FFPROBE-ALL(1)

🌸 NAME 🌸

okay bestie, this is FFPROBE-ALL(1), aka the manual page for ffprobe — literally “ffprobe media prober,” and yes it’s giving main-character energy for media inspection 💅✨

✨ SYNOPSIS ✨

ffprobe [options] input_url

🦋 DESCRIPTION 🦋

ffprobe gathers info from multimedia streams and prints it in human- and machine-readable fashion. slay for both humans and scripts 💖

for example, it can be used to check the format of the container used by a multimedia stream, and the format and type of each media stream contained in it. iykyk, it’s the vibe check for containers + streams 🔥

if a url is specified in input, ffprobe will try to open and probe the url content. if the url cannot be opened or recognized as a multimedia file, a positive exit code is returned. periodt 💀

if no output is specified as output with o, ffprobe will write to stdout. no cap, that’s the default bestie 🫶

ffprobe may be employed both as a standalone application or in combination with a textual filter, which may perform more sophisticated processing, e.g. statistical processing or plotting. it’s giving “works alone and works with your bestie” energy 🌈

options are used to list some of the formats supported by ffprobe, or for specifying which information to display, and for setting how ffprobe will show it. literally the whole control panel 🎀

ffprobe output is designed to be easily parsable by a textual filter, and consists of one or more sections of a form defined by the selected writer, which is specified by the output_format option. yes, the writer chooses the shape of the output — iconic ⭐

sections may contain other nested sections, and are identified by a name (which may be shared by other sections), and an unique name. see the output of sections. nested sections are lowkey a whole aesthetic 💅

metadata tags stored in the container or in the streams are recognized and printed in the corresponding "FORMAT", "STREAM", "STREAM_GROUP_STREAM" or "PROGRAM_STREAM" section. all the little tags get their moment 💖

💅 OPTIONS 💅

all the numerical options, if not specified otherwise, accept a string representing a number as input, which may be followed by one of the SI unit prefixes, for example: 'K', 'M', or 'G'. numbers with cute little suffixes, we love to see it ✨

if i is appended to the SI unit prefix, the complete prefix will be interpreted as a unit prefix for binary multiples, which are based on powers of 1024 instead of powers of 1000. appending B to the SI unit prefix multiplies the value by 8. this allows using, for example: 'KB', 'MiB', 'G' and 'B' as number suffixes. math but make it fashion 🧚

options which do not take arguments are boolean options, and set the corresponding value to true. they can be set to false by prefixing the option name with "no". for example using -nofoo will set the boolean option with name "foo" to false. flippity flip, no cap 🦋

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

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

will load a filtergraph description from the file named filter.script. reading args from a file is such a flex 💅

🎀 Stream specifiers 🎀

some options are applied per-stream, e.g. bitrate or codec. stream specifiers are used to precisely specify which stream(s) a given option belongs to. basically: tell ffprobe exactly which stream you’re obsessed with 💖

a stream specifier is a string generally appended to the option name and separated from it by a colon. e.g. -codec:a:1 ac3 contains the "a:1" stream specifier, which matches the second audio stream. therefore, it would select the ac3 codec for the second audio stream. targeted, specific, iconic 🎯

a stream specifier can match several streams, so that the option is applied to all of them. e.g. the stream specifier in -b:a 128k matches all audio streams. group chat for audio streams only fr fr 🫶

an empty stream specifier matches all streams. for example, -codec copy or -codec: copy would copy all the streams without reencoding. empty specifier = everyone’s invited 🌈

possible forms of stream specifiers are:

stream_index — matches the stream with this index. e.g. -threads:1 4 would set the thread count for the second stream to 4. if stream_index is used as an additional stream specifier (see below), then it selects stream number stream_index from the matching streams. stream numbering is based on the order of the streams as detected by libavformat except when a stream group specifier or program ID is also specified. in this case it is based on the ordering of the streams in the group or program. numbering lore, but make it clear ✨
stream_type[:additional_stream_specifier] — stream_type is one of 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. type matching but make it cute 💅
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. 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. program-specific, no cap 🎀
#stream_id or i:stream_id — match the stream by stream id (e.g. PID in MPEG-TS container). stream id said “this is me” 💖
m:key[:value] — matches streams with the metadata tag key having the specified value. if value is not given, matches streams that contain the given tag with any value. the colon character ':' in key or value needs to be backslash-escaped. metadata matchmaking, but be careful with that colon bestie 😭
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 '+'. disposition list said slay 🌸
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. usable config only, we don’t do mystery streams 💀

lowkey important: note that in ffmpeg, matching by metadata will only work properly for input files.

🍓 Generic options 🍓

these options are shared amongst the ff* tools. group-chat shared options, cute 🫶

-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. help me help you 💅

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.
decoder=decoder_name — print detailed information about the decoder named decoder_name. use the -decoders option to get a list of all decoders.
encoder=encoder_name — print detailed information about the encoder named encoder_name. use the -encoders option to get a list of all encoders.
demuxer=demuxer_name — print detailed information about the demuxer named demuxer_name. use the -formats option to get a list of all demuxers and muxers.
muxer=muxer_name — print detailed information about the muxer named muxer_name. use the -formats option to get a list of all muxers and demuxers.
filter=filter_name — print detailed information about the filter named filter_name. use the -filters option to get a list of all filters.
bsf=bitstream_filter_name — print detailed information about the bitstream filter named bitstream_filter_name. use the -bsfs option to get a list of all bitstream filters.
protocol=protocol_name — print detailed information about the protocol named protocol_name. use the -protocols option to get a list of all protocols.

-version — show version.
-buildconf — show the build configuration, one option per line.
-formats — show available formats (including devices).
-demuxers — show available demuxers.
-muxers — show available muxers.
-devices — show available devices.
-codecs — show all codecs known to libavcodec. note that the term 'codec' is used throughout this documentation as a shortcut for what is more correctly called a media bitstream format. codec said “technically…” 💀
-decoders — show available decoders.
-encoders — show all available encoders.
-bsfs — show available bitstream filters.
-protocols — show available protocols.
-filters — show available libavfilter filters.
-pix_fmts — show available pixel formats.
-sample_fmts — show available sample formats.
-layouts — show channel names and standard channel layouts.
-dispositions — show stream dispositions.
-colors — show recognized color names.
-sources device[,opt1=val1[,opt2=val2]...] — show autodetected sources of the input device. some devices may provide system-dependent source names that cannot be autodetected. the returned list cannot be assumed to be always complete.

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. the returned list cannot be assumed to be always complete.

ffmpeg -sinks pulse,server=192.168.0.4

-loglevel [flags+]loglevel | -v [flags+]loglevel — set logging level and flags used by the library.

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

quiet, -8 — show literally nothing. Silent mode. 🤫 Like, not a peep.
panic, 0 — only fatal errors that could actually crash the process, like an assertion failure. Ngl this one isn't currently used for anything.
fatal, 8 — only fatal errors. These are the errors where the process absolutely cannot keep going. Periodt.
error, 16 — shows all errors, even the ones it can recover from.
warning, 24 — shows all warnings and errors. Anything about a possibly incorrect or unexpected event shows up here.
info, 32 — shows informative messages while processing. This is on top of warnings and errors. This is the default value, bestie.
verbose, 40 — same as info but, like, more verbose. Obviously.
debug, 48 — shows everything, including debugging info. The full tea. 🫖
trace, 56 — yeah, that's the whole entry. 💀

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

And another example that turns on repeated log output without touching the current state of the level prefix flag or loglevel:

ffmpeg [...] -loglevel +repeat

By default the program logs to stderr. If your terminal supports coloring, colors get used to mark errors and warnings. 🌈 You can turn log coloring off by setting the environment variable AV_LOG_FORCE_NOCOLOR, or force it on by setting AV_LOG_FORCE_COLOR.

-report — dumps the full command line and the log output into a file named program-YYYYMMDD-HHMMSS.log in the current directory. This file is genuinely useful for bug reports, fr fr. 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, those 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).

These are the options it recognizes:

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, and %% 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 (the 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 they won't appear in the report either.

-hide_banner — suppresses printing the banner. All FFmpeg tools normally show a copyright notice, build options and library versions, and this option is how you shut all that off. 🙈
-cpuflags flags (global) — lets you set and clear cpu flags. Heads up: this option is intended for testing. Do not use it unless you know what you're doing, no cap.

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) — overrides detection of CPU count. Same warning as above: this option is intended for testing. Do not use it unless you know what you're doing, bestie. 💅

ffmpeg -cpucount 2

-max_alloc bytes — sets 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 it if you don't understand the full consequence of doing so. Default is INT_MAX.

✨ AVOptions ✨

These options are provided directly by the libavformat, libavdevice and libavcodec libraries. Wanna see the list of available AVOptions? Use the -help option. They come split into two categories:

generic — these 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 are specific to the given container, device or codec. Private options are listed under their corresponding containers/devices/codecs. 🎀

Example time — to write an ID3v2.3 header instead of a default ID3v2.4 into 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, so you gotta attach a stream specifier to them:

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

In that example, a multichannel audio stream 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 the absolute index of the output stream. 👑

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

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

💖 Main options 💖

-f format — force the format to use.
-unit — show the unit of the displayed values.
-prefix — use SI prefixes for the displayed values. Unless the -byte_binary_prefix option is used, all the prefixes are decimal.
-byte_binary_prefix — force the use of binary prefixes for byte values.
-sexagesimal — use sexagesimal format HH:MM:SS.MICROSECONDS for time values. ⏱️
-pretty — prettify the format of the displayed values. It corresponds to the options -unit -prefix -byte_binary_prefix -sexagesimal. It's giving "all the glow-up flags at once" ✨
-output_format, -of, -print_format writer_name[=writer_options] — set the output printing format.

writer_name specifies the name of the writer, and writer_options specifies the options to be passed to the writer.

So for example, to print the output in JSON format, specify:

-output_format json

For more details on the available output printing formats, see the Writers section below.

-sections — print sections structure and section information, then exit. Heads up: the output is not meant to be parsed by a machine.
-select_streams stream_specifier — select only the streams specified by stream_specifier. This option affects only the options related to streams (e.g. show_streams, show_packets, etc.).

For example, to show only audio streams, you can use the command:

ffprobe -show_streams -select_streams a INPUT

To show only video packets belonging to the video stream with index 1:

ffprobe -show_packets -select_streams v:1 INPUT

-show_data — show payload data, as a hexadecimal and ASCII dump (other formats can be selected using -data_dump_format). Coupled with -show_packets, it'll dump the packets' data. Coupled with -show_streams, it'll dump the codec extradata. 🧚

The dump is printed as the data field. It may contain newlines.

-show_data_hash algorithm — show a hash of payload data, for packets with -show_packets and for codec extradata with -show_streams.
-data_dump_format format — select a format used for the data dumps enabled with the -show_data option. The default is xxd, which is a hexdump format compatible with the well-known xxd program. base64 is also supported.
-show_error — show information about the error found when trying to probe the input.

The error information is printed within a section with name ERROR. 😭

-show_format — show information about the container format of the input multimedia stream.

All the container format information is printed within a section with name FORMAT.

-show_entries section_entries — set the list of entries to show.

Entries are specified according to the following syntax. section_entries contains a list of section entries separated by :. Each section entry is composed of a section name (or unique name), optionally followed by a list of entries local to that section, separated by ,.

If a section name is specified but is followed by no =, all entries are printed to output, together with all the contained sections. Otherwise only the entries specified in the local section entries list are printed. In particular, if = is specified but the list of local entries is empty, then no entries will be shown for that section.

Note that the order of specification of the local section entries is not honored in the output — the usual display order gets retained. 🙃

The formal syntax is given by:

<LOCAL_SECTION_ENTRIES> ::= <SECTION_ENTRY_NAME>[,<LOCAL_SECTION_ENTRIES>]
<SECTION_ENTRY>         ::= <SECTION_NAME>[=[<LOCAL_SECTION_ENTRIES>]]
<SECTION_ENTRIES>       ::= <SECTION_ENTRY>[:<SECTION_ENTRIES>]

For example, to show only the index and type of each stream, and the PTS time, duration time, and stream index of the packets, you can specify the argument:

packet=pts_time,duration_time,stream_index : stream=index,codec_type

To show all the entries in the section format, but only the codec type in the section stream, specify the argument:

format : stream=codec_type

To show all the tags in the stream and format sections:

stream_tags : format_tags

To show only the title tag (if available) in the stream sections:

-show_packets — bestie, this dumps info about every single packet chilling inside the input multimedia stream 📦. Each packet gets its own little solo moment in a section literally named "PACKET". Iconic behavior.
-show_frames — shows you the tea on every frame and subtitle in the input multimedia stream 🎬. Every single frame gets printed in its own dedicated section named "FRAME" or "SUBTITLE" (yep, subtitles get their own label, we love that for them).
-show_log loglevel — surfaces the logging info from the decoder for each frame, based on whatever you set in loglevel (peep -loglevel) 📜. Heads up though: this one requires -show_frames or it simply won't do anything. Each log message shows up in its own dedicated section called "LOG".
-show_streams — spills all the info about each media stream in the input multimedia stream 🌊. Every stream's info lands in a dedicated section named "STREAM".
-show_programs — shows info about the programs and the streams inside them contained in the input multimedia stream 📺. Each media stream's info gets a dedicated section named "PROGRAM_STREAM".
-show_stream_groups — same energy but for stream groups and the streams they contain 🫂. Each media stream's info is printed within a dedicated section named "STREAM_GROUP_STREAM".
-show_chapters — shows info about the chapters stored in the format 📖. Every chapter gets printed within a dedicated section named "CHAPTER".
-count_frames — literally counts how many frames each stream has and reports the number inside that stream's own section 🔢. No cap, it's just counting.
-count_packets — same vibe, but counting packets per stream and reporting it in the corresponding stream section 📊.
-read_intervals read_intervals — only read the intervals you actually want, queen 👑. read_intervals has to be a sequence of interval specs separated by ",". ffprobe will seek to the interval's starting point and keep reading from there.

Every interval is made of two optional parts, separated by "%".

Part one = the interval start position 🎯. It's read as an absolute position, or as a relative offset from the current position if it's preceded by the "+" character. If you don't specify this first part at all, then no seeking happens when reading this interval.

Part two = the interval end position 🏁. Also read as absolute, or as a relative offset from the current position if it starts with "+". But here's the plot twist: if that offset specification starts with "#", it's interpreted as the number of packets to read (and no, the flushing packets don't count) starting from the interval's start. And if you skip the second part entirely, the program just reads until the end of the input. Simple as.

Lowkey important: seeking is not accurate, so the actual interval start point may differ from the position you specified 😭. Also, when you give an interval duration, the absolute end time is computed by adding the duration to the interval start point found by seeking the file — not to the start value you specified. So don't come for her when the numbers look off.

The formal syntax is given by:

<INTERVAL>  ::= [<START>|+<START_OFFSET>][%[<END>|+<END_OFFSET>]]
<INTERVALS> ::= <INTERVAL>[,<INTERVALS>]

A few examples follow, because we're nice like that 💅

Seek to time 10, read packets until 20 seconds after the found seek point, then seek to position "01:30" (1 minute and thirty seconds) and read packets until position "01:45":

10%+20,01:30%01:45

Read only 42 packets after seeking to position "01:23":

01:23%+#42

Read only the first 20 seconds from the start:

%+20

Read from the start until position "02:30":

%02:30

-show_private_data, -private — shows private data, aka data that depends on the format of whatever particular element is being shown 🙈. This one is enabled by default, but you might need to turn it off for specific uses — for example when you're creating XSD-compliant XML output. Do with that what you will.
-show_program_version — shows info related to the program version 🏷️. That version info is printed within a section named "PROGRAM_VERSION".
-show_library_versions — shows info related to library versions 📚. Version info for each library is printed within a section named "LIBRARY_VERSION".
-show_versions — shows info related to both program and library versions, bestie ✨. It's the equivalent of setting both -show_program_version and -show_library_versions at the same time. Two birds, one slay.
-show_pixel_formats — shows info about all the pixel formats FFmpeg supports 🎨. Pixel format info for each format gets printed within a section named "PIXEL_FORMAT".
-show_optional_fields value — okay so some writers (viz. JSON and XML) just omit printing fields that have invalid or non-applicable values, while other writers always print them 🤔. This option lets you control that behaviour. Valid values are "always"/1, "never"/0, and "auto"/-1. Default is auto.
-analyze_frames — analyzes frames and/or their side data up to whatever read interval you provided, handing you extra info that may be useful at the stream level 🧠✨. It must be paired with the -show_streams option or it will have absolutely no effect (sorry, them's the rules).

Currently, the extra fields this option provides when enabled are the "closed_captions" and "film_grain" fields.

For example, to analyze the first 20 seconds and populate these fields:

ffprobe -show_streams -analyze_frames -read_intervals "%+20" INPUT

-bitexact — forces bitexact output 🎯. Super useful when you want to produce output that isn't dependent on the specific build.
-i input_url — reads input_url 📥.
-o output_url — writes output to output_url 📤. If you don't specify it, the output gets sent to stdout.
-c:media_specifier codec_name, -codec:media_specifier codec_name — forces a specific decoder implementation for the stream identified by media_specifier 🎛️. That specifier can take the values "a" (audio), "v" (video), "s" (subtitle), and "d" (data). Pick your fighter.

💖 WRITERS 💖

okay so bestie, a writer is literally the output-format main character for ffprobe — it defines the output format that ffprobe adopts, and it gets used for printing, like, all the parts of the output. slay.

A writer may accept one or more arguments, and those arguments specify the options to adopt. The options are specified as a list of key=value pairs, separated by :.

All writers support the following options:

string_validation, sv — set the string validation mode. these are the accepted values:
- fail — the writer will fail immediately in case an invalid string (UTF-8) sequence or code point is found in the input. this is especially useful to validate input metadata.
- ignore — any validation error will be ignored. this will result in possibly broken output, especially with the json or xml writer. 💀
- replace — the writer will substitute invalid UTF-8 sequences or code points with the string specified with the string_validation_replacement.

Default value is replace.

string_validation_replacement, svr — set the replacement string to use in case string_validation is set to replace.

In case the option is not specified, the writer will assume the empty string, that is it will remove the invalid sequences from the input strings. 🫶

A description of the currently available writers follows.

🧚 default 🧚

Default format.

Print each section in the form:

[SECTION]
key1=val1
...
keyN=valN
[/SECTION]

Metadata tags are printed as a line in the corresponding FORMAT, STREAM, STREAM_GROUP_STREAM or PROGRAM_STREAM section, and are prefixed by the string TAG:.

A description of the accepted options follows.

nokey, nk — if set to 1 specify not to print the key of each field. Default value is 0.
noprint_wrappers, nw — if set to 1 specify not to print the section header and footer. Default value is 0.

💅 compact, csv 💅

Compact and CSV format.

The csv writer is equivalent to compact, but supports different defaults.

Each section is printed on a single line. If no option is specified, the output has the form:

section|key1=val1| ... |keyN=valN

Metadata tags are printed in the corresponding format or stream section. A metadata tag key, if printed, is prefixed by the string tag:.

The description of the accepted options follows.

item_sep, s — specify the character to use for separating fields in the output line. It must be a single printable character, it is | by default (, for the csv writer).
nokey, nk — if set to 1 specify not to print the key of each field. Its default value is 0 (1 for the csv writer).
escape, e — set the escape mode to use, default to c (csv for the csv writer).

It can assume one of the following values:

c — perform C-like escaping. Strings containing a newline (\n), carriage return (\r), a tab (\t), a form feed (\f), the escaping character (\) or the item separator character SEP are escaped using C-like fashioned escaping, so that a newline is converted to the sequence \n, a carriage return to \r, \ to \\ and the separator SEP is converted to \SEP.
csv — perform CSV-like escaping, as described in RFC4180. Strings containing a newline (\n), a carriage return (\r), a double quote ("), or SEP are enclosed in double-quotes.
none — perform no escaping.

print_section, p — print the section name at the beginning of each line if the value is 1, disable it with value set to 0. Default value is 1.

🌈 flat 🌈

Flat format.

A free-form output where each line contains an explicit key=value, such as streams.stream.3.tags.foo=bar. The output is shell escaped, so it can be directly embedded in sh scripts as long as the separator character is an alphanumeric character or an underscore (see sep_char option).

The description of the accepted options follows.

sep_char, s — separator character used to separate the chapter, the section name, IDs and potential tags in the printed field key.

Default value is ..

hierarchical, h — specify if the section name specification should be hierarchical. If set to 1, and if there is more than one section in the current chapter, the section name will be prefixed by the name of the chapter. A value of 0 will disable this behavior.

Default value is 1.

🍓 ini 🍓

INI format output.

Print output in an INI based format.

The following conventions are adopted:

all key and values are UTF-8
. is the subgroup separator
newline, \t, \f, \b and the following characters are escaped
\ is the escape character
# is the comment indicator
= is the key/value separator
: is not used but usually parsed as key/value separator

This writer accepts options as a list of key=value pairs, separated by :.

The description of the accepted options follows.

hierarchical, h — specify if the section name specification should be hierarchical. If set to 1, and if there is more than one section in the current chapter, the section name will be prefixed by the name of the chapter. A value of 0 will disable this behavior.

Default value is 1.

✨ json ✨

JSON based format.

Each section is printed using JSON notation.

The description of the accepted options follows.

compact, c — if set to 1 enable compact output, that is each section will be printed on a single line. Default value is 0.

For more information about JSON, see <http://www.json.org/>.

🦋 xml 🦋

XML based format.

The XML output is described in the XML schema description file ffprobe.xsd installed in the FFmpeg datadir.

An updated version of the schema can be retrieved at the url <http://www.ffmpeg.org/schema/ffprobe.xsd>, which redirects to the latest schema committed into the FFmpeg development source code tree.

Note that the output issued will be compliant to the ffprobe.xsd schema only when no special global output options (unit, prefix, byte_binary_prefix, sexagesimal etc.) are specified.

The description of the accepted options follows.

fully_qualified, q — if set to 1 specify if the output should be fully qualified. Default value is 0. This is required for generating an XML file which can be validated through an XSD file.
xsd_strict, x — if set to 1 perform more checks for ensuring that the output is XSD compliant. Default value is 0. This option automatically sets fully_qualified to 1.

For more information about the XML format, see <https://www.w3.org/XML/>.

⏱️ TIMECODE ⏱️

ffprobe supports Timecode extraction:

MPEG1/2 timecode is extracted from the GOP, and is available in the video stream details (-show_streams, see timecode).
MOV timecode is extracted from tmcd track, so is available in the tmcd stream metadata (-show_streams, see TAG:timecode).
DV, GXF and AVI timecodes are available in format metadata (-show_format, see TAG:timecode).

📚 SYNTAX 📚

This section documents the syntax and formats employed by the FFmpeg libraries and tools.

🌸 Quoting and escaping 🌸

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

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

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

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

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

🌟 Examples 🌟

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

Crime d\'Amour

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

'Crime d'\''Amour'

Include leading or trailing whitespaces using quoting:

'  this string starts and ends with whitespaces  '

Escaping and quoting can be mixed together:

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

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

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

🎀 Date 🎀

The accepted syntax is:

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

If the value is now it takes the current time.

Time is local time unless Z is appended, in which case it is interpreted as UTC. If the year-month-day part is not specified it takes the current year-month-day.

💜 Time duration 💜

There are two accepted syntaxes for expressing time duration.

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

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

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

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

In both expressions, the optional - indicates negative duration.

🌈 Examples 🌈

The following examples are all valid time duration:

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

🦋 Video size 🦋

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

The following abbreviations are recognized:

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

💖 Video rate 💖

Okay so this one sets the frame rate of a 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 number, a float number, or a valid video frame rate abbreviation. slay.

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.

⚠️ heads up bestie: a ratio with infinite (1/0) or negative value is considered valid, so you should check on the returned value if you want to exclude those values. like, actually check it.

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

🌈 Color 🌈

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

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

The string random will result in a random color. chaotic icon behavior, honestly.

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
Indigo — 0x4B0082
Ivory — 0xFFFFF0
Khaki — 0xF0E68C
Lavender — 0xE6E6FA
LavenderBlush — 0xFFF0F5
LawnGreen — 0x7CFC00
LemonChiffon — 0xFFFACD
LightBlue — 0xADD8E6
LightCoral — 0xF08080
LightCyan — 0xE0FFFF
LightGoldenRodYellow — 0xFAFAD2
LightGreen — 0x90EE90
LightGrey — 0xD3D3D3
LightPink — 0xFFB6C1
LightSalmon — 0xFFA07A
LightSeaGreen — 0x20B2AA
LightSkyBlue — 0x87CEFA
LightSlateGray — 0x778899
LightSteelBlue — 0xB0C4DE
LightYellow — 0xFFFFE0
Lime — 0x00FF00
LimeGreen — 0x32CD32
Linen — 0xFAF0E6
Magenta — 0xFF00FF
Maroon — 0x800000
MediumAquaMarine — 0x66CDAA
MediumBlue — 0x0000CD
MediumOrchid — 0xBA55D3
MediumPurple — 0x9370D8
MediumSeaGreen — 0x3CB371
MediumSlateBlue — 0x7B68EE
MediumSpringGreen — 0x00FA9A
MediumTurquoise — 0x48D1CC
MediumVioletRed — 0xC71585
MidnightBlue — 0x191970
MintCream — 0xF5FFFA
MistyRose — 0xFFE4E1
Moccasin — 0xFFE4B5
NavajoWhite — 0xFFDEAD
Navy — 0x000080
OldLace — 0xFDF5E6
Olive — 0x808000
OliveDrab — 0x6B8E23
Orange — 0xFFA500
OrangeRed — 0xFF4500
Orchid — 0xDA70D6
PaleGoldenRod — 0xEEE8AA
PaleGreen — 0x98FB98
PaleTurquoise — 0xAFEEEE
PaleVioletRed — 0xD87093
PapayaWhip — 0xFFEFD5
PeachPuff — 0xFFDAB9
Peru — 0xCD853F
Pink — 0xFFC0CB
Plum — 0xDDA0DD
PowderBlue — 0xB0E0E6
Purple — 0x800080
Red — 0xFF0000
RosyBrown — 0xBC8F8F
RoyalBlue — 0x4169E1
SaddleBrown — 0x8B4513
Salmon — 0xFA8072
SandyBrown — 0xF4A460
SeaGreen — 0x2E8B57
SeaShell — 0xFFF5EE
Sienna — 0xA0522D
Silver — 0xC0C0C0
SkyBlue — 0x87CEEB
SlateBlue — 0x6A5ACD
SlateGray — 0x708090
Snow — 0xFFFAFA
SpringGreen — 0x00FF7F
SteelBlue — 0x4682B4
Tan — 0xD2B48C
Teal — 0x008080
Thistle — 0xD8BFD8
Tomato — 0xFF6347
Turquoise — 0x40E0D0
Violet — 0xEE82EE
Wheat — 0xF5DEB3
White — 0xFFFFFF
WhiteSmoke — 0xF5F5F5
Yellow — 0xFFFF00
YellowGreen — 0x9ACD32

🦋 Channel Layout 🦋

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

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

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

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

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

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

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

A standard channel layout can be specified by the following:

the name of a single channel (e.g. FL, FR, FC, LFE, etc.)
the name of a standard channel layout (e.g. mono, stereo, 4.0, quad, 5.0, etc.)
a number of channels, in decimal, followed by c, which gives you the default channel layout for that number of channels (peep the function av_channel_layout_default). Heads up tho: not every channel count has a default layout, so don't come crying to me if it's not there 💀
- a number of channels, in decimal, followed by C, which gives you an unknown channel layout with that many channels specified. Fair warning bestie: not all channel layout specification strings support unknown channel layouts, so lower your expectations accordingly 🙈
- a channel layout mask, in hexadecimal starting with 0x (see the AV_CH_* macros in libavutil/channel_layout.h.

Before libavutil version 53 the trailing c used to specify a number of channels was optional, but now it is required — glow-up era, no shortcuts — while a channel layout mask can also be specified as a decimal number (if and only if it's not followed by c or C).

Also go look at the function av_channel_layout_from_string defined in libavutil/channel_layout.h. It's giving reference material ✨

🌸 EXPRESSION EVALUATION 🌸

When evaluating an arithmetic expression, FFmpeg uses an internal formula evaluator, implemented through the libavutil/eval.h interface. Lowkey iconic infrastructure ngl.

An expression may contain unary operators, binary operators, constants, and functions. All the good stuff 💅

Two expressions expr1 and expr2 can be combined to form another expression expr1;expr2. Both expr1 and expr2 are evaluated in turn, and the new expression evaluates to the value of expr2. It's giving "second one ate, honestly" 🍓

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

The following unary operators are available: +, -.

Some internal variables can be used to store and load intermediary results. You access them using the ld and st functions with an index argument that ranges from 0 to 9, specifying which internal variable you're grabbing. Little numbered lockers, very demure 🎀

The following functions are available — buckle up, this is a long list and every single one matters:

abs(x) — Compute absolute value of x.
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.
bitand(x, y), 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). 💀 fr fr, do not trust the big numbers

ceil(expr) — Round the value of expression expr upwards to the nearest integer. For example, ceil(1.5) is 2.0.
clip(x, min, max) — Return the value of x clipped between min and max.
cos(x) — Compute 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. Undefined behaviour era, tread carefully 🫶
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. Geometry class flashback 😭
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. ✨ the retrieve half of the equation
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), 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. Yes it prints AND returns, multitasking queen 👑
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. No cap, don't play with this one 🙈

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)).
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. Sad but true 💔

tan(x) — Compute tangent of x.
tanh(x) — Compute hyperbolic tangent of x.
taylor(expr, x), taylor(expr, x, idx) — Evaluate a Taylor series at x, given an expression representing the ld(idx)-th derivative of a function at 0.

When the series does not converge the result is undefined. Convergence is not optional bestie 😤

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. Little trick, very cunty of the math 🎀

time(0) — Return the current (wallclock) time in seconds.
trunc(expr) — Round the value of expression expr towards zero to the nearest integer. For example, trunc(-1.5) is -1.0.
while(cond, expr) — Evaluate expression expr while the expression cond is non-zero, and returns the value of the last expr evaluation, or NAN if cond was always false.

The following constants are available:

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

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

* works like AND
+ works like OR

For example the construct:

if (A AND B) then C

is equivalent to:

if(A*B, C)

In your C code, you can extend the list of unary and binary functions, and define recognized constants, so that they are available for your expressions. Go build ur own little ecosystem 👑

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. Slay, suffix literacy 🌈

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. Main character options, honestly 👑

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. Don't just yeet flags around and hope for the best 😭

Also some options are meant only for decoding or encoding. Read the room, basically 💅

okay so real talk, bestie: you can set these options by specifying -option value in the FFmpeg tools, or by setting the value explicitly in the AVCodecContext options, or by using the libavutil/opt.h API for programmatic use. 💅 the full list of supported options follows — and yes, we're doing all of them.

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

Possible values:

mv4 — use four motion vector by macroblock (mpeg4).
qpel — use 1/4 pel motion compensation.
loop — use loop filter.
qscale — use fixed qscale.
pass1 — use internal 2pass ratecontrol in first pass mode.
pass2 — use 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 first decoded frame in stream. Error AVERROR_INPUT_CHANGED is returned when a frame is dropped.
ildct — use interlaced DCT.
low_delay — force low delay.
global_header — place global headers in extradata instead of every keyframe.
bitexact — only write platform-, build- and time-independent data. (except (I)DCT). This ensures that file and data checksums are reproducible and match between platforms. Its primary use is for regression testing.
aic — apply H263 advanced intra coding / mpeg4 ac prediction.
ilme — apply interlaced motion estimation.
cgop — use closed gop.
output_corrupt — output even potentially corrupted frames.

time_base — rational number: set codec time base. It 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.
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.)
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.
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.
bug — flags (decoding,video): workaround 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.

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.
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.
block_align — integer.
rc_override_count — integer.
maxrate — integer (encoding,audio,video): set max bitrate tolerance (in bits/s). Requires bufsize to be set.
minrate — integer (encoding,audio,video): set min bitrate tolerance (in bits/s). Most useful in setting up a CBR encode. It 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 a IDCT compatible with the simple one
arm
altivec
sh4
simplearm
simplearmv5te
simplearmv6
simpleneon
xvid
faani — floating point AAN IDCT

slice_count — integer.
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.
aspect — rational number (encoding,video): set sample aspect ratio.
sar — rational number (encoding,video): set sample aspect ratio. Alias to aspect.
debug — flags (decoding/encoding,audio,video,subtitles): print specific debug info.

Possible values:

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

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

Possible values:

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

subcmp integer (encoding,video) — sets the sub pel me compare function. bestie you get to pick the whole vibe of how it compares:

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) — sets the macroblock compare function, same little menu of options as above:

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

ildctcmp integer (encoding,video) — sets the interlaced dct compare function. yep, same value list, third time's the charm 💅

Possible values:

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

dia_size integer (encoding,video) — sets the diamond type & size for motion estimation. it's giving ✨shapes✨, pick your fighter:
- (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 how many motion predictors carry over from the previous frame. lowkey nostalgic 🥲
precmp integer (encoding,video) — sets the pre motion estimation compare function. yes it's the same iconic menu again:

Possible values:

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

pre_dia_size integer (encoding,video) — sets the diamond type & size for the motion estimation pre-pass. like a little warm-up round 🎀
subq integer (encoding,video) — sets the sub pel motion estimation quality.
me_range integer (encoding,video) — sets the limit for the motion vectors range (1023 for DivX player). fyi the DivX player is out here with opinions.
global_quality integer (encoding,audio,video) — no extra notes, it's just vibes and a number ✨
slice_flags integer — same, just the flag, nothing more to say 🫶
mbd integer (encoding,video) — sets the macroblock decision algorithm (high quality mode). choose your destiny:

Possible values:

simple — use mbcmp (default)
bits — use fewest bits
rd — use best rate distortion

rc_init_occupancy integer (encoding,video) — sets how many bits should already be loaded into the rc buffer before decoding even starts. pre-gaming the buffer, we love to see it 💖
flags2 flags (decoding/encoding,audio,video,subtitles) — the big flag bag. Possible values:
- fast — allow non spec compliant speedup tricks.
- noout — skip bitstream encoding.
- ignorecrop — ignore cropping information from sps.
- local_header — place global headers at every keyframe instead of in extradata.
- chunks — frame data might be split into multiple chunks.
- showall — show all frames before the first keyframe.
- export_mvs — export motion vectors into frame side-data (see "AV_FRAME_DATA_MOTION_VECTORS") for codecs that support it. See also doc/examples/export_mvs.c.
- skip_manual — do not skip samples and export skip information as frame side data.
- ass_ro_flush_noop — do not reset ASS ReadOrder field on flush.
- icc_profiles — generate/parse embedded ICC profiles from/to colorimetry tags.
export_side_data flags (decoding/encoding,audio,video,subtitles) — what extra little goodies go out as side-data. Possible values:
- mvs — export motion vectors into frame side-data (see "AV_FRAME_DATA_MOTION_VECTORS") for codecs that support it. See also doc/examples/export_mvs.c.
- prft — export encoder Producer Reference Time into packet side-data (see "AV_PKT_DATA_PRFT") for codecs that support it.
- venc_params — export video encoding parameters through frame side data (see "AV_FRAME_DATA_VIDEO_ENC_PARAMS") for codecs that support it. At present, those are H.264 and VP9.
- film_grain — export film grain parameters through frame side data (see "AV_FRAME_DATA_FILM_GRAIN_PARAMS"). Supported at present by AV1 decoders.
- enhancements — export picture enhancement metadata through frame side data, e.g. LCEVC (see "AV_FRAME_DATA_LCEVC").

threads integer (decoding/encoding,video) — sets the number of threads used, in case the codec implementation you picked actually supports multi-threading. more cores more chaos 🔥

Possible values:

auto, 0 — automatically select the number of threads to set

Default value is auto.

dc integer (encoding,video) — sets intra_dc_precision.
nssew integer (encoding,video) — sets the nsse weight.
skip_top integer (decoding,video) — sets how many macroblock rows at the top get skipped. bye top rows 👋
skip_bottom integer (decoding,video) — sets how many macroblock rows at the bottom get skipped. and bye bottom rows 👋
profile integer (encoding,audio,video) — sets the encoder codec profile. default value is unknown, and encoder specific profiles are documented in the relevant encoder documentation (go read those, bestie).
level integer (encoding,audio,video) — sets the encoder level. this level depends on the specific codec and might correspond to the profile level. it's 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. tiny mode activate 🐜
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, skip_idct, skip_frame integer (decoding,video) — these make the decoder straight up discard processing depending on the frame type you select with the option value.

skip_loop_filter skips frame loop filtering, skip_idct skips frame IDCT/dequantization, skip_frame skips decoding entirely.

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) — refines the two motion vectors used in bidirectional macroblocks. two little vectors getting a glow-up ✨
keyint_min integer (encoding,video) — sets the minimum interval between IDR-frames.
refs integer (encoding,video) — sets the reference frames to consider for motion compensation.
trellis integer (encoding,audio,video) — sets rate-distortion optimal quantization.
mv0_threshold integer (encoding,video) — no extra notes attached, just the knob 🎛️
compression_level integer (encoding,audio,video) — same, it's a level, for compression, iconic.
bits_per_raw_sample integer — that's it, that's the option 💅
channel_layout integer (decoding/encoding,audio) — see the Channel Layout section in the ffmpeg-utils(1) manual for the required syntax. go check the other page, iykyk.
rc_max_vbv_use float (encoding,video) — a float, no description, moving on 🦋
rc_min_vbv_use float (encoding,video) — same deal, float energy.
color_primaries integer (decoding/encoding,video) — the color primaries menu. 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) — the transfer characteristics menu, and yes some of these have aliases so you've got options. 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

and the color_trc menu keeps serving these last few icons, no cap:

smpte2084 — SMPTE ST 2084
smpte428, smpte428_1 — SMPTE ST 428-1
arib-std-b67 — ARIB STD-B67
colorspace integer (decoding/encoding,video) — the color space menu is giving full rainbow, bestie. 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) — if you use it as an input parameter, it’s a hint to the decoder about which color_range the input has. Possible values:
- tv, mpeg, limited — MPEG (219*2^(n-8))
- pc, jpeg, full — JPEG (2^n-1)
chroma_sample_location integer (decoding/encoding,video) — possible values are left, center, topleft, top, bottomleft, bottom.
alpha_mode integer (decoding/encoding,video) — possible values are premultiplied and straight.
log_level_offset integer — sets the log level offset.
slices integer (encoding,video) — number of slices, used in parallelized encoding.
thread_type flags (decoding/encoding,video) — pick which multithreading methods to use. Using frame will increase decoding delay by one frame per thread, so clients that cannot 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.
- frame — decode more than one frame at once.

Default value is slice+frame.

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 — literally a rational number, iykyk.
sub_charenc encoding (decoding,subtitles) — sets the input subtitles character encoding.
field_order field_order (video) — sets or overrides the field order of the video. Possible values:
- progressive — progressive video
- tt — interlaced video, top field coded and displayed first
- bb — interlaced video, bottom field coded and displayed first
- tb — interlaced video, top coded first, bottom displayed first
- bt — interlaced video, bottom coded first, top displayed first
skip_alpha bool (decoding,video) — set to 1 to disable processing alpha (transparency). This works like the gray flag in the flags option, which skips chroma information instead of alpha. Default is 0.
codec_whitelist list (input) — a ","-separated list of allowed decoders. By default all are allowed.
dump_separator string (input) — the separator used to separate the fields printed on the command line about 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. This value can be used to avoid out of memory failures due to large images.
apply_cropping bool (decoding,video) — enables cropping if the cropping parameters are multiples of the required alignment for the left and top parameters. If the alignment is not met, the cropping will be partially applied to maintain alignment. Default is 1 (enabled).

