination. allowed range is from 1 to 10000.
type — sets fractal type, can be default "carpet" or "triangle".

💜 `zoneplate`

generates a zoneplate test video pattern. very test, very video, very zoneplate 🌀

this source accepts the following options:

size, s — sets frame size. for the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. default value is "320x240".
rate, r — sets the frame rate, i.e. how many frames per second. the default is "25" ⭐
duration, d — how long the sourced video lasts. the accepted syntax is in the Time duration section of the ffmpeg-utils(1) manual, iykyk. and if you don't specify it — or the duration you give is negative — the video is supposed to be generated forever 😭💀
sar — sets the sample aspect ratio of the sourced video. cute, simple.
precision — sets the precision in bits for the look-up table used for sine calculations. default is 10, allowed range is 4 to 16 ✨
xo — horizontal axis offset for the output signal. default 0.
yo — vertical axis offset for the output signal. default 0.
to — time axis offset for the output signal. default 0.
k0 — the 0-order constant added to the signal phase. default 0.
kx — the 1-order phase factor multiplier for the horizontal axis. default 0.
ky — the 1-order phase factor multiplier for the vertical axis. default 0.
kt — the 1-order phase factor multiplier for the time axis. default 0.
kxt, kyt, kxy — phase factor multipliers for the combination of spatial and temporal axis. default 0.
kx2 — the 2-order phase factor multiplier for the horizontal axis. default 0.
ky2 — the 2-order phase factor multiplier for the vertical axis. default 0.
kt2 — the 2-order phase factor multiplier for the time axis. default 0.
ku — the constant added to the final phase to produce the chroma-blue component of the signal. default 0 💙
kv — the constant added to the final phase to produce the chroma-red component of the signal. default 0 ❤️

🎮 Commands

this source also supports some of the options above as commands. periodt.

✨ 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 on the currently available video sinks. the end-of-the-pipeline girlies 🫶

buffersink

buffers video frames and makes them available to the end of the filter graph. it's giving "the final boss of the graph".

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 AVBufferSinkContext structure — that's the thing that defines the incoming buffers' formats — which gets passed as the opaque parameter to "avfilter_init_filter" for initialization. no cap.

nullsink

null video sink: does absolutely nothing with the input video 💀 it's mainly useful as a template, and for use in analysis / debugging tools. it's the "i'm just here to vibe" of sinks.

🌈 MULTIMEDIA FILTERS

below is the description of the currently available multimedia filters. these ones do the most, fr fr 👑

a3dscope

converts input audio to 3d scope video output. it's giving audio but make it cinematic 🎀

the filter accepts the following options:

rate, r — sets the frame rate, expressed as number of frames per second. default value is "25".
size, s — specifies 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 "hd720" ✨
fov — sets the camera field of view. default is 90 degrees. 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 displayed audio waves in number of frames.

🎮 Commands

the filter supports some of the options above as commands. slay.

abitscope

converts input audio to a video output, displaying the audio bit scope. little visual of your bits, basically 💅

the filter accepts the following options:

rate, r — sets the frame rate, expressed as number of frames per second. default value is "25".
size, s — specifies 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 "1024x256".
colors — specify a list of colors separated by a space or by '|', which will be used to draw the channels. unrecognized or missing colors get replaced by white. so if your color choices are a mystery, congrats, it's white now 🙈
mode, m — sets the output mode. can be "bars" or "trace". default is "bars".

adrawgraph

draws a graph using input audio metadata. ✏️ See drawgraph.

agraphmonitor

See graphmonitor. that's it, that's the whole vibe check ✨

ahistogram

converts input audio to a video output, displaying the volume histogram. it's the "look at my range" filter 📊

the filter accepts the following options:

dmode — specifies 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, expressed as number of frames per second. default value is "25".
size, s — specifies 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 "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. setting this to -1 accumulates all frames 😤
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" — that's the mean phase of the current audio frame. a video output can also be produced and is enabled by default. the audio is passed through as the first output. main character behavior: shows up, does its job, leaves your audio intact 🫶

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. no in-between erasure 💅

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, blue contrast. default values are 2, 7 and 1. allowed range is "[0, 255]".
mpc — sets the color used for drawing the median phase. if the color is "none", which is the default, no median phase value will be drawn.
video — enables 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 set. it's literally the group chat notifying you that something's off 😭

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. allowed range is "[0, 1]".
angle, a — sets the angle threshold for out of phase detection, in degree. default is 170. allowed range is "[90, 180]".
duration, d — sets the mono or out of phase duration until notification, expressed in seconds. default is 2.

✨ Examples

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 input audio to a video output, representing the audio vector scope. 🦋

the filter is used to measure the difference between channels of a stereo audio stream. a monaural signal — identical left and right signal — results in a straight vertical line. any stereo separation is visible as a deviation from this line, creating a Lissajous figure. and if that straight line (or a deviation from it) shows up but horizontal, that indicates the left and right channels are out of phase. so like, the vibe check is entirely geometric 👑

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 — 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. 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. allowed range is "[0, 255]".
zoom — sets the zoom factor. default value is 1. allowed range is "[0, 10]". values lower than 1 will auto adjust the zoom factor to the maximal possible value. it's giving "i'll fix it for you bestie" 🫶
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 — specifies the amplitude scale of 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.
- none — no mirror.
x — mirror only the x axis 🪞
y — mirror only the y axis 🪞
xy — mirror both axes, go off king 👑

🌸 Examples

Complete example using ffplay — yes bestie, this is the whole thing, run it and watch the pretty circles:

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 all the options above as commands — every single one — except size and rate. Those two are built different and don't take orders mid-stream, sorry. 💅

⏱️ bench, abench

Lowkey a stopwatch for your filtergraph. It benchmarks part of a filtergraph, no cap.

The filter accepts the following options:

action — start or stop a timer. That's the whole vibe check.

Available values are:

start — grab the current time, stash it as frame metadata (using the key lavfi.bench.start_time), and forward that frame on to the next filter. 🫶
stop — grab the current time and fetch the lavfi.bench.start_time metadata off the input frame's metadata to work out the time difference. Then it prints the time difference, average, maximum and minimum time (respectively t, avg, max and min). The timestamps are expressed in seconds. ⏲️

✨ Examples

Benchmark the selectivecolor filter — wrap it in a bench sandwich, iconic:

bench=start,selectivecolor=reds=-.2 .12 -.49,bench=stop

🔗 concat

Concatenate audio and video streams, joining them together one after the other. It's giving "playlist but make it filters." 🎶

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. No mixing and matching, bestie — keep it consistent. ✨

The filter accepts the following options:

n — set the number of segments. Default is 2.
v — set the number of output video streams, which is also the number of video streams in each segment. Default is 1.
a — set the number of output audio streams, which 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. ⚠️ (it's called unsafe for a reason, bestie, proceed with caution)

The filter has v+a outputs: first v video outputs, then a audio outputs. 📺🔊

There are nx(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. Order matters, always. 💅

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. 🤫 (silence is golden, and also a padding strategy)

For this filter to work correctly, all segments must start at timestamp 0. Not 0.1, not "roughly zero" — zero. 🎯

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. That part's on you, bestie. 🫡

Different frame rates are acceptable but will result in variable frame rate at output; be sure to configure the output file to handle it. It's giving chaos but the manageable kind. 💀

🌈 Examples

Concatenate an opening, an episode and an ending, all in bilingual version (video in stream 0, audio in streams 1 and 2) — the full trilogy arc, obsessed:

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 — multitasking queen behavior:

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]

Heads up bestie: a desync will happen at the stitch if the audio and video streams do not have exactly the same duration in the first file. Not a maybe. A will. 😭

🎀 Commands

This filter supports the following commands:

next — close the current segment and step to the next one. On to the next era, periodt. ✨

🔊 ebur128

EBU R128 scanner filter. It takes an audio stream and analyzes its loudness level — like a vibe check but for volume. 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). 📊

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. Don't forget your conversion bestie, or you'll get something sus at the end. 🔄

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 information about the Loudness Recommendation EBU R128 on <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. Pick your fighter. 🥊
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 s 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 — visually stunning, honestly:

ffplay -f lavfi -i "amovie=input.mp3,ebur128=video=1:meter=18 [out0][out1]"

Run an analysis with ffmpeg — the no-notes version:

ffmpeg -nostats -i input.mp3 -filter_complex ebur128 -f null -

🧵 interleave, ainterleave

Temporally interleave frames from several inputs. Take turns bestie, we're sharing. 🤝

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. No time travel allowed. ⏰

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. Not me waiting forever for one input to text back. 😭

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. That's the whole mood.

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. Fair's fair. 💅

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 is 💫

Examples ✨

Interleave frames that belong to different streams, using ffmpeg:

ffmpeg -i bambi.avi -i pr0n.mkv -filter_complex "[0:v][1:v] interleave" out.avi

Add a flickering blur effect, bestie:

select='if(gt(random(0), 0.2), 1, 2)':n=2 [tmp], boxblur=2:2, [tmp] interleave

latency, alatency ⏱️

Measures how much latency the filtering has, okay? 🫶

It reports the previous filter's filtering latency — the delay 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 previously running filter in the filtergraph. cute, right? 💅

metadata, ametadata 🏷️

Manipulates frame metadata. Obsessed, honestly.

This filter accepts the following options:

mode — sets the mode of operation of the filter 🌸

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 that key in its metadata.
add — adds new metadata key and value. If the key is already there, it does nothing.
modify — modifies the value of a key that's already present.
delete — if value is set, it deletes only keys that have such value. Otherwise, it deletes the key. If key isn't set, it deletes all metadata values in the frame 💀
print — prints the key and its value if metadata was found. If key isn't set, it prints all metadata values available in the frame.

key — sets the key used with all modes. It must be set for all modes except print and delete.
value — sets 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 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 — sets 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 you can specify any writable url. The filename ``-'' is shorthand for standard output. If the file option isn't set, output goes to the log with 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 the 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 🔐

Sets read/write permissions for the output frames.

These filters are mainly aimed at developers, to test the direct path in the following filter in the filtergraph. iykyk 🧚

The filters accept the following options:

mode — selects the permissions mode.

It accepts the following values:

none — does nothing. This is the default.
ro — sets all the output frames read-only.
rw — sets all the output frames directly writable.
toggle — makes the frame read-only if writable, and writable if read-only.
random — sets each output frame read-only or writable randomly.

seed — sets the seed for the random mode, and must be an integer included between 0 and UINT32_MAX. If it's not specified, or if it's 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.

realtime, arealtime ⏳

Slows down filtering to match real time approximately.

These filters pause the filtering for a variable amount of time to match the output rate with the input timestamps. They're similar to the re option to ffmpeg.

They accept the following options:

limit — time limit for the pauses. Any pause longer than that gets considered a timestamp discontinuity and resets 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 values will slow processing down. The limit is automatically adapted accordingly. Default is 1.0.

A processing speed faster than what's possible without these filters cannot be achieved. no cap 🫶

Commands 💜

Both filters support all the above options as commands.

segment, asegment ✂️

Splits a single input stream into multiple streams.

This filter does the opposite of the concat filters.

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 runs from the beginning of the input stream. The last segment runs until the end of the input stream.
frames, samples — exact frame/sample count to split the segments.

In all cases, prefixing each segment with '+' makes 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 👑

Selects frames to pass in the output.

This filter accepts the following options:

expr, e — sets the expression, which is evaluated for each input frame.

If the expression evaluates to zero, the frame is discarded.

If the evaluation result is negative or NaN, the frame is sent to the first output; otherwise it's sent to the output with index ceil(val)-1, assuming 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. math ate 💅

outputs, n — sets the number of outputs. The output to which the selected frame gets sent 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 isn't NAN. It remains NAN if not found.
start_t — the first PTS, in seconds, in the stream which isn't 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 isn't 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's possible to skip frames generated by the concat demuxer which aren't 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. 🎀

Okay so the concatdec_select variable 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 it’s NaN if the start_time metadata is missing.

That basically means an input frame gets selected if its pts is inside the interval set by the concat demuxer. Bestie, that’s the whole vibe check. 🦋

iw (video only) — the width of the input video frame.
ih (video only) — the height of the input video frame.
view (video only) — View ID for multi-view video.

The default value of the select expression is "1".

✨ Examples ✨

Select all frames in input:

select

The example above is the same as:

select=1

Skip all frames:

select=0

Select only I-frames:

select='eq(pict_type\,I)'

Select one frame every 100:

select='not(mod(n\,100))'

Select only frames contained in the 10-20 time interval:

select=between(t\,10\,20)

Select only I-frames contained in the 10-20 time interval:

select=between(t\,10\,20)*eq(pict_type\,I)

Select frames with a minimum distance of 10 seconds:

select='isnan(prev_selected_t)+gte(t-prev_selected_t\,10)'

Use aselect to select only audio frames with samples number > 100:

aselect='gt(samples_n\,100)'

Create a mosaic of the first scenes:

ffmpeg -i video.avi -vf select='gt(scene\,0.4)',scale=160:120,tile -frames:v 1 preview.png

Comparing scene against a value between 0.3 and 0.5 is generally a sane choice.

Send even and odd frames to separate outputs, and compose them:

select=n=2:e='mod(n, 2)+1' [odd][even]; [odd] pad=h=2*ih [tmp]; [tmp][even] overlay=y=h

Select useful frames from an ffconcat file which is 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 💌

Send commands to filters in the filtergraph.

These filters read commands that get sent to other filters in the filtergraph.

sendcmd has to be inserted between two video filters, and asendcmd has to be inserted between two audio filters — but apart from that, they act the same way. Periodt. 💅

You can provide the command specification in the filter arguments with the commands option, or in a file specified by the filename option.

These filters accept the following options:

commands, c — set the commands to be read and sent to the other filters.
filename, f — set the filename of the commands to be read and sent to the other filters.

Commands syntax

A commands description is a sequence of interval specifications, and it includes a list of commands to execute when a particular event related to that interval happens. The event is typically the current frame time entering or leaving a given time interval. 🎀

An interval is specified by this syntax:

<START>[-<END>] <COMMANDS>;

The time interval is specified by the START and END times. END is optional and defaults to the maximum time.

The current frame time counts as within the specified interval if it’s included in the interval [START, END) — that is, when the time is greater than or equal to START and lesser than END.

COMMANDS is a sequence of one or more command specifications, separated by ,, relating to that interval. The syntax of a command specification is:

[<FLAGS>] <TARGET> <COMMAND> <ARG>

FLAGS is optional and specifies the type of events relating to the time interval that enable sending the specified command. It must be a non-null sequence of identifier flags separated by + or | and enclosed between [ and ]. ✨

The following flags are recognized:

enter — the command is sent when the current frame timestamp enters the specified interval. In other words, it’s sent when the previous frame timestamp was not in the given interval, and the current one is.
leave — the command is sent when the current frame timestamp leaves the specified interval. In other words, it’s sent when the previous frame timestamp was in the given interval, and the current one is not.
expr — the command ARG is interpreted as an expression and the result of the expression is passed as ARG.

The expression is evaluated through the eval API and can contain the following constants:

POS — original position in the file of the frame, or undefined if 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 is not 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 specification and another, whitespaces, or sequences of characters starting with # until the end of line, are ignored and can be used to annotate comments.

A simplified BNF description of the commands specification syntax follows:

<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 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 with:

sendcmd=f=test.cmd,drawtext=fontfile=FreeSerif.ttf:text='',hue

⏱️ setpts, asetpts ⏱️

Change the PTS (presentation timestamp) of the input frames.

setpts works on video frames, and asetpts works on audio frames. Slay, but make it timestamps. 💅

This filter accepts the following options:

expr — the expression evaluated for each frame to construct its timestamp.
strip_fps (video only) — a boolean option that determines if the original framerate and frame duration metadata is unset. If set to true, be advised that a sane frame rate should be explicitly specified if output is sent to a constant frame rate muxer. Default is "false".

The expression is evaluated through the eval API and can contain the following 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 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 command was applied or time of the first frame if no commands.

Examples

Start counting PTS from zero:

setpts=PTS-STARTPTS

Apply fast motion effect:

setpts=0.5*PTS

Apply slow motion effect:

setpts=2.0*PTS

Set 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 fixed rate 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 onto the current timebase:

setpts='(RTCTIME - RTCSTART) / (TB * 1000000)'

Generate timestamps by counting samples:

asetpts=N/SR/TB

Commands

Both filters support all above options as commands. Iconic. 👑

🌈 setrange 🌈

Force color range for the output video frame.

The setrange filter marks the color range property for the output frames. It does not change the input frame, but only sets the corresponding property, which affects how the frame is treated by following filters.

The filter accepts the following options:

`range`

okay so this one's a lil picker for the color range 💅 here's what you can hand it:

auto — keeps the exact same color range property it already had. no notes, it just vibes.
unspecified, unknown — marks the color range as unspecified.
limited, tv, mpeg — sets the color range to limited.
full, pc, jpeg — sets the color range to full.

✨ `settb`, `asettb` ✨

This one sets the timebase used for the output frames' timestamps. Honestly? It's mostly here for testing timebase config — like a little lab rat for your timestamps 🧪

It takes these parameters:

expr, tb — the expression that gets evaluated into the output timebase.

So the value you give for tb is an arithmetic expression that represents a rational number. That expression can use the constants "AVTB" (the default timebase, bestie 💖), "intb" (the input timebase) and "sr" (the sample rate — audio only, ikyk). The 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` 🎀

This is the iconic one, no cap 👑 — it 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. And yes, the manpage itself is here to humble it: the transform itself is not really constant Q — the Q factor is actually variable/clamped 🙈. It comes with a musical tone scale, from E0 to D#10.

The filter accepts these options:

size, s — specifies 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 — sets the output frame rate. Default value is 25.
bar_h — sets the bargraph height. It must be even. Default value is -1, which computes the bargraph height automatically ✨
axis_h — sets the axis height. It must be even. Default value is -1, which computes the axis height automatically.
sono_h — sets the sonogram height. It must be even. Default value is -1, which computes the sonogram height automatically.
fullhd — sets the fullhd resolution. ⚠️ This option is deprecated, use size, s instead. Default value is 1.
sono_v, volume — specifies the sonogram volume expression. It can contain these 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 these 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 — specifies the bargraph volume expression. It can contain these 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 these 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 — specifies the sonogram gamma. Lower gamma = more contrast in the spectrum, higher gamma = the spectrum gets more range. Default value is 3. Acceptable range is [1, 7].
bar_g, gamma2 — specifies the bargraph gamma. Default value is 1. Acceptable range is [1, 7].
bar_t — specifies the bargraph transparency level. Lower value makes the bargraph sharper. Default value is 1. Acceptable range is [0, 1].
timeclamp, tc — specifies the transform timeclamp. At low frequency there's a trade-off between accuracy in the time domain vs the frequency domain 🫶 If timeclamp is lower, events in the time domain get represented more accurately (think a fast bass drum); otherwise events in the frequency domain get represented more accurately (think a bass guitar). Acceptable range is [0.002, 1]. Default value is 0.17.
attack — sets the attack time in seconds. The default is 0 (disabled). Otherwise it limits future samples by applying asymmetric windowing in the time domain — super useful when you need low latency ⚡ Accepted range is [0, 1].
basefreq — specifies the transform base frequency. Default value is 20.01523126408007475, which is the frequency 50 cents below E0. Acceptable range is [10, 100000].
endfreq — specifies the transform end frequency. Default value is 20495.59681441799654, which is the frequency 50 cents above D#10. Acceptable range is [10, 100000].
coeffclamp — ⚠️ this option is deprecated and ignored. it's just sitting there. leave it alone 💀
tlength — specifies the transform length in the time domain. Use this to control the accuracy trade-off between time domain and frequency domain at every frequency sample. It can contain these 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 — specifies the transform count for every video frame. Default value is 6. Acceptable range is [1, 30].
fcount — specifies the transform count for every single pixel. Default value is 0, which makes it computed automatically. Acceptable range is [0, 10].
fontfile — specifies the font file for use with freetype to draw the axis. If it's not specified, it uses the embedded font. 📌 Note that drawing with a font file or the embedded font is not implemented with custom basefreq and endfreq — use the axisfile option instead.
font — specifies a fontconfig pattern. This has lower priority than fontfile. The ":" in the pattern may be replaced by "|" to avoid unnecessary escaping.
fontcolor — specifies the font color expression. This is an arithmetic expression that should return an integer value 0xRRGGBB. It can contain these variables:
- frequency, freq, f — the frequency where it is evaluated
- timeclamp, tc — the value of the timeclamp option

and these functions:

midi(f) — midi number of frequency f. Some midi numbers for you: 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 — specifies an image file to draw the axis. This option overrides the fontfile and fontcolor options 💥
axis, text — enable/disable drawing text to the axis. If it's set to 0, drawing to the axis is disabled, ignoring the fontfile and axisfile options. Default value is 1.
csp — set colorspace. 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 — sets the spectrogram color scheme. This is a list of floating point values with the 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 the 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 the 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 an image file:

axisfile=myaxis.png:basefreq=40:endfreq=10000

🦋 `showcwt` 🦋

Converts input audio into a video output that represents the frequency spectrum using a Continuous Wavelet Transform and Morlet wavelet 🌊 cute AND mathy.