Note: the required alignment depends on 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 configured elements in FFmpeg that let you decode multimedia streams. Okay so 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 all available decoders using the configure option --list-decoders.

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

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

🎥 VIDEO DECODERS 🎥

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

✨ av1 ✨

AOMedia Video 1 (AV1) decoder.

Options

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

🌸 hevc 🌸

HEVC (AKA ITU-T H.265 or ISO/IEC 23008-2) decoder.

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

Only the base layer is decoded by default.

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

Options

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

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

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

Same validity restrictions as for view_ids_available apply to this option.

🦋 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. Requires the presence of the libdav1d headers and library during configuration. You need to explicitly configure the build with --enable-libdav1d.

Options

The following options are supported by the libdav1d wrapper.

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

This option is deprecated and will be removed in the future. See the global option export_side_data to export Film Grain parameters instead of applying it.

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

🧚 libdavs2 🧚

AVS2-P2/IEEE1857.4 video decoder wrapper.

This decoder allows libavcodec to decode AVS2 streams with the davs2 library.

💜 libuavs3d 💜

AVS3-P2/IEEE1857.10 video decoder.

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

Options

The following option is supported by the libuavs3d wrapper.

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

🔥 libxevd 🔥

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

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

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

Options

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

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

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

👑 QSV Decoders 👑

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

Common Options

The following options are supported by all qsv decoders.

async_depth — internal parallelization depth. The higher the value, the higher the latency.
gpu_copy — a GPU-accelerated copy between video and system memory. Values:
- default
- on
- off

HEVC Options

Extra options for hevc_qsv.

load_plugin — a user plugin to load in an internal session. Values:
- 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.

Options

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

🌸 AUDIO DECODERS 🌸

okay so here's the tea on some of the audio decoders that are available rn 💖

🎧 ac3

the AC-3 audio decoder, periodt.

this decoder implements part of ATSC A/52:2010 and ETSI TS 102 366, plus the undocumented RealAudio 3 (a.k.a. dnet). yes bestie, even the mysterious undocumented stuff 💀

AC-3 Decoder Options

-drc_scale value — the Dynamic Range Scale Factor. it's the factor applied to dynamic range values coming from the AC-3 stream, and it gets applied exponentially ✨ the default value is 1. there are 3 notable scale factor ranges:
- drc_scale == 0 — DRC disabled. produces full range audio 🎶
- 0 < drc_scale <= 1 — DRC enabled. applies a fraction of the stream DRC value. audio reproduction lands somewhere between full range and full compression 🫶
- drc_scale > 1 — DRC enabled. applies drc_scale asymmetrically. loud sounds get fully compressed, soft sounds get enhanced 💅

🌈 flac

the FLAC audio decoder.

this decoder aims to implement the complete FLAC specification from Xiph, no cap 👑

FLAC Decoder options

-use_buggy_lpc — okay so the lavc FLAC encoder used to produce buggy streams with high lpc values (like the default value 🙈). this option makes it possible to decode such streams correctly by using lavc's old buggy lpc logic for decoding. literally fixing the past with the past.

🦋 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 🤫 iykyk

🎀 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 🫶 two for the price of one, slay.

🌸 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) ✨

👑 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 🔥 (yeah, both of those, together, don't forget the second one bestie)

🍓 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 fr fr.

🧚 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 as the narrowband one.

🎶 libopus

libopus decoder wrapper.

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

an FFmpeg native decoder for Opus exists, so users can decode Opus without this library 🌟 iconic, honestly.

💖 SUBTITLES DECODERS 💖

🎀 libaribb24

Okay so this one is an ARIB STD-B24 caption decoder. It implements profiles A and C of the ARIB STD-B24 standard. Slay, standardization!

✨ libaribb24 Decoder Options ✨

-aribb24-base-path path — sets the base path for the libaribb24 library. It's used for reading config files (for custom unicode conversions) and for dumping non-text symbols as images under that location. It's unset by default, bestie.
-aribb24-skip-ruby-text boolean — tells the decoder wrapper to skip text blocks that contain half-height ruby text. It's enabled by default. Not me skipping the ruby text 💅

🦋 libaribcaption

Yet another ARIB STD-B24 caption decoder, but this one uses the external libaribcaption library. 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. Iconic international range 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". Also important: if both libaribb24 and libaribcaption are enabled, libaribcaption decoder precedes. It's giving main character energy 👑

🌸 libaribcaption Decoder Options 🌸

-sub_type subtitle_type — specifies the format of the decoded subtitles. The vibes 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.

-caption_encoding encoding_scheme — specifies the encoding scheme of input subtitle text. Options:
- auto — automatically detect text encoding (default).
- jis — 8bit-char JIS encoding defined in ARIB STD B24. This encoding is used in Japan for ISDB captions.
- utf8 — UTF-8 encoding defined in ARIB STD B24. This encoding is used in Philippines for ISDB-T captions.
- latin — Latin character encoding defined in ABNT NBR 15606-1. This encoding is used in South America for SBTVD / ISDB-Tb captions.
-font font_name[,font_name2,...] — specify a comma-separated list of font family names to be used for bitmap or ass type subtitle rendering. Only the first font name is used for ass type subtitle. If not specified, it uses the internally defined default font family.
-ass_single_rect boolean — 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 supports it. This option works under FreeType or DirectWrite renderer with Adobe-Japan1 compliant fonts. e.g., IBM Plex Sans JP, Morisawa BIZ UDGothic, Morisawa BIZ UDMincho, Yu Gothic, Yu Mincho, and Meiryo. The default is true.
-canvas_size image_size — specify the resolution of the canvas to render subtitles to; usually, this should be frame size of 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 "-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 "-canvas_size" option is not specified. If "-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 values are:
- -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 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. Lowkey everywhere.

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-bits hexadecimal numbers (without 0x prefix) separated by commas, for example "0d00ee, ee450d, 101010, eaeaea, 0ce60b, ec14ed, ebff0b, 0d617a, 7b7b7b, d1d1d1, 7b2a0e, 0d950c, 0f007b, cf0dec, cfa80c, 7c127b".

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

🍓 libzvbi-teletext

Libzvbi allows libavcodec to decode DVB teletext pages and DVB teletext subtitles. It requires the presence of the libzvbi headers and library during configuration. You need to explicitly configure the build with "--enable-libzvbi".

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.
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.
txt_duration — sets the display duration of the decoded teletext pages or subtitles in milliseconds. Default value is -1 which means infinity or until the next subtitle event comes.
txt_transparent — force 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 ⭐

Okay bestie, when you configure your FFmpeg build, all the supported bitstream filters are enabled by default. You can list all available ones using the configure option "--list-bsfs".

You can disable all the bitstream filters using the configure option "--disable-bsfs", and selectively enable any bitstream filter using the option "--enable-bsf=BSF", or you can disable a particular bitstream filter using the option "--disable-bsf=BSF". Mix and match, that's the vibe 💅

The option "-bsfs" of the ff* tools will display the list of all the supported bitstream filters included in your build.

The ff* tools have a -bsf option applied per stream, taking a comma-separated list of filters, whose parameters follow the filter name after a '='.

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

ok bestie, so down below is the full rundown of every bitstream filter that's currently available, plus their params if they have any 💖

🎀 aac_adtstoasc

Turns an MPEG-2/4 AAC ADTS thingy into an MPEG-4 Audio Specific Configuration bitstream. Fancy.

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

So when do you need it? Classic case: you're copying an AAC stream from a raw ADTS AAC or an MPEG-TS container over to MP4A-LATM, or to an FLV file, or to MOV/MP4 files and related formats like 3GP or M4A. Heads up though: it gets auto-inserted for MP4A-LATM and MOV/MP4 and related formats, so it's already handling itself there 💅

🦋 av1_metadata

Modifies metadata embedded in an AV1 stream. Little glow-up for your headers.

td — insert or remove temporal delimiter OBUs in all temporal units of the stream.
- insert — drops a TD at the beginning of every TU that doesn't already have one.
- remove — deletes the TD from the beginning of every TU that has one.
color_primaries, transfer_characteristics, matrix_coefficients — set the color description fields in the stream (peep AV1 section 6.4.2 👀).
color_range — set the color range in the stream (see AV1 section 6.4.2; note that this can't 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, no cap.
- 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 isn't also set 🫶
delete_padding — deletes Padding OBUs. bye bye padding ✨

🍓 chomp

Removes zero padding at the end of a packet. That's it. Snack-sized filter, we stan.

💜 dca_core

Extracts the core from a DCA/DTS stream, dropping extensions like DTS-HD. Just the core, bestie.

👑 dovi_rpu

Manipulates Dolby Vision metadata in a HEVC/AV1 bitstream, and optionally enables metadata compression.

strip — if enabled, strips 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 was intended to be added. Like, it checks first, it's polite 💅

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

If you don't specify it, it's assumed to be k.

Example time! This ffmpeg command forces a global header (so individual packet headers get disabled) in the H.264 packets made by the "libx264" encoder, but then corrects them by adding the header stored in extradata to the key packets:

ffmpeg -i INPUT -map 0 -flags:v +global_header -c:v libx264 -bsf:v dump_extra out.ts

🧚 dv_error_marker

Blocks in DV that are marked as damaged get replaced by blocks of the specified color. Damage control but make it aesthetic.

color — the color to replace damaged blocks by.
sta — a 16 bit mask specifying 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. Core only, no extras 💖

⭐ eia608_to_smpte436m

Converts from an "EIA_608" stream to an "SMPTE_436M_ANC" data stream, wrapping the closed captions in CTA-708 CDP VANC packets. Iconic little pipeline honestly.

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

Extracts the in-band extradata. Ok so let me explain this one properly, because it slaps:

Certain codecs let the long-term headers (like MPEG-2 sequence headers, or H.264/HEVC (VPS/)SPS/PPS) be transmitted either "in-band" — meaning as part of the bitstream that contains the coded frames — or "out of band" (like, at the container level). That second form is what FFmpeg calls "extradata" ✨

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

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

🔥 filter_units

Removes units with types in — or not in — a given set from the stream. Bouncer energy, honestly.

pass_types — list of unit types or ranges of unit types to pass through while removing all others. You write it 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 are removed and all others are 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 note that if the stream contains inline parameter sets then the output may be unusable if they are removed 😭

Example time, bestie:

To remove all non-VCL NAL units from an H.264 stream:

ffmpeg -i INPUT -c:v copy -bsf:v 'filter_units=pass_types=1-5' OUTPUT

To remove all AUDs, SEI and filler from an H.265 stream:

ffmpeg -i INPUT -c:v copy -bsf:v 'filter_units=remove_types=35|38-40' OUTPUT

To remove all user data from a MPEG-2 stream, including Closed Captions:

ffmpeg -i INPUT -c:v copy -bsf:v 'filter_units=remove_types=178' OUTPUT

To remove all SEI from a H264 stream, including Closed Captions:

ffmpeg -i INPUT -c:v copy -bsf:v 'filter_units=remove_types=6' OUTPUT

To remove all prefix and suffix SEI from a HEVC stream, including Closed Captions and dynamic HDR:

ffmpeg -i INPUT -c:v copy -bsf:v 'filter_units=remove_types=39|40' OUTPUT

💅 hapqa_extract

Extracts the Rgb or Alpha part of an HAPQA file — with no recompression — so you can create an HAPQ or an HAPAlphaOnly file. Surgical 💖

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

Modifies metadata embedded in an H.264 stream. The big one, get comfy ✨

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 depends 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 zeros constraint_set4_flag and constraint_set5_flag in the SPS. These bits were reserved in a previous version of the H.264 spec, so some hardware decoders literally require them to be zero. The result of zeroing this is still a valid bitstream. 🫶
crop_left, crop_right, crop_top, crop_bottom — these set the frame cropping offsets in the SPS. These values will replace the current ones if the stream is already cropped. They’re set in pixels, and note that some sizes may not be representable if the chroma is subsampled or the stream is interlaced (see H.264 section 7.4.2.1.1). 💅
sei_user_data — inserts a string as SEI unregistered user data. The argument must be of the form UUID+string, where the UUID is as hex digits possibly separated by hyphens, and the string can be anything. For example, 086f3693-b7b3-4f2c-9653-21492feee5b8+hello will insert the string "hello" associated with the given UUID. ✨
delete_filler — deletes both filler NAL units and filler SEI messages. Bye, filler. 🫡
display_orientation — insert, extract or remove Display orientation SEI messages. See H.264 section D.1.27 and D.2.27 for syntax and semantics. The modes are pass, insert, remove, extract; default is pass. Insert mode works in conjunction with rotate and flip options. Any pre-existing Display orientation messages will be removed in insert or remove mode. extract mode attaches the display matrix to the packet as side data. 👑
rotate — set rotation in display orientation SEI (anticlockwise angle in degrees). Range is -360 to +360. Default is NaN. 🌀
flip — set flip in display orientation SEI. The values are horizontal and 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. The argument must be the name of a level (for example, 4.2), a level_idc value (for example, 42), or the special name auto indicating that the filter should attempt to guess the level from the input stream properties. 🎚️

`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"). 🌈

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. No cap. ✨

`h264_redundant_pps`

This applies a specific fixup to some Blu-ray BDMV H264 streams which contain redundant PPSs. The PPSs modify irrelevant parameters of the stream, confusing other transformations which require the correct extradata. 😭

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. 💀

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. Slay. 💅

`hevc_metadata`

Modify metadata embedded in an HEVC stream. 🦋

aud — insert or remove AUD NAL units in all access units of the stream. Values: insert, remove.
sample_aspect_ratio — set the sample aspect ratio in the stream in the VUI parameters.
video_format, video_full_range_flag — set the video format in the stream (see H.265 section E.3.1 and table E.2).
colour_primaries, transfer_characteristics, matrix_coefficients — set the colour description in the stream (see H.265 section E.3.1 and tables E.3, E.4 and E.5).
chroma_sample_loc_type — set the chroma sample location in the stream (see H.265 section E.3.1 and figure E.1).
tick_rate — set the tick rate in the VPS and VUI parameters (time_scale / num_units_in_tick). Combined with num_ticks_poc_diff_one, this can set a constant framerate in the stream. Note that it is likely to be overridden by container parameters when the stream is in a container. ⏱️
num_ticks_poc_diff_one — set poc_proportional_to_timing_flag in VPS and VUI and use this value to set num_ticks_poc_diff_one_minus1 (see H.265 sections 7.4.3.1 and E.3.1). Ignored if tick_rate is not also set. 🙈
crop_left, crop_right, crop_top, crop_bottom — set the conformance window cropping offsets in the SPS. These values will replace the current ones if the stream is already cropped. These fields are set in pixels. Note that some sizes may not be representable if the chroma is subsampled (H.265 section 7.4.3.2.1). 🎀
width, height — set width and height after crop. 📐
level — set the level in the VPS and SPS. See H.265 section A.4 and tables A.6 and A.7. The argument must be the name of a level (for example, 5.1), a general_level_idc value (for example, 153 for level 5.1), or the special name auto indicating that the filter should attempt to guess the level from the input stream properties. 🎚️

`hevc_mp4toannexb`

Convert an HEVC/H.265 bitstream from length prefixed mode to start code prefixed mode (as defined in the Annex B of the ITU-T H.265 specification). This is required by some streaming formats, typically the MPEG-2 transport stream format (muxer "mpegts"). 🌈

For example, to remux an MP4 file containing an HEVC stream to mpegts format with ffmpeg, you can use the command:

ffmpeg -i INPUT.mp4 -codec copy -bsf:v hevc_mp4toannexb OUTPUT.ts

Please note that this filter is auto-inserted for MPEG-TS (muxer "mpegts") and raw HEVC/H.265 (muxer "h265" or "hevc") output formats. ✨

`imxdump`

Modifies the bitstream to fit in MOV and to be usable by the Final Cut Pro decoder. This filter only applies to the mpeg2video codec, and is likely not needed for Final Cut Pro 7 and newer with the appropriate -tag:v. 🎬

For example, to remux 30 MB/sec NTSC IMX to MOV:

ffmpeg -i input.mxf -c copy -bsf:v imxdump -tag:v mx3n output.mov

`mjpeg2jpeg`

Convert MJPEG/AVI1 packets to full JPEG/JFIF packets. 📸

MJPEG is a video codec wherein each video frame is essentially a JPEG image. The individual frames can be extracted without loss, e.g. by

ffmpeg -i ../some_mjpeg.avi -c:v copy frames_%d.jpg

Unfortunately, these chunks are incomplete JPEG images, because they lack the DHT segment required for decoding. Quoting from http://www.digitalpreservation.gov/formats/fdd/fdd000063.shtml:

Avery Lee, writing in the rec.video.desktop newsgroup in 2001, commented that "MJPEG, or at least the MJPEG in AVIs having the MJPG fourcc, is restricted JPEG with a fixed -- and omitted -- Huffman table. The JPEG must be YCbCr colorspace, it must be 4:2:2, and it must use basic Huffman encoding, not arithmetic or progressive. . . . You can indeed extract the MJPEG frames and decode them with a regular JPEG decoder, but you have to prepend the DHT segment to them, or else the decoder won't have any idea how to decompress the data. The exact table necessary is given in the OpenDML spec."

This bitstream filter patches the header of frames extracted from an MJPEG stream (carrying the AVI1 header ID and lacking a DHT segment) to produce fully qualified JPEG images. 🔥

ffmpeg -i mjpeg-movie.avi -c:v copy -bsf:v mjpeg2jpeg frame_%d.jpg
exiftran -i -9 frame*.jpg
ffmpeg -i frame_%d.jpg -c:v copy rotated.avi

`mjpegadump`

Add an MJPEG A header to the bitstream, to enable decoding by Quicktime. 🍓

`mov2textsub`

Extract a representable text file from MOV subtitles, stripping the metadata header from each subtitle packet. See also the text2movsub filter. 📝

`mpeg2_metadata`

Modify metadata embedded in an MPEG-2 stream. 🦋

display_aspect_ratio — set the display aspect ratio in the stream. The following fixed values are supported: 4/3, 16/9, 221/100. Any other value will result in square pixels being signalled instead (see H.262 section 6.3.3 and table 6-3).
frame_rate — set the frame rate in the stream. This is constructed from a table of known values combined with a small multiplier and divisor — if the supplied value is not exactly representable, the nearest representable value will be used instead (see H.262 section 6.3.3 and table 6-4). ⏱️
video_format — set the video format in the stream (see H.262 section 6.3.6 and table 6-6).
colour_primaries, transfer_characteristics, matrix_coefficients — set the colour description in the stream (see H.262 section 6.3.6 and tables 6-7, 6-8 and 6-9). 🌈

`mpeg4_unpack_bframes`

Unpack DivX-style packed B-frames. 🧚

DivX-style packed B-frames are not valid MPEG-4 and were only a workaround for the broken Video for Windows subsystem. They use more space, can cause minor AV sync issues, require more CPU power to decode (unless the player has some decoded picture queue to compensate the 2,0,2,0 frame per packet style) and cause trouble if copied into a standard container like mp4 or mpeg-ps/ts, because MPEG-4 decoders may not be able to decode them, since they are not valid MPEG-4. 😭

For example, to fix an AVI file containing an MPEG-4 stream with DivX-style packed B-frames using ffmpeg, you can use the command:

ffmpeg -i INPUT.avi -codec copy -bsf:v mpeg4_unpack_bframes OUTPUT.avi

`noise`

Damages the contents of packets or simply drops them without damaging the container. Can be used for fuzzing or testing error resilience/concealment. 💀

Parameters:

amount — accepts an expression whose evaluation per-packet determines how often bytes in that packet will be modified. A value below 0 will result in a variable frequency. Default is 0 which results in no modification. However, if neither amount nor drop is specified, amount will be set to -1. See below for accepted variables 👇
drop — okay so this takes an expression that gets evaluated per-packet, and its value decides whether that packet gets dropped. If it evaluates to a positive value, the packet is dropped. If it evaluates to a negative value, there’s a variable chance it gets dropped, roughly inverse in proportion to the magnitude of the value. Default is 0, which results in no drops. See below for accepted variables, bestie.
dropamount — takes a non-negative integer, and that assigns a variable chance of the packet being dropped, roughly inverse in proportion to the value. Default is 0, which results in no drops. This option is kept for backwards compatibility and is equivalent to setting drop to a negative value with the same magnitude, i.e. dropamount=4 is the same as drop=-4. Ignored if drop is also specified.

Both amount and drop accept expressions containing the following variables:

n — The index of the packet, starting from zero.
tb — The timebase for packet timestamps.
pts — Packet presentation timestamp.
dts — Packet decoding timestamp.
nopts — Constant representing AV_NOPTS_VALUE.
startpts — First non-AV_NOPTS_VALUE PTS seen in the stream.
startdts — First non-AV_NOPTS_VALUE DTS seen in the stream.
duration, d — Packet duration, in timebase units.
pos — Packet position in input; may be -1 when unknown or not set.
size — Packet size, in bytes.
key — Whether packet is marked as a keyframe.
state — A pseudo random integer, primarily derived from the content of packet payload.

🎀 Examples 🎀

Apply modification to every byte but don’t drop any packets.

ffmpeg -i INPUT -c copy -bsf noise=1 output.mkv

Drop every video packet not marked as a keyframe after timestamp 30s but do not modify any of the remaining packets.

ffmpeg -i INPUT -c copy -bsf:v noise=drop='gt(pts*tb\,30)*not(key)' output.mkv

Drop one second of audio every 10 seconds and add some random noise to the rest.

ffmpeg -i INPUT -c copy -bsf:a noise=amount=-1:drop='between(mod(pts*tb\,10)\,9\,10)' output.mkv

🦋 `null` 🦋

This bitstream filter passes the packets through unchanged. Literally the chillest.

🍓 `pcm_rechunk` 🍓

Repacketize PCM audio to a fixed number of samples per packet or a fixed packet rate per second. This is similar to the asetnsamples audio filter but works on audio packets instead of audio frames.

nb_out_samples, n — Set the number of samples per each output audio packet. The number is intended as the number of samples per each channel. Default value is 1024.
pad, p — If set to 1, the filter will pad the last audio packet with silence, so that it will contain the same number of samples (or roughly the same number of samples, see frame_rate) as the previous ones. Default value is 1.
frame_rate, r — This option makes the filter output a fixed number of packets per second instead of a fixed number of samples per packet. If the audio sample rate is not divisible by the frame rate then the number of samples will not be constant but will vary slightly so that each packet will start as close to the frame boundary as possible. Using this option has precedence over nb_out_samples.

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

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

🎀 `pgs_frame_merge` 🎀

Merge 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).

💅 `prores_metadata` 💅

Modify color property metadata embedded in prores stream.

color_primaries — Set the color primaries. Available values are:
- auto — Keep the same color primaries property (default).
- unknown
- bt709
- bt470bg — BT601 625
- smpte170m — BT601 525
- bt2020
- smpte431 — DCI P3
- smpte432 — P3 D65
transfer_characteristics — Set the color transfer. Available values are:
- auto — Keep the same transfer characteristics property (default).
- unknown
- bt709 — BT 601, BT 709, BT 2020
- smpte2084 — SMPTE ST 2084
- arib-std-b67 — ARIB STD-B67
matrix_coefficients — Set the matrix coefficient. Available values are:
- auto — Keep the same colorspace property (default).
- unknown
- bt709
- smpte170m — BT 601
- bt2020nc

Set Rec709 colorspace for each frame of the file

ffmpeg -i INPUT -c copy -bsf:v prores_metadata=color_primaries=bt709:color_trc=bt709:colorspace=bt709 output.mov

Set Hybrid Log-Gamma parameters for each frame of the file

ffmpeg -i INPUT -c copy -bsf:v prores_metadata=color_primaries=bt2020:color_trc=arib-std-b67:colorspace=bt2020nc output.mov

🧚 `remove_extra` 🧚

Remove extradata from packets.

It accepts the following parameter:

freq — Set which frame types to remove extradata from.
- k — Remove extradata from non-keyframes only.
- keyframe — Remove extradata from keyframes only.
- e, all — Remove extradata from all frames.

⭐ `setts` ⭐

Set PTS and DTS in packets.

It accepts the following parameters:

ts, pts, dts — Set expressions for PTS, DTS or both.
duration — Set expression for duration.
time_base — Set output time base.
prescale — Set whether to convert time fields to user-set output time base before evaluation of expressions. Defaults to 0.

The expressions are evaluated through the eval API and can contain the following constants:

N — The count of the input packet. Starting from 0.
TS — The demux timestamp in input in case of ts or dts option or presentation timestamp in case of pts option.
POS — The original position in the file of the packet, or undefined if undefined for the current packet
DTS — The demux timestamp in input.
PTS — The presentation timestamp in input.
DURATION — The duration in input.
STARTDTS — The DTS of the first packet.
STARTPTS — The PTS of the first packet.
PREV_INDTS — The previous input DTS.
PREV_INPTS — The previous input PTS.
PREV_INDURATION — The previous input duration.
PREV_OUTDTS — The previous output DTS.
PREV_OUTPTS — The previous output PTS.
PREV_OUTDURATION — The previous output duration.
NEXT_DTS — The next input DTS.
NEXT_PTS — The next input PTS.
NEXT_DURATION — The next input duration.
TB — The timebase in which time fields are denominated. The user-set output timebase if prescale is enabled, else the input timebase.
TB_OUT — The output timebase.
SR — The sample rate of stream packet belongs.
NOPTS — The AV_NOPTS_VALUE constant.

For example, to set PTS equal to DTS (not recommended if B-frames are involved):

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

🌈 `showinfo` 🌈

Log basic packet information. Mainly useful for testing, debugging, and development.

🦋 `smpte436m_to_eia608` 🦋

Convert from a SMPTE_436M_ANC data stream to a EIA_608 stream, extracting the closed captions from CTA-708 CDP VANC packets, and ignoring all other data.

🎀 `text2movsub` 🎀

Convert text subtitles to MOV subtitles (as used by the mov_text codec) with metadata headers.

🔥 `trace_headers` 🔥

Log trace output containing all syntax elements in the coded stream headers (everything above the level of individual coded blocks). This can be useful for debugging low-level stream issues.

Supports AV1, H.264, H.265, (M)JPEG, MPEG-2 and VP9, but depending on the build only a subset of these may be available.

👑 `truehd_core` 👑

Extract the core from a TrueHD stream, dropping ATMOS data.

💜 `vp9_metadata` 💜

Modify metadata embedded in a VP9 stream.

color_space — Set the color space value in the frame header. Note that any frame set to RGB will be implicitly set to PC range and that RGB is incompatible with profiles 0 and 2.
- unknown
- bt601
- bt709
- smpte170
- smpte240
- bt2020
- rgb
color_range — Set the color range value in the frame header. Note that any value imposed by the color space will take precedence over this value.
- tv
- pc

🌸 `vp9_superframe` 🌸

Merge VP9 invisible (alt-ref) frames back into VP9 superframes. This fixes merging of split/segmented VP9 streams where the alt-ref frame was split from its visible counterpart.

🦋 `vp9_superframe_split` 🦋

Split VP9 superframes into single frames.

💅 `vp9_raw_reorder` 💅

Given a VP9 stream with correct timestamps but possibly out of order, insert additional show-existing-frame packets to correct the ordering.

🎀 FORMAT OPTIONS 🎀

okay so the libavformat library ships with some generic global options, and you can set them on, like, all the muxers and demuxers. and on top of that? each muxer or demuxer might support its own so-called private options that are specific to that one component. bestie behaviour honestly 💅

you 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 going through all of them 🫶

avioflags flags (input/output) — Possible values:
- direct — reduce buffering 🦋
probesize integer (input) — sets the probing size in bytes, like, the size of the data that gets analyzed to get stream information. a higher value = detecting more info in case it's dispersed into the stream, but it will increase latency, no cap. must be an integer not lesser than 32. it's 5000000 by default.
max_probe_packets integer (input) — the max number of buffered packets when probing a codec. default is 2500 packets ✨
packetsize integer (output) — sets packet size.
fflags flags — sets format flags. some are implemented for a limited number of formats only, iykyk.

Possible values for input files:

discardcorrupt — discard corrupted packets 🗑️
fastseek — enable fast, but inaccurate seeks for some formats (yes it's giving "we ball" energy)
genpts — generate missing PTS if DTS is present
igndts — ignore DTS if PTS is also set. in case the PTS is set, the DTS value is set to NOPTS. this is ignored when the nofillin flag is set 👀
ignidx — ignore index
nobuffer — reduce the latency introduced by buffering during initial input streams analysis
nofillin — do not fill in missing values in packet fields that can be exactly calculated
noparse — disable AVParsers, and this needs +nofillin too, bestie
sortdts — try to interleave output packets by DTS. at present, available only for AVIs with an index

Possible values for output files:

autobsf — automatically apply bitstream filters as required by the output format. enabled by default ⭐
bitexact — only write platform-, build- and time-independent data. this ensures that file and data checksums are reproducible and match between platforms. its primary use is for regression testing 🧪
flush_packets — write out packets immediately
shortest — stop muxing at the end of the shortest stream. it may be needed to increase max_interleave_delta to avoid flushing the longer streams before EOF

seek2any integer (input) — allow seeking to non-keyframes on demuxer level when supported if set to 1. default is 0.
analyzeduration integer (input) — specify how many microseconds are analyzed to probe the input. a higher value enables detecting more accurate information, but increases latency. it defaults to 5,000,000 microseconds = 5 seconds ⏳
cryptokey hexadecimal string (input) — set decryption key 🔐
indexmem integer (input) — set max memory used for timestamp index (per stream).
rtbufsize integer (input) — set max memory used for buffering real-time frames.
fdebug flags (input/output) — print specific debug info.

Possible values:

ts

max_delay integer (input/output) — set maximum muxing or demuxing delay in microseconds.
fpsprobesize integer (input) — set number of frames used to probe fps.
audio_preload integer (output) — set microseconds by which audio packets should be interleaved earlier.
chunk_duration integer (output) — set microseconds for each chunk.
chunk_size integer (output) — set size in bytes for each chunk.
err_detect, f_err_detect flags (input) — set error detection flags. heads up: f_err_detect is deprecated and should be used only via the ffmpeg tool.

Possible values:

crccheck — verify embedded CRCs
bitstream — detect bitstream specification deviations
buffer — detect improper bitstream length
explode — abort decoding on minor error detection 💥
careful — consider things that violate the spec and have not been seen in the wild as errors
compliant — consider all spec non compliancies as errors
aggressive — consider things that a sane encoder should not do as an error

max_interleave_delta integer (output) — set maximum buffering duration for interleaving. the duration is expressed in microseconds, and defaults to 10000000 (10 seconds).

so 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 🫶

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 ✨
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).

ngl there's a small typo in the original here ("delayed bt") — it 100% means delayed by, bestie 💅

format_whitelist list (input) — "," separated list of allowed demuxers. by default all are allowed.
dump_separator string (input) — separator used to separate the fields printed on the command line about the Stream parameters. for example, to separate the fields with newlines and indentation:

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

max_streams integer (input) — specifies the maximum number of streams. this can be used to reject files that would require too many resources due to a large number of streams 🚫
skip_estimate_duration_from_pts bool (input) — skip estimation of input duration if it requires an additional probing for PTS at end of file. at present, applicable for MPEG-PS and MPEG-TS.
duration_probesize integer (input) — set probing size, in bytes, for input duration estimation when it actually requires an additional probing for PTS at end of file (at present: MPEG-PS and MPEG-TS). it's aimed at users interested in better durations probing for itself, or indirectly because using the concat demuxer, for example. the typical use case is an MPEG-TS CBR with a high bitrate, high video buffering and ending cleaning with similar PTS for video and audio: in such a scenario, the large physical gap between the last video packet and the last audio packet makes it necessary to read many bytes in order to get the video stream duration. another use case is where the default probing behaviour only reaches a single video frame which is not the last one of the stream due to frame reordering, so the duration is not accurate. setting this option has a performance impact even for small files because the probing size is fixed. default behaviour is a general purpose trade-off, largely adaptive, but the probing size will not be extended to get streams durations at all costs. must be an integer not lesser than 1, or 0 for default behaviour 🦋
strict, f_strict integer (input/output) — specify how strictly to follow the standards. f_strict is deprecated and should be used only via the ffmpeg tool.

Possible values:

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

⚠️ note: experimental decoders can pose a security risk, do not use this for decoding untrusted input. like, fr fr, don't 💀

🌈 Format stream specifiers 🌈

format stream specifiers let you select one or more streams that match specific properties 👑

the exact semantics of stream specifiers is defined by the avformat_match_stream_specifier() function declared in the libavformat/avformat.h header and documented in the Stream specifiers section in the ffmpeg(1) manual. go read it, bestie, it's giving essential lore 📚

🌸 DEMUXERS 🌸

demuxers are configured elements in FFmpeg that can read the multimedia streams from a particular type of file ✨ like, each one is a lil specialist for its own file type, and honestly? iconic behaviour 🫶

okay so bestie, when you configure your FFmpeg build, every single supported demuxer is enabled by default — yes, all of them, the whole squad. You can list every available one with the configure option --list-demuxers. 💅

And when you want to flip the switch the other way? You can nuke all the demuxers with the configure option --disable-demuxers, and then selectively enable exactly one with --enable-demuxer=DEMUXER, or disable just one with --disable-demuxer=DEMUXER. Cherry-picking queen behavior. 🍓

The -demuxers option of the ff* tools will show you the list of enabled demuxers. And if you want the combined deluxe view — enabled demuxers and muxers together — use -formats. Two lists, one flex. ✨

The description of some of the currently available demuxers follows. Buckle up, it's giving buffet. 🌈

aa

Audible Format 2, 3, and 4 demuxer.

This demuxer is the one you call when you need to demux Audible Format 2, 3, and 4 (.aa) files. Audiobook girlies stay winning. 🎧

aac

Raw Audio Data Transport Stream AAC demuxer.

This demuxer is used to demux an ADTS input that contains a single AAC stream, along with any ID3v1/2 or APE tags sitting inside it. Tags and audio, holding hands. 🫶

apng

Animated Portable Network Graphics demuxer.

This demuxer is used to demux APNG files. Here's the tea: all headers, but the PNG signature, up to (but not including) the first fcTL chunk, are transmitted as extradata. Frames then get split as being all the chunks between two fcTL ones, or between the last fcTL and IEND chunks. Slice it like a cake. 🎂

-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. The default of 0 imposes no limit at all, 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 is used to demux ASF files and MMS network streams. Versatile icon. ⭐

-no_resync_search bool — do not try to resynchronize by looking for a certain optional start code.

concat

Virtual concatenation script demuxer.

This demuxer reads a list of files and other directives from a text file, then demuxes them one after the other — as if all their packets had been muxed together in the first place. It's giving "one long file, but make it a group project." 👑

The timestamps in the files get adjusted so that the first file starts at 0 and each next file starts exactly where the previous one finishes. Heads up though: this is done globally and may cause gaps if all streams do not have exactly the same length. No cap.

All files must have the same streams (same codecs, same time base, etc.). Every single one. No exceptions, bestie.

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 exists so you can override the duration stored in each file. 😌

Syntax

The script is a text file in extended-ASCII, with one directive per line. Empty lines, leading spaces and lines starting with # are ignored. The following directive is recognized:

file path — path to a file to read; special characters and spaces must be escaped with backslash or single quotes. All subsequent file-related directives apply to that file.
ffconcat version 1.0 — identify the script type and version. To make FFmpeg recognize the format automatically, this directive must appear exactly as is (no extra space or byte-order-mark) on the very first line of the script. First line. Not second. Not "close enough." 💀
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 isn't available or isn't accurate. If the duration is set for all files, then it is possible to seek in the whole concatenated video. Iconic.
inpoint timestamp — In point of the file. When the demuxer opens the file it instantly seeks to the specified timestamp. Seeking is done so that all streams can be presented successfully at In point. This directive works best with intra frame codecs, because for non-intra frame ones you will usually get extra packets before the actual In point, and the decoded content will most likely contain frames before In point too. For each file, packets before the file In point will have timestamps less than the calculated start timestamp of the file (negative in case of the first file), and the duration of the files (if not specified by the duration directive) will be reduced based on their specified In point. Because of potential packets before the specified In point, packet timestamps may overlap between two concatenated files. Watch out for that overlap drama. 🙈
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 the demuxer will not output packets with a decoding timestamp greater or equal to Out point. This directive works best with intra frame codecs and formats where all streams are tightly interleaved. For non-intra frame codecs you will usually get additional packets with presentation timestamp after Out point, therefore the decoded content will most likely contain frames after Out point too. If your streams are not tightly interleaved you may not get all the packets from all streams before Out point, and you may only will be able to decode the earliest stream until Out point. The duration of the files (if not specified by the duration directive) will be reduced based on their specified Out point.
file_packet_metadata key=value — metadata of the packets of the file. The specified metadata will be set for each file packet. You can specify this directive multiple times to add multiple metadata entries. This directive is deprecated — use file_packet_meta instead. She's retired but still on the payroll. 💀
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.
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. Trust issues, honestly. 😭
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. The default is 1. Keep this on unless you know exactly what you're doing, bestie. 🔥
auto_convert — if set to 1, try to perform automatic conversions on packet data to make the streams concatenable. The default is 1. Currently, the only conversion is adding the h264_mp4toannexb bitstream filter to H.264 streams in MP4 format. This is necessary in particular if there are resolution changes.
segment_time_metadata — if set to 1, every packet will contain the lavf.concat.start_time and the lavf.concat.duration packet metadata values, which are the start_time and the duration of the respective file segments in the concatenated output expressed in microseconds. The duration metadata is only set if it is known based on the concat file. The default is 0.

Examples

Use absolute filenames and include some comments:

# my first filename
file /mnt/share/file-1.wav
# my second filename including whitespace
file '/mnt/share/file 2.wav'
# my third filename including whitespace plus single quote
file '/mnt/share/file 3'\''.wav'

Allow for input format auto-probing, use safe filenames and set the duration of the first file:

ffconcat version 1.0

file file-1.wav
duration 20.0

file subdir/file-2.wav

dash

Dynamic Adaptive Streaming over HTTP demuxer.

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. Mirror mirror on the wall. 🪞

Options

This demuxer accepts the following option:

cenc_decryption_key — default 16-byte key, in hex, to decrypt files encrypted using ISO Common Encryption (CENC/AES-128 CTR; ISO/IEC 23001-7).
cenc_decryption_keys — dictionary of 16-byte key ID => 16-byte key, both in hex, to decrypt files encrypted using ISO Common Encryption (CENC/AES-128 CTR; ISO/IEC 23001-7).

dvdvideo

DVD-Video demuxer, powered by libdvdnav and libdvdread.

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 is still effort, okay? 💅

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

⚠️ This demuxer does NOT have decryption code of any kind. You are on your own working with encrypted DVDs, and should not expect support on the matter. Like, genuinely — none. 🙈

Underlying playback is handled by libdvdnav, and structure parsing by libdvdread. FFmpeg must be built with GPL library support available as well as the configure switches --enable-libdvdnav and --enable-libdvdread. ✨

You're gonna need either the "title number" you want or the exact PGC/PG coordinates. Tons of open-source DVD players and tools can help you hunt that info down. If you don't specify anything, the demuxer just defaults to title 1, which lowkey works for a lot of discs — but honestly, because this format is so flexible, bestie, please check manually. There are so many discs out there authored weirdly or with straight-up invalid headers. 💀

If your input is an actual real DVD drive, heads up: some drives will silently fail when reading bad sectors and just hand you random bits instead, which is effectively corrupt data. This is especially prominent on aging or rotting discs. You'd need a second pass plus integrity checks to even detect that corruption. And like — this is not an FFmpeg issue, okay? Not our problem, not our fault. 🙈

Background

DVD-Video is not a directly accessible, linear container format in the traditional sense. Nope! Instead it lets you do complex, programmatic playback of carefully muxed MPEG-PS streams stored in headerless VOB files. To the end-user these streams are just known as "titles", but the actual logical playback sequence is defined by one or more "PGCs" — Program Group Chains — inside the title. The PGC is in turn made of multiple "PGs", or "Programs", which are the actual video segments (and for a typical video feature, they're sequentially ordered). The PGC structure, along with stream layout and metadata, lives in IFO files that need to be parsed. Easiest way to think about it: PGCs are basically playlists. 🎀

A real DVD player relies on user GUI interaction via menus and an internal VM to drive the direction of demuxing. Usually the user either navigates via menus or gets automatically redirected to the PGC of their choice. During that whole process and the playback after it, the DVD player's internal VM also maintains a state and executes instructions that can create jumps to different sectors mid-playback. That's literally why libdvdnav is involved — a linear read of the MPEG-PS blobs on the disc (the VOBs) is just not enough to produce the right sequence in many cases. Periodt.

There are many other DVD structures (a whole long subject, seriously) that won't be discussed here. NAV packets in particular are handled by this demuxer to build accurate timing, but they are not emitted as a stream. For a good high-level understanding, go read: 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 meaning "end at the last possible chapter". 🏁
angle int — the video angle number, which refers to what's essentially an additional video stream 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 to a particular angle in different regions. This option will not affect the region code of a real DVD drive if one is used as input. Not applicable to menus. Default is 0, "world". 🌍
menu bool — demux menu assets instead of navigating a title. Requires the 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's a VMG menu (root-level). Default is 1, the menu of the first VTS.
pgc int — the entry PGC to start playback, in conjunction with pg. An alternative to setting title. Chapter markers are not supported at this time. Must be explicitly set for menus. Default is 0, automatically resolved from the value of title.
pg int — the entry PG to start playback, in conjunction with pgc. An alternative to setting title. Chapter markers are not supported at this time. Default is 1, the first PG of the PGC.
preindex bool — enable this to get 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. That's non-ideal extra work for real optical drives 😭 — it's 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 are many discs with filler segments at the start of the PGC, often full of junk data intended to control a real DVD player's buffering speed, with no other materially useful value whatsoever. 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. 🎮

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 that contains the alpha channel in addition to the ordinary video. Slay.

✨ imf

Interoperable Master Format demuxer.

This demuxer presents the 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. For live network streams, if you force the format you may use the live_flv option instead of flv to survive timestamp discontinuities. KUX is an flv variant used on the Youku platform.

ffmpeg -f flv -i myfile.flv ...
ffmpeg -f live_flv -i rtmp://<any.server>/anything/key ....

-flv_metadata bool — allocate the streams according to the onMetaData array content.
-flv_ignore_prevtag bool — ignore the size of previous tag value.
-flv_full_metadata bool — output all context of the onMetadata.

🍓 gif

Animated GIF demuxer. It accepts the following options:

min_delay — set the minimum valid delay between frames, in hundredths of seconds. Range is 0 to 6000. Default value is 2.
max_gif_delay — set the maximum valid delay between frames, in hundredths of seconds. Range is 0 to 65535. Default value is 65535 (nearly eleven minutes), which is 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 info 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 the overlay filter is used to end the output video at the length of the shortest input file — which in this case is input.mp4, since 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 the 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's recommended to set the whitelists correctly instead of depending on extensions. Enabled by default.
max_reload — maximum number of times an 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.

so this demuxer is out here reading from a whole list of image files that you hand it via a pattern 🎀 The syntax and the meaning of that pattern is specified by the option pattern_type.

The pattern can carry a suffix, and that suffix is what's used to automatically figure out the format of the images sitting inside the files.

Big rule, no exceptions: the size, the pixel format, and the format of each image must be the same for all the files in the sequence. all of them. same. or it's not gonna slay 💅

This demuxer accepts the following options:

framerate — sets the frame rate for the video stream. defaults to 25.
loop — if you set it to 1, it loops over the input. default value is 0.
pattern_type — selects which pattern type is used to interpret the filename you provided. it accepts one of these values:
- none — disables pattern matching, so the video will only contain the specified image. use this one if you don't want to create sequences from multiple images and your filenames might 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 that's inclusively contained between start_number and start_number+start_number_range-1, and all the following numbers must be sequential. no skipping, bestie 🙈
  - 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 doesn't necessarily have to 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 you don't specify it, the pixel format gets guessed from the first image file in the sequence.
start_number — sets the index of the file matched by the image file pattern where reading starts. 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, it will set the frame timestamp to the modification time of the image file. note that monotonity of timestamps is not provided: images go in the same order as without this option. default value is 0. if set to 2, it will set the frame timestamp to the modification time of the image file in nanosecond precision. 🕰️
video_size — sets the video size of the images to read. if not specified, the video size gets guessed from the first image file in the sequence.
export_path_metadata — if set to 1, it 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. it's giving retro arcade realness 👑

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 (sorry, greedy). track indexes start at 0. default is to pick the first track. the number of tracks is exported as the tracks metadata entry.
sample_rate — sets the sampling rate of the exported track. range is 1000 to 999999. default is 44100.
max_size (bytes) — the demuxer buffers the entire file into memory, ngl. 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. so you can have your tracker bops and your little rainbow visuals 🌈

it accepts the following options:

noise_reduction — applies a simple low-pass filter. can be 1 (on) or 0 (off). default is 0.
reverb_depth — sets the amount of reverb. range 0-100. default is 0.
reverb_delay — sets the delay in ms, clamped to 40-250 ms. default is 0.
bass_amount — applies bass expansion a.k.a. XBass or megabass. range is 0 (quiet) to 100 (loud). default is 0. turn it up fr fr 🔊
bass_range — sets the cutoff, i.e. the upper-bound for bass frequencies. range is 10-100 Hz. default is 0.
surround_depth — applies a Dolby Pro-Logic surround effect. range is 0 (quiet) to 100 (heavy). default is 0.
surround_delay — sets the surround delay in ms, clamped to 5-40 ms. default is 0.
max_size — the demuxer buffers the entire file into memory. adjust this value to set the maximum buffer size, which in turn acts as a ceiling for the size of files that can be read. range is 0 to 100 MiB. 0 removes the buffer size limit (not recommended 🙈). default is 5 MiB.
video_stream_expr — a string which is evaluated using the eval API to assign colors to the generated video stream. variables which can be used are x, y, w, h, t, speed, tempo, order, pattern and row.
video_stream — generates the video stream. can be 1 (on) or 0 (off). default is 0.
video_stream_w — sets the video frame width in 'chars' where one char indicates 8 pixels. range 20-512. default is 30.
video_stream_h — sets the video frame height in 'chars' where one char indicates 8 pixels. range 20-512. default is 30.
video_stream_ptxt — prints metadata on the video stream. includes speed, tempo, order, pattern, row and ts (time in ms). can be 1 (on) or 0 (off). default is 1.

🎼 libopenmpt 🦋

libopenmpt based module demuxer ✨

see https://lib.openmpt.org/libopenmpt/ for more information.

some files have multiple subsongs (tracks), and you can set that 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).

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

Options 🎀

this demuxer accepts the following options:

enable_drefs — enables loading of external tracks, disabled by default. enabling this can theoretically leak information in some use cases 😬
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. like, known known. do not play yourself 🙈
seek_streams_individually — when seeking, it identifies the closest point in each stream individually and demuxes 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 — ignores 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 — modifies 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 — bestie, just don't parse the chapters. That includes GoPro 'HiLight' tags/moments, so those little highlight moments get skipped too. Heads up tho: chapters only get parsed when the input is seekable. Default is false.
use_mfra_for — for seekable fragmented input, this sets the fragment's starting timestamp from the media fragment random access box — but only if it's actually there. The options are giving:
- auto — auto-detect whether to set mfra timestamps as PTS or DTS (default)
- dts — set mfra timestamps as DTS
- pts — set mfra timestamps as PTS
- 0 — don't use the mfra box to set timestamps at all
use_tfdt — for fragmented input, set the fragment's starting timestamp to "baseMediaDecodeTime" from the "tfdt" box. Default is enabled, which prefers using the "tfdt" box to set DTS. Disable it and it'll use the "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 — export 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 — export the entire contents of the XMP_ box and uuid box as a string with the key "xmp". Small lil gotcha: if "export_all" is set and this option isn't, the contents of the XMP_ box still get exported — just under the key "XMP_" instead. Default is false.
activation_bytes — the 4-byte key you need to decrypt Audible AAX and AAX+ files. Peep the Audible AAX subsection below 👇
audible_fixed_key — a fixed key used for handling Audible AAX/AAX+ files. It's already been pre-set, so you shouldn't need to specify it. It's handled, bestie 💅
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, bestie) to decrypt files encrypted using ISO Common Encryption (CENC/AES-128 CTR; ISO/IEC 23001-7).
max_stts_delta — okay so very high sample deltas written in a trak's stts box might occasionally be intentional, but usually they're written in error, or they're being used to store a negative value for dts correction when treated as signed 32-bit integers. This option lets you set an upper limit — anything beyond it 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 "uint32" range. Math slay 🔥

interleaved_read — interleave packets from multiple tracks at demuxer level. For badly interleaved files, this prevents playback issues caused by large gaps between packets in different tracks, because MOV/MP4 genuinely do not have packet placement requirements. However — and this is a real one — it can cause excessive seeking on very badly interleaved files, since it's seeking between tracks. So disabling it may prevent I/O issues, at the expense of playback. Pick your struggle 🫠

🎧 Audible AAX 🎧

Audible AAX files are encrypted M4B files, and you can decrypt them by specifying a 4 byte activation secret. Literally that's it:

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 — set the size limit for looking up a new synchronization. Default value is 65536.
skip_unknown_pmt — skip PMTs for programs not defined in the PAT. Default value is 0.
fix_teletext_pts — override teletext packet PTS and DTS values with the timestamps calculated from the PCR of the first program which the teletext stream is part of and is not discarded. Default value is 1. Set this option to 0 if you want your teletext packet PTS and DTS values left untouched, periodt.
ts_packetsize — output option carrying the raw packet size in bytes. It shows the detected raw packet size — and no, you can't set it yourself, the user has no power here 🙈
scan_all_pmts — scan and combine all PMTs. 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 — reuse existing streams when a PMT's version is updated and elementary streams move to different PIDs. Default value is 0.
max_packet_size — set the maximum size, in bytes, of the 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 demuxer lets you read MJPEG where each frame is represented as a part of a multipart/x-mixed-replace stream. Cute.

strict_mime_boundary — the default implementation applies a relaxed standard to multi-part MIME boundary detection, to prevent regression with the numerous existing endpoints that aren't generating a proper MIME MJPEG stream. Turning this option on by setting it to 1 gives you a stricter check of the boundary value. Strict queen behavior 👑

📼 rawvideo 📼

Raw video demuxer. This one lets you read raw video data. Since there's no header specifying the assumed video parameters, you have to specify them yourself so the data can be decoded correctly. No header, no hand-holding, you're the main character now.

This demuxer accepts the following options:

framerate — set input video frame rate. Default value is 25.
pixel_format — set the input video pixel format. Default value is "yuv420p".
stride — set frame line size in bytes; only required if there's extra padding. For multiplane formats, stride is a list of line size for each plane.
video_size — set the input video size. This value must be specified explicitly. No default, no vibes-based guessing.

So like, 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, you'd run:

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", 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 information on the format, see . (Yep, the manpage literally just ends there with a period — that's all the tea we get 💀)

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

🌸 Examples 🌸

Render CC to ASS using the built-in decoder:

ffmpeg -i CC.rcwt.bin CC.ass

Quick note bestie: if your output looks empty, you may have to manually set the decoder's data_field option to pick the desired CC substream. Don't panic, it's not broken, it's just shy.

Convert an RCWT backup to Scenarist (SCC) format:

ffmpeg -i CC.rcwt.bin -c:s copy CC.scc

Heads up tho: the SCC format does not support all of the possible CC extensions that can be stored in RCWT (such as EIA-708). So some of that data just doesn't make the trip 😭

🧠 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 this:

-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

Okay so 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 — easy slay. But 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 moment the script is read, and the script layout gets frozen according to that reference. What does that mean in practice? If the script is played directly, the actual times will match the absolute timestamps up to the sound controller's clock accuracy — but if the user somehow pauses playback or seeks, all the times get shifted accordingly. Forewarned is forearmed ✨

🎤 tedcaptions 🎤

JSON captions used for <http://www.ted.com/>. TED doesn't 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. Sneaky but effective 💅

This demuxer accepts the following option:

start_time — set 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 those include a 15s intro. Gotta account for the intro, periodt.

Example: convert the captions to a format most players understand:

ffmpeg -i http://www.ted.com/talks/subtitles/id/1/lang/en talk1-en.srt

🦋 vapoursynth 🦋

Vapoursynth wrapper. Due to security concerns, Vapoursynth scripts will not be autodetected, so the input format has to be forced. For ff* CLI tools, add "-f vapoursynth" before the input "-i yourscript.vpy". Security girlies stay winning 🔒

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. (Yes, it's pointing at wav. Go look at wav. We're getting to wav. Patience bestie 😌)

🎀 wav 🎀

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

max_size — specify 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. Let the demuxer cook 🔥

✨ 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 using the metadata muxer/demuxer. It's giving "metadata glow-up" and honestly? Obsessed. 🎀

The file format is as follows:

A file consists of a header and a number of metadata tags divided into sections, each on its own line.
The header is a ;FFMETADATA string, followed by a version number (now 1).
Metadata tags are of the form key=value
Immediately after header follows global metadata
After global metadata there may be sections with per-stream/per-chapter metadata.
A section starts with the section name in uppercase (i.e. STREAM or CHAPTER) in brackets ([, ]) and ends with next section or end of file.
At the beginning of a chapter section there may be an optional timebase to be used for start/end values. It must be in form TIMEBASE=num/den, where num and den are integers. If the timebase is missing then start/end times are assumed to be in nanoseconds.

Next a chapter section must contain chapter start and end times in form START=num, END=num, where num is a positive integer.

Empty lines and lines starting with ; or # are ignored.
Metadata keys or values containing special characters (=, ;, #, \ and a newline) must be escaped with a backslash \.
Note that whitespace in metadata (e.g. foo = bar) is considered to be a part of the tag (in the example above key is foo , value is bar).

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. No cap, this is the metadata pipeline of all time. 💅

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. Very "main character energy, but make it configurable." ✨

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

The list of supported options follows:

protocol_whitelist list (input) — Set a ,-separated list of allowed protocols. ALL matches all protocols. Protocols prefixed by - are disabled. All protocols are allowed by default but protocols used by an another protocol (nested protocols) are restricted to a per protocol subset. 🫶

🌐 PROTOCOLS 🌐

Protocols are configured elements in FFmpeg that enable access to resources that require specific protocols. They're lowkey the little doorways into all the data. 🚪✨

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. Choose your own protocol adventure, bestie. 🦋

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. Literally "don't leave me hanging" energy. 😭

A description of the currently available protocols follows.

amqp

Advanced Message Queueing Protocol (AMQP) version 0-9-1 is a broker based publish-subscribe communication protocol. It's giving "messages through the group chat, but enterprise." 💬

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 /. Slay, defaults. 👑

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. That's the whole routing situation, fr fr. 🎀

The following options are supported:

exchange — Sets the exchange to use on the broker. RabbitMQ has several predefined exchanges: amq.direct is the default exchange, where the publisher and subscriber must have a matching routing_key; amq.fanout is the same as a broadcast operation (i.e. the data is forwarded to all queues on the fanout exchange independent of the routing_key); and amq.topic is similar to amq.direct, but allows for more complex pattern matching (refer to the RabbitMQ documentation).
routing_key — Sets the routing key. The default value is amqp. The routing key is used on the amq.direct and amq.topic exchanges to decide whether packets are written to the queue of a subscriber.
pkt_size — Maximum size of each packet sent/received to the broker. Default is 131072. Minimum is 4096 and max is any large value (representable by an int). When receiving packets, this sets an internal buffer size in FFmpeg. It should be equal to or greater than the size of the published packets to the broker. Otherwise the received message may be truncated causing decoding errors. ⚠️
connection_timeout — The timeout in seconds during the initial connection to the broker. The default value is rw_timeout, or 5 seconds if rw_timeout is not set.
delivery_mode mode — Sets the delivery mode of each message sent to broker. The following values are accepted:
- persistent — Delivery mode set to persistent (2). This is the default value. Messages may be written to the broker's disk depending on its setup.
- non-persistent — Delivery mode set to non-persistent (1). Messages will stay in broker's memory unless the broker is under memory pressure.

async

Asynchronous data filling wrapper for input stream. Fill data in a background thread, to decouple I/O operation from demux thread. It's the "bestie handles the icky I/O so you can keep vibing" protocol. ✨

async:<URL>
async:http://host/resource
async:cache:http://host/resource

bluray

Read BluRay playlist. Fancy disc energy. 💿

The accepted options are:

angle — BluRay angle
chapter — Start chapter (1...N)
playlist — Playlist to read (BDMV/PLAYLIST/?????.mpls)

Examples:

Read longest playlist from BluRay mounted to /mnt/bluray:

bluray:/mnt/bluray

Read angle 2 of playlist 4 from BluRay mounted to /mnt/bluray, start from chapter 2:

-playlist 4 -angle 2 -chapter 2 bluray:/mnt/bluray

cache

Caching wrapper for input stream. Cache the input stream to temporary file. It brings seeking capability to live streams. Literally giving live streams a little memory, iconic. 🧠✨

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. It's the "line up all the URLs and walk in formation" protocol. 🎀

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. The shell said "not today" and you gotta backslash it. 💀

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. Secret little stream, we see you. 🔐

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 literally data just chilling inside the link. 💅

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. Lowkey powerful, highkey useful. 🫶

The accepted syntax is:

fd: -fd <file_descriptor>

If fd is not specified, by default the stdout file descriptor will be used for writing, stdin for reading. Unlike the pipe protocol, fd protocol has seek support if it corresponding to a regular file. fd protocol doesn't support pass file descriptor via URL for security. 🛡️

This protocol accepts the following options:

blocksize — sets the max block size for I/O ops, in bytes. Default is INT_MAX, which basically means "no limit, go off bestie". Setting it kinda low makes the user's cancel request get noticed faster — which is a whole vibe when data transmission is slow 😭
fd — sets the file descriptor. That's it. That's the option. 💅

🦋 `file` 🦋

The file access protocol, babes. You read from a file or write to a file. Iconic, lowkey the blueprint.

A file URL looks like this:

file:<filename>

where filename is the path of the file you're reading. Easy.

An URL with no protocol prefix is assumed to be a file URL, no questions asked. And depending on how it's built, an URL that looks like a Windows path with the drive letter chilling at the front also gets assumed to be a file URL — tho that's usually not the case in builds for unix-like systems, iykyk.

So for example, to read from a file input.mpeg with ffmpeg you'd run:

ffmpeg -i file:input.mpeg output.mpeg

This protocol accepts these options, bestie:

truncate — truncates existing files on write, if set to 1. A value of 0 means no truncating, hands off. Default value is 1.
blocksize — sets the max block size for I/O ops, in bytes. Default is INT_MAX, which means the requested block size isn't limited at all. Setting it reasonably low improves user termination request reaction time, and that's genuinely valuable when your files are on a slow medium 🐌
follow — if set to 1, the protocol will retry reading at the end of the file, so you can read files that are still being written to. Not me reading a file mid-write ✨ For this to actually terminate, you either need to use the rw_timeout option, or use the interrupt callback (that one's for API users).
seekable — controls whether seekability is advertised on the file. 0 means non-seekable, -1 means auto (seekable for normal files, non-seekable for named pipes).

So many demuxers handle seekable vs non-seekable resources differently, and overriding this might speed up opening certain files — but at the cost of losing some features (e.g. accurate seeking). Trade-offs, bestie 💅

pkt_size — sets 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 you explicitly set it, 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). We're reading from or writing to remote resources using the FTP protocol, bestie 🌈

The syntax is required to be this:

ftp://[user[:password]@]server[:port]/path/to/remote/resource.mpeg

This protocol accepts these options:

timeout — sets the timeout in microseconds 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. Vibes: unspecified.
ftp-user — sets a user to be used for authenticating to the FTP server. This is overridden by the user in the FTP URL. (URL stays winning 👑)
ftp-password — sets 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 — the password used when logging in as an anonymous user. Typically you should use an e-mail address here. fr fr.
ftp-write-seekable — controls seekability of the connection during encoding. If set to 1 the resource is supposed to be seekable, if set to 0 it's assumed not to be seekable. Default value is 0.

NOTE: the protocol can be used as output, but it's recommended that you don't, unless special care is taken (tests, customized server configuration etc.). Different FTP servers behave in different ways during seek operations, and ff* tools may produce incomplete content due to server limitations. So like... be careful, bestie 😭

🧚 `gopher` 🧚

Gopher protocol. That's the whole entry. Short and iconic, we love a minimal queen 💅

🌸 `gophers` 🌸

Gophers protocol.

It's the Gopher protocol with TLS encapsulation. Glow-up secured 🔥

💜 `http` 💜

HTTP (Hyper Text Transfer Protocol). The main character, honestly.

This protocol accepts these options:

seekable — controls seekability of the connection. If set to 1 the resource is supposed to be seekable, if set to 0 it's assumed not to be seekable, and if set to -1 it'll try to autodetect whether it's seekable. Default value is -1.
chunked_post — if set to 1, use chunked Transfer-Encoding for posts. Default is 1.
http_proxy — sets the HTTP proxy to tunnel through, e.g. http://example.com:1234
headers — sets custom HTTP headers, and can override built in default headers. The value must be a string encoding the headers.
content_type — sets a specific content type for the POST messages or for listen mode.
user_agent — overrides the User-Agent header. If not specified, the protocol will use a string describing the libavformat build (Lavf/<version>).
referer — sets the Referer header. Includes a 'Referer: URL' header in the HTTP request.
multiple_requests — uses persistent connections if set to 1. Default is 0.
request_size — limits the size of requests made. This is useful for some pathological servers that throttle unbounded range requests, and also when you're expecting to seek frequently. Disabled (set to 0) by default.

Note that if you enable this option, it's strongly recommended that you also enable the multiple_requests option, as well as setting short_seek_size to the same value or higher. Doing so lets FFmpeg reuse a single HTTP connection wherever possible. Efficient queen 👑

initial_request_size — limits the size of initial requests. Similar to request_size, but only used during initial format parsing. Useful for formats like MXF or MOV that require frequent seeks during header parsing. It lasts until the demuxer makes a read request larger than this size (with no seek in between), after which the implementation will continue using requests as usual. Disabled (set to 0) by default.

Note that if you enable this option, it's strongly recommended that you also enable the multiple_requests option, as well as setting short_seek_size to the same value or higher.

post_data — sets custom HTTP post data.
mime_type — exports the MIME type.
http_version — exports the HTTP response version number. Usually "1.0" or "1.1".
cookies — sets the cookies to be sent in future requests. The format of each cookie is the same as the value of a Set-Cookie HTTP response field. Multiple cookies can be delimited by a newline character.
icy — if set to 1, request ICY (SHOUTcast) metadata from the server. If the server supports this, the metadata has to be retrieved by the application by reading the icy_metadata_headers and icy_metadata_packet options. The default is 1.
icy_metadata_headers — if the server supports ICY metadata, this contains the ICY-specific HTTP reply headers, separated by newline characters.
icy_metadata_packet — if the server supports ICY metadata, and icy was set to 1, this contains the last non-empty metadata packet sent by the server. Applications interested in mid-stream metadata updates should poll it in regular intervals. Polling era, we love to see it ✨
metadata — sets an exported dictionary containing Icecast metadata from the bitstream, if present. Only useful with the C API.
auth_type — sets the HTTP authentication type. There's no option for Digest, since this method requires getting nonce parameters from the server first and can't be used straight away like Basic. Rude of Digest tbh 💀
- none — choose the HTTP authentication type automatically. This is the default.
- basic — choose the HTTP basic authentication.

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. Like genuinely bestie, do not do it 🙈

send_expect_100 — sends an Expect: 100-continue header for POST. If set to 1 it will send, if set to 0 it won't, and if set to -1 it'll try to send if it's applicable. Default value is -1.
location — an exported dictionary containing the content location. Only useful with the C API.
offset — sets the initial byte offset.
end_offset — tries to limit the request to bytes preceding this offset.
method — when used as a client option, it sets the HTTP method for the request.

When used as a server option, it sets the HTTP method that is going to be expected from the client(s). If the expected and the received HTTP method don't match, the client will be given a Bad Request response. When unset, the HTTP method is not checked for now. This will be replaced by autodetection in the future — watching this space 👀

reconnect — reconnects 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. Never-ending stream girlies rise 🫶
reconnect_on_network_error — reconnects 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 — sets the maximum delay in seconds after which to give up reconnecting.
reconnect_max_retries — sets the maximum number of times to retry a connection. Default unset.
reconnect_delay_total_max — sets 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. Polite and correct 💅

`listen`

Okay so if you set this bad boy to 1, you get an experimental HTTP server 💅. It's giving "I'm the server now" — you can 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... except this is not yet implemented in ffmpeg.c and thus must not be used as a command line option 😭. Like, the vibe is there but the code isn't, yk?

# Server side (sending):
ffmpeg -i somefile.ogg -c copy -listen 1 -f ogg http://<server>:<port>

# Client side (receiving):
ffmpeg -i http://<server>:<port> -c copy somefile.ogg

# Client can also be done with wget:
wget http://<server>:<port> -O somefile.ogg

# Server side (receiving):
ffmpeg -listen 1 -i http://<server>:<port> -c copy somefile.ogg

# Client side (sending):
ffmpeg -i somefile.ogg -chunked_post 0 -c copy -f ogg http://<server>:<port>

# Client can also be done with wget:
wget --post-file=somefile.ogg http://<server>:<port>

resource — the resource requested by a client, when the experimental HTTP server is in use ✨
reply_code — the HTTP code returned to the client, when the experimental HTTP server is in use 🎀
short_seek_size — sets the threshold, in bytes, for when a readahead should be preferred over a seek and a new HTTP request. Genuinely useful, fr — like, it makes sure the same connection gets used for reading large video packets with small audio packets in between 🫶

🍪 HTTP Cookies

Some HTTP requests will straight-up get denied unless cookie values are passed in with the request 😤. The cookies option lets you specify those cookies. At the very least, each cookie must specify a value along with a path and domain. HTTP requests that match both the domain and path will automatically include the cookie value in the HTTP Cookie header field. Multiple cookies can be delimited by a newline, slay 💅.

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

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

Icecast

Icecast protocol — think "stream to Icecast servers" 🎙️✨

This protocol accepts the following options:

ice_genre — set the stream genre 🎵
ice_name — set the stream name 🏷️
ice_description — set the stream description 📝
ice_url — set the stream website URL 🔗
ice_public — set if the stream should be public. The default is 0 (not public) 🙈
user_agent — override the User-Agent header. If not specified, a string of the form "Lavf/<version>" will be used 🤖
password — set the Icecast mountpoint password 🔐
content_type — set the stream content type. This must be set if it is different from audio/mpeg ⚠️
legacy_icecast — this enables support for Icecast versions < 2.4.0, that do not support the HTTP PUT method but the SOURCE method 👵
tls — establish a TLS (HTTPS) connection to Icecast 🔒

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

ipfs

InterPlanetary File System (IPFS) protocol support 🪐✨. You can access files stored on the IPFS network through so-called gateways — these are http(s) endpoints. This protocol wraps the IPFS native protocols (ipfs:// and ipns://) to be sent to such a gateway. Users can (and should) host their own node, which means this protocol will use your local gateway to access files on the IPFS network 🧚

This protocol accepts the following options:

gateway — defines the gateway to use. When not set, the protocol will first try locating the local gateway by looking at $IPFS_GATEWAY, $IPFS_PATH and "$HOME/.ipfs/", in that order 👑

One can use this protocol in 2 ways. Using IPFS:

ffplay ipfs://<hash>

Or the IPNS protocol (IPNS is mutable IPFS):

ffplay ipns://<hash>

mmst

MMS (Microsoft Media Server) protocol over TCP 🖥️

mmsh

MMS (Microsoft Media Server) protocol over HTTP 🌐

The required syntax is:

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

md5

MD5 output protocol 🔢

Computes the MD5 hash of the data to be written, and on close writes this to the designated output — or stdout if none is specified. It can be used to test muxers without writing an actual file, which is honestly iconic for testing 🧪✨

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 bestie 💀 — some formats (typically MOV) require the output protocol to be seekable, so they will fail with the MD5 output protocol. No cap, that's just how it is.

pipe

UNIX pipe access protocol 🚰

Read and write from UNIX pipes. 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 is not specified, by default the stdout file descriptor will be used for writing, 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 — 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 — lowkey valuable if data transmission is slow 🐌
fd — set file descriptor 🎯

Same warning as above, bestie 😭 — some formats (typically MOV) require the output protocol to be seekable, so they will fail with the pipe output protocol.

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. Very "protect the packets" of it 💅

This protocol must be used in conjunction with the "rtp_mpegts" muxer and the "rtp" protocol 🫶

The required syntax is:

-f rtp_mpegts -fec prompeg=<option>=<val>... rtp://<hostname>:<port>

The destination UDP ports are "port + 2" for the column FEC stream and "port + 4" for the row FEC stream 👑

This protocol accepts the following options:

l=n — the number of columns (4-20, LxD <= 100) 📊
d=n — the number of rows (4-20, LxD <= 100) 📈

Example usage:

-f rtp_mpegts -fec prompeg=l=8:d=4 rtp://<hostname>:<port>

rist

Reliable Internet Streaming Transport protocol 🚀

The accepted options are:

rist_profile — supported values:
- simple
- main — this one is default 👑
- advanced
buffer_size — set internal RIST buffer size in milliseconds for retransmission of data. Default value is 0, which means the librist default (1 sec). Maximum value is 30 seconds ⏱️
fifo_size — size of the librist receiver output fifo in number of packets. This must be a power of 2. Defaults to 8192 (vs the librist default of 1024) 📦
overrun_nonfatal=1|0 — survive in case of librist fifo buffer overrun. Default value is 0 🙈
pkt_size — set maximum packet size for sending data. 1316 by default 📨
log_level — set loglevel for RIST logging messages. You only need to set this if you explicitly want to enable debug level messages or packet loss simulation; otherwise the regular loglevel is respected 🔍
secret — set override of encryption secret, by default is unset 🔐
encryption — set encryption type, by default is disabled. Acceptable values are 128 and 256 🔒

rtmp

Real-Time Messaging Protocol 📺✨

The Real-Time Messaging Protocol (RTMP) is used for streaming multimedia content across a TCP/IP network 🌐

The required syntax is:

rtmp://[<username>:<password>@]<server>[:<port>][/<app>][/<instance>][/<playpath>]

The accepted parameters are:

username — an optional username (mostly for publishing) 👤
password — an optional password (mostly for publishing) 🔑
server — the address of the RTMP server 🖥️
port — the number of the TCP port to use (by default is 1935) 🔢
app — it is the name of the application to access. It usually corresponds to the path where the application is installed on the RTMP server (e.g. /ondemand/, /flash/live/, etc.). You can override the value parsed from the URI through the "rtmp_app" option, too 🎯
playpath — it is the path or name of the resource to play with reference to the application specified in app, may be prefixed by "mp4:". You can override the value parsed from the URI through the "rtmp_playpath" option, too 🎬
listen — act as a server, listening for an incoming connection 👂
timeout — maximum time to wait for the incoming connection. Implies listen ⏳

Additionally, the following parameters can be set via command line options (or in code via "AVOption"s):

rtmp_app — name of application to connect on the RTMP server. This option overrides the parameter specified in the URI 🏷️
rtmp_buffer — set the client buffer time in milliseconds. The default is 3000 🧊
rtmp_conn — extra arbitrary AMF connection parameters, parsed from a string, e.g. like "B:1 S:authMe O:1 NN:code:1.23 NS:flag:ok O:0". Each value is prefixed by a single character denoting the type: B for Boolean, N for number, S for string, O for object, or Z for null, followed by a colon. For Booleans the data must be either 0 or 1 for FALSE or TRUE, respectively. Likewise for Objects the data must be 0 or 1 to end or begin an object, respectively. Data items in subobjects may be named, by prefixing the type with 'N' and specifying the name before the value (i.e. "NB:myFlag:1"). This option may be used multiple times to construct arbitrary AMF sequences 🧩
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 (RTMPT only, bestie). The default is 10.
rtmp_live — tells it the media is a live stream. Heads up: no resuming or seeking in live streams is possible, so don't even try. The default value is "any", which lowkey means the subscriber first tries to play the live stream named in the playpath, and if a live stream with that name isn't found, it plays the recorded stream instead. The other possible values are "live" and "recorded". Slay.
rtmp_pageurl — the 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, so it wins, periodt.
rtmp_subscribe — the name of the live stream to subscribe to. By default no value will be sent; it is only sent if the option is specified or if rtmp_live is set to live. iykyk.
rtmp_swfhash — SHA256 hash of the decompressed SWF file (32 bytes). Don't ask, just pass it. 💅
rtmp_swfsize — the size of the decompressed SWF file, which is required for SWFVerification. No swfsize, no verification, bestie.
rtmp_swfurl — the URL of the SWF player for the media. By default no value will be sent.
rtmp_swfverify — URL to the player swf file, and it computes hash/size automatically for you. auto-magic ✨
rtmp_tcurl — the 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 reduces the efficiency / effect of TCP_NODELAY. So like… it's giving "we tried" 😭 but that's literally what the docs say.

tcp_keepalive=1|0 — turn on 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. Soz. 🫶

For example, to read with ffplay a multimedia resource named "sample" from the application "vod" from an RTMP server "myserver":

ffplay rtmp://myserver/vod/sample

To publish to a password protected server, passing the playpath and app names separately:

ffmpeg -re -i <input> -f flv -rtmp_playpath some/long/path -rtmp_app long/app/name rtmp://username:password@myserver/

rtmpe

Encrypted Real-Time Messaging Protocol. 👑

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.

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, 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. Encrypted and sneaky, we love a multitasking icon. 💜

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. Networking but make it Windows-flavored.

Following syntax is required.

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

This protocol accepts the following options.

timeout — set the timeout in milliseconds 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.
truncate — truncate existing files on write, if set to 1. A value of 0 prevents truncating. Default value is 1.
workgroup — set the workgroup used for making connections. By default 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. Bestie, it's file transfer but secure.

Following syntax is required.

sftp://[user[:password]@]server[:port]/path/to/remote/resource.mpeg

This protocol accepts the following options.

timeout — set the timeout 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.
truncate — truncate existing files on write, if set to 1. A value of 0 prevents truncating. Default value is 1.
private_key — specify the path of the file containing the private key to use during authorization. By default libssh searches for keys in the ~/.ssh/ directory.

Example: play a file stored on a remote server.

ffplay sftp://user:password@server_address:22/home/user/resource.mpeg

librtmp rtmp, rtmpe, rtmps, rtmpt, rtmpte

Real-Time Messaging Protocol and its variants supported through librtmp. Yes, all of them, the whole squad. 👯‍♀️

Requires the presence of the librtmp headers and library during configuration. You need to explicitly configure the build with "--enable-librtmp". If enabled, this will replace the native RTMP protocol — like, fully swap it out, no cap.

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). It does the most, honestly.

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. Go read it, it's giving bonus lore. 📖

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. Cute and URL-safe, we stan. 🧚

Options can also can be specified via command line options (or in code via "AVOption"s).

The list of supported options follows.

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. Like, it still works but don't, bestie. 🙈

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. Blocked blocked blocked. 💀
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 the timeout (in microseconds) of socket I/O operations to n.

Important notes, and yes these are important, don't skip them:

If rtcpport is not set, the RTCP port will be set to the RTP port value plus 1.
If localrtpport (the local RTP port) is not set, any available port will be used for the local RTP and RTCP ports.
If localrtcpport (the local RTCP port) is not set, it will be set to the local RTP port value plus 1.

rtsp

Real-Time Streaming Protocol. 🎬

RTSP is not technically a protocol handler in libavformat — plot twist — it 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. Bye bestie 👋

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. Sometimes you just need a second to breathe, y'know? 🌸
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 — super handy for sneaking through proxies and honestly just the move for security reasons. 💅

You can stack up multiple lower transport protocols in one go — they get tried one at a time (if setting one up fails, it just moves onto the next, no drama, no cap). Tiny catch tho: for the muxer, only the tcp and udp options work. 🫶

rtsp_flags — sets the RTSP flags. Here's the accepted value list, bestie:
- filter_src — accept packets only from the negotiated peer address and port. Picky on purpose. 🙈
- listen — act as a server and listen for an incoming connection. main character energy only. 👑
- prefer_tcp — try TCP for RTP transport first, if TCP is actually available as RTSP RTP transport.
- satip_raw — export the raw MPEG-TS stream instead of demuxing it. This flag literally just writes out the raw stream with the original PAT/PMT/PIDs intact. ✨

Default value is none. (yes, boring default, I know 💀)

allowed_media_types — set which media types you'll accept from the server. The accepted flags:
- video
- audio
- data
- subtitle

By default it accepts all media types. generous queen 🌸

min_port — minimum local UDP port. Default value is 5000.
max_port — maximum local UDP port. Default value is 65000.
listen_timeout — maximum timeout (in seconds) to establish an initial connection. Heads up: setting listen_timeout > 0 also sets rtsp_flags to listen. Default is -1, which means an infinite timeout when listen mode is set. 🕰️
reorder_queue_size — number of packets to buffer for handling reordered packets.
timeout — set the socket TCP I/O timeout, in microseconds.
user_agent — override the User-Agent header. If you don't specify it, it defaults to the libavformat identifier string. iykyk 🎀
buffer_size — set the maximum socket buffer size in bytes.

When receiving data over UDP, the demuxer tries to reorder received packets (because yeah, they can arrive out of order, or packets can get lost totally — rude 😭). 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, you can choose which streams to display with -vst n and -ast n for video and audio respectively — and you can switch them on the fly by pressing v and a. iconic little party trick 🦋

🎬 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). Ngl, this isn't technically a protocol handler in libavformat — it's a muxer and a demuxer. It's used for signalling RTP streams by announcing the SDP for those streams on a separate port, regularly, like clockwork. ⭐

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, announcements go 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. Meanwhile the RTP stack in libavformat for receiving requires all streams to be sent on unique ports. picky picky 💀

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. 🎧

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 AVOptions).

The list of supported options follows.

listen — if set to any value, listen for an incoming connection. Outgoing connection is done by default.
max_streams — set the maximum number of streams. By default no limit is set.

srt

Haivision Secure Reliable Transport Protocol via libsrt.

The supported syntax for a SRT URL is:

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

options contains a list of &-separated options of the form key=val. Standard percent-encoding (and using the plus sign for space) can be used to escape keys and values.

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

The list of supported options follows. buckle up bestie, it's a long one 💖

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).
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). 🔐

so real — write-only means write-only, don't come crying when you can't read it back 😭

enforced_encryption= 1|0 — if true, both connection parties must have the same password set (including empty, that is, with no encryption). If the password doesn't match or only one side is unencrypted, the connection is rejected. Default is true.
kmrefreshrate= packets — the number of packets to be transmitted after which the encryption key is switched to a new key. Default is -1. -1 means auto (0x1000000 in srt library). The range for this option is integers in the 0 - INT_MAX.
kmpreannounce=packets — the gap between when a brand-new encryption key gets sent and when the actual switchover goes down. This value also applies to the next interval, like, the one between switchover happening and the old encryption key getting decommissioned. Default is -1, and -1 means auto (that's 0x1000 in the srt library). Range is integers in the 0 – "INT_MAX" zone ✨
snddropdelay=microseconds — the sender's little extra delay before it drops packets. This delay gets added onto the default drop delay time interval value.