The filter accepts these options:

size, s — specifies 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 — sets the output frame rate. Default value is 25.
scale — sets the frequency scale used. Allowed values are:
- linear
- log
- bark
- mel
- erbs
- sqrt
- cbrt
- qdrt
- fm

Default value is "linear".

iscale — sets the intensity scale used. Allowed values are:
- linear
- log
- sqrt
- cbrt
- qdrt

Default value is "log".

min — sets the minimum frequency that will be used in output. Default is 20 Hz.
max — sets the maximum frequency that will be used in output. Default is 20000 Hz. Heads up bestie: the real frequency upper limit depends on the input audio's sample rate, and that limit will be enforced on this value when it's set to something greater than the Nyquist frequency 🚨
imin — sets the minimum intensity that will be used in output.
imax — sets the maximum intensity that will be used in output.
logb — sets the logarithmic basis for how brightness strength gets mapped when we translate calculated magnitude values into pixel values, bestie. Allowed range is 0 to 1. Default value is 0.0001. 💅
deviation — sets the frequency deviation. Values lower than 1 are more frequency oriented, values higher than 1 are more time oriented (iykyk). Allowed range is 0 to 10. Default value is 1.
pps — sets the number of pixel output per each second in one row. Allowed range is from 1 to 1024. Default value is 64.
mode — sets the output visual mode. 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 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 🌈

Converts input audio to video output representing the audio power spectrum. Audio amplitude is on the Y-axis while frequency is on the X-axis. It's giving science fair but make it cute 🧚

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. She picks for you, queen 👑
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 okay 💀
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 to a video output, representing the spatial relationship between two channels. Just the two of them, main character energy fr fr.

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 to a video output, representing the audio frequency spectrum. Obsessed with her, no cap 💖

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 to a single video frame, representing the audio frequency spectrum. One frame, one moment, immaculate vibes ✨

The filter accepts the following options:

size, s — sets the video size for the output 💖 for the syntax of this one, peep the "Video size" section in the ffmpeg-utils manual. Default is 4096x2048 (yes, big girl size).
mode — the display mode, bestie. Here's what it takes:
- combined — all channels displayed in the same row
- separate — all channels displayed in separate rows

Default is combined.

color — picks the display color mode. The options are genuinely giving a whole palette:
- channel — each channel gets its own separate color
- intensity — each channel uses the same color scheme
- rainbow — rainbow color scheme
- moreland — moreland color scheme
- nebulae — nebulae color scheme
- fire — fire color scheme
- fiery — fiery color scheme
- fruit — fruit color scheme
- cool — cool color scheme
- magma — magma color scheme
- green — green color scheme
- viridis — viridis color scheme
- plasma — plasma color scheme
- cividis — cividis color scheme
- terrain — terrain color scheme

Default is intensity.

scale — the scale used for calculating intensity color values. takes:
- lin — linear
- sqrt — square root, default
- cbrt — cubic root
- log — logarithmic
- 4thrt — 4th root
- 5thrt — 5th root

Default is log. (yeah the page says both, that's literally what it says, don't come for me 🫶)

fscale — sets the frequency scale. takes:
- lin — linear
- log — logarithmic

Default is lin.

saturation — sets the saturation modifier for displayed colors 🌈 negative values give you an alternative color scheme, 0 is no saturation at all. Must be in the [-10.0, 10.0] range. Default is 1.
win_func — sets the window function. The full roster, no skips:
- 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.

orientation — sets the orientation of the time vs frequency axis. Can be vertical or horizontal. Default is vertical.
gain — scale gain for calculating intensity color values. Default is 1.
legend — draws time and frequency axes and legends. Enabled by default ✨
rotation — sets color rotation, must be in the [-1.0, 1.0] range. Default 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 pixel format output with an alpha component 💅

Examples

Extract an audio spectrogram of a whole audio track in a 1024x1024 picture using ffmpeg:

ffmpeg -i audio.flac -lavfi showspectrumpic=s=1024x1024 spectrogram.png

`showvolume` 🍓

Turns input audio volume into a video output. Iconic behavior honestly.

The filter takes these 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. That expression can use these variables:
- VOLUME — current max volume of channel in dB
- PEAK — current peak
- CHANNEL — current channel number, starting from 0
t — if set, displays channel names. Enabled by default.
v — if set, displays volume values. Enabled by default.
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 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 when the dm option is set to > 0. default is: orange

`showwaves` 🌊

Turns input audio into a video output representing the sample waves. So cute.

The filter takes these options:

size, s — sets the video size for the output. For the syntax, check the "Video size" section in the ffmpeg-utils manual. Default is 600x240.
mode — sets the display mode. Available values:
- point — draw a point for each sample
- line — draw a vertical line for each sample
- p2p — draw a point for each sample and a line between them
- cline — draw a centered vertical line for each sample

Default 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. 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 is 25.
split_channels — sets if channels should be drawn separately or overlap. Default is 0.
colors — sets colors separated by '|' which are going to be used for drawing of 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 — scale pixel values for each drawn sample
- full — draw every sample directly

Default 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` 🦋

Turns input audio into a single video frame, representing the sample waves. One frame, all the vibes.

The filter takes these options:

size, s — sets the video size for the output. For the syntax, check the "Video size" section in the ffmpeg-utils manual. Default is 600x240.
split_channels — sets if channels should be drawn separately or overlap. Default is 0.
colors — sets colors separated by '|' which are going to be used for drawing of 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 — scale pixel values for each drawn sample
- full — draw every sample directly

Default is scale.

filter — sets the filter mode. Available values:
- average — use average samples values for each drawn sample
- peak — use peak samples values for each drawn sample

Default is average.

Examples

Extract a channel split representation of the wave form of a whole audio track in 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. Chaotic but useful.

This filter takes these options:

mode — sets the mode of operation of the filter. Can be one of these:
- select — select every frame with side data of type
- delete — delete side data of type. If type is not set, delete 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 frequency domain as displayed in videos back to time domain as presented in audio output. Literally reverse-engineering the visual into sound, we love a glow-up 💜

This filter is literally built to reverse processed showspectrum filter outputs — but bestie, it can also synthesize sound from other spectrograms too. ✨ Just know that in that case the results are gonna be poor if the phase data isn't available, because then the phase data has to be recreated, and ngl it's usually just recreated from random noise. 😭

For best results:

use gray only output (that's "channel" color mode in the showspectrum filter)
use "log" scale for magnitude video
use "lin" scale for phase video
to produce phase, for the 2nd video, use the "data" option
input videos should generally use "fullframe" slide mode, since that saves the resources needed for decoding video 🫶

The filter accepts the following options:

sample_rate — set the sample rate of the output audio. Fr fr, the sample rate of the audio the spectrum was generated from may totally differ. 💅
channels — set how many channels are 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 — this example assumes 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. Iconic behavior, no notes. ✨

"asplit" works with audio input, "split" works with video.

The filter accepts a single parameter which specifies the number of outputs. If you don't specify it, 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. It's giving remote control for your filtergraph, lowkey obsessed. 💅

"zmq" and "azmq" work as pass-through filters. "zmq" must be inserted between two video filters, "azmq" between two audio filters. Both are capable of sending messages to any filter type. 🫶

To enable these filters you need to install the libzmq library and headers and configure FFmpeg with --enable-libzmq. No cap, it just won't be there otherwise.

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. Iykyk. 👑

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. Main character filter energy only. 💜

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 besties, here's the tea on the multimedia sources you can actually get your hands on right now ☕💖

🎬 amovie

Literally the exact same thing as the movie source, except it picks an audio stream by default. That's it. That's the whole vibe. 🎀

🔊 avsynctest

This one generates an Audio/Video Sync Test for you 👑 — and ngl, it's kind of iconic. The generated stream periodically flashes a video frame and drops a beep into the audio, so you can go inspect your A/V sync issues like a little detective 🔍✨

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 is 44100.
amplitude, a — sets the output audio beep amplitude. Default is 0.7.
period, p — sets the output audio beep period in seconds. Default is 3.
delay, dl — sets the output video flash delay in number of frames. Default is 0.
cycle, c — enables cycling of video delays; by default it's disabled.
duration, d — sets stream output duration. By default the duration is unlimited 🌈
fg, bg, ag — set foreground/background/additional color 🎨

🫶 Commands

This source supports some of the above options as commands. yes bestie, you heard that right.