Special value -1: don't drop packets on the sender at all, periodt 💅

payload_size=bytes — sets the maximum declared size of a packet moved 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 vibe, different name 🫶
peerlatency=microseconds — the latency value (as described over in rcvlatency) that the sender side sets as a minimum value for the receiver.
pbkeylen=bytes — sender encryption key length, in bytes. You can only set it to 0, 16, 24 and 32. Sender encryption gets enabled if it's not 0. Not required on the receiver (set it to 0), because the key size is obtained from the sender in the HaiCrypt handshake. Default value is 0.
rcvlatency=microseconds — the time that should elapse between the moment a packet was sent and the moment it's delivered to the receiver application in the receiving function. This time should be a buffer time big enough to cover the time spent sending, an unexpectedly extended RTT time, and the time needed to retransmit the lost UDP packet. The effective latency value ends up being the max of this option's value and the peerlatency value set by the peer side. Before version 1.3.0 this option was only available as latency — history, bestie 📜
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 the raise-error timeouts for read, write and connect operations. Heads up tho: 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 it's enabled on the receiver, it skips missing packets that haven't been delivered in time and delivers the following packets to the application once their time-to-play has come. It also sends a fake ACK to the sender. When it's 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 gets 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.

Careful bestie: 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, the packet loss report gets delayed until that number of packets come in. Reorder Tolerance increases every time a "belated" packet has come in but it wasn't due to retransmission (basically when UDP packets tend to come out of order), by 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 the mechanism is turned off and the loss report is always sent immediately upon experiencing a "gap" in sequences 💀
minversion — the minimum SRT version that's required from the peer. A connection to a peer that doesn't satisfy the minimum version requirement will be rejected, no exceptions.

The version format in hex is 0xXXYYZZ for x.y.z in human readable form.

streamid=string — a string limited to 512 characters that can be set on the socket before connecting. This stream ID will be able to be retrieved by the listener side from the socket that's returned from srt_accept and was connected by a socket with that set stream ID. SRT does not enforce any special interpretation of the contents of this string, it's just vibes. 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 a matter of luck which one would win 🙈
srt_streamid=string — alias for streamid, here purely to avoid conflict with the ffmpeg command line option 🎀
smoother=live|file — the type of Smoother used for the transmission for that socket, which is the thing responsible for the transmission and congestion control. The Smoother type must be exactly the same on both connecting parties, otherwise the connection gets rejected. No matching fits, no connection 👑
messageapi=1|0 — when set, this socket uses the Message API; otherwise it uses the Buffer API. Note that in live mode (see transtype) there's only the message API available. In File mode you can choose between two modes:

Stream API (the 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 takes care of any speed and congestion control. When receiving, you can also receive as many data as desired — the data that isn't extracted will be waiting for the next call. There's no boundary between data portions in 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. In particular, setting this option sets multiple other parameters to their default values as required for a particular transmission type.

live: set options as for live transmission. In this mode you should send, by one sending instruction, only so many data that fit in one UDP packet, limited to the value defined first in payload_size (1316 is the default in this mode). There's no speed control in this mode, only the bandwidth control — if configured — so you don't 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 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). Range is integers in the 0 – "INT_MAX" zone.
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. Yes, the secure one, she's serving encryption 💅

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 the master key, and the following 14 bytes are used as the master salt. Cute little split, honestly 🎀

📼 subfile

Virtually extract a segment of a file or another stream. The underlying stream must be seekable, bestie — no exceptions, it needs to be able to jump around ✨

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 the end of the 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 | — one little pipeline bar and you're cloning your output, main character energy fr fr 👑

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

🌐 tcp

Transmission Control Protocol. The reliable one, the one that checks in on you 💖

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 the tcp socket connect.
local_port=port — local port used for the tcp socket connect.
timeout=microseconds — set the 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 the listen timeout, expressed in milliseconds.
recv_buffer_size=bytes — set the receive buffer size, expressed in bytes.
send_buffer_size=bytes — set the 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 that reduces the efficiency / effect of TCP_NODELAY. Just so you know, bestie 💀

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

Like, the classic listen-y example — setting up a listening TCP connection with ffmpeg, and then bopping over to 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) 💖 the locked-up, encrypted bestie of the protocols.

The required syntax for a TLS/SSL url is:

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

options contains a list of &-separated options of the form key=val. Standard percent-encoding (and using the plus sign for space) can be used to escape keys and values.

Options can also can be specified via command line options (or in code via "AVOption"s).

The list of supported options follows.

ca_file, cafile=filename — a file containing certificate authority (CA) root certificates to treat as trusted. If the linked TLS library contains a default this might not need to be specified for verification to work, but not all libraries and setups have defaults built in. The file must be in OpenSSL PEM format.
tls_verify, verify=1|0 — if enabled, try to verify the peer that we are communicating with. Note, if using OpenSSL, this currently only makes sure that the peer certificate is signed by one of the root certificates in the CA database, but it does not validate that the certificate actually matches the host name we are trying to connect to. (With other backends, the host name is validated as well.) This is disabled by default since it requires a CA database to be provided by the caller in many cases.
cert_file, cert=filename — a file containing a certificate to use in the handshake with the peer. (When operating as server, in listen mode, this is more often required by the peer, while client certificates only are mandated in certain setups.) 🎀
key_file, key=filename — a file containing the private key for the certificate.
listen=1|0 — if enabled, listen for connections on the provided port, and assume the server role in the handshake instead of the client role.
http_proxy — the HTTP proxy to tunnel through, e.g. "http://example.com:1234". The proxy must support the CONNECT method.

Example command lines:

To create a TLS/SSL server that serves an input stream.

ffmpeg -i <input> -f <format> tls://<hostname>:<port>?listen&cert=<server.crt>&key=<server.key>

To play back a stream from the TLS/SSL server using ffplay:

ffplay tls://<hostname>:<port>

dtls

Datagram Transport Layer Security (DTLS) — same vibe as TLS, except it's riding on UDP because it literally cannot sit still. 🦋

The required syntax for a DTLS URL is:

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

options contains a list of &-separated options of the form key=val. Standard percent-encoding (and using the plus sign for space) can be used to escape keys and values.

Options can also can be specified via command line options (or in code via "AVOption"s).

DTLS shares most options with TLS, but operates over UDP instead of TCP.

The list of supported options follows.

ca_file, cafile=filename — a file containing certificate authority (CA) root certificates to treat as trusted. If the linked TLS library contains a default this might not need to be specified for verification to work, but not all libraries and setups have defaults built in. The file must be in OpenSSL PEM format.
tls_verify, verify=1|0 — if enabled, try to verify the peer that we are communicating with. Note, if using OpenSSL, this currently only makes sure that the peer certificate is signed by one of the root certificates in the CA database, but it does not validate that the certificate actually matches the host name we are trying to connect to. This is disabled by default since it requires a CA database to be provided by the caller in many cases.
cert_file, cert=filename — a file 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 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 one that just fires packets into the void and hopes for the best.

The required syntax for an UDP URL is:

udp://<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).

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.

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.
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.
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 — no network, just the machine talking to itself, main character energy honestly. ✨

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 "AVOption"s):

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.

The required syntax for streaming or connecting to a stream is:

zmq:tcp://ip-address:port

Example: Create a localhost stream on port 5555:

ffmpeg -re -i input -f mpegts zmq:tcp://127.0.0.1:5555

Multiple clients may connect to the stream using:

ffplay zmq:tcp://127.0.0.1:5555

Streaming to multiple clients is implemented using a ZeroMQ Pub-Sub pattern. The server side binds to a port and publishes data. Clients connect to the server (via IP address/port) and subscribe to the stream. The order in which the server and client start generally does not matter.

ffmpeg must be compiled with the --enable-libzmq option to support this protocol.

Options can be set on the ffmpeg/ffplay command line. The following options are supported:

pkt_size — this one forces the max packet size for sending/receiving data. Default is 131,072 bytes, bestie. On the server side it sets the biggest packet you can send via ZeroMQ; on the clients it sets an internal buffer size for receiving packets. ⚠️ and listen — pkt_size on the clients should be equal to or greater than pkt_size on the server, otherwise your received message may get truncated and that causes decoding errors. no cap, keep them matched. 💅

🎀 DEVICE OPTIONS 🎀

Okay so the libavdevice library gives you literally the exact same interface as libavformat. Like, an input device is treated like a demuxer and an output device like a muxer, and the interface + generic device options are the same ones libavformat provides (go peek at the ffmpeg-formats manual, iykyk).

On top of that, each input or output device may support its own little private options that 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 if you're doing it programmatically. 🫶

🦋 INPUT DEVICES 🦋

Input devices are these configured elements in FFmpeg that let you access data coming from a multimedia device attached to your system. main character energy honestly.

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 input devices with the configure option --disable-indevs, and selectively enable one with --enable-indev=INDEV, or you can disable a particular input device with --disable-indev=INDEV. choose your own adventure, bestie.

The -devices option of the ff* tools will display the list of supported input devices. 🌈

A description of the currently available input devices follows. let's gooo 💖

🌸 alsa 🌸

ALSA (Advanced Linux Sound Architecture) input device. iconic.

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

This device lets you capture from an ALSA device. The name of the device to capture has to be an ALSA card identifier.

An ALSA identifier has the syntax:

hw:<CARD>[,<DEV>[,<SUBDEV>]]

where the DEV and SUBDEV components are optional. 🎀

The three arguments (in order: CARD, DEV, SUBDEV) specify card number or identifier, device number and subdevice number (-1 means any).

To see the list of cards currently recognized by your system, check the files /proc/asound/cards and /proc/asound/devices. 🔍

For example, to capture with ffmpeg from an ALSA device with card id 0, you may run the command:

ffmpeg -f alsa -i hw:0 alsaout.wav

For more information see: <http://www.alsa-project.org/alsa-doc/alsa-lib/pcm.html>

Options

sample_rate — set the sample rate in Hz. Default is 48000. 💫
channels — set the number of channels. Default is 2.

🧚 android_camera 🧚

Android camera input device.

This input device uses the Android Camera2 NDK API, which is available on devices with API level 24+. The availability of android_camera is autodetected during configuration. ✨

This device lets you capture from all the cameras on an Android device that are integrated into the Camera2 NDK API.

The available cameras are enumerated internally and can be selected with the camera_index parameter. The input file string is discarded, literally just thrown away. 💀

Generally the back facing camera has index 0 while the front facing camera has index 1. (front cam = selfie cam, you're welcome 📸)

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 🍓

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 send 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. (not me having to explain widl to you, just trust the process 😭)

DeckLink is very picky about the formats it supports. fr fr, it's giving diva. The 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: DeckLink SDK 11.0 have replaced the duplex property by a profile property. For the DeckLink Duo 2 and DeckLink Quad 2, a profile is shared between any 2 sub-devices that utilize the same connectors. For the DeckLink 8K Pro, a profile is shared between all 4 sub-devices. So DeckLink 8K Pro support four profiles.

Valid profile modes for DeckLink 8K Pro (with DeckLink SDK >= 11.0): one_sub_device_full, one_sub_device_half, two_sub_device_full, four_sub_device_half

Valid profile modes for DeckLink Quad 2 and DeckLink Duo 2: half, full

timecode_format — the timecode type to include in the frame and video stream metadata. Has to be one of none, rp188vitc, rp188vitc2, rp188ltc, rp188hfr, rp188any, vitc, vitc2, or serial. Defaults to none, aka literally not included at all 💅

Fr fr, 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. And note that this is slightly different to the ordering used by the DeckLink API, which is HFR, VITC1, LTC, VITC2 — so don't @ me, that's just how it is 🫠

video_input — sets the video input source. Must be unset, sdi, hdmi, optical_sdi, component, composite or s_video. Defaults to unset.
audio_input — sets the audio input source. Must be unset, embedded, aes_ebu, analog, analog_xlr, analog_rca or microphone. Defaults to unset.
video_pts — sets the video packet timestamp source. Must be video, audio, reference, wallclock or abs_wallclock. Defaults to video.
audio_pts — sets the audio packet timestamp source. Must be video, audio, reference, wallclock or abs_wallclock. Defaults to audio.
draw_bars — if set to true, color bars are drawn in the event of a signal loss. Defaults to true. ⚠️ This option is deprecated, please use the signal_loss_action option instead!
signal_loss_action — sets the action to take in the event of a signal loss. Accepts one of the following values:
- 1, none — do nothing on signal loss. This usually results in black frames 🖤
- 2, bars — draw color bars on signal loss. Only supported for 8-bit input signals.
- 3, repeat — repeat the last video frame on signal loss 🔁

Defaults to bars.

queue_size — sets maximum input buffer size in bytes. If the buffering reaches this value, incoming frames will be dropped. Defaults to 1073741824.
audio_depth — sets the audio sample bit depth. Must be 16 or 32. Defaults to 16.
decklink_copyts — if set to true, timestamps are forwarded as they are without removing the initial offset. Defaults to false.
timestamp_align — capture start time alignment in seconds. If set to nonzero, input frames are dropped till the system timestamp aligns with configured value. Alignment difference of up to one frame duration is tolerated. This is useful for maintaining input synchronization across N different hardware devices deployed for 'N-way' redundancy. The system time of different hardware devices should be synchronized with protocols such as NTP or PTP, before using this option.

⚠️ Note that this method is not foolproof. In some border cases input synchronization may not happen due to thread scheduling jitters in the OS. Either sync could go wrong by 1 frame or in a rarer case timestamp_align seconds. Defaults to 0.

wait_for_tc (bool) — drop frames till a frame with timecode is received. Sometimes serial timecode isn't received with the first input frame, and 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, iconic behaviour honestly:

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, we're on Windows now, buckle up 🪟

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 👯

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. That's it, that's the deal 🫶

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 selected device's options and exit.
video_device_number — set video device number for devices with the same name (starts at 0, defaults to 0).
audio_device_number — set audio device number for devices with the same name (starts at 0, defaults to 0).
pixel_format — select 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 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 video capture pin to use by name or alternative name.
audio_pin_name — select audio capture pin to use by name or alternative name.
crossbar_video_input_pin_number — select video input pin number for crossbar device. This will be routed to the crossbar device's Video Decoder output pin. Note that changing this value can affect future invocations (sets a new default) until system reboot occurs. ⚠️ Like, it literally sticks around, bestie.
crossbar_audio_input_pin_number — select audio input pin number for 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 system reboot occurs.
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 avoiding green bars at the bottom, flickering scan lines, etc. Note that with some devices, changing these properties can also affect future invocations (sets new defaults) until 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. To use this an audio capture source has to be specified, but it can be anything even fake one 👀
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.
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. To use this a video capture source has to be specified, but it can be anything even 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 allows working around devices that provide unreliable timestamps 🕰️

✨ 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:
Open video device Camera with the second video device number:

$ ffmpeg -f dshow -video_device_number 1 -i video="Camera"

Open video device Camera and audio device Microphone at the same time, iconic combo:

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

Print every supported option on the selected device and then dip out:

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

Name your pins by name or alternative name, and pick a device by its alternative name too (yes the string is that long, don't @ me):

$ 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"

Set up a crossbar device, pick the crossbar pins, and let the user tweak video capture properties at startup:

$ ffmpeg -f dshow -show_video_device_dialog true -crossbar_video_input_pin_number 0
     -crossbar_audio_input_pin_number 3 -i video="AVerMedia BDA Analog Capture":audio="AVerMedia BDA Analog Capture"

🖼️ fbdev

The Linux framebuffer input device 💖

So the Linux framebuffer is this graphic hardware–independent abstraction layer that puts graphics on your monitor — usually on the console. You get to it through a file device node, and that's usually /dev/fb0.

Want the full lore? Go read the file Documentation/fb/framebuffer.txt in the Linux source tree. It's giving documentation realness.

Also check out <http://linux-fbdev.sourceforge.net/>, and fbset(1).

To record from that framebuffer device /dev/fb0 with ffmpeg:

ffmpeg -f fbdev -framerate 10 -i /dev/fb0 out.avi

And you can absolutely grab a single screenshot image with:

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

Options

framerate — sets the frame rate. Default is 25 ✨

🪟 gdigrab

The Win32 GDI–based screen capture device 🖥️

This one lets you capture a region of your display on Windows. Love that for us.

The input filename can be a few different things, like:

desktop

title=<window_title>

hwnd=<window_hwnd>

The first option grabs the entire desktop, or a fixed region of the desktop. The second and third options instead capture the contents of a single window — doesn't matter where it is on screen, it's getting grabbed 👑

For example, to grab the whole desktop using ffmpeg:

ffmpeg -f gdigrab -framerate 6 -i desktop out.mpg

Grab a 640x480 region at position "10,20":

ffmpeg -f gdigrab -framerate 6 -offset_x 10 -offset_y 20 -video_size vga -i desktop out.mpg

Grab the contents of the window named "Calculator":

ffmpeg -f gdigrab -framerate 6 -i title=Calculator out.mpg

Options

draw_mouse — specifies whether to draw the mouse pointer. Use the value 0 to not draw the pointer. Default value is 1 🖱️
framerate — sets the grabbing frame rate. Default value is "ntsc", which corresponds to a frame rate of "30000/1001".
show_region — shows the grabbed region on screen.

If show_region is specified with 1, then the grabbing region will be indicated on screen. With this option it's super easy to see what's actually 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. 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 that the offset calculation is from the top left corner of the primary monitor on Windows. If you've got a monitor positioned to the left of your primary monitor, you'll need to use a negative offset_x value to move the region over to that monitor 💅

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

Note that the offset calculation is from the top left corner of the primary monitor on Windows. If you've got a monitor positioned above your primary monitor, you'll need to use a negative offset_y value to move the region up to that monitor 🫶

🔥 iec61883

FireWire DV/HDV input device using libiec61883 🎥

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 got removed. Rip old stack 🪦

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

Options

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

Real talk: treating a DV device as HDV (or vice versa) will not work and will result in undefined behavior 😭

The values auto, dv and hdv are supported.

dvbuffer — sets the maximum size of the buffer for incoming data, in frames. For DV this is an exact value. For HDV it's not frame exact, since HDV doesn't have a fixed frame size. It's giving "close enough".
dvguid — selects the capture device by specifying its GUID. Capturing will only happen from the specified device, and it fails if no device with the given GUID is found. This is useful for selecting the input when 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 🎶

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 the name client_name:input_N, where client_name is the name provided by the application, and N is a number that identifies the channel. Each writable client sends 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

Want more? Read: <http://jackaudio.org/>

Options

channels — sets the number of channels. Default is 2 🎚️

🖥️ kmsgrab

The KMS video input device ✨

Captures the KMS scanout framebuffer associated with a specified CRTC or plane as a DRM object that can be passed to other hardware functions. Very main character energy.

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 🙈

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 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. 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

Want 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 🦋

This input device reads data from the open output pads of a libavfilter filtergraph.

For each filtergraph open output, the input device will create a corresponding stream which is mapped to the generated output. The filtergraph is specified through the option graph.

Options

`graph`

Okay so this is the big one — bestie, this is how you tell lavfi exactly what filtergraph to use as input. Every video open output has 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? It automatically gets the "out0" label. But all the others — every single one — have to be written out explicitly. No shortcuts, sorry babe 💅

And there's this extra little sprinkle: you can tack the suffix "+subcc" onto an output label to create an extra stream with the closed captions packets attached to that output (this one's experimental, fyi — only for EIA-608 / CEA-708 for now). The subcc streams get created after all the normal streams, in the order of the corresponding stream. So like, 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. Math class but make it video 🧮✨

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 exactly the same as the one you'd specify with the graph option. Same vibes, different delivery method 🎀

`dumpgraph`

Dumps the graph to stderr. That's it. That's the whole thing. Diagnostic queen 💅

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 to specify the graph description instead, and omit the "out0" label:

ffplay -f lavfi color=c=pink

Create three different video test filtered sources and play them all:

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 both 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, built on top of libcdio. Retro tech, we stan a physical media moment 💿

To enable this input device during configuration you need libcdio installed on your system — it requires the configure option --enable-libcdio. No libcdio, no CD party, sorry 🫶

This device lets you play and grab from an Audio-CD, which is honestly iconic.

So like, for example, to copy the entire Audio-CD in /dev/sr0 with ffmpeg, you'd run:

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

Options

`speed`

Set the drive reading speed. Default value is 0.

The speed is specified in CD-ROM speed units. It gets set through the libcdio cdio_cddap_speed_set function. Heads up though: on many CD-ROM drives, if you specify a value that's too large, you'll just end up using the fastest speed. The drive said "I'll do what I want" 😭

`paranoia_mode`

Set the paranoia recovery mode flags. It accepts one of these values:

disable
verify
overlap
neverskip
full

Default value is disable.

For more info about the available recovery modes, go consult the paranoia project documentation. (Yes the mode is literally called paranoia, and no we're not being dramatic for once 💀)

🦋 libdc1394

IIDC1394 input device, based on libdc1394 and libraw1394. FireWire girlies rise 🔥

Requires the configure option --enable-libdc1394.

Options

`framerate`

Set the frame rate. Default is "ntsc", which corresponds to a frame rate of "30000/1001".

`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 that have a working OpenAL 1.1 implementation. Cross-platform audio capture, we love to see it ✨

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 come 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 --extra-cflags and --extra-ldflags so the build system can actually locate the OpenAL headers and libraries. Basically you gotta help the compiler find its glasses 🙈

Here's an incomplete list of OpenAL implementations:

Creative — the official Windows implementation, providing hardware acceleration with supported devices plus a software fallback. See <http://openal.org/>.
OpenAL Soft — a 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 lets you capture from an audio input device that's handled through OpenAL.

You need to specify the name of the device to capture in the filename you provide. If you give it the empty string, the device will automatically select the default device. You can get the list of supported devices by using the list_devices option. Slay 🎀

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 the filename):

$ ffmpeg -f openal -i '' out.ogg

Capture from two devices simultaneously, writing to two different files, all 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. It's giving "works on my machine" but make it audio 😭

🌸 oss

Open Sound System input device. Old school, but it's still here and it's still serving 🫶

The filename you provide to the input device is the device node representing the OSS input device, and it's usually set to /dev/dsp.

So like, 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. The people's audio server, no cap ✨

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

The filename you provide to the input device is a source device or the string "default".

To list the PulseAudio source devices and their properties, you can invoke the command pactl list sources. Peek behind the curtain, bestie 👑

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. The default server is used when this isn't provided.

`name`

Specify the application name PulseAudio will use when showing active clients. By default it's the "LIBAVFORMAT_IDENT" string.

`stream_name`

Specify the stream name PulseAudio will use when showing active streams. By default it's "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. It's giving ghost of options past 💀

`fragment_size`

Specify the size in bytes of the minimal buffering fragment in PulseAudio — it will affect the audio latency. By default it's set to 50 ms amount of data.

`wallclock`

Set the initial PTS using the current time. Default is 1.

Examples

Record a stream from the default device:

ffmpeg -f pulse -i default /tmp/pulse.wav

🍓 sndio

sndio input device. BSD besties, this one's for you 💖

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

The filename you provide to the input device is the device node representing the sndio input device, and it's usually set to /dev/audio0.

So like, 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. Webcam girlies, this is your moment 📷✨

"v4l2" can be used as an alias for "video4linux2". Cute nickname, same device.

If FFmpeg is built with v4l-utils support (by using the --enable-libv4l2 configure option), it's 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. a USB webcam) is plugged into the system, and it has a name of the kind /dev/videoN, where N is a number associated with the device.

Video4Linux2 devices usually support a limited set of widthx*height* sizes and frame rates. You can check which ones are supported using -list_formats all for Video4Linux2 devices. Some devices, like TV cards, support one or more standards too, and it's 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. Timezone drama but make it kernel-level 😭

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, leaving the frame rate and size as previously set:

ffmpeg -f video4linux2 -input_format mjpeg -i /dev/video0 out.mpeg

For more information about Video4Linux, check <http://linuxtv.org/>.

Options

standard — sets the standard, bestie. Has to be the name of a standard that's actually supported, so don't just make one up 💀. Wanna see the whole menu of supported standards? Hit up the list_standards option.
channel — the input channel number. Defaults to -1, which lowkey just means "use whatever channel we already picked" 🫶.
video_size — sets the video frame size. The argument's gotta be a string shaped like WIDTHxHEIGHT, or one of those valid size abbreviations ✨.
pixel_format — picks the pixel format. Heads up: this only works for raw video input 👑.
input_format — sets the preferred pixel format (for raw video) or a codec name. It's giving "let me choose", because this option lets you select the input format when several of them are available. Options, options 💅.
framerate — sets the preferred video frame rate. Simple, iconic, no notes.
list_formats — lists all the available formats (that's supported pixel formats, codecs, and frame sizes, bestie) and then dips out — exit. The values it takes are:
- all — show all the available formats, compressed and non-compressed alike 🌈.
- raw — show only raw video (non-compressed) formats.
- compressed — show only the compressed formats. Periodt.
list_standards — lists the supported standards and exits, bye 👋. Values:
- all — show all supported standards.
timestamps, ts — sets the type of timestamps for grabbed frames ⏰. Values:
- default — use the timestamps straight from the kernel.
- abs — absolute timestamps (wall clock, like, actual time o'clock 🕰️).
- mono2abs — forces a conversion from monotonic to absolute timestamps.

The default value is "default", obviously 💖.

use_libv4l2 — uses libv4l2 (v4l-utils) conversion functions. Default is 0, so it's off unless you say otherwise.

🎀 vfwcap

The VfW (Video for Windows) capture input device. Retro vibes 🦋.

The filename you pass as input is literally the capture driver number, ranging from 0 to 9. And yes, you may use "list" as the filename to print out a list of drivers — so helpful of it 😭. Any other filename gets interpreted as device number 0, no cap.

Options

video_size — sets the video frame size. That's it, that's the whole job 💅.
framerate — sets the grabbing frame rate. Default value is "ntsc", which corresponds to a frame rate of "30000/1001". Math girlies rise 🧚.

🌸 x11grab

The X11 video input device — screen capture icon behavior 👑

To even enable this input device during configuration you need libxcb installed on your system. Good news: it gets automatically detected during configuration, so she'll find it herself ✨.

This device lets you capture a region of an X11 display. Main character energy fr fr.

The filename you pass as input has this syntax:

[<hostname>]:<display_number>.<screen_number>[+<x_offset>,<y_offset>]

hostname:display_number.screen_number specifies the X11 display name of the screen you're grabbing from. hostname can be omitted, and it defaults to "localhost". Also, the environment variable DISPLAY contains the default display name, iykyk 💖.

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 (like, man X) for more detailed info — she has lore 📖.

Use the xdpyinfo program for getting the basic info about the properties of your X11 display (e.g. grep for "name" or "dimensions"). Lowkey a lifesaver.

For example, to grab from :0.0 using ffmpeg:

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

And to 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. (She's a little bossy but we love her 💅)
draw_mouse — specify whether to draw the mouse pointer. A value of 0 says nope, don't draw the pointer. Default value is 1.
follow_mouse — makes the grabbed area follow the mouse 🐭. The argument can be "centered" or a number of pixels, PIXELS.

When you specify it with "centered", the grabbing region follows the mouse pointer and keeps the pointer at the center of the region. Otherwise, the region only follows when the mouse pointer reaches within PIXELS (greater than zero) of the edge of the region.

For example:

ffmpeg -f x11grab -follow_mouse centered -framerate 25 -video_size cif -i :0.0 out.mpg

To follow only when the mouse pointer reaches within 100 pixels to the edge:

ffmpeg -f x11grab -follow_mouse 100 -framerate 25 -video_size cif -i :0.0 out.mpg

framerate — sets the grabbing frame rate. Default value is "ntsc", corresponding to a frame rate of "30000/1001" ⭐.
show_region — 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's super easy to know what's being grabbed if you're only grabbing a portion of the screen. Obsessed with the transparency honestly.

region_border — sets the region border thickness if -show_region 1 is used. Range is 1 to 128 and the default is 3 (XCB-based x11grab only, bestie — don't @ me).

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 the options -tree and -root. Detective work but make it cute 🔍.

If the window is later enlarged, the new area is not recorded — she said what she said. Video ends when the window is closed, unmapped (i.e. iconified), or shrunk beyond the video size (which defaults to the initial window size). So like, don't minimize it, babe 😭.

Heads up! This option disables the options follow_mouse and select_region. Pick a lane 💀.

video_size — sets the video frame size. Default is the full desktop or window.
grab_x
grab_y — set the grabbing region coordinates. They're expressed as an 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 so the audio resampler has a whole little family of named options and yes you're gonna learn all of them 💅

Here's how you set them: you can pass -option value in the FFmpeg tools, or option=value for the aresample filter, or set the value explicitly in the SwrContext options, or use the libavutil/opt.h API if you're doing it programmatically. Three ways to the same slay, basically 🫶

uchl, used_chlayout — sets the used input channel layout. Default is unset, and it's only used for that special remapping situation ✨
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 — specify the input sample format. It gets set by default to none.
osf, out_sample_fmt — specify the output sample format. Also defaulting to none.
tsf, internal_sample_fmt — set the internal sample format. Default value is none. If you don't set it explicitly, it'll get chosen automatically for you, no cap 🧚
ichl, in_chlayout and ochl, out_chlayout — set the input/output channel layout. For the required syntax, go peek at the Channel Layout section in the ffmpeg-utils(1) manual 📖
clev, center_mix_level — set the center mix level. It's a value in deciBel, and it must be in the interval [-32,32]. Like, must must.
slev, surround_mix_level — set the surround mix level. Also in deciBel, also has to live inside [-32,32].
lfe_mix_level — sets the LFE mix into non-LFE level. It's used when there's an LFE input but no LFE output. deciBel again, interval [-32,32] again. Consistency is iconic.
rmvol, rematrix_volume — set the rematrix volume. Default value is 1.0.
rematrix_maxval — set the maximum output value for rematrixing. It's the whole "prevent clipping vs. prevent volume reduction" tradeoff, so pick your fighter. A value of 1.0 prevents clipping.
flags, swr_flags — set the flags used by the converter. Default value is 0. It supports the following individual flags:
- res — force resampling. This flag makes resampling happen even when the input and output sample rates match 🔥
dither_scale — set the dither scale. Default value is 1.
dither_method — set the dither method. Default value is 0. Supported values are:
- 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 — set the resampling engine. Default value is swr. Supported values are:
- swr — selects the native SW Resampler; the filter options precision and cheby are not applicable in this case
- soxr — selects the SoX Resampler (where available); compensation, and the filter options filter_size, phase_shift, exact_rational, filter_type & kaiser_beta, are not applicable in this case
filter_size — for swr only, set the resampling filter size. Default value is 32.
phase_shift — for swr only, set the resampling phase shift. Default value is 10, and it must be in the interval [0,30].
linear_interp — use linear interpolation when enabled (that's the default btw). Disable it if you'd rather preserve speed instead of quality when exact_rational fails. Your call 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 falls back to 1 << phase_shift. Default is enabled.
cutoff — set the 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 (which, with a sample-rate of 44100, preserves the entire audio band to 20kHz — obsessed 😌).
precision — for soxr only, the precision in bits to which the resampled signal gets 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' 👑
cheby — for soxr only, selects passband rolloff none (Chebyshev) & higher-precision approximation for 'irrational' ratios. Default value is 0.
async — for swr only, a simple 1-parameter audio sync to timestamps using 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 is applied to make the samples match the audio timestamps.
first_pts — for swr only, assume the first pts should be this value. The time unit is 1 / sample rate. This allows for 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 happens. 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 caused by encoder delay. genuinely such a useful little guy 🥺
min_comp — for swr only, set the minimum difference between timestamps and audio data (in seconds) needed to trigger stretching/squeezing/filling or trimming of the data to match the timestamps. The default is that stretching/squeezing/filling and trimming is disabled (min_comp = FLT_MAX). So yeah, off by default.
min_hard_comp — for swr only, set the minimum difference between timestamps and audio data (in seconds) to trigger adding/dropping samples to make it match the timestamps. This option is basically a threshold choosing between hard (trim/fill) and soft (squeeze/stretch) compensation. Note that all compensation is disabled by default through min_comp. The default here is 0.1.
comp_duration — for swr only, set the duration (in seconds) over which data gets stretched/squeezed to match the timestamps. Must be a non-negative double float value. Default value is 1.0.
max_soft_comp — for swr only, set the maximum factor by which data is stretched/squeezed to match the timestamps. Must be a non-negative double float value. Default value is 0.
matrix_encoding — select the 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, select the resampling filter type. 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, set 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, set 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 🦋

✨ SCALER OPTIONS ✨

okay so the video scaler has its own little collection of named options and bestie, they're lowkey important 👑

You can set these by passing -option value in the FFmpeg tools — with a few lil exceptions that are API-only (I'll flag those, don't worry 💖). And if you're doing that programmatic thing, you can set them explicitly in the SwsContext options or through the libavutil/opt.h API. iykyk 🫶

🌸 sws_flags

Set the scaler flags, periodt. This is also what you use to pick the scaling algorithm — and ngl, only a single algorithm should be selected at a time. Default value is bicubic 💅

It accepts the following values:

fast_bilinear — selects the fast bilinear scaling algorithm ⚡
bilinear — selects the bilinear scaling algorithm
bicubic — selects the bicubic scaling algorithm (the default, main character energy 👑)
experimental — selects the experimental scaling algorithm 🧪
neighbor — selects nearest neighbor rescaling algorithm
area — selects averaging area rescaling algorithm
bicublin — selects bicubic scaling for the luma component, bilinear for the chroma components 🎀
gauss — selects Gaussian rescaling algorithm
sinc — selects sinc rescaling algorithm
lanczos — selects Lanczos rescaling algorithm. The default width (alpha) is 3 and can be changed by setting param0 🌈
spline — selects natural bicubic spline rescaling algorithm
print_info — enable printing/debug logging 🖨️
accurate_rnd — enable accurate rounding
full_chroma_int — enable full chroma interpolation
full_chroma_inp — select full chroma input
bitexact — enable bitexact output
unstable — allow the use of experimental new code. May subtly affect the output or even produce wrong results. For testing only 🚨

bestie pls note: unstable is literally right there in the name — testing only, no cap 💀

💜 srcw (API only)

Set source width.

💜 srch (API only)

Set source height.

💜 dstw (API only)

Set destination width.

💜 dsth (API only)

Set destination height.

💜 src_format (API only)

Set source pixel format (must be expressed as an integer 🔢).

💜 dst_format (API only)

Set destination pixel format (must be expressed as an integer 🔢).

🦋 src_range (boolean)

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

🦋 dst_range (boolean)

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

🦋 gamma (boolean)

If value is set to 1, enable gamma correct scaling. Default value is 0.

🎀 param0, param1

Set scaling algorithm parameters. The specified values are specific of some scaling algorithms and ignored by others — so like, some algorithms care, some literally don't 🙈 The specified values are floating point number values.

⭐ sws_dither

Set the dithering algorithm. Accepts one of the following values. Default value is auto.

auto — automatic choice 🪄
none — no dithering
bayer — bayer dither
ed — error diffusion dither
a_dither — arithmetic dither, based using addition
x_dither — arithmetic dither, based using xor (more random/less apparent patterning that a_dither) 🔀

🌸 alphablend

Set the alpha blending to use when the input has alpha but the output does not. Default value is none.

uniform_color — blend onto a uniform background color
checkerboard — blend onto a checkerboard 🏁
none — no blending

🎬 FILTERING INTRODUCTION 🎬

Filtering in FFmpeg is enabled through the libavfilter library. slay 💅

In libavfilter, a filter can have multiple inputs and multiple outputs. 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. iconic behavior honestly 🪞✨

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.

🌈 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.

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 🖼️✨

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 🧚

💅 FILTERGRAPH DESCRIPTION ✨

okay bestie, picture this: a filtergraph is literally a directed graph of connected filters, no cap. it can contain cycles (yes, cycles, it's giving chaos but make it legal), and there can be multiple links between a pair of filters. every single link has one input pad on one side — connecting it to one filter from which it takes its input — and one output pad on the other side, connecting it to one filter accepting its output. that's the whole anatomy, periodt.

each filter in a filtergraph is an instance of a filter class registered in the application, and that class is what defines the features and the number of input and output pads the filter gets. it's the blueprint, the filter is the fit.

a filter with no input pads is called a "source", and a filter with no output pads is called a "sink". main character energy vs. exit-the-chat energy, honestly.

✨ Filtergraph syntax ✨

a filtergraph has a textual representation, and it's recognized by the -filter/-vf/-af and -filter_complex options in ffmpeg, and by -vf/-af in ffplay, and also by the avfilter_graph_parse_ptr() function defined in libavfilter/avfilter.h. so yeah, it gets around 🦋

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 it has to be the name of one of the filter classes registered in the program — optionally followed by @id. and the name of the filter class is optionally followed by the string =arguments.

arguments is a string containing the parameters used to initialize the filter instance. it can take 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. so then the parameter list in:0:30 means the value in is assigned to the option type, 0 to start_frame and 30 to nb_frames. iykyk 💅
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. slay, flexible.

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 by prepending a slash / to the option name, then the supplied value is interpreted as a path from which the actual value is loaded. e.g.

ffmpeg -i <INPUT> -vf drawtext=/text=/tmp/some_text <OUTPUT>

will load the text to be drawn from /tmp/some_text. API users wishing to implement a similar feature should use the avfilter_graph_segment_*() functions together with custom IO code. fancy!

the name and arguments of the filter are optionally preceded and followed by a list of link labels. a link label lets you name a link and associate it to a filter output or input pad. the preceding labels in_link_1 ... in_link_N are associated to the filter input pads, the following labels out_link_1 ... out_link_M are associated to the output pads.

when two link labels with the same name are found in the filtergraph, a link between the corresponding input and output pad is created. match made in heaven 💖

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. see how that works? unlabelled ones just find each other 🫶

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.

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 floating pads, no exceptions, bestie — that's the vibe check.

leading and trailing whitespaces (space, tabs, or line feeds) separating tokens in the filtergraph specification are ignored. which means you can express the filtergraph using empty lines and spaces to improve readability. aesthetic AND functional, we love to see it 🌸

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. and it's possible to specify swscale flags for those automatically inserted scalers by prepending sws_flags=flags; to the filtergraph description. auto-scaling, but make it your call.

here is a BNF description of the filtergraph syntax:

<NAME>             ::= sequence of alphanumeric characters and '_'
<FILTER_NAME>      ::= <NAME>["@"<NAME>]
<LINKLABEL>        ::= "[" <NAME> "]"
<LINKLABELS>       ::= <LINKLABEL> [<LINKLABELS>]
<FILTER_ARGUMENTS> ::= sequence of chars (possibly quoted)
<FILTER>           ::= [<LINKLABELS>] <FILTER_NAME> ["=" <FILTER_ARGUMENTS>] [<LINKLABELS>]
<FILTERCHAIN>      ::= <FILTER> [,<FILTERCHAIN>]
<FILTERGRAPH>      ::= [sws_flags=<flags>;] <FILTERCHAIN> [;<FILTERGRAPH>]

🎀 Notes on filtergraph escaping 🎀

ok this is the part where it gets spicy 😭 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.

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). yeah. look at all those backslashes. it's giving ancient runes 💀

finally an additional level of escaping is needed when writing the filtergraph description in a shell command, which depends on the escaping rules of the adopted shell. for example, assuming that \ is special and needs to be escaped with another \, the previous string will finally result in:

-vf "drawtext=text=this is a \\\\\\'string\\\\\\'\\\\: may contain one\\, or more\\, special characters"

in order to avoid cumbersome escaping when using a commandline tool accepting a filter specification as input, it is advisable to avoid direct inclusion of the filter or options specification in the shell.

for example, in case of the drawtext filter, you might prefer to use the textfile option in place of text to specify the text to render. seriously bestie, save yourself the emotional damage 🫶

⏰ 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. like, the filter just steps aside, respectfully.