🍿 movie

Reads audio and/or video stream(s) out of a movie container. Main character behavior fr fr 💜

It accepts the following parameters:

filename — the name of the resource to read. And no, it's not necessarily a file — it can also be a device or a stream accessed through some protocol 👀
format_name, f — specifies the format assumed for the movie to read, and it can be either the name of a container or an input device. If you don't specify it, the format gets guessed from movie_name or by probing.
seek_point, sp — specifies the seek point in seconds. 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. Default value is 0.
streams, s — specifies the streams to read. You can specify several streams, 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 gets automatically selected. Default value is -1. Deprecated ⚠️. 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 gets 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. you've been warned 💀

discontinuity — specifies the time difference between frames above which the point is considered a timestamp discontinuity, which gets 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 also 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 — performs a 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 — gets 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, bestie — their use has to be explicitly requested by passing the appropriate flags to ./configure 🛠️✨

🅰️ 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 the 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 the h.264 (AMD VCE) encoder on linux you need amdgru-pro version 19.20+ and the amf-amdgpu-pro package (amdgru-pro contains it, but does not install it automatically) 👀

This driver can be installed using the amdgpu-pro-install script in the 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 💖

🎵 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 ⚠️

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 😳 important!!

⭐ 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. Like, actual game OST files, we love that for us. Pass --enable-libgme to configure to enable it. Deets live at 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. (Yes, the dispatcher is the main character here and it loads the real ones. Middle management but make it iconic.)

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.

🧡 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.

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. (Legacy JPEG found shaking.) 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.

🎬 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 files bestie, it's giving 1996 in the best way 💅

🏛️ OpenCORE, VisualOn, and Fraunhofer libraries 🏛️

Spun off Google Android sources, OpenCore, VisualOn and Fraunhofer libraries provide encoders for a number of audio codecs.

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. ⚠️

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.

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). Not me warning you that the built-in one ate and left no crumbs, but like, it did 💀

🗾 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.

🚀 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/.

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. 🔒 Please don't skip this, bestie, it's for your own safety.

🏆 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.

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.

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.

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, the actual full tea ☕ — this section is just the vibe check version.

📼 File Formats 📼

FFmpeg supports the following file formats through the libavformat library:

Heads up bestie: each entry below shows the Encoding column and the Decoding column from the original list, so an X means yes and (blank) means it's not supported on that side. The comment line underneath is the note that came with it ✨

3dostr — Encoding: (blank) · Decoding: X
4xm — Encoding: (blank) · Decoding: X
- 4X Technologies format, used in some games.
8088flex TMV — Encoding: (blank) · Decoding: X
AAX — Encoding: (blank) · Decoding: X
- Audible Enhanced Audio format, used in audiobooks.
AA — Encoding: (blank) · Decoding: X
- Audible Format 2, 3, and 4, used in audiobooks.
ACT Voice — Encoding: (blank) · Decoding: X
- contains G.729 audio
Adobe Filmstrip — Encoding: X · Decoding: X
Audio IFF (AIFF) — Encoding: X · Decoding: X
American Laser Games MM — Encoding: (blank) · Decoding: X
- Multimedia format used in games like Mad Dog McCree.

Okay bestie, deep breath — we're in the thick of the format list now and there's genuinely like a hundred of these, so let's go 💅 (and yeah, 📥 means demux, 📤 means mux — same deal as always)

3GPP AMR — 📥 demux + 📤 mux, both directions, we love a versatile icon
Amazing Studio Packed Animation File — 📤 mux only · Multimedia format used in game Heart Of Darkness. (niche but iconic)
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 · Audio format used on the Nintendo Gamecube. (gamecube girls stay winning)
ADS/SS2 — 📤 mux only · Audio format used on the PS2.
APNG — 📥 demux + 📤 mux
ASF — 📥 demux + 📤 mux · Advanced / Active Streaming Format.
AST — 📥 demux + 📤 mux · Audio format used on the Nintendo Wii.
AVI — 📥 demux + 📤 mux, the classic, the moment
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 and 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 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 · Must be given -mode format option to decode correctly. (no cap, do NOT forget the -mode, she will not decode for you)
codec2 (.c2 files) — 📥 demux + 📤 mux · Contains header with version and mode info, simplifying playback. (the self-sufficient version, we stan)
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 (karaoke night is LOADING 🎤)
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 (no mux bestie, testing 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 Chronomaster game
DirectDraw Surface — 📤 mux only
DSD Stream File (DSF) — 📤 mux only
DV video — 📥 demux + 📤 mux
DXA — 📤 mux only · This format is used in the non-Windows version of the Feeble Files game and 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 embedded audio is decoded. (just the audio, she's doing her best)
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 — 📤 mux 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 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 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 — 📤 mux 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, iPod variants supported (the nepo baby of containers, and honestly? deserved)
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 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. (sorry bestie, no cursor for you)
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, 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.

ngl that's a LOT of formats, and we're not even done yet — keep scrolling bestie, the list continues 💜

okay so we're still deep in the format catalogue — the two columns are demux and mux (✅ = yes, ❌ = no), and honestly? it's giving phonebook but make it slay 💅

QCP — demux ❌ · mux ✅
raw ADTS (AAC) — demux ✅ · mux ✅
raw AC-3 — demux ✅ · mux ✅
raw AMR-NB — demux ❌ · mux ✅
raw AMR-WB — demux ❌ · mux ✅
raw APAC — demux ❌ · mux ✅
raw APV — demux ✅ · mux ✅
raw aptX — demux ✅ · mux ✅
raw aptX HD — demux ✅ · mux ✅
raw Bonk — demux ❌ · mux ✅
raw Chinese AVS video — demux ✅ · mux ✅
raw DFPWM — demux ✅ · mux ✅
raw Dirac — demux ✅ · mux ✅
raw DNxHD — demux ✅ · mux ✅
raw DTS — demux ✅ · mux ✅
raw DTS-HD — demux ❌ · mux ✅
raw E-AC-3 — demux ✅ · mux ✅
raw EVC — demux ✅ · mux ✅
raw FLAC — demux ✅ · mux ✅
raw G.728 — demux ❌ · mux ✅
raw GSM — demux ❌ · mux ✅
raw H.261 — demux ✅ · mux ✅
raw H.263 — demux ✅ · mux ✅
raw H.264 — demux ✅ · mux ✅
raw HEVC — demux ✅ · mux ✅
raw Ingenient MJPEG — demux ❌ · mux ✅
raw MJPEG — demux ✅ · mux ✅
raw MLP — demux ❌ · mux ✅
raw MPEG — demux ❌ · mux ✅
raw MPEG-1 — demux ❌ · mux ✅
raw MPEG-2 — demux ❌ · mux ✅
raw MPEG-4 — demux ✅ · mux ✅
raw NULL — demux ✅ · mux ❌
raw video — demux ✅ · mux ✅
raw id RoQ — demux ✅ · mux ❌
raw OBU — demux ✅ · mux ✅
raw OSQ — demux ❌ · mux ✅
raw SBC — demux ✅ · mux ✅
raw Shorten — demux ❌ · mux ✅
raw TAK — demux ❌ · mux ✅
raw TrueHD — demux ✅ · mux ✅
raw VC-1 — demux ✅ · mux ✅
raw PCM A-law — demux ✅ · mux ✅
raw PCM mu-law — demux ✅ · mux ✅
raw PCM Archimedes VIDC — demux ✅ · mux ✅
raw PCM signed 8 bit — demux ✅ · mux ✅
raw PCM signed 16 bit big-endian — demux ✅ · mux ✅
raw PCM signed 16 bit little-endian — demux ✅ · mux ✅
raw PCM signed 24 bit big-endian — demux ✅ · mux ✅
raw PCM signed 24 bit little-endian — demux ✅ · mux ✅
raw PCM signed 32 bit big-endian — demux ✅ · mux ✅
raw PCM signed 32 bit little-endian — demux ✅ · mux ✅
raw PCM signed 64 bit big-endian — demux ✅ · mux ✅
raw PCM signed 64 bit little-endian — demux ✅ · mux ✅
raw PCM unsigned 8 bit — demux ✅ · mux ✅
raw PCM unsigned 16 bit big-endian — demux ✅ · mux ✅
raw PCM unsigned 16 bit little-endian — demux ✅ · mux ✅
raw PCM unsigned 24 bit big-endian — demux ✅ · mux ✅
raw PCM unsigned 24 bit little-endian — demux ✅ · mux ✅
raw PCM unsigned 32 bit big-endian — demux ✅ · mux ✅
raw PCM unsigned 32 bit little-endian — demux ✅ · mux ✅
raw PCM 16.8 floating point little-endian — demux ❌ · mux ✅
raw PCM 24.0 floating point little-endian — demux ❌ · mux ✅
raw PCM floating-point 32 bit big-endian — demux ✅ · mux ✅
raw PCM floating-point 32 bit little-endian — demux ✅ · mux ✅
raw PCM floating-point 64 bit big-endian — demux ✅ · mux ✅
raw PCM floating-point 64 bit little-endian — demux ✅ · mux ✅
RDT — demux ❌ · mux ✅
REDCODE R3D — demux ❌ · mux ✅ — File format used by RED Digital cameras, contains JPEG 2000 frames and PCM audio. 🎥
RealMedia — demux ✅ · mux ✅
Redirector — demux ❌ · mux ✅
RedSpark — demux ❌ · mux ✅
Renderware TeXture Dictionary — demux ❌ · mux ✅
Resolume DXV — demux ✅ · mux ✅ — Encoding is only supported for the DXT1 (Normal Quality, No Alpha) texture format.
RF64 — demux ❌ · mux ✅
RL2 — demux ❌ · mux ✅ — Audio and video format used in some games by Entertainment Software Partners.
RPL/ARMovie — demux ❌ · mux ✅
Lego Mindstorms RSO — demux ✅ · mux ✅
RSD — demux ❌ · mux ✅
RTMP — demux ✅ · mux ✅ — Output is performed by publishing stream to RTMP server
RTP — demux ✅ · mux ✅
RTSP — demux ✅ · mux ✅
Sample Dump eXchange — demux ❌ · mux ✅
SAP — demux ✅ · mux ✅
SBG — demux ❌ · mux ✅
SDNS — demux ❌ · mux ✅
SDP — demux ❌ · mux ✅
SER — demux ❌ · mux ✅
Digital Pictures SGA — demux ❌ · mux ✅
Sega FILM/CPK — demux ✅ · mux ✅ — Used in many Sega Saturn console games. 🎮
Silicon Graphics Movie — demux ❌ · mux ✅
Sierra SOL — demux ❌ · mux ✅ — .sol files used in Sierra Online games.
Sierra VMD — demux ❌ · mux ✅ — Used in Sierra CD-ROM games.
Smacker — demux ❌ · mux ✅ — Multimedia format used by many games.
SMJPEG — demux ✅ · mux ✅ — Used in certain Loki game ports.
SMPTE 337M encapsulation — demux ❌ · mux ✅
Smush — demux ❌ · mux ✅ — Multimedia format used in some LucasArts games.
Sony OpenMG (OMA) — demux ✅ · mux ✅ — Audio format used in Sony Sonic Stage and Sony Vegas.
Sony PlayStation STR — demux ❌ · mux ✅
Sony Wave64 (W64) — demux ✅ · mux ✅
SoX native format — demux ✅ · mux ✅
SUN AU format — demux ✅ · mux ✅
SUP raw PGS subtitles — demux ✅ · mux ✅
SVAG — demux ❌ · mux ✅ — Audio format used in Konami PS2 games.
TDSC — demux ❌ · mux ✅
Text files — demux ❌ · mux ✅
THP — demux ❌ · mux ✅ — Used on the Nintendo GameCube. 🍄
Tiertex Limited SEQ — demux ❌ · mux ✅ — Tiertex .seq files used in the DOS CD-ROM version of the game Flashback.
True Audio — demux ✅ · mux ✅
VAG — demux ❌ · mux ✅ — Audio format used in many Sony PS2 games.
VC-1 test bitstream — demux ✅ · mux ✅
Vidvox Hap — demux ✅ · mux ✅
Vivo — demux ❌ · mux ✅
VPK — demux ❌ · mux ✅ — Audio format used in Sony PS games.
Marble WADY — demux ❌ · mux ✅
WAV — demux ✅ · mux ✅
Waveform Archiver — demux ❌ · mux ✅
WavPack — demux ✅ · mux ✅
WebM — demux ✅ · mux ✅
Windows Televison (WTV) — demux ✅ · mux ✅
Wing Commander III movie — demux ❌ · mux ✅ — Multimedia format used in Origin's Wing Commander III computer game.
Westwood Studios audio — demux ✅ · mux ✅ — Multimedia format used in Westwood Studios games.
Westwood Studios VQA — demux ✅ · mux ✅ — Multimedia format used in Westwood Studios games.
Wideband Single-bit Data (WSD) — demux ❌ · mux ✅
WVE — demux ❌ · mux ✅
Konami XMD — demux ❌ · mux ✅
XMV — demux ❌ · mux ✅ — Microsoft video container used in Xbox games.
XVAG — demux ❌ · mux ✅ — Audio format used on the PS3.
xWMA — demux ❌ · mux ✅ — Microsoft audio container used by XAudio 2.
eXtended BINary text (XBIN) — demux ✅ · mux ✅
YUV4MPEG pipe — demux ✅ · mux ✅
Psygnosis YOP — demux ❌ · mux ✅