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 ⚠️ like fr fr, don't
w, h — width and height of the input frame if video

additionally, these filters support an enable command that can be used to re-define the expression.

like any other filtering option, the enable option follows the same rules.

for example, to enable a blur filter (smartblur) from 10 seconds to 3 minutes, and a curves filter starting at 3 seconds:

smartblur = enable='between(t,10,3*60)',
curves    = enable='gte(t,3)' : preset=cross_process

see ffmpeg -filters to view which filters have timeline support. go check, it's free 👑

🎮 CHANGING OPTIONS AT RUNTIME WITH A COMMAND 🎮

some options can be changed during the operation of the filter using a command. these options are marked T on the output of ffmpeg -h filter=<name of filter>. the name of the command is the name of the option, and the argument is the new value. live tweaking, no restart needed 💅

✨ OPTIONS FOR FILTERS WITH SEVERAL INPUTS ✨

okay so some filters have several inputs and they all share a lil common set of options 🫶 and here's the tea: you can only set these by name — the short notation simply does not work for them, no cap 💅

eof_action — what to do when EOF hits the secondary input. it takes one of these values:
- repeat — repeat the last frame. this is the default, she's the safe queen 👑
- endall — end both streams, everybody goes home 💀
- pass — pass the main input through
shortest — set to 1 and the output is forced to terminate when the shortest input terminates. default is 0
repeatlast — set to 1 and the filter is forced to extend the last frame of the secondary streams alllll the way to the end of the primary stream 🧚 a value of 0 disables this behavior. default is 1
ts_sync_mode — how strictly to sync streams based on secondary input timestamps. the values are:
- 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 with --disable-filters ✨ and the configure output will show you the audio filters that made it into your build, so you know exactly who's invited to the party 💖

below is a description of the currently available audio filters. buckle up bestie, it's a long list 🎀

🦋 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. the Affine projection algorithm can make trade-offs between computation complexity and convergence speed — it's literally giving balance 💅

here's the description of the accepted options:

order — set the filter order
projection — set the projection order
mu — set the filter mu
delta — set the coefficient to initialize the internal covariance matrix
out_mode — set the filter output samples. it accepts these 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 — lowkey it's a whole journey 🧚

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 as 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 these 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 the above options as commands, periodt 💅

🎀 acontrast

simple audio dynamic range compression/expansion filter.

the filter accepts these 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, she's literally just standing there being a vibe check 🫶

🌈 acrossfade

apply cross fade from one input audio stream to another input audio stream ✨ the cross fade is applied for the specified duration near the end of the first stream.

the filter accepts these 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 the 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 — icon behavior, honestly 👑

the filter accepts these 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 a 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 an 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 literally a bit crusher with enhanced functionality, and she is serving 💅 A bit crusher is used to audibly reduce the number of bits an audio signal is sampled with. It does not 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 and the upper half of the signal. And there's an Anti-Aliasing setting that can produce "softer" crushing sounds. iconic.

Another feature 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. obsessed.

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 all the above options as commands. slay.

🎀 acue

Delay audio filtering until a given wallclock timestamp. See the cue filter. no cap.

🦋 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 the size of the 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 the 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 the 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 also be 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 the size of the 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 the number of given delays is smaller than the number of channels, all remaining channels will not be delayed. If you want to delay an exact number of samples, append 'S' to the number. If you want instead to delay in seconds, append 's' to the 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 all the above options as commands. 💖

🧚 aderivative, aintegral

Compute derivative/integral of audio stream.

Applying both filters one after another produces original audio. literally math but make it audio.

🔥 adrc

Apply spectral dynamic range controller filter to input audio stream.

A description of the accepted options follows.

transfer — Set the transfer expression.
- 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 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 all the above options as commands. periodt.

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.
cutbelow — cuts the frequencies below the detection threshold. gone, ✂️ bye.
cutabove — cuts the frequencies above the detection threshold.
boostbelow — boosts the frequencies below the detection threshold instead of cutting them 🔥
boostabove — boosts the frequencies above the detection threshold.

Default mode is cutbelow, bestie. That's the starting vibe.

dftype — sets the type of detection filter. pick your fighter, one of these:
- bandpass
- lowpass
- highpass
- peak

Default type is bandpass.

tftype — sets the type of target filter. the options are giving very specific energy:
- bell
- lowshelf
- highshelf

Default type is bell.

auto — automatically grabs the threshold from the detection filter. by default it's disabled. this one is lowkey iconic when you want to detect the threshold inside a certain time frame of the input audio stream — in that case the option value literally changes at runtime, no cap. the available values are:
- disabled — disables using the automatically gathered threshold value.
- off — stop picking the threshold value.
- on — start picking the threshold value.
- adaptive — adaptively picks the threshold value by calculating sliding window entropy. big brain moment 🧠✨
precision — sets which precision to use when processing samples:
- auto — auto picks the internal sample format depending on other filters.
- float — always uses the single-floating point precision sample format.
- double — always uses the double-floating point precision sample format.

Commands

This filter supports all the above options as commands. yes, all of them. live tweaking, slay 💅

🦋 adynamicsmooth 🦋

Applies dynamic smoothing to the input audio stream. it's giving glow-up but for your waveform ✨

Here's what the accepted options are about:

sensitivity — sets how sensitive it is to frequency fluctations. default is 2. allowed range is from 0 to 1e+06.
basefreq — sets a base frequency for the smoothing. default value is 22050. allowed range is from 2 to 1e+06.

Commands

This filter supports all the above options as commands 💖

🌈 aecho 🌈

Applies echoing to the input audio. we're literally in a cave now, bestie 🏔️

So — echoes are reflected sound. they can happen naturally amongst mountains (and sometimes large buildings) when you talk or shout, and digital echo effects emulate that whole behaviour. people use them constantly to help fill out the sound of a single instrument or vocal, it's a classic. 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 each have their own different delays and decays. obsessed.

Here are the accepted parameters:

in_gain — sets the input gain of the reflected signal. default is 0.6.
out_gain — sets the output gain of the reflected signal. default is 0.3.
delays — sets the list of time intervals in milliseconds between the original signal and the reflections, separated by '|'. allowed range for each "delay" is (0 - 90000.0]. default is 1000.
decays — sets the list of loudness of the 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 🎀

The audio emphasis filter creates or restores material that was taken directly from LPs or emphased CDs with different filter curves. like, to store music on vinyl the signal has to be altered by a filter first to even out the disadvantages of that recording medium. then once the material is played back, the inverse filter has to be applied to restore the distortion of the frequency response. it's a whole cycle, fr fr 💿

The filter accepts the following options:

level_in — sets the input gain.
level_out — sets the output gain.
mode — sets the filter mode. for restoring material use "reproduction" mode, otherwise use "production" mode. default is "reproduction" mode.
type — sets the filter type. this selects the medium. can be one of the following:
- col — select Columbia.
- emi — select EMI.
- bsi — select BSI (78RPM).
- riaa — select RIAA.
- cd — select Compact Disc (CD).
- 50fm — select 50µs (FM).
- 75fm — select 75µs (FM).
- 50kf — select 50µs (FM-KF).
- 75kf — select 75µs (FM-KF).

Commands

This filter supports all the above options as commands 🫶

⭐ aeval ⭐

Modifies an audio signal according to the specified expressions. math but make it audio 🧮✨

This filter accepts one or more expressions (one for each channel), which get evaluated and used to modify a corresponding audio signal.

It accepts the following parameters:

exprs — sets 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. so the last one is the main character for everyone left over 👑
channel_layout, c — sets the 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 the input channel with number CH.

Note: this filter is slow. For faster processing you should use a dedicated filter. we respect honesty here 💀

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 then 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 — the whole point is to create a more "crisp" or "brilliant" sound. it's giving ✨sparkle✨

The filter accepts the following options:

level_in — sets the input level prior to processing of the signal. allowed range is from 0 to 64. default value is 1.
level_out — sets the output level after processing of the signal. allowed range is from 0 to 64. default value is 1.
amount — sets the amount of harmonics added to the original signal. allowed range is from 0 to 64. default value is 1.
drive — sets the amount of newly created harmonics. allowed range is from 0.1 to 10. default value is 8.5.
blend — sets the octave of the newly created harmonics. allowed range is from -10 to 10. default value is 0.
freq — sets the lower frequency limit of producing harmonics, in Hz. allowed range is from 2000 to 12000 Hz. default is 7500 Hz.
ceil — sets the upper frequency limit of producing harmonics. allowed range is from 9999 to 20000 Hz. If the value is lower than 10000 Hz, no limit is applied.
listen — mutes the original signal and outputs only the added harmonics. by default it's disabled. solo mode, basically 🙈

Commands

This filter supports all the above options as commands 💜

🌸 afade 🌸

Applies a fade-in/out effect to the input audio. the classic, the icon, the moment 🎬

Here are the accepted parameters:

type, t — specifies the effect type. either "in" for a fade-in, or "out" for a fade-out effect. default is "in".
start_sample, ss — specifies the number of the start sample for starting to apply the fade effect. default is 0.
nb_samples, ns — specifies 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, and at the end of the fade-out transition the output audio will be silence. default is 44100.
start_time, st — specifies 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 — specifies 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, and 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 — sets the curve for the fade transition. prepare yourself, there's a lot of them:
- 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 — sets the initial gain for fade-in, or the final gain for fade-out. default value is 0.0.
unity — sets the initial gain for fade-out, or the final gain for fade-in. default value is 1.0.

Commands

This filter supports all the above options as commands ✨

Examples

Fade in the first 15 seconds of audio:

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

Fade out the last 25 seconds of a 900 seconds audio:

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

🧚 afftdn 🧚

Denoise audio samples with FFT. we love a clean-sounding bestie 🫧

okay so here's the full rundown of every parameter this thing takes. buckle up 🎀

noise_reduction, nr — sets how much noise reduction you get, 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 — picks what kind of noise we're dealing with. 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. literally the classic 💅

band_noise, bn — set a custom band noise profile for every one of 15 bands. bands are separated by ' ' or '|'. yep, fifteen, count 'em.
residual_floor, rf — sets the residual floor in dB. allowed range is -80 to -20. default is -38 dB.
track_noise, tn — turns on noise floor tracking. by default it's disabled. with this enabled, the noise floor gets automatically adjusted. 🦋
track_residual, tr — turns on residual tracking. by default it's disabled.
output_mode, om — sets the output mode. it accepts these values:
- input, i — pass input unchanged.
- output, o — pass noise filtered out.
- noise, n — pass only noise.

default value is output.

adaptivity, ad — sets the adaptivity factor, which is used for how fast to adapt gains adjustments per each frequency bin. value 0 enables instant adaptation, while higher values react much slower. allowed range is from 0 to 1. default value is 0.5.
floor_offset, fo — sets the noise floor offset factor. this option is used to adjust offset applied to measured noise floor. it's only effective when noise floor tracking is enabled. allowed range is from -2.0 to 2.0. default value is 1.0.
noise_link, nl — sets the noise link used for multichannel audio. it accepts these values:
- none — use unchanged channel's noise floor.
- min — use measured min noise floor of all channels.
- max — use measured max noise floor of all channels.
- average — use measured average noise floor of all channels.

default value is min.

band_multiplier, bm — sets the band multiplier factor, used for how much to spread bands across frequency bins. allowed range is from 0.2 to 5. default value is 1.25.
sample_noise, sn — toggles capturing and measurement of the noise profile from the input audio. it accepts these values:
- start, begin — start sample noise capture.
- stop, end — stop sample noise capture and measure new noise band profile.

default value is "none".

gain_smooth, gs — sets the gain smooth spatial radius, used to smooth gains applied to each frequency bin. useful to reduce random music noise artefacts. higher values increase smoothing of gains. allowed range is from 0 to 50. default value is 0. 🧚

Commands

this filter supports some of the above-mentioned options as commands. yes bestie, runtime control, we love to see it 👑

Examples

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

afftdn=nr=10:nf=-40

reduce white noise by 10dB, also set initial noise floor to -80dB and enable automatic tracking of noise floor so noise floor will gradually change during processing:

afftdn=nr=10:nf=-80:tn=1

reduce noise by 20dB, using noise floor of -40dB and using commands to take noise profile of first 0.4 seconds of input audio:

asendcmd=0.0 afftdn sn start,asendcmd=0.4 afftdn sn stop,afftdn=nr=20:nf=-40

afftfilt

apply arbitrary expressions to samples in the frequency domain. this is the "do whatever math you want in the frequency domain" filter and it's honestly iconic 💜

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 these 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 window size. allowed range is from 16 to 131072. default is 4096
win_func — sets the window function. it accepts these 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". that's a lot of windows, and not one of them is the kind you look through 🪟
overlap — sets window overlap. if 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

apply an arbitrary Finite Impulse Response filter. 🌀

this filter is designed for applying long FIR filters, up to 60 seconds long.

it can be used as a component for digital crossover filters, room equalization, cross talk cancellation, wavefield synthesis, auralization, ambiophonics, ambisonics and spatialization. like, it does a lot.

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 same as the number of channels in the first stream.

it accepts these 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 whole IR is processed.
gtype — ⚠️ this option is deprecated, and does nothing. she's retired, leave her alone 💀
irnorm — set norm to be applied to IR coefficients before filtering. allowed range is from -1 to 2. IR coefficients are normalized with the calculated vector norm set by this option. for negative values, no norm is calculated, and IR coefficients are not modified at all. default is 1.
irlink — for multichannel IR, if this option is set to true, all IR channels will be normalized with the maximal measured gain of all IR channels coefficients as set by the irnorm option. when disabled, all IR coefficients in each IR channel 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.

(four for four on the do-nothing squad, no cap 🙈)

minp — set minimal partition size used for convolution. default is 8192. allowed range is from 1 to 65536. lower values decrease latency at the cost of higher CPU usage.
maxp — set maximal partition size used for convolution. default is 8192. allowed range is from 8 to 65536. lower values may increase CPU usage.
nbirs — set the number of input impulse response streams which will be switchable at runtime. allowed range is from 1 to 32. default is 1.
ir — set the IR stream which will be used for convolution, starting from 0, and it should always be lower than the value supplied by the nbirs option. default is 0. this option can be changed at runtime via commands. main character energy fr 🔥
precision — set which precision to use when processing samples. it accepts these values:
- auto — auto pick internal sample format depending on other filters.
- float — always use single-floating point precision sample format.
- double — always use double-floating point precision sample format.

default value is auto.

irload — set when to load the IR stream. can be "init" or "access". the first one loads and prepares all IRs on initialization, the second one once on first access of a specific IR. default is "init".

Examples

apply reverb to stream using 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, where 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 an 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"

okay real talk, that last one literally says irnom=-1 — yes, irnom, not irnorm. that's straight from the source bestie, copy it exactly as shown and don't come for me about the spelling 😭

aformat

set output format constraints for the input audio. the framework will negotiate the most appropriate format to minimize conversions. she's a negotiator, a diplomat, a peacemaker 🫶

it accepts these parameters:

sample_fmts, f — a '|'-separated list of the sample formats you're asking for 🎀
sample_rates, r — a '|'-separated list of the sample rates you want ✨
channel_layouts, cl — a '|'-separated list of channel layouts, bestie 💅

Check the Channel Layout section in the ffmpeg-utils(1) manual for the exact syntax you need, iykyk.

If you just... don't mention a parameter? Then literally all values are allowed. No cap. 🦋

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 your input audio samples. Iconic.

The filter accepts these options:

shift — sets the frequency shift. Allowed range is -INT_MAX to INT_MAX. Default value is 0.0 🎚️
level — sets the output gain applied to the final output. Allowed range is from 0.0 to 1.0. Default value is 1.0 🔥
order — sets the filter order used for filtering. Allowed range is from 1 to 16. Default is 8 ⭐

Commands

This filter supports all of the above options as commands, fr fr 👑

🧚 afwtdn 🧚

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

Here's what the accepted options do:

sigma — sets the noise sigma. Allowed range is from 0 to 1. Default value is 0. This option controls the strength of the denoising applied to your input samples. The most useful way to set this is via decibels, e.g. -45dB 💜
levels — sets the number of wavelet levels of decomposition. Allowed range is from 1 to 12. Default is 10. Heads up: setting this too low makes the denoising performance very poor 😭
wavet — sets the wavelet type used for decomposition of the input frame. They're sorted by number of coefficients, from lowest to highest. More coefficients means worse filtering speed, but overall better quality. The available wavelets are:
- sym2
- sym4
- rbior68
- deb10
- sym10
- coif5
- bl3
percent — sets the percent of full denoising. Allowed range is from 0 to 100 percent. Default value is 85 percent, aka partial denoising ✨
profile — if enabled, the first input frame will be used as the noise profile. Fair warning: 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 from 512 to 65536. Default frame size is 8192 samples 🎀
softness — sets the softness applied inside the thresholding function. Allowed range is from 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 your useful signals. Main character energy: only the loud stuff gets to pass.

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. And 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, bestie.

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. Signals shorter than the chosen attack time will be left untouched 🫶

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 you set "upward" mode, the higher parts of the signal get amplified, expanding the dynamic range in the upward direction. Otherwise, in the "downward" case, the 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 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 by 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 — 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. Very math, very slay.

It accepts these parameters:

zeros, z — sets B/numerator/zeros/reflection coefficients 🔢
poles, p — sets A/denominator/poles/ladder coefficients 🔢
gains, k — sets channels gains 🎚️
dry_gain — sets input gain 🎀
wet_gain — sets output gain ✨
format, f — sets coefficients format:
- ll — lattice-ladder function
- sf — analog transfer function
- tf — digital transfer function
- zp — Z-plane zeros/poles, cartesian (default)
- pr — Z-plane zeros/poles, polar radians
- pd — Z-plane zeros/poles, polar degrees
- sp — S-plane zeros/poles
process, r — sets type of processing:
- d — direct processing
- s — serial processing
- p — parallel processing
precision, e — sets filtering precision:
- dbl — double-precision floating-point (default)
- flt — single-precision floating-point
- i32 — 32-bit integers
- i16 — 16-bit integers
normalize, n — normalize filter coefficients, enabled by default. Enabling it will normalize the magnitude response at DC to 0dB 💅
mix — how much to use the filtered signal in the output. Default is 1. Range is between 0 and 1 🎛️
response — show IR frequency response, magnitude (magenta), phase (green) and group delay (yellow) in an additional video stream. By default it's disabled 🌈
channel — set for which IR channel to display the frequency response. By default the first channel is displayed. This option is only used when response is enabled 👀
size — set video stream size. This option is only used when response is enabled 📐

Coefficients in "tf" and "sf" format are separated by spaces and are in ascending order. Periodt.

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, iykyk 🧮

Different coefficients and gains can be provided for every channel — in that case use '|' to separate the coefficients or gains. The last provided coefficients will be 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 the analog transfer function format:

aiir=z=1.3057 0 0 0:p=1.3057 2.3892 2.1860 1:f=sf:r=d

💀 alimiter 💀

The limiter prevents an input signal from rising over a desired threshold. This limiter uses lookahead technology to keep your signal from distorting. Which means... there's a small delay after the signal is processed. Keep in mind that the delay it produces is the attack time you set, bestie — don't say I didn't warn you 😭

The filter accepts these options:

level_in — sets input gain. Default is 1 🎚️
level_out — sets output gain. Default is 1 🎚️
limit — don't let signals above this level pass the limiter. Default is 1 🛑
attack — the limiter will reach its attenuation level in this amount of time, in milliseconds. Default is 5 milliseconds ⏱️
release — come back from limiting to attenuation 1.0 in this amount of milliseconds. Default is 50 milliseconds ⏳
asc — when gain reduction is always needed, ASC takes care of releasing to an average reduction level rather than reaching a reduction of 0 in the release time 🎀
asc_level — select how much the release time is affected by ASC. 0 means nearly no changes in release time, while 1 produces higher release times 📈
level — auto level output signal. Default is enabled. This normalizes audio back to 0dB if enabled ✨
latency — compensate the delay introduced by using the lookahead buffer set with the attack parameter. Also flush the valid audio data in the lookahead buffer when the stream hits EOF 🫶

Depending on the setting you pick, it's recommended to upsample the input 2x or 4x times with aresample before applying this filter. No cap, take the advice 🔥

⭐ 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. It's giving "same energy, different vibe" 💫

The filter accepts these 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 — which channels are we even filtering bestie? default is all available ones. slay.
normalize, n — normalizes the biquad coefficients. default? disabled 🙈. turn it on and the magnitude response at DC gets normalized to 0dB. cute.
order, o — sets the filter order. it can be 1 or 2. default is 2. that's it that's the option 💅
transform, a — sets the transform type of the IIR filter. the options are:
- 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, bestie:

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

🌀 aloop

loops audio samples. living that replay life, no cap.

the filter accepts these options:

loop — sets the number of loops. setting it to -1 gives you infinite loops (main character energy, it never ends 💀). 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 the 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. friendship bracelet behavior 🫶.

all inputs must have the same sample rate, and format. non-negotiable.

if the inputs don't have the same duration, the output will stop with the shortest one. sad but true 😭.

the filter accepts the following options:

inputs — sets the number of inputs. default is 2.
layout_mode — controls how the output channel layout is determined and whether the audio channels get reordered during the merge. three modes, all valid, all iconic:
- legacy — this is the mode where the filter behaved historically, 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 this 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. lowkey a gamble 🙈.

reset — this mode ignores the input channel layouts and does no channel reordering. the output will have all the channels of the first input, then all the channels of the second input, in that order, and so on.

the output channel layout will only specify the total channel count.

normal — this mode keeps channel name and designation information from the input channels and does no channel reordering. the output will have all the channels of the first input, then all the channels of the second input, in that order, and so on.

Examples ✨

merge two mono files into a stereo stream:

amovie=left.wav [l] ; amovie=right.mp3 [r] ; [l] [r] amerge

multiple merges assuming 1 video stream and 6 audio streams in input.mkv:

ffmpeg -i input.mkv -filter_complex "[0:1][0:2][0:3][0:4][0:5][0:6] amerge=inputs=6" -c:a pcm_s16le output.mkv

🎧 amix

mixes multiple audio inputs into a single output. the group chat merge feature fr fr.

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. it's giving helpful.

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. 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 to or after filtering by this filter if this option is disabled. by default it's enabled. protect your ears 💀.

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. easy.

✖️ 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.

with this element-wise multiplication you can create amplitude fades and amplitude modulations. math but make it fashion 💅.

🎚️ anequalizer

high-order parametric multiband equalizer for each channel. the glow-up filter, literally.

it accepts the following parameters:

params — this option string is in the format: "cchn f=cf w=w g=g t=f | ..." each equalizer band is separated by '|'.
- chn — set the channel number to which equalization will be applied. if the input doesn't have that channel the entry is ignored. rude but okay.
- 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, 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. visualization slay 📈.
size — set the video stream size. only useful if the curves option is activated.
mgain — set 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 — set the frequency scale used to draw the frequency response in the video output. can be linear or logarithmic. default is logarithmic.
colors — set 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. default? white. always white. 🤍

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. 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.

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. bye background hiss, you will not be missed 💀.

each sample is adjusted by looking for other samples with similar contexts. this context similarity is defined by comparing their surrounding patches of size p. patches are searched in an area of r around the sample.

the filter accepts the following options:

strength, s — set the denoising strength. allowed range is from 0.00001 to 10000. default value is 0.00001.
patch, p — set the patch radius duration. allowed range is from 1 to 100 milliseconds. default value is 2 milliseconds.
research, r — set the research radius duration. allowed range is from 2 to 300 milliseconds. default value is 6 milliseconds.
output, o — set the output mode. it accepts the following values:
- i — pass input unchanged.
- o — pass noise filtered out.
- n — pass only noise.

default value is o.

smooth, m — set the smooth factor. default value is 11. allowed range is from 1 to 1000.

Commands 💬

this filter supports all the above options as commands. every single one. generous queen 👑.

🧠 anlmf, anlms

apply the Normalized Least-Mean-(Squares|Fourth) algorithm to the first audio stream using the second audio stream.

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 (the difference between the desired, 2nd input audio stream, and the actual signal, the 1st input audio stream). it's giving "learn from your mistakes" but make it DSP 🎀.

a description of the accepted options follows.

order — set filter order.
mu — set filter mu.
eps — sets the filter's eps. that's it, that's the whole option ✨
leakage — sets the filter's leakage.
out_mode — picks which samples actually come out the other side. it takes these values:
- i — pass the 1st input. straightforward icon 💅
- d — pass the 2nd input.
- o — pass the difference between desired (2nd input) and the error signal estimate.
- n — pass the difference between input (1st input) and the error signal estimate.
- e — pass the error signal estimated samples.

default value is o. yes bestie, o is the main character here 👑

precision — tells the filter which precision to use when processing samples:
- auto — auto pick the internal sample format depending on other filters. the lowkey reliable one 🫶
- float — always use single-floating point precision sample format.
- double — always use double-floating point precision sample format.

🌸 Examples 🌸

one of many usages of this filter is noise reduction — input audio gets filtered with the same samples delayed by a fixed amount. one such example for stereo audio is:

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, excluding option "order". so don't even try it, she's not taking requests on that one 💀

anull

pass the audio source unchanged to the output. literally does nothing and honestly? respect. no notes, no drama, zero processing ✨

apad

pad the end of an audio stream with silence. like adding a lil quiet outro so nothing ends abruptly 🦋

this can be used together with ffmpeg -shortest to extend audio streams to the same length as the video stream. iconic combo fr fr.

here's a description of the accepted options:

packet_size — set 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. they cannot both be the chosen one 🙈
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. used only if set to 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. used only if set to 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. girl will just keep going 😭 like no cap, forever.

ngl, note that for ffmpeg 4.4 and earlier a zero pad_dur or whole_dur also caused the filter to add silence indefinitely. old behaviour, be aware 💀

💖 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, 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

add a phasing effect to the input audio. it's giving swoosh ✨

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. obsessed with the sweep, honestly 🔥

here's a description of the accepted parameters:

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 the vibe, literally 🌈

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. all of them. every single one. periodt 👑

apsnr

measure Audio Peak Signal-to-Noise Ratio. it's a measurement girlie, not a filter girlie 📊

this filter takes two audio streams for input, and outputs first audio stream. results are in dB per channel at end of either input. so you feed her two streams and she tells you the numbers 💅

apsyclip

apply Psychoacoustic clipper to input audio stream. clipping, 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.
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 psychoacoustic clipper. allowed range is from 1 to 20. default value is 10.
level — auto level output signal. default is disabled. this normalizes audio back to 0dBFS if enabled.

⭐ Commands ⭐

this filter supports all the above options as commands. yes, all of them, go off 💜

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 have 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 — 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) — pulsator acts as an autopanner. at 1 both curves match again. every setting in between moves the phase shift gapless 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. it's giving whoosh left, whoosh right 🔥🎀

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 waveform shape the LFO will use. can be one of: sine, triangle, square, sawup or sawdown. default is sine.
amount — set modulation. define how much of 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. she's flexible like that 💅

this filter is also able to stretch/squeeze the audio data to make it match the timestamps, or to inject silence / cut out audio to make it match the timestamps, do a combination of both, or do neither. 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. iykyk 🎀

🌈 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, like backwards. the drama of it all 😭

Warning: this filter requires memory to buffer the entire clip, so trimming is suggested. fr fr, trim first or your RAM is gonna have a bad time 💀

🍓 Examples 🍓

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

atrim=end=5,areverse

arls

apply Recursive Least Squares algorithm to the first audio stream using the second audio stream. big math energy 🧚✨

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 (difference between the desired, 2nd input audio stream and the actual signal, the 1st input audio stream). yes that's a whole sentence, i read it twice so you didn't have to 🫶

a description of the accepted options follows:

order — set the filter order.
lambda — set the forgetting factor.
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. again. o stays winning 👑

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.

arnndn

reduce noise from speech using Recurrent Neural Networks. the neural net is literally doing the "shhh" for you 🤫✨

this filter accepts the following options:

model, m — set train model file to load. this option is always required. like, non-negotiable, she will not run without it 💀
mix — sets how much of the filtered samples get mixed into the final output 🫶 Allowed range is -1 to 1, default is 1. Ngl the negative values are the special little gremlins here: they set how much filtered noise to keep in the final filter output. Want to actually hear the noise that got removed from your input signal? Set this to -1 and enjoy the chaos 💀

Commands

This filter supports all the above options as commands ✨

🎧 asdr

Measures the Audio Signal-to-Distortion Ratio fr fr.

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.

🎚️ asetnsamples

Sets the number of samples per each output audio frame.

Heads up bestie: 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 🌸

The filter accepts the following options:

nb_out_samples, n — the number of frames per each output audio frame. This number is meant as the number of samples per each channel. Default is 1024.
pad, p — if set to 1, the filter pads the last audio frame with zeroes so the last frame has the same number of samples as all the previous ones. Default is 1.

For example, to set the number of per-frame samples to 1234 and disable padding for the last frame, do:

asetnsamples=n=1234:p=0

🚀 asetrate

Sets the sample rate without altering the PCM data. The way this works, it'll cause a change of speed and pitch — yes, both, that's the vibe 😭

It accepts these options:

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

🔍 ashowinfo

Shows a line containing all kinds of info for each input audio frame 👀 The input audio itself is not modified, she's just observing.

The shown line is a sequence of key/value pairs of the form key:value.

These are the values 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 Audio Scaled-Invariant Signal-to-Distortion Ratio ✨

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 gets saturated along a smooth curve, rather than the abrupt shape of hard-clipping — think cozy rounded edges instead of a jagged little cliff 🔥

This filter accepts the following options:

type — set the type of soft-clipping.

It accepts the following values:

hard
tanh
atan
cubic
exp
alg
quintic
sin
erf

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

Commands

This filter supports all the 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 — set the window length in samples. Default is 2048. Allowed range is 32 to 65536.
win_func — set the window function.

It accepts the following values:

rect
bartlett
hann, hanning
hamming
blackman
welch
flattop
bharris
bnuttall
bhann
sine
nuttall
lanczos
gauss
tukey
dolph
cauchy
parzen
poisson
bohman
kaiser

Default is "hann".

overlap — set window overlap. Allowed range is 0 to 1. Default is 0.5.
measure — select the parameters which are measured. The metadata keys can be used as flags; default is all, which measures everything. none disables all measurement.

A list of each metadata key follows:

mean
variance
centroid
spread
skewness
kurtosis
entropy
flatness
crest
flux
slope
decrease
rolloff

🗣️ asr

Automatic Speech Recognition 🎤

This filter uses PocketSphinx for speech recognition. To enable compilation of this filter, you need to configure FFmpeg with --enable-pocketsphinx.

It accepts the following options:

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

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

📈 astats

Displays time domain statistical information about the audio channels. Statistics are calculated and displayed for each audio channel and, where applicable, an overall figure is also given 👑

It accepts the following option:

length — short window length in seconds, used for peak and through RMS measurement. Default is 0.05 (50 milliseconds). Allowed range is [0 - 10].
metadata — set metadata injection. All metadata keys are prefixed with lavfi.astats.X, where X is the channel number starting from 1 or the string Overall. Default is disabled.

Available keys for each channel are: Bit_depth Crest_factor DC_offset Dynamic_range Entropy Flat_factor Max_difference Max_level Mean_difference Min_difference Min_level Noise_floor Noise_floor_count Number_of_Infs Number_of_NaNs Number_of_denormals Peak_count Abs_Peak_count Peak_level RMS_difference RMS_peak RMS_trough Zero_crossings Zero_crossings_rate

and for Overall: Bit_depth DC_offset Entropy Flat_factor Max_difference Max_level Mean_difference Min_difference Min_level Noise_floor Noise_floor_count Number_of_Infs Number_of_NaNs Number_of_denormals Number_of_samples Peak_count Abs_Peak_count Peak_level RMS_difference RMS_level RMS_peak RMS_trough

For example, a full key looks like lavfi.astats.1.DC_offset or lavfi.astats.Overall.Peak_count ✨

Read below for the description of the keys.

reset — set the number of frames over which cumulative stats are calculated before being reset. Default is disabled.
measure_perchannel — select the parameters which are measured per channel. The metadata keys can be used as flags; default is all, which measures everything. none disables all per-channel measurement.
measure_overall — select the parameters which are measured overall. The metadata keys can be used as flags; default is all, which measures everything. none disables all overall measurement.

A description of the measure keys follows:

none — no measures.
all — all measures.
Bit_depth — overall bit depth of audio, i.e. number of bits used for each sample.
Crest_factor — standard ratio of peak to RMS level (note: not in dB).
DC_offset — mean amplitude displacement from zero.
Dynamic_range — measured dynamic range of audio in dB.
Entropy — entropy measured across 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 🎶

The filter accepts the following options:

dry — set dry gain, i.e. how much of the original signal is kept. Allowed range is 0 to 1. Default is 1.0.
wet — set wet gain, i.e. how much of the filtered signal is kept. Allowed range is 0 to 1. Default is 1.0.
boost — set max boost factor. Allowed range is 1 to 12. Default is 2.
decay — set delay line decay gain value. Allowed range is 0 to 1. Default is 0.0.
feedback — set delay line feedback gain value. Allowed range is 0 to 1. Default is 0.9.
cutoff — set cutoff frequency in Hertz. Allowed range is 50 to 900. Default is 100.
slope — sets how much slope the cutoff frequency gets. Allowed range is 0.0001 to 1, and the default is 0.5. ✨
delay — sets the delay, bestie. Allowed range is from 1 to 100, default 20.
channels — which channels get processed. Default is literally all the available ones, no notes. 💅

Commands 💬

This filter supports the all above options as commands — yes, every single one of those, go off.

asubcut 🎀

Cuts subwoofer frequencies. Like, bye to the rumble.

The whole point of this one: it lets you set a custom, steeper roll off than the highpass filter does, so it's way better at attenuating frequency content in the stop-band. It's giving surgical. 🦋

The filter accepts the following options:

cutoff — sets the cutoff frequency in Hertz. Allowed range is 2 to 200, default 20.
order — sets the filter order. Available values are from 3 to 20, default 10.
level — sets the input gain level. Allowed range is from 0 to 1, default 1.

Commands 💬

This filter supports the all above options as commands. periodt.

asupercut 👑

Cuts super frequencies. High-pitch girlies, your time is up.

The filter accepts the following options:

cutoff — sets the cutoff frequency in Hertz. Allowed range is 20000 to 192000, default 20000.
order — sets the filter order. Available values are from 3 to 20, default 10.
level — sets the input gain level. Allowed range is from 0 to 1, default 1.

Commands 💬

This filter supports the all above options as commands. slay.

asuperpass 🌸

Applies a high order Butterworth band-pass filter. Fancy filter, fancy name, we love her.

The filter accepts the following options:

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

Commands 💬

This filter supports the all above options as commands. no cap.

asuperstop 💀

Applies a high order Butterworth band-stop filter. Same energy as above but it's the stop one, iykyk.

The filter accepts the following options:

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

Commands 💬

This filter supports the all above options as commands. 💜

atempo 🔥

Adjusts audio tempo. The vibes-speed dial.

The filter accepts exactly one parameter: the audio tempo. If you don't specify it, the filter will assume a nominal 1.0 tempo. Tempo must be in the [0.5, 100.0] range. That's the whole range, that's the whole rule.

Heads up bestie 🫶 — note that a tempo greater than 2 will skip some samples rather than blend them in. If for any reason that's a concern, it's always possible to daisy-chain several instances of atempo to achieve the desired product tempo. Chain her up.

Examples ✨

Slow down audio to 80% tempo:

atempo=0.8

To speed up audio to 300% tempo:

atempo=3

To speed up audio to 300% tempo by daisy-chaining two atempo instances:

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

Commands 💬

This filter supports the following commands:

tempo — change filter tempo scale factor. Syntax for the command is : "tempo".

atilt 🦋

Applies a spectral tilt filter to the audio stream. It's giving main character audio glow-up.

This filter applies any spectral roll-off slope over any specified frequency band. Any slope, any band, we're flexible like that.

The filter accepts the following options:

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

Commands 💬

This filter supports the all above options as commands. 💖

atrim ✂️

Trims the input so the output contains one continuous subpart of the input. One clean little piece, nothing extra.

It accepts the following parameters:

start — timestamp (in seconds) of the start of the section to keep. I.e. the audio sample with the timestamp start will be the first sample in the output.
end — specify the time of the first audio sample that will be dropped, i.e. the audio sample immediately preceding the one with the timestamp end will be the last sample in the output.
start_pts — same as start, except this option sets the start timestamp in samples instead of seconds.
end_pts — same as end, except this option sets the end timestamp in samples instead of seconds.
duration — the maximum duration of the output in seconds.
start_sample — the number of the first sample that should be output.
end_sample — the number of the first sample that should be dropped.

start, end, and duration are expressed as time duration specifications; see the Time duration section in the ffmpeg-utils(1) manual. 📖

Okay so here's the part that matters 😭 — note that the first two sets of the start/end options and the duration option look at the frame timestamp, while the _sample options simply count the samples that pass through the filter. So start/end_pts and start/end_sample will give you different results when the timestamps are wrong, inexact, or don't start at zero. Also note that this filter does not modify the timestamps. If you want the output timestamps to start at zero, insert the asetpts filter after the atrim filter. That's the move. 💅

If multiple start or end options are set, this filter tries to be greedy and keep all samples that match at least one of the specified constraints. To keep only the part that matches all the constraints at once, chain multiple atrim filters. Greedy queen.

The defaults are such that all the input is kept. So it's possible to set, e.g., just the end values to keep everything before the specified time. Obsessed with a low-effort config.

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. Two streams, one glow-up.

Resulted samples are always between -1 and 1 inclusive, fr fr:

If the result is 1, it means two input samples are highly correlated in that selected segment.
Result 0 means they are not correlated at all.
If the result is -1, it means two input samples are out of phase, which means they cancel each other. 💀

The filter accepts the following options:

size — sets the size of the segment over which cross-correlation is calculated. Default is 256. Allowed range is from 2 to 131072.
algo — sets the algorithm for cross-correlation. Can be "slow" or "fast" or "best". Default is "best". The fast algorithm assumes mean values over any given segment are always zero and thus needs much less calculation to make. This is generally not true, but it's valid for typical audio streams. (Valid ≠ always true, bestie, that's the whole tea. 🍓)

Examples ✨

Calculate correlation between channels in 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 (20dB per decade). Iconic rolloff behavior honestly.

The filter accepts the following 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 to specify the band-width of the filter.
- h — Hz
- q — Q-Factor
- o — octave
- s — slope
- k — kHz
width, w — specifies the band-width of a filter in width_type units.
mix, m — how much to use the filtered signal in the output. Default is 1. Range is between 0 and 1.
channels, c — specifies which channels to filter; by default all available 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 — sets the transform type of the IIR filter.
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — sets the precision of filtering.
- auto — pick automatic sample format depending on surround filters.
- s16 — always use signed 16-bit.
- s32 — always use signed 32-bit.
- f32 — always use float 32-bit.
- f64 — always use float 64-bit.
block_size, b — sets the block size used for reverse IIR processing. If this value is set to a high enough value (higher than the impulse response length truncated when it reaches near zero values), filtering will become linear phase; otherwise, if it's not big enough, it 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 bandpass frequency. Syntax for the command is : "frequency".
width_type, t — change bandpass width_type. Syntax for the command is : "width_type".
width, w — change bandpass width. Syntax for the command is : "width".
mix, m — change bandpass mix. Syntax for the command is : "mix".

bandreject 💜

Applies a two-pole Butterworth band-reject filter with central frequency frequency, and (3dB-point) band-width width. The filter rolls off at 6dB per octave (20dB per decade). Same math, opposite attitude — she rejects. 🙈

The filter accepts the following options:

frequency, f — sets the filter's central frequency. Default is 3000.
width_type, t — sets the method to specify the band-width of the filter.
- h — Hz
- q — Q-Factor
- o — octave
- s — slope
- k — kHz
width, w — specifies the band-width of a filter in width_type units.
mix, m — how much to use the filtered signal in the output. Default is 1. Range is between 0 and 1.
channels, c — specifies which channels to filter; by default all available 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 — 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 for you, depending on the surround filters ✨
s16 — always signed 16-bit, no cap
s32 — always signed 32-bit
f32 — always float 32-bit
f64 — always float 64-bit
block_size, b — sets the block size used for reverse IIR processing 🧚 if you set this high enough (higher than the impulse response length truncated when it reaches near-zero values) the filtering goes linear phase — otherwise, if it's not big enough, you just get nasty artifacts, and nobody wants that 💀

Heads up bestie: the filter delay will be exactly this many samples whenever you set it to a non-zero value. Like, exactly. 🎀

Commands

This filter supports the following commands, periodt:

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

bass, lowshelf

Boosts or cuts the bass (lower) frequencies of your 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), iykyk.