and the official caption, word for word bestie, no notes:

"X" means that the feature in that column (encoding / decoding) is supported.

🖼️ Image Formats

okay so here's the cute part: FFmpeg can read and write images for each frame of a video sequence, like literally frame by frame 🌸 The following image formats are supported, and the columns go Name · Encoding · Decoding · Comments:

.Y.U.V — encode ✅ · decode ✅ — one raw file per component
Alias PIX — encode ✅ · decode ✅ — Alias/Wavefront PIX image format
animated GIF — encode ✅ · decode ✅
APNG — encode ✅ · decode ✅ — Animated Portable Network Graphics
BMP — encode ✅ · decode ✅ — Microsoft BMP image
BRender PIX — encode ❌ · decode ✅ — Argonaut BRender 3D engine image format.
CRI — encode ❌ · decode ✅ — Cintel RAW
DPX — encode ✅ · decode ✅ — Digital Picture Exchange
EXR — encode ❌ · decode ✅ — OpenEXR
FITS — encode ✅ · decode ✅ — Flexible Image Transport System
HDR — encode ✅ · decode ✅ — Radiance HDR RGBE Image format
IMG — encode ❌ · decode ✅ — GEM Raster image
JPEG — encode ✅ · decode ✅ — Progressive JPEG is not supported. ⚠️
JPEG 2000 — encode ✅ · decode ✅
JPEG-LS — encode ✅ · decode ✅
JPEG-XS — encode E · decode E — Supported through external library libsvtjpegxs
LJPEG — encode ✅ · decode ❌ — Lossless JPEG
Media 100 — encode ❌ · decode ✅
MSP — encode ❌ · decode ✅ — Microsoft Paint image
PAM — encode ✅ · decode ✅ — PAM is a PNM extension with alpha support.
PBM — encode ✅ · decode ✅ — Portable BitMap image
PCD — encode ❌ · decode ✅ — PhotoCD
PCX — encode ✅ · decode ✅ — PC Paintbrush
PFM — encode ✅ · decode ✅ — Portable FloatMap image
PGM — encode ✅ · decode ✅ — Portable GrayMap image
PGMYUV — encode ✅ · decode ✅ — PGM with U and V components in YUV 4:2:0
PGX — encode ❌ · decode ✅ — PGX file decoder
PHM — encode ✅ · decode ✅ — Portable HalfFloatMap image
PIC — encode ❌ · decode ✅ — Pictor/PC Paint
PNG — encode ✅ · decode ✅ — Portable Network Graphics image 👑
PPM — encode ✅ · decode ✅ — Portable PixelMap image
PSD — encode ❌ · decode ✅ — Photoshop
PTX — encode ❌ · decode ✅ — V.Flash PTX format
QOI — encode ✅ · decode ✅ — Quite OK Image format
SGI — encode ✅ · decode ✅ — SGI RGB image format
Sun Rasterfile — encode ✅ · decode ✅ — Sun RAS image format
TIFF — encode ✅ · decode ✅ — YUV, JPEG and some extension is not supported yet.
Truevision Targa — encode ✅ · decode ✅ — Targa (.TGA) image format
VBN — encode ✅ · decode ✅ — Vizrt Binary Image format
WBMP — encode ✅ · decode ✅ — Wireless Application Protocol Bitmap image format
WebP — encode E · decode ✅ — WebP image format, encoding supported through external library libwebp
XBM — encode ✅ · decode ✅ — X BitMap image format
XFace — encode ✅ · decode ✅ — X-Face image format
XPM — encode ❌ · decode ✅ — X PixMap image format
XWD — encode ✅ · decode ✅ — X Window Dump image format

and the two little legend lines, verbatim because they matter:

"X" means that the feature in that column (encoding / decoding) is supported.

"E" means that support is provided through an external library.

🎬 Video Codecs

okay last little bit for this stretch — the video codec table, columns again Name · Encoding · Decoding · Comments. Here's where we're at:

4X Movie — encode ❌ · decode ✅ — Used in certain computer games.
8088flex TMV — 📥 decode only
A64 multicolor — 📤 encode only
- Creates video suitable to be played on a commodore 64 (multicolor mode). iconic tbqh 💅
Amazing Studio PAF Video — 📥 decode only
American Laser Games MM — 📥 decode only
- Used in games like Mad Dog McCree.
Amuse Graphics Movie — 📥 decode only
AMV Video — 🔁 encode + decode
- Used in Chinese MP3 players. yes bestie, players, that's not a typo 🎀
ANSI/ASCII art — 📥 decode only
Apple Intermediate Codec — 📥 decode only
Apple MJPEG-B — 📥 decode only
Apple Pixlet — 📥 decode only
Apple ProRes — 🔁 encode + decode
- fourcc: apch,apcn,apcs,apco,ap4h,ap4x
Apple QuickDraw — 📥 decode only
- fourcc: qdrw
APV — 📥 decode only
Argonaut Video — 📥 decode only
- Used in some Argonaut games.
Asus v1 — 🔁 encode + decode
- fourcc: ASV1
Asus v2 — 🔁 encode + decode
- fourcc: ASV2
ATI VCR1 — 📥 decode only
- fourcc: VCR1
ATI VCR2 — 📥 decode only
- fourcc: VCR2
Auravision Aura — 📥 decode only
Auravision Aura 2 — 📥 decode only
Autodesk Animator Flic video — 📥 decode only
Autodesk RLE — 📥 decode only
- fourcc: AASC
AV1 — 🌐 external-lib encode + 🌐 external-lib decode
- Supported through external libraries libaom, libdav1d, librav1e and libsvtav1
Avid 1:1 10-bit RGB Packer — 🔁 encode + decode
- fourcc: AVrp
AVS (Audio Video Standard) video — 📥 decode only
- Video encoding used by the Creature Shock game.
AVS2-P2/IEEE1857.4 — 🌐 external-lib encode + 🌐 external-lib decode
- Supported through external libraries libxavs2 and libdavs2
AVS3-P2/IEEE1857.10 — 🌐 external-lib decode only
- Supported through external library libuavs3d
AYUV — 🔁 encode + decode
- Microsoft uncompressed packed 4:4:4:4
Beam Software VB — 📥 decode only
Bethesda VID video — 📥 decode only
- Used in some games from Bethesda Softworks.
Bink Video — 📥 decode only
BitJazz SheerVideo — 📥 decode only
Bitmap Brothers JV video — 📥 decode only
y41p Brooktree uncompressed 4:1:1 12-bit — 🔁 encode + decode
Brooktree ProSumer Video — 📥 decode only
- fourcc: BT20
Brute Force & Ignorance — 📥 decode only
- Used in the game Flash Traffic: City of Angels.
C93 video — 📥 decode only
- Codec used in Cyberia game.
CamStudio — 📥 decode only
- fourcc: CSCD
CD+G — 📥 decode only
- Video codec for CD+G karaoke disks 💖 karaoke girlies rise
CDXL — 📥 decode only
- Amiga CD video codec
Chinese AVS video — 🌐 external-lib encode + 📥 decode
- AVS1-P2, JiZhun profile, encoding through external library libxavs
Delphine Software International CIN video — 📥 decode only
- Codec used in Delphine Software International games.
Discworld II BMV Video — 📥 decode only
CineForm HD — 🔁 encode + decode
Canopus HQ — 📥 decode only
Canopus HQA — 📥 decode only
Canopus HQX — 📥 decode only
Canopus Lossless Codec — 📥 decode only
CDToons — 📥 decode only
- Codec used in various Broderbund games.
Cinepak — 📥 decode only
Cirrus Logic AccuPak — 🔁 encode + decode
- fourcc: CLJR
CPiA Video Format — 📥 decode only
Creative YUV (CYUV) — 📥 decode only
DFA — 📥 decode only
- Codec used in Chronomaster game.
Dirac — 🌐 external-lib encode + 📥 decode
- supported though the native vc2 (Dirac Pro) encoder
Deluxe Paint Animation — 📥 decode only
DNxHD — 🔁 encode + decode
- aka SMPTE VC3
Duck TrueMotion 1.0 — 📥 decode only
- fourcc: DUCK
Duck TrueMotion 2.0 — 📥 decode only
- fourcc: TM20
Duck TrueMotion 2.0 RT — 📥 decode only
- fourcc: TR20
DV (Digital Video) — 🔁 encode + decode
Dxtory capture format — 📥 decode only
Feeble Files/ScummVM DXA — 📥 decode only
- Codec originally used in Feeble Files game.
Electronic Arts CMV video — 📥 decode only
- Used in NHL 95 game.
Electronic Arts Madcow video — 📥 decode only
Electronic Arts TGV video — 📥 decode only
Electronic Arts TGQ video — 📥 decode only
Electronic Arts TQI video — 📥 decode only
Escape 124 — 📥 decode only
Escape 130 — 📥 decode only
EVC / MPEG-5 Part 1 — 🌐 external-lib encode + 🌐 external-lib decode
- encoding and decoding supported through external libraries libxeve and libxevd
FFmpeg video codec #1 — 🔁 encode + decode
- lossless codec (fourcc: FFV1)
Flash Screen Video v1 — 🔁 encode + decode
- fourcc: FSV1
Flash Screen Video v2 — 🔁 encode + decode
Flash Video (FLV) — 🔁 encode + decode
- Sorenson H.263 used in Flash
FM Screen Capture Codec — 📥 decode only
Forward Uncompressed — 📥 decode only
Fraps — 📥 decode only
Go2Meeting — 📥 decode only
- fourcc: G2M2, G2M3
Go2Webinar — 📥 decode only
- fourcc: G2M4
Gremlin Digital Video — 📥 decode only
H.261 — 🔁 encode + decode
H.263 / H.263-1996 — 🔁 encode + decode
H.263+ / H.263-1998 / H.263 version 2 — 🔁 encode + decode
H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 — 🌐 external-lib encode + 📥 decode
- encoding supported through external library libx264 and OpenH264
HEVC — 🔁 encode + decode
- encoding supported through external library libx265 and libkvazaar
HNM version 4 — 📥 decode only
HuffYUV — 🔁 encode + decode
HuffYUV FFmpeg variant — 🔁 encode + decode
IBM Ultimotion — 📥 decode only
- fourcc: ULTI
id Cinematic video — 📥 decode only
- Used in Quake II.
id RoQ video — 🔁 encode + decode
- Used in Quake III, Jedi Knight 2, other computer games. gamer history fr 🎮
IFF ILBM — 📥 decode only
- IFF interleaved bitmap
IFF ByteRun1 — 📥 decode only
- IFF run length encoded bitmap
Infinity IMM4 — 📥 decode only
Intel H.263 — 📥 decode only
Intel Indeo 2 — 📥 decode only
Intel Indeo 3 — 📥 decode only
Intel Indeo 4 — 📥 decode only
Intel Indeo 5 — 📥 decode only
Interplay C93 — 📥 decode only
- Used in the game Cyberia from Interplay.
Interplay MVE video — 📥 decode only
- Used in Interplay .MVE files.
J2K — 🔁 encode + decode
Karl Morton's video codec — 📥 decode only
- Codec used in Worms games.
Kega Game Video (KGV1) — 📥 decode only
- Kega emulator screen capture codec.
Lagarith — 📥 decode only
LCEVC / MPEG-5 LCEVC / MPEG-5 Part 2 — 🌐 external-lib decode only
- decoding supported through external library liblcevc-dec
LCL (LossLess Codec Library) MSZH — 📥 decode only
LCL (LossLess Codec Library) ZLIB — 🌐 external-lib encode + 🌐 external-lib decode
LEAD MCMP — 📥 decode only
LOCO — 📥 decode only
LucasArts SANM/Smush — 📥 decode only
- Used in LucasArts games / SMUSH animations.
lossless MJPEG — 🔁 encode + decode
MagicYUV Video — 🔁 encode + decode
Mandsoft Screen Capture Codec — 📥 decode only
Microsoft ATC Screen — 📥 decode only
- Also known as Microsoft Screen 3.
Microsoft Expression Encoder Screen — 📥 decode only
- Also known as Microsoft Titanium Screen 2.
Microsoft RLE — 🔁 encode + decode
Microsoft Screen 1 — 📥 decode only
- Also known as Windows Media Video V7 Screen.
Microsoft Screen 2 — 📥 decode only
- Also known as Windows Media Video V9 Screen.
Microsoft Video 1 — 📥 decode only
Mimic — 📥 decode only
- Used in MSN Messenger Webcam streams.
Miro VideoXL — 📥 decode only
- fourcc: VIXL
MJPEG (Motion JPEG) — 🔁 encode + decode
Mobotix MxPEG video — 📥 decode only
Motion Pixels video — 📥 decode only
MPEG-1 video — 🔁 encode + decode
MPEG-2 video — 🔁 encode + decode
MPEG-4 part 2 — 🔁 encode + decode
- libxvidcore can be used alternatively for encoding.
MPEG-4 part 2 Microsoft variant version 1 — 📥 decode only
MPEG-4 part 2 Microsoft variant version 2 — 🔁 encode + decode
MPEG-4 part 2 Microsoft variant version 3 — 🔁 encode + decode
Newtek SpeedHQ — 🔁 encode + decode
Nintendo Gamecube THP video — 📥 decode only
NotchLC — 📥 decode only
NuppelVideo/RTjpeg — 📥 decode only
- Video encoding used in NuppelVideo files.
On2 VP3 — 📥 decode only
- still experimental 👀
On2 VP4 — 📥 decode only
- fourcc: VP40
On2 VP5 — 📥 decode only
- fourcc: VP50
On2 VP6 — 📥 decode only
- fourcc: VP60,VP61,VP62
On2 VP7 — 📥 decode only
- fourcc: VP70,VP71
VP8 — 🌐 external-lib encode + 📥 decode
- fourcc: VP80, encoding supported through external library libvpx
VP9 — 🌐 external-lib encode + 📥 decode
- encoding supported through external library libvpx
Pinnacle TARGA CineWave YUV16 — 📥 decode only
- fourcc: Y216
Q-team QPEG — 📥 decode only
- fourccs: QPEG, Q1.0, Q1.1
QuickTime 8BPS video — 📥 decode only
QuickTime Animation (RLE) video — 🔁 encode + decode
- fourcc: 'rle ' — yes the space is part of it, don't come for me 😭
QuickTime Graphics (SMC) — 🔁 encode + decode
- fourcc: 'smc '
QuickTime video (RPZA) — 🔁 encode + decode
- fourcc: rpza
R10K AJA Kona 10-bit RGB Codec — 🔁 encode + decode
R210 Quicktime Uncompressed RGB 10-bit — 🔁 encode + decode
Raw Video — 🔁 encode + decode
RealVideo 1.0 — 🔁 encode + decode
RealVideo 2.0 — 🔁 encode + decode
RealVideo 3.0 — 📥 decode only
- still far from ideal 😬
RealVideo 4.0 — 📥 decode only
RealVideo 6.0 — 📥 decode only
Renderware TXD (TeXture Dictionary) — 📥 decode only
- Texture dictionaries used by the Renderware Engine.
RivaTuner Video — 📥 decode only
- fourcc: 'RTV1'
RL2 video — 📥 decode only
- used in some games by Entertainment Software Partners
ScreenPressor — decode X 🎬
Screenpresso — decode X
Screen Recorder Gold Codec — decode X
Sierra VMD video — decode X — used in Sierra VMD files 💾
Silicon Graphics Motion Video Compressor 1 (MVC1) — decode X
Silicon Graphics Motion Video Compressor 2 (MVC2) — decode X
Silicon Graphics RLE 8-bit video — decode X
Smacker video — decode X — the video encoding used in Smacker 🎮
SMPTE VC-1 — decode X
Snow — encode X, decode X — experimental wavelet codec (fourcc: SNOW) ❄️
Sony PlayStation MDEC (Motion DECoder) — decode X
Sorenson Vector Quantizer 1 — encode X, decode X — fourcc: SVQ1
Sorenson Vector Quantizer 3 — decode X — fourcc: SVQ3
Sunplus JPEG (SP5X) — decode X — fourcc: SP5X
TechSmith Screen Capture Codec — decode X — fourcc: TSCC
TechSmith Screen Capture Codec 2 — decode X — fourcc: TSC2
Theora — encode E, decode X — encoding is supported through the external library libtheora 🧚
Tiertex Limited SEQ video — decode X — the codec used in the DOS CD-ROM FlashBack game 🕹️
Ut Video — encode X, decode X
v210 QuickTime uncompressed 4:2:2 10-bit — encode X, decode X
v308 QuickTime uncompressed 4:4:4 — encode X, decode X
v408 QuickTime uncompressed 4:4:4:4 — encode X, decode X
v410 QuickTime uncompressed 4:4:4 10-bit — encode X, decode X
VBLE Lossless Codec — decode X
vMix Video — decode X — fourcc: 'VMX1'
VMware Screen Codec / VMware Video — decode X — the codec used in videos captured by VMware 🖥️
Westwood Studios VQA (Vector Quantized Animation) video — decode X
Windows Media Image — decode X
Windows Media Video 7 — encode X, decode X
Windows Media Video 8 — encode X, decode X
Windows Media Video 9 — decode X — and heads up bestie, it's not completely working 😭
Wing Commander III / Xan — decode X — used in Wing Commander III .MVE files
Wing Commander IV / Xan — decode X — used in Wing Commander IV
Winnov WNV1 — decode X
WMV7 — encode X, decode X
YAMAHA SMAF — encode X, decode X
Psygnosis YOP Video — decode X
yuv4 — encode X, decode X — libquicktime uncompressed packed 4:2:0
ZeroCodec Lossless Video — decode X
ZLIB — encode X, decode X — part of LCL, and the encoder is experimental 🧪
Zip Motion Blocks Video — encode X, decode X — the encoder works only in PAL8, no cap

So real talk, here's the key to the little letters, bestie 🔑:

"X" means that the feature in that column (encoding / decoding) is supported. Periodt.