The filter accepts the following options:

gain, g — gives the gain at 0 Hz. Useful range is about -20 (big cut) to +20 (big boost) 💅

Beware of clipping when you're using a positive gain, bestie. That's a jump scare you don't want. 😭

frequency, f — sets the filter's central frequency, so you can extend or reduce the frequency range being boosted or cut. Default value is 100 Hz.
width_type, t — sets the method for specifying 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 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 — normalizes biquad coefficients. By default it's disabled. Enabling it will normalize the magnitude response at DC to 0dB.
transform, a — sets the transform type of the IIR filter:
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — sets the precision of filtering:
- auto — pick automatic sample format depending on surround filters
- s16 — always use signed 16-bit
- s32 — always use signed 32-bit
- f32 — always use float 32-bit
- f64 — always use float 64-bit
block_size, b — sets the block size used for reverse IIR processing. If this value is set high enough (higher than the impulse response length truncated when it reaches near-zero values) the filtering will become linear phase; otherwise, if it's not big enough, it'll just produce nasty artifacts 🙈

Again for the people in the back: the filter delay will be exactly this many samples when set to a non-zero value.

Commands

This filter supports the following commands:

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

biquad

Applies a biquad IIR filter with the coefficients you give it 🦋 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 channels get filtered.

Commands

This filter supports the following commands:

a0
a1
a2
b0
b1
b2 — change the biquad parameter. Syntax for the command is: "value"

And also:

mix, m — how much of the filtered signal goes 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 normalizes 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 — 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. 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 you get nasty artifacts instead 💀

And yeah — the filter delay will be exactly this many samples when set to a non-zero value.

bs2b

Bauer stereo to binaural transformation, which lowkey makes listening to stereo audio records on headphones way better 🎧

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

Remaps input channels to new locations, iconic behaviour honestly 🌈

It accepts the following parameters:

map — maps channels from input to output. The argument is a '|'-separated list of mappings, each in the in_channel-out_channel or in_channel form. in_channel can be either the name of the input channel (e.g. FL for front left) or its index in the input channel layout. out_channel is the name of the output channel or its index in the output channel layout. If out_channel isn't given, then it's implicitly an index, starting with zero and increasing by one for each mapping. Mixing different types of mappings is not allowed and will result in a parse error 🚫
channel_layout — the channel layout of the output stream. If it's not specified, 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, bestie. Just so you know. 🙈

If no mapping is present, the filter implicitly maps input channels to output channels, preserving indices.

Examples

So for example, assuming a 5.1+downmix input MOV file:

ffmpeg -i in.mov -filter 'channelmap=map=DL-FL|DR-FR' out.wav

will create an output WAV file tagged as stereo from the downmix channels of the input ✨

To fix a 5.1 WAV improperly encoded in AAC's native channel order:

ffmpeg -i in.wav -filter 'channelmap=1|2|0|5|3|4:5.1' out.wav

channelsplit

Splits each channel from an input audio stream into its own separate output stream 🎀

It accepts the following parameters:

channel_layout — the channel layout of the input stream. Default is "stereo".
channels — a channel layout describing the channels to be extracted as separate output streams, or "all" to extract each input channel as a separate stream. Default is "all".

Choosing channels that aren't present in the channel layout of the input will result in an error. No cap. 💅

Examples

For example, assuming a stereo input MP3 file:

ffmpeg -i in.mp3 -filter_complex channelsplit out.mkv

will create an output Matroska file with two audio streams — one containing only the left channel and the other the right channel 💖

Split a 5.1 WAV file into per-channel files:

ffmpeg -i in.wav -filter_complex
'channelsplit=channel_layout=5.1[FL][FR][FC][LFE][SL][SR]'
-map '[FL]' front_left.wav -map '[FR]' front_right.wav -map '[FC]'
front_center.wav -map '[LFE]' lfe.wav -map '[SL]' side_left.wav -map '[SR]'
side_right.wav

Extract only LFE from a 5.1 WAV file:

ffmpeg -i in.wav -filter_complex 'channelsplit=channel_layout=5.1:channels=LFE[LFE]'
-map '[LFE]' lfe.wav

chorus

Adds a chorus effect to the audio 🦋 Can make a single vocal sound like a whole chorus, but you can also throw it on instrumentation.

Chorus is like an echo effect with a short delay, but where echo's delay is constant, chorus's delay is varied using sinusoidal or triangular modulation. The modulation depth defines the range the modulated delay is played before or after the delay. So the delayed sound will sound slower or faster — like the delayed sound is tuned around the original one, exactly like a chorus where some vocals are slightly off key 🎤

It accepts the following parameters:

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

Examples

A single delay:

chorus=0.7:0.9:55:0.4:0.25:2

Two delays:

chorus=0.6:0.9:50|60:0.4|0.32:0.25|0.4:2|1.3

Fuller sounding chorus with three delays:

chorus=0.5:0.9:50|60|40:0.4|0.32|0.3:0.25|0.4|0.3:2|2.3|1.3

compand

Compresses or expands the audio's dynamic range, which is genuinely so useful 🫶

It accepts the following parameters:

attacks
decays

These are a list of times in seconds, per channel, over which the instantaneous level of the input signal is averaged to determine its volume. attacks refers to an increase in volume and decays refers to a decrease in volume. For most situations the attack time (the response to 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 number of attacks & decays you specify is lower than the number of channels, the last set attack/decay will be used for all remaining channels 💅

points — the actual transfer function, given as a list of points in dB relative to the max possible signal amplitude. Every key-point list follows this syntax: "x0/y0|x1/y1|x2/y2|...." or "x0/y0 x1/y1 x2/y2 ...." — yes bestie, either the pipe or the space, both slay.

The input values must be in strictly increasing order, but the transfer function does not have to rise monotonically. The point "0/0" is assumed by default but you can override it (with "0/out-dBn"). Typical values for the transfer function are "-70/-70|-60/-20|1/0".

soft-knee — sets the curve radius in dB for all joints. Defaults to 0.01.
gain — extra gain in dB applied at every point on the transfer function, so you can tweak the overall gain without rewriting the whole curve. Defaults to 0.
volume — an initial volume in dB assumed for each channel when filtering starts. This lets you supply a nominal level up front so that, for example, a huge gain isn't slammed onto the initial signal levels before the companding has even started doing its thing. For audio that's initially quiet, a typical value is -90 dB. Defaults to 0.
delay — a delay in seconds. The input audio gets analyzed immediately, but the audio itself is delayed before it's fed to the volume adjuster. Pick a delay roughly equal to the attack/decay times and the filter gets to operate in predictive mode instead of reactive mode. Defaults to 0. 💅

✨ Examples ✨

Make music with both quiet and loud passages, suitable for listening in a noisy environment:

compand=.3|.3:1|1:-90/-60|-60/-40|-40/-30|-20/-20:6:0:-90:0.2

Another example, this one for audio with whisper parts and explosion parts:

compand=0|0:1|1:-90/-900|-70/-70|-30/-9|0/-3:6:0:0:0

A noise gate for when the noise is at a lower level than the signal:

compand=.1|.1:.2|.2:-900/-900|-50.1/-900|-50/-50:.01:0:-90:.1

Here's another noise gate, this time for when the noise is at a higher level than the signal (which makes it, in some ways, similar to squelch):

compand=.1|.1:.1|.1:-45.1/-45.1|-45/-900|0/-900:.01:45:-90:.1

2:1 compression starting at -6dB:

compand=points=-80/-80|-6/-6|0/-3.8|20/3.5

2:1 compression starting at -9dB:

compand=points=-80/-80|-9/-9|0/-5.3|20/2.9

2:1 compression starting at -12dB:

compand=points=-80/-80|-12/-12|0/-6.8|20/1.9

2:1 compression starting at -18dB:

compand=points=-80/-80|-18/-18|0/-9.8|20/0.7

3:1 compression starting at -15dB:

compand=points=-80/-80|-15/-15|0/-10.8|20/-5.2

Compressor/Gate:

compand=points=-80/-105|-62/-80|-15.4/-15.4|0/-12|20/-7.6

Expander:

compand=attacks=0:points=-80/-169|-54/-80|-49.5/-64.6|-41.1/-41.1|-25.8/-15|-10.8/-4.5|0/0|20/8.3

Hard limiter at -6dB:

compand=attacks=0:points=-80/-80|-6/-6|20/-6

Hard limiter at -12dB:

compand=attacks=0:points=-80/-80|-12/-12|20/-12

Hard noise gate at -35 dB:

compand=attacks=0:points=-80/-115|-35.1/-80|-35/-35|20/20

Soft limiter:

compand=attacks=0:points=-80/-80|-12.4/-12.4|-6/-8|0/-6.8|20/-2.8

🎀 compensationdelay 🎀

The Compensation Delay Line is a metric-based delay — its whole personality is compensating for microphones or speakers sitting in different positions.

Picture it: you recorded guitar with two mics placed in different spots. Because the front of a sound wave has a fixed speed under normal conditions, the phasing of the mics can vary depending on where they are and how they're positioned relative to each other. The best mix happens when those mics are in phase (synchronized). And heads up — a distance of ~30 cm between mics makes one mic capture the signal in antiphase to the other, which makes the final mix sound moody. 😭 This filter fixes phasing problems by adding different delays to each mic track and syncing them up.

You get the best result when you take one track as your base and synchronize the other tracks to it one by one. Also remember that the synchronization/delay tolerance depends on the sample rate too — higher sample rates give you more tolerance. ✨

The filter accepts these parameters:

mm — set the distance in millimeters. This is compensation distance for fine tuning. Default is 0.
cm — set the distance in cm. This is compensation distance for tightening up the distance setup. Default is 0.
m — set the distance in meters. This is compensation distance for hard distance setup. Default is 0.
dry — set the dry amount, aka how much unprocessed (dry) signal you want. Default is 0.
wet — set the wet amount, aka how much processed (wet) signal you want. Default is 1.
temp — set the temperature in degrees Celsius. This is the temperature of the environment. Default is 20.

Commands

This filter supports all of the above options as commands. 👑

🦋 crossfeed 🦋

Applies a headphone crossfeed filter.

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

The whole intent? Produce a more speaker-like sound for the listener. 🎧

The filter accepts these options:

strength — set the strength of the crossfeed. Default is 0.2. Allowed range is from 0 to 1. This sets the gain of the low shelf filter for the side part of the stereo image. Default is -6dB. Max allowed is -30db when strength is set to 1.
range — set the soundstage wideness. Default is 0.5. Allowed range is from 0 to 1. This sets the cut off frequency of the low shelf filter. Default cut off is 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 the input gain. Default is 0.9.
level_out — set the output gain. Default is 1.
block_size — set the block size used for reverse IIR processing. Set this to a high enough value (higher than the impulse response length truncated when it reaches near-zero values) and the filtering becomes linear phase — otherwise, if it's not big enough, it'll just produce nasty artifacts. 💀

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

Commands

This filter supports all of 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 these 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). Want inverse filtering? Use a negative value.
c — enable clipping. Enabled by default.

Commands

This filter supports all of the above options as commands.

⭐ dcshift ⭐

Applies a DC shift to the audio.

Super useful for removing a DC offset (which could be, like, a hardware problem in your recording chain) from the audio. The effect of a DC offset is reduced headroom, and therefore reduced volume. Pro tip: the astats filter can tell you if a signal has a DC offset. 🔍

shift — set the DC shift. Allowed range is [-1, 1]. It says how much to shift the audio by.
limitergain — optional. It should have a value much less than 1 (e.g. 0.05 or 0.02) and it's used to prevent clipping.

💅 deesser 💅

Applies de-essing to the audio samples. Bye bye, harsh ssss sounds.

i — set the intensity for triggering de-essing. Allowed range is from 0 to 1. Default is 0.
m — set the amount of ducking on the treble part of the sound. Allowed range is from 0 to 1. Default is 0.5.
f — how much of the original frequency content to keep when de-essing. Allowed range is from 0 to 1. Default is 0.5.
s — set the output mode.

It accepts these values:

i — pass input unchanged.
o — pass ess filtered out.
e — pass only ess.

Default value is o.

👑 dialoguenhance 👑**

Enhances dialogue in stereo audio.

This filter takes stereo input and produces surround (3.0) channel output. The newly produced front center channel carries enhanced speech dialogue that was originally available in both stereo channels. The front left and front right channels come out the same as they were in the stereo input. 🎬

The filter accepts these 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 of the above options as commands.

🍓 drmeter 🍓

Measures audio dynamic range.

For reference: DR values of 14 and higher show up in very dynamic material. DR of 8 to 13 shows up in transition material. And anything less than 8 has very poor dynamics and is very compressed. 📉

The filter accepts this option:

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

🌈 dynaudnorm 🌈

Dynamic Audio Normalizer. Say hi to the main character of normalization. ✨

This filter applies a certain amount of gain to the input audio in order to bring its peak magnitude up to a target level (e.g. 0 dBFS). But — unlike more "simple" normalization algorithms — the Dynamic Audio Normalizer dynamically re-adjusts the gain factor to the input audio. That means it can apply extra gain to the "quiet" sections of the audio while avoiding distortions or clipping the "loud" sections. In other words: the Dynamic Audio Normalizer "evens out" the volume of quiet and loud sections, in the sense that the volume of each section gets brought to the same target level. Note, though, that it pulls this off without applying "dynamic range compressing" — it retains 100% of the dynamic range within each section of the audio file. 🔥

`framelen`, `f`

Sets the frame length in milliseconds — range is 10 to 8000 milliseconds, default is 500 milliseconds. 💅

Okay so here's the tea on why frames even exist: the Dynamic Audio Normalizer chews through your input audio in little chunks called frames. Why? Because a peak magnitude is literally meaningless for a single lonely sample value 😭 — you have to look at a contiguous sequence of samples to figure out the peak. A "standard" normalizer would just grab the peak of the whole entire file, but our girl here determines the peak magnitude individually for each frame, one at a time, no cap. 🌸

You spec the frame length in milliseconds, and the default 500 has been found to give good results with most files (it just works, iykyk). Heads up though: the exact frame length in number of samples gets determined automatically, based on 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. 🫶

Honestly? This is probably the most important parameter of the whole Dynamic Audio Normalizer — it's the "window size" of the Gaussian smoothing filter. The window is specified in frames, centered around the current frame, and for simplicity's sake it has to be odd. So the default 31 takes into account the current frame, plus the 15 frames before it and the 15 frames after it. Bestie math. 🧚

Bigger window → stronger smoothing effect → less gain variation → slower gain adaptation
Smaller window → weaker smoothing effect → more gain variation → faster gain adaptation

In other words: the more you crank this value up, the more the Dynamic Audio Normalizer starts acting like a "traditional" normalization filter. Flip it around and the more you decrease it, the more she behaves like a dynamic range compressor. Choose your fighter. 💀

`peak`, `p`

Sets the target peak value — the highest permissible magnitude level for the normalized audio input. 🌈

The filter will try to approach that target peak magnitude as closely as possible, but at the same time it makes sure the normalized signal will never exceed the peak magnitude. A frame's maximum local gain factor is imposed directly by the target peak magnitude. Default value is 0.95, which leaves a headroom of 5%*. 🎀

It is not recommended to go above this value. Respect the headroom, bestie.

`maxgain`, `m`

Sets the maximum gain factor. Range 1.0 to 100.0, default 10.0. 🔥

The Dynamic Audio Normalizer works out the maximum possible (local) gain factor for each input frame — meaning the biggest gain factor that won't cause clipping or distortion. That number is determined by the frame's highest magnitude sample. But she also bounds each frame's maximum gain factor by a predetermined (global) maximum gain factor, specifically so we don't get unhinged gain factors in "silent" or almost-silent frames. 🙈

Default 10.0 is enough for most inputs, and it's usually not recommended to raise it. That said, if your input has an extremely low overall volume level, you might need to allow higher gain factors. 👑

Important caveat though: this is not a simple "hard" threshold (i.e. it doesn't just cut off values above the threshold). Instead a "sigmoid" threshold function is applied, so the gain factors smoothly approach the threshold value and never exceed it. Smooth like butter. 🦋

`targetrms`, `r`

Sets the target RMS. Range 0.0 to 1.0. Default is 0.0 — disabled. 💜

By default, the Dynamic Audio Normalizer does "peak" normalization, meaning the maximum local gain factor for each frame is defined (only) by that frame's highest magnitude sample. That way the samples can be amplified as much as possible without exceeding the maximum signal level — i.e. without clipping. ✨

Optionally, though, she can also take into account the frame's root mean square, abbreviated RMS. 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 looking at the signal's peak magnitude. So by adjusting all frames to a constant RMS value, you can establish a uniform "perceived loudness." 🎧

If you specify a target RMS value, a frame's local gain factor becomes the factor that would result in exactly that RMS value. Note, however, that the maximum local gain factor is still restricted by the frame's highest magnitude sample, in order to prevent clipping. Safety first. 🫶

`coupling`, `n`

Turn on channels coupling — and yes, it's enabled by default. 🌸

By default the Dynamic Audio Normalizer amplifies all channels by the same amount, meaning the same gain factor gets applied to every channel, so the maximum possible gain factor is determined by the "loudest" channel. Main character channel energy. 💅

But sometimes recordings have uneven volume across channels — e.g. one channel might be "quieter" than the other one(s). In that case you can use this option to disable channel coupling. Then the gain factor is determined independently for each channel, depending only on that individual channel's highest magnitude sample. Which is lowkey perfect for harmonizing the volume of the different channels. 🎀

`correctdc`, `c`

Turns on 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 values are represented in the -1.0 to 1.0 range, regardless of the original input format. Normally the audio signal, or "waveform", should be centered around the zero point. Meaning: if we calculate the mean value 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 value from 0.0, in either the positive or negative direction, that's called a DC bias or DC offset. And since a DC bias is clearly undesirable, this filter provides 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 makes sure those samples are centered around 0.0 again. 💖

Also, to avoid "gaps" at the frame boundaries, the DC correction offset values get interpolated smoothly between neighbouring frames. No awkward seams. ✨

`altboundary`, `b`

Turns on alternative boundary mode — disabled by default. 🧚

The Dynamic Audio Normalizer takes into account a certain neighbourhood around each frame: the preceding frames and the subsequent frames. But for the "boundary" frames sitting at the very beginning and the very end of the audio file, not all neighbouring frames are available. Specifically, for the first few frames the preceding frames aren't known, and for the last few frames the subsequent frames aren't known. 😭

So the question is: which gain factors should we assume for the missing frames in that "boundary" region? She implements two modes to deal with this. The default boundary mode assumes a gain factor of exactly 1.0 for the missing frames, resulting 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 won't be pruned, so the full dynamic range is retained within each local neighbourhood. But sometimes you do want to combine her 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 she doesn't just slap a fixed threshold on there — the threshold value gets adjusted for each individual frame. 👑

In general, smaller parameters result in stronger compression, and vice versa. And values below 3.0 are not recommended, because audible distortion may appear. You've been warned bestie. 💀

`threshold`, `t`

Sets the target threshold value — the lowest permissible magnitude level for the audio input that will actually get normalized. 🌈

If an input frame's volume is above this value, the frame will be normalized. Otherwise, the frame may not be normalized at all. Default is 0, which means all input frames get normalized. This option is mostly useful when you don't want digital noise getting amplified. 🙈

`channels`, `h`

Specify which channels to filter. By default, all available channels get filtered. 🎀

`overlap`, `o`

Specify overlap for frames. If set to 0 (the default), no frame overlapping is done. 🦋

Using values >0 and <1 will make less conservative gain adjustments — like when the framelen option is set to a smaller value. And if the framelen option value is compensated for non-zero overlap, then the gain adjustments will be smoother across time compared to the zero overlap case. Smooth queen. ✨

`curve`, `v`

Specify the peak mapping curve expression, which is going to be 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. yep, all of them, go wild ✨

🎧 earwax

Makes audio easier to listen to on headphones. literally what it says on the tin 🫶

This filter adds `cues' to 44.1kHz stereo (i.e. audio CD format) audio so that when listened to on headphones the stereo image is moved from inside your head (standard for headphones) to outside and in front of the listener (standard for speakers). think main character energy but for your ears 👑

Ported from SoX.

💜 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 (unlike bandpass and bandreject filters) that at all other frequencies is unchanged. it's giving "surgical glow-up for one frequency band" 💅

In order to produce complex equalisation curves, this filter can be given several times, each with a different central frequency. stack 'em, bestie 🎀

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.
- 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. ⚠️ fr fr, watch that ceiling
mix, m — How much to use filtered signal in output. Default is 1. Range is between 0 and 1.
channels, c — Specify which channels to filter, by default all available are filtered.
normalize, n — Normalize biquad coefficients, by default is disabled. Enabling it will normalize magnitude response at DC to 0dB.
transform, a — Set transform type of IIR filter.
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — Set precision of filtering.
- auto — Pick automatic sample format depending on surround filters.
- s16 — Always use signed 16-bit.
- s32 — Always use signed 32-bit.
- f32 — Always use float 32-bit.
- f64 — Always use float 64-bit.
block_size, b — Set block size used for reverse IIR processing. If this value is set to high enough value (higher than impulse response length truncated when reaches near zero values) filtering will become linear phase otherwise if not big enough it will just produce nasty artifacts. 😭

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

✨ Examples

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. hi-fi girlies stay winning 💖

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. channels doing a whole swap arc 😭
c — Enable clipping. By default is enabled.

💬 Commands

This filter supports the all above options as commands.

🦋 firequalizer

Apply FIR Equalization using arbitrary frequency response. the way this filter just does whatever curve you dream up 👑

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. that classic whoosh, iconic behavior 💫

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. mono girlies, this one's for you 🫶

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 make it bit depth ✨

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. Use something like -acodec pcm_s24le after the filter to get 24-bit PCM output. don't let wav snatch your bits 😭

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.

ngl if you skip this one you might get converted on without even knowing 💀

process_stereo — deals with the stereo channels together, as a unit. If target_gain doesn't match between the two channels, the filter treats that as invalid and just keeps the last valid target_gain instead. bestie behaviour honestly, it's not gonna trust a mismatched value 💅
cdt_ms — sets the code detect timer period, in ms. that's it, that's the whole thing ✨
force_pe — always extend peaks above -3dBFS, even when PE isn't signaled. like, pushing the peak extend through regardless 🙈
analyze_mode — replaces the audio with a solid tone and adjusts its amplitude to signal some specific part of the decoding process. the output file can be loaded in an audio editor next to the original to help with analysis. literally a visualisation hack and it ate 🔥

"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 build virtual loudspeakers around the listener, for binaural listening via headphones — it's giving "surround sound but make it your earbuds" ✨ the HRIRs come in via additional streams, and for each channel you need one stereo input stream 🎀

the filter accepts the following options:

map — sets the mapping of input streams for the 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. number of input streams must be not less than number of channels in the first stream plus one. no cap.
gain — sets the gain applied to audio. value is in dB. default is 0.
type — sets the processing type. can be time or freq. time processes audio in the time domain, which is slow 🐢 freq processes audio in the frequency domain, which is fast ⚡ default is freq. slay.
lfe — sets a custom gain for LFE channels. value is in dB. default is 0.
size — sets the size of the frame, in number of samples, processed at once. default value is 1024. allowed range is from 1024 to 96000. periodt.
hrir — sets the format of the hrir stream. default value is stereo. alternative value is multich. if it's set to stereo, the number of additional streams should be greater than or equal to the number of input channels in the first input stream — and 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 that additional stream should be equal to or greater than twice the number of channels of the first input stream. iykyk.

Examples

Full example using wav files as coefficients with amovie filters for a 7.1 downmix — each amovie filter uses a stereo file with IR coefficients as input. the files give coefficients for each position of the virtual loudspeaker 🌈

ffmpeg -i input.wav
-filter_complex "amovie=azi_270_ele_0_DFC.wav[sr];amovie=azi_90_ele_0_DFC.wav[sl];amovie=azi_225_ele_0_DFC.wav[br];amovie=azi_135_ele_0_DFC.wav[bl];amovie=azi_0_ele_0_DFC.wav,asplit[fc][lfe];amovie=azi_35_ele_0_DFC.wav[fl];amovie=azi_325_ele_0_DFC.wav[fr];[0:a][fl][fr][fc][lfe][bl][br][sl][sr]headphone=FL|FR|FC|LFE|BL|BR|SL|SR"
output.wav

Full example using wav files as coefficients with amovie filters for 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). the filter rolls off at 6dB per pole per octave (20dB per pole per decade). iconic behaviour, honestly 🌸

the filter accepts the following 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 a filter, in width_type units. applies only to double-pole filters. the default is 0.707q and 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 — specifies which channels to filter — by default, all available channels are filtered.
normalize, n — normalizes the biquad coefficients, disabled by default. enabling it will normalize the magnitude response at DC to 0dB.
transform, a — sets the transform type of the IIR filter:
- di
- dii
- tdi
- tdii
- latt
- svf
- zdf
precision, r — sets the precision of filtering:
- auto — pick 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 high enough (higher than the impulse response length, truncated when it reaches near-zero values), the filtering becomes linear phase — otherwise, if it isn't 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. no cap, that's literally the warning.

Commands 🫶

this filter supports the following commands:

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 accepts the following parameters:

inputs — the number of input streams. defaults to 2.
channel_layout — the desired output channel layout. defaults to stereo.
map — maps channels from the inputs to the output. the argument is a '|'-separated list of mappings, each in the "input_idx.in_channel-out_channel" form. input_idx is the 0-based index of the input stream. in_channel can be either the name of the input channel (e.g. FL for front left) or its index in the specified input stream. out_channel is the name of the output channel 🎀

the filter will attempt to guess the mappings when they're not specified explicitly — it does so by first trying to find an unused matching input channel, and if that fails, it picks the first unused input channel. bestie is doing her best with the info she has 💅

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. the way this one opens up a whole plugin ecosystem, obsessed 🌟

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 in each one of the directories specified by the colon-separated list in LADSPA_PATH; otherwise it's searched 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 (for example delay, threshold or gain). controls need to be defined using the following syntax: c0=value0|c1=value1|c2=value2|..., where valuei is the value set on the i-th control. alternatively they can also be defined using this syntax: value0|value1|value2|..., where valuei is the value set on the i-th control. if controls is set to "help", all available controls and their valid ranges are printed 👑
sample_rate, s — specifies the sample rate, default 44100. only used if the plugin has zero inputs.
nb_samples, n — sets the number of samples per channel per 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 longer than the specified duration, because 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.
latency, l — enables latency compensation, disabled by default. 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' "SWH Plugins" collection:

ladspa=fast_lookahead_limiter_1913:fastLookaheadLimiter:20|0|2

Attenuate low frequencies using the Multiband EQ from Steve Harris' "SWH Plugins" collection:

ladspa=mbeq_1197:mbeq:-24|-24|-24|0|0|0|0|0|0|0|0|0|0|0|0

Reduce the stereo image using "Narrower" from the "C* Audio Plugin Suite" (CAPS) library:
ladspa=caps:Narrower ✨
Okay so next up, ANOTHER white noise moment — this time repping the "C* Audio Plugin Suite" (CAPS) library:

ladspa=caps:White:.2

And some fractal noise, also from CAPS, because we're fancy like that:

ladspa=caps:Fractal:c=c1=1

Dynamic volume normalization via the "VLevel" plugin — chef's kiss for making quiet bits audible without blowing up the loud bits:

ladspa=vlevel-ladspa:vlevel_mono

🌸 Commands 🌸

This filter is a chatty one, it takes commands:

cN — modify the N-th control value. 💅 And if the value you pass isn't valid? Bestie, it's just ignored and the previous one stays. No drama, no crash.

🎀 loudnorm 🎀

EBU R128 loudness normalization, and it's giving main character energy in the audio world. It does both dynamic and linear normalization modes, and it supports both single pass (livestreams, files) and double pass (files) modes. This algorithm can target IL, LRA, and maximum true peak. In dynamic mode, to accurately detect true peaks, the audio stream will be upsampled to 192 kHz — so use the -ar option or the aresample filter to explicitly set an output sample rate. 🦋

It accepts these options:

I, i — set integrated loudness target. Range is -70.0 – -5.0. Default value is -24.0.
LRA, lra — set loudness range target. Range is 1.0 – 50.0. Default value is 7.0.
TP, tp — set maximum true peak. Range is -9.0 – +0.0. Default value is -2.0.
measured_I, measured_i — measured IL of 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 offset gain. The gain is applied before the true-peak limiter. Range is -99.0 – +99.0. Default is +0.0.
linear — normalize by linearly scaling the source audio. 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 which exceeds the target TP. If any of these conditions aren't met, normalization mode will revert to dynamic. Options are true or false. Default is true.
dual_mono — treat mono input files as "dual-mono". If a mono file is intended for playback on a stereo system, its EBU R128 measurement will be perceptually incorrect, and this option compensates for that effect when set to true. Multi-channel input files are not affected by this option. Options are true or false. Default is false. 🙈
print_format — set print format for stats. Options are summary, json, or none. Default value is none.
stats_file — write stats to 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 (that's the default, bestie). The filter rolls off at 6dB per pole per octave (20dB per pole per decade). Lowkey iconic.

It accepts these 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 used 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 the filter in width_type units. Applies only to the double-pole filter. The default is 0.707q and 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 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 — set the precision of filtering:
- auto — pick the 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 this value is set 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 will just produce nasty artifacts. 😭

Heads up: the filter delay will be exactly this many samples when set to a non-zero value. No cap.

🌈 Examples 🌈

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

lowpass=c=LFE

🌸 Commands 🌸

This filter supports these commands:

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

🦋 lv2 🦋

Load an LV2 (LADSPA Version 2) plugin. Plugin girlies rise.

To enable compilation of this filter you need to configure FFmpeg with --enable-lv2.

plugin, p — specifies the plugin URI. You may need to escape ':'. 💅
controls, c — set the '|'-separated list of controls, which are zero or more floating point values that determine the behavior of the loaded plugin (for example delay, threshold or gain). If controls is set to help, all available controls and their valid ranges are printed. Super handy.
sample_rate, s — specify the sample rate, default is 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 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 — yes, the crackle is on purpose, that's the vibe:

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. Slay.

🔥 mcompand 🔥

Multiband compress or expand the audio's dynamic range. It's giving loudness warpaint.

The input audio is divided into bands using 4th order Linkwitz-Riley IIRs. This is akin to the crossover of a loudspeaker, and results in a flat frequency response when there's no compander action happening.

It accepts this parameter:

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.

This filter is also designed to efficiently remap the channels of an audio stream — so like, it's not just a mixer, it's also a glow-up artist for your channel layout. ✨

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. 🫶

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. Big brain move.

💖 Mixing examples 💖

For example, if you want to down-mix from stereo to mono, but with a bigger factor for the left channel:

pan=1c|c0=0.9*c0+0.1*c1

A customized down-mix to stereo that works automatically for 3-, 4-, 5- and 7-channels surround:

pan=stereo| FL < FL + 0.5*FC + 0.6*BL + 0.6*SL | FR < FR + 0.5*FC + 0.6*BR + 0.6*SR

Note that ffmpeg integrates a default down-mix (and up-mix) system that should be preferred (see the -ac option) unless you have very specific needs. Like, don't do this manually if you don't have to, bestie.

🦋 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. Literally the best case scenario.

For example, if you have a 5.1 source and want a stereo audio stream by dropping the extra channels:

pan="stereo| c0=FL | c1=FR"

Given the same source, you can also switch front left and front right channels and keep the input channel layout:

pan="5.1| c0=c1 | c1=c0 | c2=c2 | c3=c3 | c4=c4 | c5=c5"

If the input is a stereo audio stream, you can mute the front left channel (and still keep the stereo channel layout) with:

pan="stereo|c1=c1"

Still with a stereo audio stream input, you can copy the right channel into both front left and right:

pan="stereo| c0=FR | c1=FR"

🍓 replaygain 🍓

ReplayGain scanner filter. This filter takes an audio stream as input and outputs it unchanged — like, it's just here to observe and judge, no notes. At the end of filtering it displays track_gain and track_peak.

It accepts these 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. Like, this is backstage crew, not the main character — let the other filters call it.

🌈 rubberband 🌈

Apply time-stretching and pitch-shifting with librubberband. Stretchy, pitchy, iconic.

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

The filter accepts these options:

tempo — sets the tempo scale factor 🎚️ (how fast the vibes go).
pitch — sets the pitch scale factor, so you can yeet the key up or down 💅.
transients — picks the transients detector. The possible values are literally:
- crisp
- mixed
- smooth
detector — picks the detector. Possible values:
- compound
- percussive
- soft
phase — sets the phase. Possible values:
- laminar
- independent
window — sets the processing window size. Possible values:
- standard
- short
- long
smoothing — sets smoothing. Possible values:
- off
- on
formant — enables formant preservation when shift pitching 🎤. Possible values:
- 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, bestie — you can poke it live while it's running 👀:

tempo — change the filter tempo scale factor. Syntax for the command is : "tempo"
pitch — change the filter pitch scale factor. Syntax for the command is : "pitch"

🎛️ `sidechaincompress`

Okay so this one is genuinely iconic ✨ — it acts like a normal compressor but it has the ability to compress the detected signal using a second input signal. It needs two input streams and returns one output stream. The first input stream gets processed depending on the second stream's signal. The filtered signal can then be filtered with other filters in later stages of processing. Go look at pan and amerge filter, iykyk 🫶.

The filter accepts the following options:

level_in — set input gain. Default is 1. Range is between 0.015625 and 64.
mode — set 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 is 0.125. Range is between 0.00097563 and 1.
ratio — set a ratio about which the signal is reduced. 1:2 means that if the level raised 4dB above the threshold, it will be only 2dB above after the reduction. Default is 2. Range is between 1 and 20.
attack — amount of milliseconds the signal has to rise above the threshold before gain reduction starts. Default is 20. Range is between 0.01 and 2000.
release — amount of milliseconds the signal has to fall below the threshold before reduction is decreased again. Default is 250. Range is between 0.01 and 9000.
makeup — 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 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 — set sidechain gain. Default is 1. Range is between 0.015625 and 64.
mix — how much to use the compressed signal in output. Default is 1. Range is between 0 and 1.

Commands

This filter supports all the above options as commands 💀 (yes, all of them, go off).

Examples

Full ffmpeg example taking 2 audio inputs, 1st input to be compressed depending on the signal of 2nd input and later compressed signal to be merged with 2nd input:

ffmpeg -i main.flac -i sidechain.flac -filter_complex "[1:a]asplit=2[sc][mix];[0:a][sc]sidechaincompress[compr];[compr][mix]amerge"

🚪 `sidechaingate`

A sidechain gate acts like a normal (wideband) gate but it has the ability to filter the detected signal before sending it to the gain reduction stage. Normally a gate uses the full range signal to detect a level above the threshold. For example: if you cut all lower frequencies from your sidechain signal, the gate will decrease the volume of your track only if not enough highs appear. With this technique you are 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 — set input level before filtering. Default is 1. Allowed range is from 0.015625 to 64.
mode — set the mode of operation. Can be "upward" or "downward". Default is "downward". If set to "upward" mode, higher parts of the signal will be amplified, expanding dynamic range in the upward direction. Otherwise, in the case of "downward", lower parts of the signal will be reduced.
range — set the level of gain reduction when the signal is below the threshold. Default is 0.06125. Allowed range is from 0 to 1. Setting this to 0 disables reduction and then the filter behaves like an expander 👀.
threshold — if a signal rises above this level the gain reduction is released. Default is 0.125. Allowed range is from 0 to 1.
ratio — set a ratio about which the signal is reduced. Default is 2. Allowed range is from 1 to 9000.
attack — amount of milliseconds the signal has to rise above the threshold before gain reduction stops. Default is 20 milliseconds. Allowed range is from 0.01 to 9000.
release — amount of milliseconds the signal has to fall below the threshold before the reduction is increased again. Default is 250 milliseconds. Allowed range is from 0.01 to 9000.
makeup — set amount of amplification of signal after processing. Default is 1. Allowed range is from 1 to 64.
knee — curve the sharp knee around the threshold to enter gain reduction more softly. Default is 2.828427125. Allowed range is from 1 to 8.
detection — choose if the exact signal should be taken for detection or an RMS-like one. Default is rms. Can be peak or rms.
link — choose if the average level between all channels or the louder channel affects the reduction. Default is average. Can be average or maximum.
level_sc — set sidechain gain. Default is 1. Range is from 0.015625 to 64.

Commands

This filter supports all the above options as commands 🫶.

🤫 `silencedetect`

Detects silence in an audio stream. Like, it literally clocks the quiet parts for you 💜.

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 — set noise tolerance. Can be specified in dB (in case "dB" is appended to the specified value) or amplitude ratio. Default is -60dB, or 0.001.
duration, d — set 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 is disabled.

Examples

Detect 5 seconds of silence with -50dB noise tolerance:

silencedetect=n=-50dB:d=5

Complete example with ffmpeg to detect silence with 0.0001 noise tolerance in silence.mp3:

ffmpeg -i silence.mp3 -af silencedetect=noise=0.0001 -f null -

✂️ `silenceremove`

Remove silence from the beginning, middle or end of the audio. The glow-up potential is unreal, ngl 🌟.

The filter accepts the following options:

start_periods — this value is used to indicate if audio should be trimmed at the beginning of the audio. A value of zero indicates no silence should be trimmed from the beginning. When specifying a non-zero value, it trims audio up until it finds non-silence. Normally, when trimming silence from the beginning of audio the start_periods will be 1, but it can be increased to higher values to trim all audio up to a specific count of non-silence periods. Default value is 0.
start_duration — specify the amount of time that non-silence must be detected before it stops trimming audio. By increasing the duration, bursts of noise can be treated as silence and trimmed off. Default value is 0.
start_threshold — this indicates what sample value should be treated as silence. For digital audio, a value of 0 may be fine, but for audio recorded from analog, you may wish to increase the value to account for background noise. Can be specified in dB (in case "dB" is appended to the specified value) or amplitude ratio. Default value is 0.
start_silence — specify max duration of silence at the beginning that will be kept after trimming. Default is 0, which is equal to trimming all samples detected as silence.
start_mode — specify 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, fr.
stop_periods — set the count for trimming silence from the end of audio. When specifying a positive value, it trims audio after it finds the specified silence period. To remove silence from the middle of a file, specify a stop_periods that is negative. This value is then treated as a positive value and is used to indicate the effect should restart processing as specified by stop_periods, making it suitable for removing periods of silence in the middle of the audio. Default value is 0.
stop_duration — specify a duration of silence that must exist before audio is not copied any more. By specifying a higher duration, silence that is wanted can be left in the audio. Default value is 0.
stop_threshold — okay so this is literally start_threshold's twin sister, except she's the one who trims the silence at the end of your audio instead of the start 💅. You can give it dB (just slap a "dB" on the end of your value) or an amplitude ratio. Default is 0. Slay.
stop_silence — sets the max duration of silence at the end that gets kept after all the trimming goes down 🎀. Default is 0, which literally means "trim every single sample we detected as silence", no survivors, periodt.
stop_mode — picks the detection mode for where silence starts, after your multi-channel audio has already begun. Values are any or all, default is all.
- any — literally any sample from any channel that reads as silence will trigger the start of silence trimming after the audio stream starts. Lowkey limited usage, ngl.
- all — only when every sample from every channel is detected as silence does the trimming start after the audio stream starts. Strict queen behavior 👑.
detection — tells the filter how to detect silence, bestie:
- avg — mean of absolute values of the samples in the moving window.
- rms — root squared mean of absolute values of the samples in the moving window.
- peak — maximum of absolute values of the samples in the moving window.
- median — median of absolute values of the samples in the moving window.
- ptp — absolute of the max peak to min peak difference of the samples in the moving window.
- dev — standard deviation of the values of the samples in the moving window.