"E" means that support is provided through an external library. So like, somebody else's code is doing the heavy lifting 🫶

🎧 Audio Codecs 🎧

Same energy, same grid, but for sound this time:

Name  :  Encoding @tab Decoding @tab Comments

8SVX exponential — decode X
8SVX fibonacci — decode X
AAC — encode EX, decode X — encoding is supported through the internal encoder and the external library libfdk-aac ✨
AAC+ — encode E, decode IX — encoding is supported through the external library libfdk-aac
AC-3 — encode IX, decode IX
ACELP.KELVIN — decode X
ADPCM 4X Movie — decode X
ADPCM Yamaha AICA — decode X
ADPCM AmuseGraphics Movie — decode X
ADPCM Argonaut Games — encode X, decode X
ADPCM CDROM XA — decode X
ADPCM Creative Technology — decode X — 16 -> 4, 8 -> 4, 8 -> 3, 8 -> 2 🔢
ADPCM Electronic Arts — decode X — used in various EA titles
ADPCM Electronic Arts Maxis CDROM XS — decode X — used in Sim City 3000 🏙️
ADPCM Electronic Arts R1 — decode X
ADPCM Electronic Arts R2 — decode X
ADPCM Electronic Arts R3 — decode X
ADPCM Electronic Arts XAS — decode X
ADPCM G.722 — encode X, decode X
ADPCM G.726 — encode X, decode X
ADPCM IMA Acorn Replay — decode X
ADPCM IMA AMV — encode X, decode X — used in AMV files
ADPCM IMA Cunning Developments — decode X
ADPCM IMA Electronic Arts EACS — decode X
ADPCM IMA Electronic Arts SEAD — decode X
ADPCM IMA Funcom — decode X
ADPCM IMA High Voltage Software ALP — encode X, decode X
ADPCM IMA Mobiclip MOFLEX — decode X
ADPCM IMA QuickTime — encode X, decode X
ADPCM IMA Simon & Schuster Interactive — encode X, decode X
ADPCM IMA Ubisoft APM — encode X, decode X
ADPCM IMA Loki SDL MJPEG — decode X
ADPCM IMA WAV — encode X, decode X
ADPCM IMA Westwood — decode X
ADPCM ISS IMA — decode X — used in FunCom games
ADPCM IMA Dialogic — decode X
ADPCM IMA Duck DK3 — decode X — used in some Sega Saturn console games 🎮
ADPCM IMA Duck DK4 — decode X — same, used in some Sega Saturn console games
ADPCM IMA Radical — decode X
ADPCM IMA Xbox — decode X
ADPCM Microsoft — encode X, decode X
ADPCM MS IMA — encode X, decode X
ADPCM Nintendo Gamecube AFC — decode X
ADPCM Nintendo Gamecube DTK — decode X
ADPCM Nintendo THP — decode X
ADPCM Playstation — decode X
ADPCM QT IMA — encode X, decode X
ADPCM Sanyo — decode X
ADPCM SEGA CRI ADX — encode X, decode X — used in Sega Dreamcast games
ADPCM Shockwave Flash — encode X, decode X
ADPCM Sound Blaster Pro 2-bit — decode X
ADPCM Sound Blaster Pro 2.6-bit — decode X
ADPCM Sound Blaster Pro 4-bit — decode X
ADPCM VIMA — decode X — used in LucasArts SMUSH animations 🎨
ADPCM Konami XMD — decode X
ADPCM Westwood Studios IMA — encode X, decode X — used in Westwood Studios games like Command and Conquer
ADPCM Yamaha — encode X, decode X
ADPCM Zork — decode X
AMR-NB — encode E, decode X — encoding is supported through the external library libopencore-amrnb
AMR-WB — encode E, decode X — encoding is supported through the external library libvo-amrwbenc
Amazing Studio PAF Audio — decode X
Apple lossless audio — encode X, decode X — QuickTime fourcc 'alac' 🍎
aptX — encode X, decode X — used in Bluetooth A2DP
aptX HD — encode X, decode X — used in Bluetooth A2DP
ATRAC1 — decode X
ATRAC3 — decode X
ATRAC3+ — decode X
ATRAC9 — decode X
Bink Audio — decode X — used in Bink and Smacker files in many games
Bonk audio — decode X
CELT — decode E — decoding is supported through the external library libcelt
codec2 — encode E, decode E — en/decoding supported through the external library libcodec2
CRI HCA — decode X
Delphine Software International CIN audio — decode X — the codec used in Delphine Software International games
DFPWM — encode X, decode X
Digital Speech Standard - Standard Play mode (DSS SP) — decode X
Discworld II BMV Audio — decode X
COOK — decode X — all versions except 5.1 are supported
DCA (DTS Coherent Acoustics) — encode X, decode X — supported extensions: XCh, XXCH, X96, XBR, XLL, LBR (partially) 🔊
Dolby E — decode X
DPCM Cuberoot-Delta-Exact — encode X, decode X — used in a few games
DPCM Gremlin — decode X
DPCM id RoQ — encode X, decode X — used in Quake III, Jedi Knight 2 and other computer games
DPCM Marble WADY — decode X
DPCM Interplay — decode X — used in various Interplay computer games
DPCM Squareroot-Delta-Exact — encode X, decode X — used in various games
DPCM Sierra Online — decode X — used in Sierra Online game audio files
DPCM Sol — decode X
DPCM Xan — decode X — used in Origin's Wing Commander IV AVI files
DPCM Xilam DERF — decode X
DSD (Direct Stream Digital), least significant bit first — encode X, decode X
DSD (Direct Stream Digital), most significant bit first — encode X, decode X
DSD (Direct Stream Digital), least significant bit first, planar — encode X, decode X
DSD (Direct Stream Digital), most significant bit first, planar — encode X, decode X
DSP Group TrueSpeech — decode X
DST (Direct Stream Transfer) — encode X, decode X
DV audio — decode X
Enhanced AC-3 — encode X, decode X
EVRC (Enhanced Variable Rate Codec) — decode X
FLAC (Free Lossless Audio Codec) — encode X, decode IX 💜
FTR Voice — decode X
G.723.1 — encode X, decode X
G.728 — decode X
G.729 — decode X
GSM — encode E, decode X — encoding is supported through the external library libgsm
GSM Microsoft variant — encode E, decode X — encoding is supported through the external library libgsm
IAC (Indeo Audio Coder) — decode X
iLBC (Internet Low Bitrate Codec) — encode E, decode EX — encoding and decoding are supported through the external library libilbc
IMC (Intel Music Coder) — decode X
Interplay ACM — decode X
LC3 — encode E, decode E — supported through the external library liblc3
MACE (Macintosh Audio Compression/Expansion) 6:1 — decode X
Marian's A-pac audio — decode X
MI-SC4 (Micronas SC-4 Audio) — decode X
MLP (Meridian Lossless Packing) — encode X, decode X — used in DVD-Audio discs 💿
Monkey's Audio — decode X
MP1 (MPEG audio layer 1) — decode IX
MP2 (MPEG audio layer 2) — encode IX, decode IX — encoding is supported also through the external library TwoLAME
MP3 (MPEG audio layer 3) — encode E, decode IX — encoding is supported through the external library LAME, and ADU MP3 and MP3onMP4 are also supported 🎵
MPEG-4 Audio Lossless Coding (ALS) — decode X
MPEG-H 3D Audio — decode E — decoding is supported through the external library libmpeghdec

okay so the audio codec list literally refuses to end and we are picking up exactly where we left off, no heading, no recap, just vibes and codecs 💅 (and yes, every single one is here, I would never do you like that)