Default value is rms ✨.

window — duration in seconds used to calculate the window size (in samples) for detecting silence. Setting it to 0 effectively disables all windowing and uses only a single sample per channel for detection — and in that case you may need to also set start_silence and/or stop_silence to nonzero values and start_duration and/or stop_duration to nonzero values. Default is 0.02, allowed range is 0 to 10. 🦋
timestamp — processing mode for every audio frame's output timestamp:
- write — full timestamps rewrite, keeps only the start time for the first output frame.
- copy — non-dropped frames keep the same timestamp as the input audio frame.

Default value is "write" 💖.

Examples 🌸

Bestie, this one lets you start a recording that doesn't have that awkward delay at the start — you know, the gap between pressing record and actually performing:

silenceremove=start_periods=1:start_duration=5:start_threshold=0.02

Trim all the silence from beginning to end where there's more than 1 second of it:

silenceremove=stop_periods=-1:stop_duration=1:stop_threshold=-90dB

Trim every digital silence sample using peak detection, beginning to end, where there's more than 0 samples of digital silence and it's detected in all channels at the same positions in the stream:

silenceremove=window=0:detection=peak:stop_mode=all:start_mode=all:stop_periods=-1:stop_threshold=0

Trim every 2nd silence period you encounter, beginning to end, where there's more than 1 second of silence per period:

silenceremove=stop_periods=-2:stop_duration=1:stop_threshold=-90dB

Same vibe as above but keep a maximum of 0.5 seconds of silence from each trimmed period:

silenceremove=stop_periods=-2:stop_duration=1:stop_threshold=-90dB:stop_silence=0.5

Same as above again, but keep a maximum of 1.5 seconds of silence from the start of the audio:

silenceremove=stop_periods=-2:stop_duration=1:stop_threshold=-90dB:stop_silence=0.5:start_periods=1:start_duration=1:start_silence=1.5:stop_threshold=-90dB

Commands 🎀

This filter supports some of the above options as commands. Periodt.

sofalizer 🎧

SOFAlizer uses head-related transfer functions (HRTFs) to build virtual loudspeakers around the listener so you get binaural listening through headphones — and yes, audio formats up to 9 channels are supported 💜. The HRTFs live in SOFA files (there's a whole database at <http://www.sofacoustics.org/>). SOFAlizer is developed at the Acoustics Research Institute (ARI) of the Austrian Academy of Sciences. Iconic institution behavior, honestly.

To enable compilation of this filter you need to configure FFmpeg with --enable-libmysofa ⭐

The filter accepts these options:

sofa — set the SOFA file used for rendering.
gain — set the gain applied to audio. Value is in dB. Default is 0.
rotation — set the rotation of the virtual loudspeakers in deg. Default is 0.
elevation — set the elevation of the virtual speakers in deg. Default is 0.
radius — set the distance in meters between the loudspeakers and the listener, using near-field HRTFs. Default is 1.
type — set the processing type. Values are time or freq. time processes audio in the time domain, which is slow 🐌. freq processes audio in the frequency domain, which is fast ⚡. Default is freq.
speakers — set custom positions of the virtual loudspeakers. The syntax is: <CH> <AZIM> <ELEV>[|<CH> <AZIM> <ELEV>|...]. Each virtual loudspeaker is described by a short channel name followed by azimuth and elevation in degrees, and each description is separated by '|'. For example, to override the front left and front right channel positions use: 'speakers=FL 45 15|FR 345 15'. Descriptions with unrecognised channel names are just ignored, no drama 💅.
lfegain — set a custom gain for LFE channels. Value is in dB. Default is 0.
framesize — set a custom frame size in number of samples. Default is 1024. Allowed range is from 1024 to 96000. Only used if the type option is set to freq.
normalize — should all IRs be normalized when importing the SOFA file? By default this is enabled.
interpolate — should the nearest IRs be interpolated with neighbor IRs if the exact position doesn't match? By default this is disabled.
minphase — minphase all IRs when loading the SOFA file. By default this is disabled.
anglestep — set the neighbor search angle step. Only used if the interpolate option is enabled.
radstep — set the neighbor search radius step. Only used if the interpolate option is enabled.

Examples 🌈

Using the ClubFritz6 sofa file:

sofalizer=sofa=/path/to/ClubFritz6.sofa:type=freq:radius=1

Using the ClubFritz12 sofa file with a bigger radius and a small rotation:

sofalizer=sofa=/path/to/ClubFritz12.sofa:type=freq:radius=2:rotation=5

Similar to above but with custom speaker positions for front left, front right, back left and back right, plus a custom gain:

"sofalizer=sofa=/path/to/ClubFritz6.sofa:type=freq:radius=2:speakers=FL 45|FR 315|BL 135|BR 225:gain=28"

speechnorm 🗣️

Speech Normalizer. It's giving audio glow-up ✨

This filter expands or compresses each half-cycle of audio samples — that's a local set of samples all above zero or all below zero, sitting between the two nearest zero crossings — depending on the threshold value, so the audio reaches the target peak value under the conditions controlled by the options below.

It accepts these options:

peak, p — set the expansion target peak value. This specifies the highest allowed absolute amplitude level for the normalized audio input. Default is 0.95, allowed range is 0.0 to 1.0.
expansion, e — set the maximum expansion factor. Allowed range is 1.0 to 50.0, default is 2.0. This option controls the maximum local half-cycle expansion of samples. The max expansion would be such that the local peak value reaches the target peak value but never surpasses it, and the ratio between the new and previous peak value does not surpass this option's value. Highkey important, ngl.
compression, c — set the maximum compression factor. Allowed range is 1.0 to 50.0, default is 2.0. This controls the maximum local half-cycle compression of samples. It's only used if the threshold option is set to a value greater than 0.0 — in those cases, when the local peak is lower than or the same as the value set by threshold, all samples belonging to that peak's half-cycle get compressed by the current compression factor.
threshold, t — set the threshold value. Default is 0.0, allowed range is 0.0 to 1.0. This specifies which half-cycles get compressed and which get expanded: any half-cycle whose local peak value is below or the same as this value gets compressed by the current compression factor; otherwise, if it's greater than the threshold value, it gets expanded with the expansion factor so it could reach the peak target value — but never surpass it 💖.
raise, r — set the expansion raising amount per each half-cycle of samples. Default is 0.001, allowed range is 0.0 to 1.0. This controls how fast the expansion factor is raised per each new half-cycle until it reaches the expansion value.

Setting this one too high may lead to distortions 😭 careful bestie.

fall, f — set the compression raising amount per each half-cycle of samples. Default is 0.001, allowed range is 0.0 to 1.0. This controls how fast the compression factor is raised per each new half-cycle until it reaches the compression value.
channels, h — specify which channels to filter. By default all available channels are filtered.
invert, i — enable inverted filtering, disabled by default. This flips the interpretation of the threshold option: when enabled, any half-cycle of samples whose local peak value is below or the same as the threshold option gets expanded, otherwise it gets compressed 🔄.
link, l — link channels when calculating the gain applied to each filtered channel sample, disabled by default. When disabled, each filtered channel's gain calculation is independent; when 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 is 0.0, which means disabled. Allowed range is 0.0 to 1.0.

Commands 🎀

This filter supports all the above options as commands. All of them. Obsessed.

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 the stereo swiss-army knife — it's got all the handy little utilities for managing stereo signals, like converting M/S stereo recordings to L/R while keeping full control over the parameters, or spreading the stereo image of your master track. Iconic behaviour honestly.

The filter accepts the following options:

level_in — sets the input level before filtering, for both channels. Default is 1. Allowed range is from 0.015625 to 64.
level_out — sets the output level after filtering, for both channels. Default is 1. Allowed range is from 0.015625 to 64.
balance_in — input balance between both channels. Default is 0. Allowed range is from -1 to 1.
balance_out — output balance between both channels. Default is 0. Allowed range is from -1 to 1.
softclip — enables softclipping, which gives you analog distortion instead of that harsh digital 0dB clipping. Disabled by default.
mutel — mutes the left channel. Disabled by default.
muter — mutes the right channel. Disabled by default.
phasel — flips the phase of the left channel. Disabled by default.
phaser — flips the phase of the right channel. Disabled by default.
mode — sets the stereo mode. Available values are:
- lr>lr — Left/Right to Left/Right, this is the default, no notes.
- lr>ms — Left/Right to Mid/Side.
- ms>lr — Mid/Side to Left/Right.
- lr>ll — Left/Right to Left/Left.
- lr>rr — Left/Right to Right/Right.
- lr>l+r — Left/Right to Left + Right.
- lr>rl — Left/Right to Right/Left.
- ms>ll — Mid/Side to Left/Left.
- ms>rr — Mid/Side to Right/Right.
- ms>rl — Mid/Side to Right/Left.
- lr>l-r — Left/Right to Left - Right.
slev — sets the level of the side signal. Default is 1. Allowed range is from 0.015625 to 64.
sbal — sets the balance of the side signal. Default is 0. Allowed range is from -1 to 1.
mlev — sets the level of the middle signal. Default is 1. Allowed range is from 0.015625 to 64.
mpan — sets the middle signal pan. Default is 0. Allowed range is from -1 to 1.
base — sets the stereo base between mono and inversed channels. Default is 0. Allowed range is from -1 to 1.
delay — sets the delay in milliseconds, i.e. how much to delay the left from the right channel and vice versa. Default is 0. Allowed range is from -20 to 20.
sclevel — sets the S/C level. Default is 1. Allowed range is from 1 to 100.
phase — sets the stereo phase in degrees. Default is 0. Allowed range is from 0 to 360.
bmode_in, bmode_out — set the balance mode for the balance_in/balance_out option. Can be one of the following:
- 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. Slay.

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 that's common to both channels, and by delaying the signal of the left into the right and vice versa — thereby widening the stereo effect. It's giving "bigger than your headphones," fr fr.

The filter accepts the following options:

delay — time in milliseconds of the delay of the left signal into the right and vice versa. Default is 20 milliseconds.
feedback — amount of gain in the delayed signal into the right and vice versa. This gives a delay effect of the left signal in the right output and vice versa, which gives the widening effect. Default is 0.3.
crossfeed — cross feed of the left into the right with inverted phase. This helps suppress the mono. If the value is 1 it will cancel all the signal common to both channels. Default is 0.3.
drymix — sets the level of the input signal of the original channel. Default is 0.8.

Commands 💬

This filter supports all of the above options except delay as commands. (Not delay, bestie — don't @ me.)

superequalizer 🎀

Applies an 18 band equalizer. Eighteen. Bands. Obsessed.

The filter accepts the following 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 🌈

Applies the audio surround upmix filter. This filter lets you produce multichannel output from an audio stream — mono-ish in, full cinematic situation out. Main character energy.

The filter accepts the following options:

chl_out — sets the output channel layout. By default, this is 5.1. See the Channel Layout section in the ffmpeg-utils(1) manual for the required syntax.
chl_in — sets the input channel layout. By default, this is stereo. See the Channel Layout section in the ffmpeg-utils(1) manual for the required syntax.
level_in — sets the input volume level. By default, this is 1.
level_out — sets the output volume level. By default, this is 1.
lfe — enable LFE channel output if the output channel layout has it. By default, this is enabled.
lfe_low — sets the LFE low cut off frequency. By default, this is 128 Hz.
lfe_high — sets the LFE high cut off frequency. By default, this is 256 Hz.
lfe_mode — sets the LFE mode, can be add or sub. Default is add. In add mode, the LFE channel is created from the input audio and added to the output. In sub mode, the LFE channel is created from the input audio and added to the output, but also all non-LFE output channels are subtracted with the output LFE channel.
smooth — sets the temporal smoothness strength, used to gradually change factors when transforming stereo sound in time. Allowed range is from 0.0 to 1.0. Useful for improving output quality with focus option values greater than 0.0. Default is 0.0. Only values inside this range and without edges are effective.
angle — sets the angle of the stereo surround transform. Allowed range is from 0 to 360. Default is 90.
focus — sets the focus of the stereo surround transform. Allowed range is from -1 to 1. Default is 0.
fc_in — sets the front center input volume. By default, this is 1.
fc_out — sets the front center output volume. By default, this is 1.
fl_in — sets the front left input volume. By default, this is 1.
fl_out — sets the front left output volume. By default, this is 1.
fr_in — sets the front right input volume. By default, this is 1.
fr_out — sets the front right output volume. By default, this is 1.
sl_in — sets the side left input volume. By default, this is 1.
sl_out — sets the side left output volume. By default, this is 1.
sr_in — sets the side right input volume. By default, this is 1.
sr_out — sets the side right output volume. By default, this is 1.
bl_in — sets the back left input volume. By default, this is 1.
bl_out — sets the back left output volume. By default, this is 1.
br_in — sets the back right input volume. By default, this is 1.
br_out — sets the back right output volume. By default, this is 1.
bc_in — sets the back center input volume. By default, this is 1.
bc_out — sets the back center output volume. By default, this is 1.
lfe_in — sets the LFE input volume. By default, this is 1.
lfe_out — sets the LFE output volume. By default, this is 1.
allx — sets the spread usage of the stereo image across the X axis for all channels. Allowed range is from -1 to 15. By default this value is negative -1, and thus unused.
ally — sets the spread usage of the stereo image across the Y axis for all channels. Allowed range is from -1 to 15. By default this value is negative -1, and thus unused.
fcx, flx, frx, blx, brx, slx, srx, bcx — set the spread usage of the stereo image across the X axis for each channel. Allowed range is from 0.06 to 15. By default this value is 0.5.
fcy, fly, fry, bly, bry, sly, sry, bcy — set the spread usage of the stereo image across the Y axis for each channel. Allowed range is from 0.06 to 15. By default this value is 0.5.
win_size — sets the window size. Allowed range is from 1024 to 65536. Default size is 4096.
win_func — sets the window function. It accepts the following values:

rect
bartlett
hann, hanning
hamming
blackman
welch
flattop
bharris
bnuttall
bhann
sine
nuttall
lanczos
gauss
tukey
dolph
cauchy
parzen
poisson
bohman
kaiser

Default is "hann".

overlap — sets the window overlap. If set to 1, the recommended overlap for the selected window function will be picked. Default is 0.5.

tiltshelf 🌸

Boosts or cuts the lower frequencies and cuts or boosts 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). Very much the "vintage warmth" vibe.

The filter accepts the following options:

gain, g — gives 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 — sets 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 — 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 is used in the output. Default is 1. Range is between 0 and 1.
channels, c — specifies which channels to filter; by default all available 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 filtering precision ✨

Pick your sample format energy:

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 💅

So here's the tea: if you set this 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 just produces nasty artifacts. No cap, it's giving cursed.

Heads up bestie: the filter delay will be exactly this many samples when set to a non-zero value. Every. Single. Sample. 😭

Commands

This filter supports some options as commands. 🫶

treble, highshelf 👑

Boosts or cuts the treble (upper) frequencies of the audio, using a two-pole shelving filter with a response similar to a standard hi-fi's tone-controls. This is also known as shelving equalisation (EQ). 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 (for a large cut) to +20 (for a large boost). ⚠️ Beware of clipping when using a positive gain — she will clip, and she will not apologize.

`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 for specifying 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. Slay the transition, basically. 🔥

`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 are filtered. 💖

`normalize`, `n`

Normalizes the biquad coefficients — by default this is 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. Set it high enough (higher than the impulse response length truncated when it reaches near-zero values) and the filtering will become linear phase; 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 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. It's giving underwater vibes.

The filter accepts the following options:

f — modulation frequency in Hertz. Modulation frequencies in the subharmonic range (20 Hz or lower) will result in a tremolo effect. This filter may also be used 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. Tremolo's more mysterious cousin fr fr.

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 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 as available in the stereo input. Main character energy: keeps the originals, adds a whole new icon. 👑

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 literal volume knob of the filter world.

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 input sample format to U8, S16, and S32.
float — 32-bit floating-point; this limits input sample format to FLT. (default)
double — 64-bit floating-point; this limits input sample format to DBL.

`replaygain`

Chooses the behaviour on encountering 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

Note that when eval is set to once, only the sample_rate and tb variables are available — all other variables will evaluate to NAN. 🙈

Commands

This filter supports the following commands:

volume — modify the volume 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. (She said "no thanks" to your nonsense expression. 💅)

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, for example like in:

volume=0.5

Increase input audio power by 6 decibels using fixed-point precision:

volume=volume=6dB:precision=fixed

Fade volume after time 10 with an annihilation period of 5 seconds:

volume='if(lt(t,10),1,max(1-(t-10)/5,0))':eval=frame

volumedetect 🔍

Detects the volume of the input video. Yes, video — that's what the page says, we're keeping it. 🎀

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 will be printed in the log when the input stream end is reached.

In particular it will show the mean volume (root mean square), 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 is 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 does not cause any clipping, raising it by +5 dB causes clipping for 6 samples, etc. ⚡

whisper 🧚

It runs automatic speech recognition using OpenAI's Whisper model. Literally the coolest thing on this page, ngl.

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). You HAVE to set this one bestie, no skipping. 💜
language — the language to use for transcription ('auto' for auto-detect). Default value: "auto"
queue — how big the queue into the filter can get before the audio actually gets processed by whisper. Small value = your audio stream gets processed more often, but ngl the transcription quality drops and it eats way more processing power. Large value (like 10-20s) = more accurate results using less CPU (basically like using the whisper-cli tool), but the transcription latency goes up, so it's not useful for real-time streams. Lowkey consider using the vad_model option together with a large queue value. Default value: "3" ✨
use_gpu — whether GPU support should be on. Default value: "true"
gpu_device — which GPU device index to use. Default value: "0"
destination — if you set this, the transcription output goes to the file or URL you name (use one of the FFmpeg AVIO protocols); otherwise it just gets logged as info messages. The output is also dropped into the "lavfi.whisper.text" frame metadata. Heads up: if the destination is a file and it already exists, it gets overwritten, no cap. 💅
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 you set it to something greater than 0, transcription segments get split so they don't go over this length. Super handy for making subtitles with shorter lines. Default value: "0"
vad_model — path to the VAD model file. If you set it, the filter loads an extra voice activity detection module (https://github.com/snakers4/silero-vad) that gets used to fragment the audio queue; use this option by setting a valid path grabbed from the whisper.cpp repository (e.g. "../whisper.cpp/models/ggml-silero-v5.1.2.bin") and bump the queue parameter up higher (e.g. 20).
vad_threshold — the VAD threshold to use. Default value: "0.5"
vad_min_speech_duration — the minimum VAD speaking duration. Default value: "0.1"
vad_min_silence_duration — the minimum VAD silence duration. Default value: "0.5"

Examples 🎀

Run a transcription with srt file generation:

ffmpeg -i input.mp4 -vn -af "whisper=model=../whisper.cpp/models/ggml-base.en.bin\
:language=en\
:queue=3\
:destination=output.srt\
:format=srt" -f null -

Run a transcription and send the output in JSON format to an HTTP service:

ffmpeg -i input.mp4 -vn -af "whisper=model=../whisper.cpp/models/ggml-base.en.bin\
:language=en\
:queue=3\
:destination=http\\://localhost\\:3000\
:format=json' -f null -

Transcribe the microphone input using the VAD option:

ffmpeg -loglevel warning -f pulse -i default \
-af 'highpass=f=200,lowpass=f=3000,whisper=model=../whisper.cpp/models/ggml-medium.bin\
:language=en\
:queue=10\
:destination=-\
:format=json\
:vad_model=../whisper.cpp/models/ggml-silero-v5.1.2.bin' -f null -

🌈 AUDIO SOURCES 🌈

Okay so below is the whole rundown of the audio sources that are available rn. Buckle up, bestie. 🦋

abuffer

Buffers audio frames and makes them available to the filter chain. Iconic little middleman behavior.

This source is mainly meant for programmatic use, in particular through the interface defined in libavfilter/buffersrc.h.

It accepts these parameters:

time_base — the timebase used for timestamps of submitted frames. It's gotta 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 matching 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 matching 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 have to be consistent with each other. No contradicting yourself, periodt.

Examples

abuffer=sample_rate=44100:sample_fmt=s16p:channel_layout=stereo

will tell the source to accept planar 16bit signed stereo at 44100Hz. Since the sample format named "s16p" corresponds to the number 6 and the "stereo" channel layout corresponds to the value 0x3, this is equivalent to:

abuffer=sample_rate=44100:sample_fmt=6:channel_layout=0x3

aevalsrc

Generates an audio signal specified by an expression. Math but make it sound. 🧮✨

This source takes one or more expressions as input (one for each channel), which get evaluated and used to generate the matching audio signal.

This source accepts the following options:

exprs — sets the '|'-separated expressions list for each separate channel. If the channel_layout option isn't specified, the selected channel layout depends on how many expressions you provided. Otherwise the last specified expression gets applied to the remaining output channels.
channel_layout, c — sets the channel layout. The number of channels in the specified layout must equal the number of specified expressions.
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 it's not specified, or the expressed duration is negative, the audio is supposed to be generated forever.
nb_samples, n — sets the number of samples per channel per each output frame, defaults to 1024.
sample_rate, s — specifies the sample rate, defaults to 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, specify the channel layout (Front Center + Back Center) explicitly:

aevalsrc="sin(420*2*PI*t)|cos(430*2*PI*t):c=FC|BC"

Generate white noise:

aevalsrc="-2+random(0)"

Generate an amplitude modulated signal:

aevalsrc="sin(10*2*PI*t)*sin(880*2*PI*t)"

Generate 2.5 Hz binaural beats on a 360 Hz carrier:

aevalsrc="0.1*sin(2*PI*(360-2.5/2)*t) | 0.1*sin(2*PI*(360+2.5/2)*t)"

afdelaysrc

Generates fractional delay FIR coefficients. So specific, so niche, we love her. 💜

The resulting stream can be used with the afir filter for filtering the audio signal.

The filter accepts these options:

delay, d — sets the fractional delay. Default is 0.
sample_rate, r — sets the sample rate, default is 44100.
nb_samples, n — sets the number of samples per each frame. Default is 1024.
taps, t — sets 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. Yes she's the equalizer baddie. 🎚️

The resulting stream can be used with the afir filter for filtering the audio signal.

The filter accepts the following options:

preset, p — sets the equalizer preset. Default preset is "flat". The 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 — sets 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 — sets the custom bands from where the custom equalizer gains are set. This must be in strictly increasing order. Only used if the preset option is set to "custom". Bands are separated by white spaces and each band represents frequency in Hz. Default is "25 40 63 100 160 250 400 630 1000 1600 2500 4000 6300 10000 16000 24000".
taps, t — sets the number of filter coefficients in the output audio stream. Default value is 4096.
sample_rate, r — sets the sample rate of the output audio stream, default is 44100.
nb_samples, n — sets the number of samples per each frame in the output audio stream. Default is 1024.
interp, i — sets the interpolation method for the FIR equalizer coefficients. Can be "linear" or "cubic".
phase, h — sets the phase type of the FIR filter. Can be "linear" or "min": minimum-phase. Default is minimum-phase filter.

afirsrc

Generates FIR coefficients using the frequency sampling method. Frequency sampling is lowkey a flex, ngl. 🔥

The resulting stream can be used with the afir filter for filtering the audio signal.

The filter accepts the following options:

taps, t — sets the number of filter coefficients in the output audio stream. Default value is 1025.
frequency, f — sets the frequency points from where magnitude and phase are 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 — sets 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 — sets 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 — sets the sample rate, default is 44100.
nb_samples, n — sets the number of samples per each frame. Default is 1024.
win_func, w — sets the window function. Default is blackman.

anullsrc

The null audio source — it just hands you back unprocessed audio frames, like, the most nothing-ever generator. It's mainly useful as a template, and for analysis / debugging tools, or as the source for filters that literally ignore their input data (the sox synth filter is the classic example, iykyk).

This source accepts the following options:

channel_layout, cl — sets the channel layout. Can be an integer or a string that represents a channel layout. Default is "stereo". If you wanna see the full mapping between strings and channel layout values, check the channel_layout_map definition in libavutil/channel_layout.c. 🎀
sample_rate, r — sets the sample rate, defaults to 44100.
nb_samples, n — set the number of samples per requested frames.
duration, d — set the duration of the sourced audio. See the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. If it's not specified, or the expressed duration is negative, the audio is supposed to be generated forever. ♾️

Examples

Set the sample rate to 48000 Hz and the channel layout to AV_CH_LAYOUT_MONO:

anullsrc=r=48000:cl=4

Do the same operation with a more obvious syntax:

anullsrc=r=48000:cl=mono

ok important bestie: all the parameters need to be explicitly defined. no vibes-based defaults here, define them or perish 💀

flite

Synthesize a voice utterance using the libflite library. it literally talks, iconic.

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. keep that in mind before you come for me.

The filter accepts the following options:

list_voices — if set to 1, list the names of the available voices and exit immediately. Default value is 0.
nb_samples, n — set the maximum number of samples per frame. Default value is 512.
textfile — set the filename containing the text to speak.
text — set the text to speak.
voice, v — set the voice to use for the speech synthesis. Default value is "kal". See also the list_voices option.

Examples

Read from file speech.txt, and synthesize the text using the standard flite voice:

flite=textfile=speech.txt

Read the specified text selecting the "slt" voice:

flite=text='So fare thee well, poor devil of a Sub-Sub, whose commentator I am':voice=slt

Input text to ffmpeg:

ffmpeg -f lavfi -i flite=text='So fare thee well, poor devil of a Sub-Sub, whose commentator I am':voice=slt

Make ffplay speak the specified text, using "flite" and the "lavfi" device:

ffplay -f lavfi flite=text='No more be grieved for which that thou hast done.'

For more information about libflite, check: <http://www.festvox.org/flite/> ✨

anoisesrc

Generate a noise audio signal. yes. just noise. the ambient sound of my brain at 3am 💀

The filter accepts the following options:

sample_rate, r — specify the sample rate. Default value is 48000 Hz.
amplitude, a — specify the amplitude (0.0 - 1.0) of the generated audio stream. Default value is 1.0.
duration, d — specify the duration of the generated audio stream. Not specifying this option results in noise with an infinite length. 🌈
color, colour, c — specify the color of noise. Available noise colors are white, pink, brown, blue, violet and velvet. Default color is white.
seed, s — specify a value used to seed the PRNG.
nb_samples, n — set the number of samples per each output frame, default is 1024.
density — set the density (0.0 - 1.0) for the velvet noise generator, default is 0.05.

Examples

Generate 60 seconds of pink noise, with a 44.1 kHz sampling rate and an amplitude of 0.5:

anoisesrc=d=60:c=pink:r=44100:a=0.5

hilbert

Generate odd-tap Hilbert transform FIR coefficients. fancy math, no cap.

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. 🧚

The filter accepts the following options:

sample_rate, s — set sample rate, default is 44100.
taps, t — set length of FIR filter, default is 22051.
nb_samples, n — set number of samples per each frame.
win_func, w — set window function to be used when generating FIR coefficients.

sinc

Generate a sinc kaiser-windowed low-pass, high-pass, band-pass, or band-reject FIR coefficients. that's a mouthful and honestly? she ate. 💅

The resulting stream can be used with the afir filter for filtering the audio signal.

The filter accepts the following options:

sample_rate, r — set sample rate, default is 44100.
nb_samples, n — set number of samples per each frame. Default is 1024.
hp — set high-pass frequency. Default is 0.
lp — set low-pass frequency. Default is 0. If high-pass frequency is lower than low-pass frequency and low-pass frequency is higher than 0, then the filter will create band-pass filter coefficients, otherwise band-reject filter coefficients. keep that logic in your notes bestie 🫶
phase — set filter phase response. Default is 50. Allowed range is from 0 to 100.
beta — set Kaiser window beta.
att — set stop-band attenuation. Default is 120dB, allowed range is from 40 to 180 dB.
round — enable rounding, by default is disabled.
hptaps — set number of taps for high-pass filter.
lptaps — set number of taps for low-pass filter.

sine

Generate an audio signal made of a sine wave with amplitude 1/8.

The audio signal is bit-exact. ⭐

The filter accepts the following options:

frequency, f — set the carrier frequency. Default is 440 Hz.
beep_factor, b — enable a periodic beep every second with frequency beep_factor times the carrier frequency. Default is 0, meaning the beep is disabled.
sample_rate, r — specify the sample rate, default is 44100.
duration, d — specify the duration of the generated audio stream.
samples_per_frame — set the number of samples per output frame.

The expression can contain the following constants:

n — the (sequential) number of the output audio frame, starting from 0.
pts — the PTS (Presentation TimeStamp) of the output audio frame, expressed in TB units.
t — the PTS of the output audio frame, expressed in seconds.
TB — the timebase of the output audio frames.

Default is 1024.

Examples

Generate a simple 440 Hz sine wave:

sine

Generate a 220 Hz sine wave with a 880 Hz beep each second, for 5 seconds:

sine=220:4:d=5
sine=f=220:b=4:d=5
sine=frequency=220:beep_factor=4:duration=5

Generate a 1 kHz sine wave following "1602,1601,1602,1601,1602" NTSC pattern:

sine=1000:samples_per_frame='st(0,mod(n,5)); 1602-not(not(eq(ld(0),1)+eq(ld(0),3)))'

🌸 AUDIO SINKS 🌸

Below is a description of the currently available audio sinks. (the place where your audio goes to rest 😌)

abuffersink

Buffer audio frames, and make them available to the end of 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; do absolutely nothing with the input audio. it's giving "i'm just here for the vibes" — 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. let's get into it 🎬

addroi

Mark a region of interest in a video frame.

The frame data is passed through unchanged, but metadata is attached to the frame indicating regions of interest which can affect the behaviour of later encoding. Multiple regions can be marked by applying the filter multiple times.

x — region distance in pixels from the left edge of the frame.
y — region distance in pixels from the top edge of the frame.
w — region width in pixels.
h — region height in pixels.

The parameters x, y, w and h are expressions, and may contain the following variables:

iw — width of the input frame.
ih — height of the input frame.
qoffset — quantisation offset to apply within the region.

This must be a real value in the range -1 to +1. A value of zero indicates no quality change. A negative value asks for better quality (less quantisation), while a positive value asks for worse quality (greater quantisation).

The range is calibrated so that the extreme values indicate the largest possible offset — if the rest of the frame is encoded with the worst possible quality, an offset of -1 indicates that this region should be encoded with the best possible quality anyway. Intermediate values are then interpolated in some codec-dependent way.

For example, in 10-bit H.264 the quantisation parameter varies between -12 and 51. A typical qoffset value of -1/10 therefore indicates that this region should be encoded with a QP around one-tenth of the full range better than the rest of the frame. So, if most of the frame were to be encoded with a QP of around 30, this region would get a QP of around 24 (an offset of approximately -1/10 * (51 - -12) = -6.3). An extreme value of -1 would indicate that this region should be encoded with the best possible quality regardless of the treatment of the rest of the frame — that is, should be encoded at a QP of -12. 🤓 math girlies rise

clear — if set to true, remove any existing regions of interest marked on the frame before adding the new one.

Examples

Mark the centre quarter of the frame as interesting.

addroi=iw/4:ih/4:iw/2:ih/2:-1/10

Mark the 100-pixel-wide region on the left edge of the frame as very uninteresting (to be encoded at much lower quality than the rest of the frame).

addroi=0:0:100:ih:+1/5

alphaextract

Extract the alpha component from the input as a grayscale video. This is especially useful with the alphamerge filter. ✨

alphamerge

Add or replace the alpha component of the primary input with the grayscale value of a second input. This is intended for use with alphaextract to allow the transmission or storage of frame sequences that have alpha in a format that doesn't support an alpha channel. 🦋

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

Okay so this one literally turns up the volume on the differences between the current pixel and the pixels sitting in the same spot on neighboring frames. Little differences get louder ✨

It takes these options:

radius — sets the frame radius. Default is 2. Allowed range is 1 to 63. Bestie, for example a radius of 3 tells the filter to average 7 frames together 🫶
factor — sets the factor used to amplify the difference. Default is 2. Allowed range is 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 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 0 to 65535.
low — sets the lower limit for changing the source pixel. Default is 65535. Allowed range is 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 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 0 to 15.

Commands

This filter supports the following commands, and each one corresponds to the option of the same name:

factor
threshold
tolerance
low
high
planes

ass

Same thing as the subtitles filter, except it does not need libavcodec and libavformat to work. The tradeoff? It's limited to ASS (Advanced Substation Alpha) subtitle files. iykyk 💅

On top of the common options it inherits from the subtitles filter, it takes this one:

shaping — sets the shaping engine.

Available values are:

auto — the default libass shaping engine, which is the best available. 👑
simple — a fast, font-agnostic shaper that can do only substitutions.
complex — a 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, bestie 🦋

The filter accepts these options:

0a — sets threshold A for the 1st plane. Default is 0.02. Valid range is 0 to 0.3.
0b — sets threshold B for the 1st plane. Default is 0.04. Valid range is 0 to 5.
1a — sets threshold A for the 2nd plane. Default is 0.02. Valid range is 0 to 0.3.
1b — sets threshold B for the 2nd plane. Default is 0.04. Valid range is 0 to 5.
2a — sets threshold A for the 3rd plane. Default is 0.02. Valid range is 0 to 0.3.
2b — sets threshold B for the 3rd plane. Default is 0.04. Valid range is 0 to 5.

Threshold A is designed to react on abrupt changes in the input signal, and threshold B is designed to react on continuous changes in the input signal. Different jobs, both slaying 💖

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, but the other way around is never true. Parallel will abort early on the first change being greater than the thresholds, while serial will keep processing the other side of frames if they are equal to or below the thresholds. Ngl that's kind of iconic 🔥

0s, 1s, 2s — sets sigma for the 1st plane, 2nd plane or 3rd plane. Default is 32767. Valid range is 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 as the corresponding option.

avgblur

Applies an average blur filter. Soft focus moment ✨

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, and 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 chaos, no drama 💅

backgroundkey</backgroundkey>

Turns a static background into transparency. Poof, gone, bye 👋

Takes this 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. Detective mode 🔍

It 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.

Takes this 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. Smooth and sharp, she contains multitudes 🎀

Options:

sigmaS — sets sigma of the gaussian function used to calculate spatial weight. Allowed range is 0 to 512. Default is 0.1.
sigmaR — sets sigma of the gaussian function used to calculate range weight. Allowed range is 0 to 1. Default is 0.1.
planes — sets planes to filter. Default is first only.

Commands

This filter supports all of the above options as commands.

bitplanenoise</bitplanenoise>

Shows and measures bit plane noise. Numbers time, bestie 🧚

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. Super useful for spotting chapter transitions, commercials, or invalid recordings. 🙈

The filter outputs its detection analysis to both the log and the 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". On top of that, a log line with level "debug" is printed per frame showing the black amount detected for that frame.

It also attaches metadata to the first frame of a black segment with key "lavfi.black_start", and to the first frame after the black segment ends with key "lavfi.black_end". The value is the frame's timestamp. This metadata is added regardless of the minimum duration specified — no cap, it always tags them 💅

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. This detects frames that are (almost) transparent, instead of frames that are almost black. Spiritually the same vibe check, different channel 👻

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 for detecting chapter transitions or commercials. The 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 forget or you'll see nothing and be confused, and I refuse to let that be your story 😭

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 a 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:

addition
and
average
bleach
burn
darken
difference
divide
dodge
exclusion
extremity
freeze
geometric
glow
grainextract
grainmerge
hardlight
hardmix
hardoverlay
harmonic
heat
interpolate
lighten
linearlight
multiply
multiply128
negation
normal
or
overlay
phoenix
pinlight
reflect
screen
softdifference
softlight
stain
subtract
vividlight
xor

okay that's the whole mode menu, bestie, we're done with that list 💀 moving on!!

c0_opacity, c1_opacity, c2_opacity, c3_opacity, all_opacity — set the blend opacity for one specific pixel component, or like, for all of them if you use all_opacity 🫶 these only do anything when you're using the pixel-component blend modes, fyi.
c0_expr, c1_expr, c2_expr, c3_expr, all_expr — set the blend expression for a specific pixel component, or all of them with all_expr ✨ tiny heads up tho: if you set these, the related mode options just get ignored, lowkey ghosted.

The expressions get to use these variables:

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 — width and height scale for the plane being filtered. It's the ratio between the dimensions of the current plane and the luma plane, so like, for a yuv420p frame the values are 1,1 for the luma plane and 0.5,0.5 for the chroma planes.
T — time of the current frame, expressed in seconds.
TOP, A — value of the pixel component at the current location for the first video frame (top layer).
BOTTOM, B — value of the pixel component at the current location for the second video frame (bottom layer).

The blend filter also supports the framesync options 💅

Examples

Apply transition from bottom layer to top layer in first 10 seconds:

blend=all_expr='A*(if(gte(T,10),1,T/10))+B*(1-(if(gte(T,10),1,T/10)))'

Apply linear horizontal transition from top layer to bottom layer:

blend=all_expr='A*(X/W)+B*(1-X/W)'

Apply 1x1 checkerboard effect:

blend=all_expr='if(eq(mod(X,2),mod(Y,2)),A,B)'

Apply uncover left effect:

blend=all_expr='if(gte(N*SW+X,W),A,B)'

Apply uncover down effect:

blend=all_expr='if(gte(Y-N*SH,0),A,B)'

Apply 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 diagonally video and shows top and bottom layer on each side:

blend=all_expr='if(gt(X,Y*(W/H)),A,B)'

Display differences between the current and the previous frame:

tblend=all_mode=grainextract

Commands

This filter supports same commands as options 😌 yes bestie, you can fiddle with it live.

🧱 blockdetect

Determines the blockiness of frames without altering the input frames — it's purely a measuring girlie, no touchy 🙅‍♀️

Based on Remco Muijs and Ihor Kirenko: "A no-reference blocking artifact measure for adaptive video processing." 2005 13th European signal processing conference.

The filter accepts the following options:

period_min, period_max — set the minimum and maximum values for determining pixel grids (periods). Default values are [3,24].
planes — set 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

Determines the blurriness of frames without altering the input frames. Based on Marziliano, Pina, et al. "A no-reference perceptual blur metric." It's got a block-based abbreviation too, fancy.

The filter accepts the following options:

low, high — set the low and high threshold values used by the Canny thresholding algorithm.

So the high threshold picks out the "strong" edge pixels, and then those get connected through 8-connectivity with the "weak" edge pixels that the low threshold selected 🔗

low and high have to be chosen in the range [0,1], and low should be lesser or equal to high (don't be chaotic about it).