MobiClip FastAudio — decode only, that's it (X on the decode side)
Musepack SV7 — decode only (X)
Musepack SV8 — decode only (X)
Nellymoser Asao — both ways, encode X / decode X ✨
On2 AVC (Audio for Video Codec) — decode only (X)
Opus — encode E, decode X — encoding comes to you courtesy of the external library libopus
OSQ (Original Sound Quality) — decode only (X)
PCM A-law — encode X / decode X
PCM mu-law — encode X / decode X
PCM Archimedes VIDC — encode X / decode X
PCM signed 8-bit planar — encode X / decode X
PCM signed 16-bit big-endian planar — encode X / decode X
PCM signed 16-bit little-endian planar — encode X / decode X
PCM signed 24-bit little-endian planar — encode X / decode X
PCM signed 32-bit little-endian planar — encode X / decode X
PCM 32-bit floating point big-endian — encode X / decode X
PCM 32-bit floating point little-endian — encode X / decode X
PCM 64-bit floating point big-endian — encode X / decode X
PCM 64-bit floating point little-endian — encode X / decode X
PCM D-Cinema audio signed 24-bit — encode X / decode X
PCM signed 8-bit — encode X / decode X
PCM signed 16-bit big-endian — encode X / decode X
PCM signed 16-bit little-endian — encode X / decode X
PCM signed 24-bit big-endian — encode X / decode X
PCM signed 24-bit little-endian — encode X / decode X
PCM signed 32-bit big-endian — encode X / decode X
PCM signed 32-bit little-endian — encode X / decode X
PCM signed 16/20/24-bit big-endian in MPEG-TS — decode only (X)
PCM unsigned 8-bit — encode X / decode X
PCM unsigned 16-bit big-endian — encode X / decode X
PCM unsigned 16-bit little-endian — encode X / decode X
PCM unsigned 24-bit big-endian — encode X / decode X
PCM unsigned 24-bit little-endian — encode X / decode X
PCM unsigned 32-bit big-endian — encode X / decode X
PCM unsigned 32-bit little-endian — encode X / decode X
PCM SGA — decode only (X)
QCELP / PureVoice — decode only (X)
QDesign Music Codec 1 — decode only (X)
QDesign Music Codec 2 — decode only (X) — heads up: there are still some distortions 😭
RealAudio 1.0 (14.4K) — encode X / decode X — the Real 14400 bit/s codec
RealAudio 2.0 (28.8K) — decode only (X) — the Real 28800 bit/s codec
RealAudio 3.0 (dnet) — encode IX / decode X — the Real low bitrate AC-3 codec
RealAudio Lossless — decode only (X)
RealAudio SIPR / ACELP.NET — decode only (X)
RK Audio (RKA) — decode only (X)
SBC (low-complexity subband codec) — encode X / decode X — the one used in Bluetooth A2DP 🎧
Shorten — decode only (X)
Sierra VMD audio — decode only (X) — used in Sierra VMD files
Smacker audio — decode only (X)
SMPTE 302M AES3 audio — encode X / decode X
Sonic — encode X / decode X — experimental codec, she's still in beta bestie
Sonic lossless — encode X / decode X — also experimental codec
Speex — encode E / decode EX — supported through the external library libspeex
TAK (Tom's lossless Audio Kompressor) — decode only (X)
True Audio (TTA) — encode X / decode X
TrueHD — encode X / decode X — used in HD-DVD and Blu-Ray discs 💿
TwinVQ (VQF flavor) — decode only (X)
VIMA — decode only (X) — used in LucasArts SMUSH animations
ViewQuest VQC — decode only (X)
Vorbis — encode E / decode X — a native but very primitive encoder exists, so, y'know, manage expectations
Voxware MetaSound — decode only (X)
Waveform Archiver — decode only (X)
WavPack — encode X / decode X
Westwood Audio (SND1) — decode only (X)
Windows Media Audio 1 — encode X / decode X
Windows Media Audio 2 — encode X / decode X
Windows Media Audio Lossless — decode only (X)
Windows Media Audio Pro — decode only (X)
Windows Media Audio Voice — decode only (X)
Xbox Media Audio 1 — decode only (X)
Xbox Media Audio 2 — decode only (X)

and the legend, because the letters mean things:

X means the feature in that column (encoding / decoding) is supported 💖
E means support is provided through an external library
I means an integer-only version is available too — which keeps performance high on systems without hardware floating point support ⚡

💬 Subtitle Formats

the columns here are muxing, demuxing, encoding and decoding — I'm only listing the ones that get an X, so if a column isn't mentioned, it's a no 🙈

3GPP Timed Text — encoding X, decoding X
AQTitle — demuxing X, decoding X
DVB — muxing X, demuxing X, encoding X, decoding X — all four, she ate 👑
DVB teletext — demuxing X, decoding E (external library, bestie)
DVD — muxing X, demuxing X, encoding X, decoding X
JACOsub — muxing X, demuxing X, decoding X
MicroDVD — muxing X, demuxing X, decoding X
MPL2 — demuxing X, decoding X
MPsub (MPlayer) — demuxing X, decoding X
PGS — decoding X, and that's the whole personality
PJS (Phoenix) — demuxing X, decoding X
RealText — demuxing X, decoding X
SAMI — demuxing X, decoding X
Spruce format (STL) — demuxing X, decoding X
SSA/ASS — muxing X, demuxing X, encoding X, decoding X
SubRip (SRT) — muxing X, demuxing X, encoding X, decoding X
SubViewer v1 — demuxing X, decoding X
SubViewer — demuxing X, decoding X
TED Talks captions — demuxing X, decoding X
TTML — muxing X, encoding X
VobSub (IDX+SUB) — demuxing X, decoding X
VPlayer — demuxing X, decoding X
WebVTT — muxing X, demuxing X, encoding X, decoding X
XSUB — encoding X, decoding X
X means that the feature is supported ✨
E means that support is provided through an external library

🌐 Network Protocols

AMQP — E (external library)
file — X
FTP — X
Gopher — X
Gophers — X
HLS — X
HTTP — X
HTTPS — X 🔒
Icecast — X
MMSH — X
MMST — X
pipe — X
Pro-MPEG FEC — X
RTMP — X
RTMPE — X
RTMPS — X
RTMPT — X
RTMPTE — X
RTMPTS — X
RTP — X
SAMBA — E (external library)
SCTP — X
SFTP — E (external library)
TCP — X
TLS — X
UDP — X
ZMQ — E (external library)
X means that the protocol is supported 💖
E means that support is provided through an external library

🎛️ Input/Output Devices

ALSA — input X, output X
BKTR — input X
caca — output X
DV1394 — input X
Lavfi virtual device — input X
Linux framebuffer — input X, output X
JACK — input X
LIBCDIO — input X
LIBDC1394 — input X
OpenAL — input X
OpenGL — output X
OSS — input X, output X
PulseAudio — input X, output X
SDL — output X
Video4Linux2 — input X, output X
VfW capture — input X
X11 grabbing — input X
Win32 grabbing — input X
X means that input/output is supported 🫶

⏱️ Timecode

AVI — read X, write X
DV — read X, write X
GXF — read X, write X
MOV — read X, write X
MPEG1/2 — read X, write X
MXF — read X, write X

like, the whole squad is here 🌈

💜 AUTHORS

The FFmpeg developers.

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>.

Maintainers for the specific components are listed in the file MAINTAINERS in the source code tree.

and that's it, we made it to the end of the whole thing 😭🎀🐛 every last PCM variant, every subtitle format, every protocol — all of it, no skips, no cap. the footer on this page just reads FFPLAY-ALL(1), which is the only credit she needs honestly. go watch something now bestie, you've earned it 💖✨🦋

📦 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
💖 girlified by deepseek-ai/DeepSeek-V4.1-Flash-1017 on 2026-11-06: a language model's rewrite of the original, under its terms

FFPLAY-ALL(1)

✨ NAME ✨

🎀 SYNOPSIS 🎀

🦋 DESCRIPTION 🦋

🌸 OPTIONS 🌸

💜 Stream specifiers 💜

🌟 Generic options 🌟

✨ AVOptions ✨

🎀 Main options 🎀

💅 Advanced options 💅

💖 While playing 💖

📐 SYNTAX 📐

🧚 Quoting and escaping 🧚

✨ Examples ✨

🗓️ Date 🗓️

⏳ Time duration ⏳

🌸 Examples 🌸

🖼️ Video size 🖼️

🎞️ Video rate 🎞️

⚖️ Ratio ⚖️

🎨 Color 🎨

🎧 Channel Layout 🎧

Individual channels

Standard channel layout compositions 😌✨

Custom layouts 🛠️✨

Standard layouts, the short way

✨ EXPRESSION EVALUATION ✨

🧮 The Functions 🧮

abs(x)

acos(x)

asin(x)

atan(x)

atan2(y, x)

between(x, min, max)

bitand(x, y) and bitor(x, y)

ceil(expr)

clip(x, min, max)

cos(x)

cosh(x)

eq(x, y)

exp(x)

floor(expr)

gauss(x)

gcd(x, y)

gt(x, y)

gte(x, y)

hypot(x, y)

if(x, y)

if(x, y, z)

ifnot(x, y)

ifnot(x, y, z)

isinf(x)

isnan(x)

ld(idx)

lerp(x, y, z)

log(x)

lt(x, y)

lte(x, y)

max(x, y)

min(x, y)

mod(x, y)

not(expr)

pow(x, y)

print(t) and print(t, l)

random(idx)

randomi(idx, min, max)

root(expr, max)

round(expr)

sgn(x)

sin(x)

sinh(x)

sqrt(expr)

squish(x)

st(idx, expr)

tan(x)

tanh(x)

taylor(expr, x) and taylor(expr, x, idx)

time(0)

trunc(expr)

while(cond, expr)

`abs(x)`

`acos(x)`

`asin(x)`

`atan(x)`

`atan2(y, x)`

`between(x, min, max)`

`bitand(x, y)` and `bitor(x, y)`

`ceil(expr)`

`clip(x, min, max)`

`cos(x)`

`cosh(x)`

`eq(x, y)`

`exp(x)`

`floor(expr)`

`gauss(x)`

`gcd(x, y)`

`gt(x, y)`

`gte(x, y)`

`hypot(x, y)`

`if(x, y)`

`if(x, y, z)`

`ifnot(x, y)`

`ifnot(x, y, z)`

`isinf(x)`

`isnan(x)`

`ld(idx)`

`lerp(x, y, z)`

`log(x)`

`lt(x, y)`

`lte(x, y)`

`max(x, y)`

`min(x, y)`

`mod(x, y)`

`not(expr)`

`pow(x, y)`

`print(t)` and `print(t, l)`

`random(idx)`

`randomi(idx, min, max)`

`root(expr, max)`

`round(expr)`

`sgn(x)`

`sin(x)`

`sinh(x)`

`sqrt(expr)`

`squish(x)`

`st(idx, expr)`

`tan(x)`

`tanh(x)`

`taylor(expr, x)` and `taylor(expr, x, idx)`

`time(0)`

`trunc(expr)`

`while(cond, expr)`

`av1` ✨

`hevc` 👑

`rawvideo` 🎥

`libdav1d` 🧚

`libdavs2` 🌈

`libuavs3d` ⭐

`libxevd` 💜

`v210` 🍓

`zero_new_constraint_set_flags` 🎀

`crop_left`, `crop_right`, `crop_top`, `crop_bottom` 🦋

`sei_user_data` ⭐

`delete_filler` 💀

`display_orientation` 🌸

`rotate` 🌀

`flip` 🪞

`level` 👑

`h264_mp4toannexb` 🧚

`h264_redundant_pps` 🎀

`hevc_metadata` 🌈

`aud`

`sample_aspect_ratio`

`video_format`, `video_full_range_flag`

`colour_primaries`, `transfer_characteristics`, `matrix_coefficients`