Default value for low is 20/255, and default value for high is 50/255.

radius — define the radius to search around an edge pixel for local maxima.
block_pct — determine blurriness only for the most significant blocks, given in percentage.
block_width — determine blurriness for blocks of width block_width. If set to any value smaller 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 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 most significant 32x32 blocks:

blurdetect=block_width=32:block_height=32:block_pct=80

🧊 bm3d

Denoise frames using the Block-Matching 3D algorithm — it's giving fancy-science-girlie energy fr fr 🧚

The filter accepts the following options.

sigma — set denoising strength. Default value is 1. Allowed range is from 0 to 999.9. Heads up tho: the denoising algorithm is very sensitive to sigma, so adjust it according to the source, okay? 😭
block — set local patch size. This sets dimensions in 2D.
bstep — set 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 (iconic tradeoff 💅).
group — set maximal number of similar blocks for 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 radius for search block matching. Default is 9. Allowed range is from 1 to INT32_MAX.
mstep — set step between two search locations for block matching. Default is 1. Allowed range is from 1 to 64. Smaller is slower.
thmse — set threshold of mean square error for block matching. Valid range is 0 to INT32_MAX.
hdthr — set thresholding parameter for hard thresholding in 3D transformed domain. Larger values result in stronger hard-thresholding filtering in the frequency domain.
estim — set 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 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 nlmeans filter instead:

split[a][b],[a]nlmeans=s=3:r=7:p=3[a],[b][a]bm3d=sigma=3:block=4:bstep=2:group=16:estim=final:ref=1

🌸 boxblur

Apply a boxblur algorithm to the input video. Soft-focus main character vibes ✨

It accepts the following parameters:

luma_radius, lr
luma_power, lp
chroma_radius, cr
chroma_power, cp
alpha_radius, ar
alpha_power, ap

A description of the accepted options follows.

luma_radius, lr / chroma_radius, cr / alpha_radius, ar — set an expression for the box radius in pixels used for blurring the corresponding input plane.

The radius value must be a non-negative number, and must not be greater than the value of the expression min(w,h)/2 for the luma and alpha planes, and of min(cw,ch)/2 for the chroma planes. Respect the limits bestie 🙏

Default value for luma_radius is 2. If not specified, chroma_radius and alpha_radius default to the corresponding value set for luma_radius.

The expressions can contain the following constants:

w, h — the input width and height in pixels.
cw, ch — the input chroma image width and height in pixels.
hsub, vsub — the horizontal and vertical chroma subsample values. For example, for the pixel format yuv422p, hsub is 2 and vsub is 1.

luma_power, lp / chroma_power, cp / alpha_power, ap — specify how many times the boxblur filter is applied to the corresponding plane.

Default value for luma_power is 2. If not specified, chroma_power and alpha_power default to the corresponding value set for luma_power.

A value of 0 will disable the effect (so like, zero blur, ghost mode 👻).

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 alpha and chroma radius to 0:

boxblur=2:1:cr=0:ar=0

Set the luma and chroma radii to a fraction of the video dimension:

boxblur=luma_radius=min(h\,w)/10:luma_power=1:chroma_radius=min(cw\,ch)/10:chroma_power=1

🦋 bwdif

Deinterlace the input video (bwdif stands for "Bob Weaver Deinterlacing Filter" — yes, that's a real name and yes it's iconic).

Motion adaptive deinterlacing based on yadif with the use of w3fdif and cubic interpolation algorithms. It accepts the following parameters:

mode — the interlacing mode to adopt. It accepts one of the following values:
- 0, send_frame — output one frame for each frame.
- 1, send_field — output one frame for each field.

The default value is send_field.

parity — the picture field parity assumed for the input interlaced video. It accepts one of the following values:
- 0, tff — assume the top field is first.
- 1, bff — assume the bottom field is first.
- -1, auto — enable automatic detection of field parity.

The default value is auto. If the interlacing is unknown or the decoder 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`

Repack CEA-708 closed captioning side data.

Okay so this filter fixes various issues seen with commercial encoders related to upstream malformed CEA-708 payloads — specifically incorrect number of tuples (wrong cc_count for the target FPS), and incorrect ordering of tuples (i.e. the CEA-608 tuples are not at the first entries in the payload). Basically it's the cleanup crew for messy caption data fr fr 🧹✨

🦋 `cas`

Apply Contrast Adaptive Sharpen filter to video stream.

The filter accepts the following options:

strength — set the sharpening strength. Default value is 0.
planes — set planes to filter. Default value is to filter all planes except alpha plane.

💬 Commands

This filter supports same commands as options.

🌸 `chromahold`

Remove all color information for all colors except for certain one. The way this one just keeps one color and greys out everything else — it's giving main character energy ngl 👑

The filter accepts the following options:

color — the color which will not be replaced with neutral chroma.
similarity — similarity percentage with the above color. 0.01 matches only the exact key color, while 1.0 matches everything.
blend — blend percentage. 0.0 makes pixels either fully gray, or not gray at all. Higher values result in more preserved color.
yuv — signals that the color passed is already in YUV instead of RGB.

Literal colors like "green" or "red" don't make sense with this enabled anymore. This can be used to pass exact YUV values as hexadecimal numbers. 🙈

💬 Commands

This filter supports same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value.

💜 `chromakey`

YUV colorspace color/chroma keying. Greenscreen energy, we love to see it 💚

The filter accepts the following options:

color — the color which will be replaced with transparency.
similarity — similarity percentage with the key color. 0.01 matches only the exact key color, while 1.0 matches everything.
blend — blend percentage. 0.0 makes pixels either fully transparent, or not transparent at all. Higher values result in semi-transparent pixels, with a higher transparency the more similar the pixels color is to the key color.
yuv — signals that the color passed is already in YUV instead of RGB.

Literal colors like "green" or "red" don't make sense with this enabled anymore. This can be used to pass exact YUV values as hexadecimal numbers. 🙈

💬 Commands

This filter supports same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value.

✨ Examples

Make every green pixel in the input image transparent:

ffmpeg -i input.png -vf chromakey=green out.png

Overlay a greenscreen-video on top of a static black background.

ffmpeg -f lavfi -i color=c=black:s=1280x720 -i video.mp4 -shortest -filter_complex "[1:v]chromakey=0x70de77:0.1:0.2[ckout];[0:v][ckout]overlay[out]" -map "[out]" output.mkv

🍓 `chromanr`

Reduce chrominance noise. Lowkey the glow-up filter for grainy footage ✨

The filter accepts the following options:

thres — set threshold for averaging chrominance values. Sum of absolute difference of Y, U and V pixel components of current pixel and neighbour pixels lower than this threshold will be used in averaging. Luma component is left unchanged and is copied to output. Default value is 30. Allowed range is from 1 to 200.
sizew — set horizontal radius of rectangle used for averaging. Allowed range is from 1 to 100. Default value is 5.
sizeh — set vertical radius of rectangle used for averaging. Allowed range is from 1 to 100. Default value is 5.
stepw — set horizontal step when averaging. Default value is 1. Allowed range is from 1 to 50. Mostly useful to speed-up filtering.
steph — set vertical step when averaging. Default value is 1. Allowed range is from 1 to 50. Mostly useful to speed-up filtering.
threy — set Y threshold for averaging chrominance values. Set finer control for max allowed difference between Y components of current pixel and neighbour pixels. Default value is 200. Allowed range is from 1 to 200.
threu — set U threshold for averaging chrominance values. Set finer control for max allowed difference between U components of current pixel and neighbour pixels. Default value is 200. Allowed range is from 1 to 200.
threv — set V threshold for averaging chrominance values. Set finer control for max allowed difference between V components of current pixel and neighbour pixels. Default value is 200. Allowed range is from 1 to 200.
distance — set distance type used in calculations.
- manhattan — absolute difference.
- euclidean — difference squared. Default distance type is manhattan.

💬 Commands

This filter supports same commands as options. The command accepts the same syntax of the corresponding option.

🌈 `chromashift`

Shift chroma pixels horizontally and/or vertically. Just… scootching the chroma around, no biggie 💫

The filter accepts the following options:

cbh — set amount to shift chroma-blue horizontally.
cbv — set amount to shift chroma-blue vertically.
crh — set amount to shift chroma-red horizontally.
crv — set amount to shift chroma-red vertically.
edge — set edge mode, can be smear, default, or warp.

💬 Commands

This filter supports the all above options as commands.

👑 `ciescope`

Display CIE color diagram with pixels overlaid onto it. Science-y and aesthetic, iconic combo 🧪🎀

The filter accepts the following options:

system — set color system.
- ntsc, 470m
- ebu, 470bg
- smpte
- 240m
- apple
- widergb
- cie1931
- rec709, hdtv
- uhdtv, rec2020
- dcip3
cie — set CIE system.
- xyy
- ucs
- luv
gamuts — set what gamuts to draw. See system option for available values.
size, s — set ciescope size, by default set to 512.
intensity, i — set intensity used to map input pixel values to CIE diagram.
contrast — set contrast used to draw tongue colors that are out of active color system gamut.
corrgamma — correct gamma displayed on scope, by default enabled.
showwhite — show white point on CIE diagram, by default disabled.
gamma — set input gamma. Used only with XYZ input color space.
fill — fill with CIE colors. By default is enabled.

🦋 `codecview`

Visualize information exported by some codecs. Ngl this one is secretly so cool 😭🔥

Some codecs can export information through frames using side-data or other means. For example, some MPEG based codecs export motion vectors through the export_mvs flag in the codec flags2 option.

The filter accepts the following option:

block — display block partition structure using the luma plane.
mv — set motion vectors to visualize. Available flags for mv are:
- pf — forward predicted MVs of P-frames
- bf — forward predicted MVs of B-frames
- bb — backward predicted MVs of B-frames
qp — display quantization parameters using the chroma planes.
mv_type, mvt — set motion vectors type to visualize. Includes MVs from all frames unless specified by frame_type option. Available flags for mv_type are:
- fp — forward predicted MVs
- bp — backward predicted MVs
frame_type, ft — set frame type to visualize motion vectors of. Available flags for frame_type are:
- if — intra-coded frames (I-frames)
- pf — predicted frames (P-frames)
- bf — bi-directionally predicted frames (B-frames)

✨ Examples

Visualize forward predicted MVs of all frames using ffplay:

ffplay -flags2 +export_mvs input.mp4 -vf codecview=mv_type=fp

Visualize multi-directionals MVs of P and B-Frames using ffplay:

ffplay -flags2 +export_mvs input.mp4 -vf codecview=mv=pf+bf+bb

🌸 `colorbalance`

Modify intensity of primary colors (red, green and blue) of input frames. Color grading era 💅🎨

The filter allows an input frame to be adjusted in the shadows, midtones or highlights regions for the red-cyan, green-magenta or blue-yellow balance.

A positive adjustment value shifts the balance towards the primary color, a negative value towards the complementary color.

The filter accepts the following options:

rs, gs, bs — adjust red, green and blue shadows (darkest pixels).
rm, gm, bm — adjust red, green and blue midtones (medium pixels).
rh, gh, bh — adjust red, green and blue highlights (brightest pixels). Allowed ranges for options are [-1.0, 1.0]. Defaults are 0.
pl — preserve lightness when changing color balance. Default is disabled.

✨ Examples

Add red color cast to shadows:

colorbalance=rs=.3

💬 Commands

This filter supports the all above options as commands.

🧚 `colorcontrast`

Adjust color contrast between RGB components.

The filter accepts the following options:

rc — set the red-cyan contrast. Defaults is 0.0. Allowed range is from -1.0 to 1.0.
gm — set the green-magenta contrast. Defaults is 0.0. Allowed range is from -1.0 to 1.0.
by — set the blue-yellow contrast. Defaults is 0.0. Allowed range is from -1.0 to 1.0.
rcw, gmw, byw — set the weight of each rc, gm, by option value. Default value is 0.0. Allowed range is from 0.0 to 1.0. If all weights are 0.0 filtering is disabled.
pl — set the amount of preserving lightness. Default value is 0.0. Allowed range is from 0.0 to 1.0.

💬 Commands

This filter supports the all above options as commands.

💜 `colorcorrect`

Adjust color white balance selectively for blacks and whites. This filter operates in YUV colorspace. Bestie it's literally a white-balance slider for your shadows and highlights, obsessed 🫶

The filter accepts the following options:

rl — set the red shadow spot. Allowed range is from -1.0 to 1.0. Default value is 0.
bl — set the blue shadow spot. Allowed range is from -1.0 to 1.0. Default value is 0.
rh — set the red highlight spot. Allowed range is from -1.0 to 1.0. Default value is 0.
bh — set the blue highlight spot. Allowed range is from -1.0 to 1.0. Default value is 0.
saturation — set the amount of saturation. Allowed range is from -3.0 to 3.0. Default value is 1.
analyze — if set to anything other than manual it will analyze every frame and use derived parameters for filtering output frame. Possible values are:
- manual
- average
- minmax
- median Default value is "manual".

💬 Commands

This filter supports the all above options as commands.

colorchannelmixer

Okay so this one is literally the color remix station 💅 It takes your video frames and re-mixes the color channels by hand — as in, you get to say "hey, how much of the blue channel should leak into my red output?" and it just… does it. Iconic.

Technically: this filter modifies a color channel by adding the values associated to the other channels of the same pixels. So if the value you're modifying is red, the output value is gonna be:

<red>=<red>*<rr> + <blue>*<rb> + <green>*<rg> + <alpha>*<ra>

The filter accepts the following options:

rr, rg, rb, ra — adjust contribution of input red, green, blue and alpha channels for output red channel. Default is 1 for rr, and 0 for rg, rb and ra.
gr, gg, gb, ga — adjust contribution of input red, green, blue and alpha channels for output green channel. Default is 1 for gg, and 0 for gr, gb and ga.
br, bg, bb, ba — adjust contribution of input red, green, blue and alpha channels for output blue channel. Default is 1 for bb, and 0 for br, bg and ba.
ar, ag, ab, aa — adjust contribution of input red, green, blue and alpha channels for output alpha channel. Default is 1 for aa, and 0 for ar, ag and ab.

Allowed ranges for options are [-2.0, 2.0]. 🫶

pc — set preserve color mode. The accepted values are:
- none — disable color preserving, this is 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 from [0.0, 1.0]. Default is 0.0, thus disabled.

Examples

Convert source to grayscale 🩶

colorchannelmixer=.3:.4:.3:0:.3:.4:.3:0:.3:.4:.3

Simulate sepia tones (vintage filter energy, no cap) 📷

colorchannelmixer=.393:.769:.189:0:.349:.686:.168:0:.272:.534:.131

Commands

This filter supports the all above options as commands. ✨

colordetect

This one's the detective of the filter family 🕵️‍♀️ It analyzes the video frames to figure out the effective value range and alpha mode. Literally doing vibe checks on your pixels.

The filter accepts the following options:

mode — set of properties to detect. Unavailable properties, such as alpha mode for an input image without an alpha channel, will be ignored automatically.

Accepts a combination of the following 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 giving "put a filter on it" but make it literal 🌈

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. Think green screen, but make it chroma-key chic 💚 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 "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. (the classic, the icon 👑)

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 certain one. So like… everything goes neutral gray except your main character color, which gets to keep its whole personality 🌸

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. Levels! Like the thing in Photoshop your artsy friend is obsessed with 🎨

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), 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 allows manual selection of a constrained output level range.

preserve — set preserve color mode. The accepted values are:
- none — disable color preserving, this is 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, bestie — it needs three input video streams. First stream is video stream that is going to be filtered out. Second and third video stream specify color patches for source color to target color mapping. It's giving "match this color to that color" and honestly? Slay.

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 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, clean, does exactly what it says on the tin ✨

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. Input video needs to have an even size — no cap, that's a requirement, don't fight it 🦋

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

iec61966-2-1 — literally iec61966-2-1, no notes, that's the whole vibe 🎀
iec61966-2-4 — iec61966-2-4, again, exactly what it says on the tin
xvycc — xvycc (iykyk)
bt2020-10 — BT.2020 but for 10-bits content ✨
bt2020-12 — BT.2020 for 12-bits content ✨

`primaries`

Set the output color primaries, bestie. The accepted values are giving:

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`

Set the output color range. Accepted values:

tv — TV (restricted) range
mpeg — MPEG (restricted) range
pc — PC (full) range
jpeg — JPEG (full) range

`format`

Set the output color format. Accepted values, and yes there are a lot, we love a thorough queen 💅:

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`

Do a fast conversion — it skips gamma/primary correction. Sounds cute, right? It uses significantly less CPU, but it will be mathematically incorrect 😭. If you want output that's compatible with whatever the colormatrix filter produces, use fast=1.

`dither`

Set the dithering mode. Accepted values:

none — no dithering
fsb — Floyd-Steinberg dithering

`wpadapt`

Whitepoint adaptation mode. Accepted values:

bradford — Bradford whitepoint adaptation
vonkries — von Kries whitepoint adaptation
identity — identity whitepoint adaptation (i.e. no whitepoint adaptation)

`clipgamut`

Controls how to clip out-of-gamut colors that 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`

Override all input properties at once. Same accepted values as all.

`ispace`

Override the input colorspace. Same accepted values as space.

`iprimaries`

Override the input color primaries. Same accepted values as primaries.

`itrc`

Override the input transfer characteristics. Same accepted values as trc.

`irange`

Override the input color range. Same accepted values as range.

Okay so here's the lore 🫶: the filter converts the transfer characteristics, color space and color primaries to whatever user values you specified. The output value, if you didn't specify it, gets set to a default value based on the all property. And if that property also isn't specified? The filter will log an error, periodt. The output color range and format default to the same value as the input color range and format. The input transfer characteristics, color space, color primaries and color range should be set on the input data — if any of these are missing, the filter logs an error and no conversion will take place 💀.

So like, for example, to convert the input to SMPTE-240M, use:

colorspace=smpte240m

`colortemperature` 🌡️

Adjust the color temperature in video to simulate variations in ambient color temperature. Cozy vibes only.

This filter accepts the following options:

temperature — sets the temperature in Kelvin. Allowed range is from 1000 to 40000. Default value is 6500 K.
mix — set mixing with filtered output. Allowed range is from 0 to 1. Default value is 1.
pl — set the amount of preserving lightness. Allowed range is from 0 to 1. Default value is 0.

Commands

This filter supports the same commands as options. Slay 💅.

`convolution` 🌀

Apply convolution of 3x3, 5x5, 7x7 or horizontal/vertical up to 49 elements. It's giving math class but make it iconic.

The filter accepts the following options:

0m
1m
2m
3m — set the matrix for each plane. The matrix is a sequence of 9, 25 or 49 signed integers in square mode, and from 1 to 49 odd number of signed integers in row mode.
0rdiv
1rdiv
2rdiv
3rdiv — set the multiplier for the calculated value for each plane. If unset or 0, it will be 1/sum of all matrix elements.
0bias
1bias
2bias
3bias — set the 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 the matrix mode for each plane. Can be square, row or column. Default is square.

Commands

This filter supports all the above options as commands. Obsessed.

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` ✨

Apply 2D convolution of the video stream in the frequency domain, using the second stream as the impulse. Fancy, I know.

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.

The convolve filter also supports the framesync options. Bestie behaviour fr fr 🫶.

`copy` 📋

Copy the input video source unchanged to the output. Mainly useful for testing purposes — it's the "no thoughts, head empty" filter and honestly? Respect.

`coreimage` 🍎

Video filtering on GPU using Apple's CoreImage API on OSX. A whole aesthetic.

Hardware acceleration is based on an OpenGL context — usually that means it's processed by video hardware. But software-based OpenGL implementations exist too, which means there's no guarantee of hardware processing 😭. It depends on the respective OSX.

There are many filters and image generators provided by Apple 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 along with the default values.

list_filters=true

filter — specify all filters by their respective name and options. Use list_filters to determine all valid filter names and options. Numerical options are specified by a float value and are automatically clamped to their respective value range. Vector and color options have to be specified by a list of space separated float values. Character escaping has to be done. A special option name default is available to use default options for a filter.

It is required to specify either default or at least one of the filter options. All omitted options are used with their default values. The syntax of the filter string is as follows:

filter=<NAME>@<OPTION>=<VALUE>[@<OPTION>=<VALUE>][@...][#<NAME>@<OPTION>=<VALUE>[@<OPTION>=<VALUE>][@...]][#...]

output_rect — specify a rectangle where the output of the filter chain is copied into the input image. It's given by a list of space separated float values:

output_rect=x\ y\ width\ height

If not given, the output rectangle equals the dimensions of the input image. The output rectangle is automatically cropped at the borders of the input image. Negative values are valid for each component. 👑

output_rect=25\ 25\ 100\ 100

Several filters can be chained for successive processing without GPU-HOST transfers, which 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.

Some filters generate output images with additional padding depending on the respective filter kernel. The padding is automatically removed to ensure the filter output has the same size as the input image. Clean.

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 do not use the pixel information of this image to generate their output. However, the generated output is blended onto this image, resulting in partial or complete coverage of the output image.

The coreimagesrc video source can be used for generating input images which are directly fed into the filter chain. By using it, providing input images by another video source or an input video is not required. Convenient much? 💅

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` 🔍

Obtain the correlation between two input videos. This filter takes two input videos, so bring a friend 🫶.

Both input videos have to match in resolution and pixel format for this filter to actually work, bestie — no exceptions. And it assumes both inputs have the same number of frames, which get compared one by one. 🫶

The per-component, average, min and max correlation it figures out gets printed through the logging system.

The filter also stashes the calculated correlation of each frame in the frame metadata.

It also supports the framesync options, iykyk.

In the example below, the input file main.mpg being processed gets compared with the reference file ref.mpg:

ffmpeg -i main.mpg -i ref.mpg -lavfi corr -f null -

cover_rect

Cover a rectangular object 💅

It accepts the following options:

cover — filepath of the optional cover image. Heads up: it needs to be in yuv420.
mode — set the 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. Classic, iconic, still serving. ✂️

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. Same deal: this expression is evaluated only once during 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 one's evaluated per-frame too.
keep_aspect — if set to 1, it forces the output display aspect ratio to be the same as the input, by changing the output sample aspect ratio. It defaults to 0.
exact — enable exact cropping. If enabled, subsampled videos get cropped at the exact width/height/x/y as specified and will not be rounded to the 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're evaluated for each new frame.
in_w, in_h — the input width and height.
iw, ih — these are the same as in_w and in_h.
out_w, out_h — the output (cropped) width and height.
ow, oh — these are the same as out_w and out_h.
a — same as iw / ih
sar — input sample aspect ratio
dar — input display aspect ratio, it's the same as (iw / ih) * sar
hsub, vsub — horizontal and vertical chroma subsample values. For example, for the pixel format yuv422p, hsub is 2 and vsub is 1.
n — the number of the input frame, starting from 0.
pos — the position in the file of the input frame, NAN if unknown; deprecated, do not use 💀
t — the timestamp expressed in seconds. It's NAN if the input timestamp is unknown.

Now the dependency tea ☕: the expression for out_w may depend on the value of out_h, and the expression for out_h may depend on out_w — but they cannot depend on x and y, because x and y are evaluated after out_w and out_h. Order matters, no cap.

The x and y parameters specify the expressions for the position of the top-left corner of the output (non-cropped) area, and they're evaluated for each frame. If the evaluated value isn't valid, it gets approximated to the nearest valid value. Close enough is genuinely fine here.

The expression for x may depend on y, and the expression for y may depend on x. Mutual, your honor. 🫶

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

These set the width/height of the output video and the horizontal/vertical position in the input video. The command accepts the same syntax as the corresponding option.

If the specified expression isn't valid, it just stays at its current value. No crash, no drama, we love a graceful non-update. ✨

cropdetect

Auto-detect the crop size. It does the math so you don't have to. 🧠✨

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, depending on mode.

It accepts the following parameters:

mode — depending on the mode, crop detection is based on either the mere black value of surrounding pixels, or a combination of motion vectors and edge pixels.
- black — detect black pixels surrounding the playing video. For fine control use 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 a higher black value threshold, which can be optionally specified from nothing (0) to everything (255 for 8-bit based formats). Any intensity value greater than 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 the low and high threshold values used by the Canny thresholding algorithm.

The high threshold selects the "strong" edge pixels, which are then connected through 8-connectivity with the "weak" edge pixels selected by the low threshold.

low and high threshold values must be chosen in the range [0,1], and low should be lesser or equal to high. Respect the ordering, bestie.

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 as the corresponding option. If the specified expression isn't valid, it's kept at its current value.

cue

Delay video filtering until a given wallclock timestamp. Main character energy: it waits. ⏰

The filter first passes on preroll amount of frames, then it buffers at most buffer amount of frames and waits for the cue. After reaching the cue it forwards the buffered frames and also any subsequent frames coming in on its input.

The filter can be used to synchronize the output of multiple ffmpeg processes for realtime output devices like decklink. By putting the delay in the filtering chain and pre-buffering frames, the process can pass 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. It's giving "close enough bestie". 😭

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. This one is so pretty, wait till you see it. 🌈

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. Identity curve, zero notes, slay. 💅

Okay so this filter lets you redefine those two points and add some more — yes bestie, you get to draw your own curve ✨ A brand new curve gets defined so it passes smoothly through all your new coordinates. Two rules though, no cap: the new defined points have to be strictly increasing over the x-axis, and their x and y values must be in the [0;1] interval. The curve itself is formed using a natural or monotonic cubic spline interpolation, depending on the interp option (default: "natural"). The "natural" spline produces a smoother curve in general, while the monotonic ("pchip") spline guarantees the transitions between the specified points are monotonic. And if the computed curves happen to go outside the vector spaces? The values get clipped accordingly. 🙈

The filter accepts the following options:

preset — pick one of the available color presets 💅 This option can be used in addition to the r, g, b parameters; in that case, the later options take priority over the preset values. Available presets are:
- none
- color_negative
- cross_process
- darker
- increase_contrast
- lighter
- linear_contrast
- medium_contrast
- negative
- strong_contrast
- vintage

Default is "none".

master, m — set the master key points. These points define a second pass mapping. It's sometimes called a "luminance" or "value" mapping. You can use it 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 in addition to the other key points component options. In that case, the unset component(s) will fallback on this all setting.
psfile — specify a Photoshop curves file (".acv") to import the settings from.
plot — save Gnuplot script of the curves in specified file.
interp — specify the kind of interpolation. Available algorithms are:
- natural — Natural cubic spline using a piece-wise cubic polynomial that is twice continuously differentiable.
- pchip — Monotonic cubic spline using a piecewise cubic Hermite interpolating polynomial (PCHIP).

To avoid some filtergraph syntax conflicts, each key points list needs to be defined using the following syntax: "x0/y0 x1/y1 x2/y2 ...". ikyk fr fr 🫶

Commands

This filter supports 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 component:

red "(0;0.11) (0.42;0.51) (1;0.95)"
green "(0;0) (0.50;0.48) (1;1)"
blue "(0;0.22) (0.49;0.44) (1;0.80)"

The previous example can also be achieved with the associated built-in preset:

curves=preset=vintage

Or simply:

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. It's giving microscope energy.

The filter accepts the following options:

size, s — set output video size.
x — set x offset from where to pick pixels.
y — set y offset from where to pick pixels.
mode — set scope mode, can be one of the following:
- mono — draw hexadecimal pixel values with white color on black background.
- color — draw hexadecimal pixel values with input video pixel color on black background.
- color2 — draw hexadecimal pixel values on color background picked from input video, the text color is picked in such a way so it's always visible. 💅
axis — draw rows and columns numbers on left and top of video.
opacity — set background opacity.
format — set display number format. Can be "hex", or "dec". Default is "hex".
components — set pixel components to display. By default all pixel components are displayed.

Commands

This filter supports same commands as options excluding the size option.

dblur

Apply Directional blur filter. 🌬️

The filter accepts the following options:

angle — set angle of directional blur. Default is 45.
radius — set radius of directional blur. Default is 5.
planes — set which planes to filter. By default all planes are filtered.

Commands

This filter supports same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value.

dctdnoiz

Denoise frames using 2D DCT (frequency domain filtering). 🧚

This filter is not designed for real time. Not me warning you twice 💀

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 according borders won't be denoised.

Default value is blocksize-1, which is the best possible setting.

expr, e — set the coefficient factor expression.

For each coefficient of a DCT block, this expression will be evaluated as a multiplier value for the coefficient.

If this option is set, the sigma option will be ignored.

The absolute value of the coefficient can be accessed through the c variable.

n — set the blocksize using the number of bits. "1<<n" defines the blocksize, which is the width and height of the processed blocks.

The default value is 3 (8x8) and can be raised to 4 for a blocksize of 16x16. Note that changing this setting has huge consequences on the speed processing. Also, a larger block size does not necessarily mean a better de-noising.

Examples

Apply a denoise with a sigma of 4.5:

dctdnoiz=4.5

The same operation can be achieved using the expression system:

dctdnoiz=e='gte(c, 4.5*3)'

Violent denoise using a block size of "16x16":

dctdnoiz=15:n=4

deband

Remove banding artifacts from input video. It works by replacing banded pixels with the average value of referenced pixels. slay 🌈

The filter accepts the following options:

1thr, 2thr, 3thr, 4thr — set banding detection threshold for each plane. Default is 0.02. Valid range is 0.00003 to 0.5. If the difference between current pixel and reference pixel is less than threshold, it will be considered as banded.
range, r — banding detection range in pixels. Default is 16. If positive, a random number in range 0 to set value will be used. If negative, exact absolute value will be used. The range defines square of four pixels around current pixel.
direction, d — set direction in radians from which four pixels will be compared. If positive, a random direction from 0 to set direction will be picked. If negative, exact absolute value will be picked. For example direction 0, -PI or -2*PI radians will pick only pixels on same row and -PI/2 will pick only pixels on same column.
blur, b — if enabled, current pixel is compared with average value of all four surrounding pixels. The default is enabled. If disabled, current pixel is compared with all four surrounding pixels — and the pixel is considered banded if only all four differences with surrounding pixels are less than threshold.
coupling, c — if enabled, current pixel is changed if and only if all pixel components are banded, e.g. banding detection threshold is triggered for all color components. The default is disabled.

Commands

This filter supports the all above options as commands.

deblock

Remove blocking artifacts from input video. 🧱✨

The filter accepts the following options:

filter — set filter type, can be weak or strong. Default is strong. This controls what kind of deblocking is applied.
block — set size of block, allowed range is from 4 to 512. Default is 8.
alpha, beta, gamma, delta — set blocking detection thresholds. Allowed range is 0 to 1. Defaults are: 0.098 for alpha and 0.05 for the rest. Using a higher threshold gives more deblocking strength. Setting alpha controls threshold detection at 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. 🗑️ bye bestie

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 — this is your duplicate-detection threshold, bestie. if the difference metric for a frame comes in at or below this number, the filter's like "yeah that's a dupe" and calls it. default is 1.1 ✨
scthresh — scene change threshold, iykyk. default is 15 🎀
blockx, blocky — the size of the x-axis and y-axis blocks used during metric calculations. bigger blocks = better noise suppression but worse detection of small movements — classic tradeoff, no cap. must be a power of two. default is 32 💅
ppsrc — marks the main input as a pre-processed input and activates the clean source input stream. the whole point: you can pre-process the input 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 🦋
chroma — sets whether or not chroma is considered in the metric calculations. default is 1 🌸
mixed — sets whether or not the input only partially contains content to be decimated. default is "false". if you enable it, the video output stream will be in variable frame rate 👑

🧚 deconvolve

applies a 2D deconvolution of the video stream in the frequency domain, using the second stream as the impulse. fancy math hours 💜

the filter accepts the following options:

planes — set which planes to process.
impulse — set which impulse video frames will be processed. can be first or all. default is all.
noise — set the noise used when doing divisions. default is 0.0000001. genuinely useful when width and height are not the same and not a power of 2, or if the stream prior to convolving had noise 🫶

the deconvolve filter also supports the framesync options.

🎀 dedot

reduces cross-luminance (that's the dot-crawl, bestie) and cross-color (the rainbows 🌈) from video. iconic little cleanup crew.

it accepts the following options:

m — set the mode of operation. can be a combination of dotcrawl for cross-luminance reduction and/or rainbows for cross-color reduction.
lt — set the spatial luma threshold. lower values increase reduction of cross-luminance.
tl — set the tolerance for temporal luma. higher values increase reduction of cross-luminance.
tc — set the tolerance for chroma temporal variation. higher values increase reduction of cross-color.
ct — set the temporal chroma threshold. lower values increase reduction of cross-color.

⭐ deflate

applies the deflate effect to the video.

this filter replaces the pixel with the local (3x3) average, taking into account only values lower than the pixel. it's giving "shrink the bright bits" energy.

it accepts the following options:

threshold0, threshold1, threshold2, threshold3 — limit the maximum change for each plane. default is 65535. if 0, that plane will remain unchanged 🙈

✨ Commands

this filter supports all of the above options as commands. slay.

🌸 deflicker

removes temporal frame luminance variations — the annoying brightness flicker, gone.

it accepts the following options:

size, s — set the moving-average filter size in frames. default is 5. allowed range is 2 - 129.
mode, m — set the averaging mode for smoothing those temporal luminance variations.

the available values are:

am — arithmetic mean
gm — geometric mean
hm — harmonic mean
qm — quadratic mean
cm — cubic mean
pm — power mean
median — median

bypass — do not actually modify the frame. useful when you only want the metadata 🫶

💅 deinterlace_d3d12

deinterlaces the input video using D3D12 hardware-accelerated video processing. gpu goes brrr.

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.

the filter automatically queries the hardware for the required number of temporal reference frames (if needed) and manages a frame queue to supply them. considerate queen behavior 👑

it accepts the following options:

method — the deinterlacing method. accepts one of the following values:
- default — select the best deinterlacing method available on the hardware. if the driver supports custom, that method is used; otherwise it falls back to bob. this is the default.
- bob — bob deinterlacing. each field is independently scaled to full frame height. simple and fast, but 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. 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.

the default value is frame.

deint — specify which frames to deinterlace. 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.

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. periodt.

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 watch something even uglier appear, your mileage may vary 😭

it accepts the following parameters:

x, y — specify the top left corner coordinates of the logo. they must be specified.
w, h — specify the width and height of the logo to clear. they must be specified.
show — when set to 1, a green rectangle is drawn on the screen 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. highkey clever 🧚

🎀 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 derain methods based on convolutional neural networks. yes bestie, we're doing ML in our video filter. supported models:

Recurrent Squeeze-and-Excitation Context Aggregation Net (RESCAN). see http://openaccess.thecvf.com/content_ECCV_2018/papers/Xia_Li_Recurrent_Squeeze-and-Excitation_Context_ECCV_2018_paper.pdf.

training as well as model generation scripts are provided in the repository at https://github.com/XueweiMeng/derain_filter.git. 🫶

the filter accepts the following options:

filter_type — specify which filter to use. accepts the following values:
- derain — derain filter. to conduct a derain filter, you need to use a derain model.
- dehaze — dehaze filter. to conduct a dehaze filter, you need to use a dehaze model.

default value is derain.

dnn_backend — specify which DNN backend to use for model loading and execution. accepts the following values:
- tensorflow — TensorFlow backend. to enable this backend you need to install the TensorFlow for C library (see https://www.tensorflow.org/install/lang_c) and configure FFmpeg with --enable-libtensorflow
model — set the path to the model file specifying the network architecture and its parameters. note that different backends use different file formats. TensorFlow can load files for only its format. 💀

to get full functionality (such as async execution), please use the dnn_processing filter instead.

⭐ deshake

attempts to fix small changes in horizontal and/or vertical shift. this filter helps remove camera shake from hand-holding a camera, bumping a tripod, moving on a vehicle, etc. shaky-cam girlies rise 🙈

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. iykyk 👑

If any or all of x, y, w and h are set to -1, the whole frame gets used 💖 — which is lowkey the move, because then you can set the later options without having to spell out the bounding box for the motion vector search. Default? Search the whole frame. periodt.

rx, ry — the max extent of movement in the x and y directions, range 0-64 pixels. Default 16. 🎀
edge — tells the filter how to generate pixels to fill the blanks at the edge of the frame. The values you can pick:
- blank, 0 — fill zeroes at blank locations
- original, 1 — original image at blank locations
- clamp, 2 — extruded edge value at blank locations
- mirror, 3 — mirrored edge at blank locations

Default value is mirror. 🪞

blocksize — the blocksize used for motion search. Range 4-128 pixels, default 8. ✨
contrast — the contrast threshold for blocks, bestie. Only blocks with more than the specified contrast (that's the difference between the darkest and the lightest pixels) get considered at all. Range 1-255, default 125. 👑
search — the search strategy. Available values:
- exhaustive, 0 — set exhaustive search
- less, 1 — set less exhaustive search.

Default value is exhaustive. 💅

filename — if you set this, a detailed log of the motion search gets written to the file you name. 🍓

despill 🧚

Removes the unwanted contamination of foreground colors that happens because of reflected color from a greenscreen or bluescreen. iconic cleanup behavior, honestly.

This filter accepts the following options:

type — sets what type of despill you're using.
mix — sets how the spillmap is going to be generated.
expand — sets how much to get rid of the still-remaining spill.
red — controls the amount of red in the spill area.
green — controls the amount of green in the spill area. Should be -1 for greenscreen. 💚
blue — controls the amount of blue in the spill area. Should be -1 for bluescreen. 💙
brightness — controls brightness of the spill area while preserving colors.
alpha — modify alpha from the generated spillmap.

Commands 💀

This filter supports the all above options as commands.

detelecine 🎀

Applies an exact inverse of the telecine operation. It requires a predefined pattern, specified with the pattern option, and that pattern must be the same one you passed to the telecine filter. no cap, they have to match.

This filter accepts the following options:

first_field
- top, t — top field first
- bottom, b — bottom field first

The default value is "top". ⭐

pattern — a string of numbers representing the pulldown pattern you want to apply. The default value is 23. 🔥
start_frame — a number representing the position of the first frame with respect to the telecine pattern. Use this if the stream is cut. The default value is 0. ✂️

dilation 🌸

Applies a dilation effect to the video. Basically this filter replaces the pixel with the local (3x3) maximum. main character pixel energy.

It accepts the following options:

threshold0, threshold1, threshold2, threshold3 — limit the maximum change for each plane, default is 65535. If it's 0, that plane will remain unchanged. 🙈
coordinates — a flag that specifies which pixel to refer to. Default is 255, i.e. all eight pixels are used.

The flags map to the local 3x3 coordinates like this:

1 2 3
4   5
6 7 8

Commands 💀

This filter supports the all above options as commands.

displace 🌀

Displaces pixels as indicated by the second and third input streams.

It takes three input streams and outputs one stream. The first input is the source, and the second and third inputs are displacement maps.

The second input specifies how much to displace pixels along the x-axis, while the third input specifies how much to displace pixels along the y-axis. If one of the displacement map streams terminates, the last frame from that displacement map will be used. 🫶

Note that once they're generated, displacement maps can be reused over and over again. obsessed with the efficiency honestly.

A description of the accepted options follows.

edge — sets the displace behavior for pixels that are out of range. Available values:
- blank — missing pixels are replaced by black pixels.
- smear — adjacent pixels will spread out to replace missing pixels.
- wrap — out of range pixels are wrapped so they point to pixels of the other side.
- mirror — out of range pixels will be replaced with mirrored pixels.

Default is smear. ✨

Examples 💖

Add a ripple effect to an 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 a wave effect to an rgb input of video size hd720:

ffmpeg -i INPUT -f lavfi -i nullsrc=hd720,geq='r=128+80*(sin(sqrt((X-W/2)*(X-W/2)+(Y-H/2)*(Y-H/2))/220*2*PI+T)):g=128+80*(sin(sqrt((X-W/2)*(X-W/2)+(Y-H/2)*(Y-H/2))/220*2*PI+T)):b=128+80*(sin(sqrt((X-W/2)*(X-W/2)+(Y-H/2)*(Y-H/2))/220*2*PI+T))' -lavfi '[1]split[x][y],[0][x][y]displace' OUTPUT

dnn_classify 🧠

Does classification with deep neural networks, based on bounding boxes. slay.

The filter accepts the following options:

dnn_backend — specify which DNN backend to use for model loading and execution. This option accepts only openvino right now; tensorflow backends will be added later.
model — set the path to the model file specifying network architecture and its parameters. Note that different backends use different file formats.
input — set the input name of the dnn network.
output — set the output name of the dnn network.
confidence — set the confidence threshold (default: 0.5).
labels — set the path to the label file specifying the mapping between label id and name. Each label name is written on one line; tailing spaces and empty lines are skipped. The first line is the name of label id 0, the second line is the name of label id 1, etc. The label id is considered as the name if the label file isn't provided.
backend_configs — set the configs to be passed into the backend.

For the tensorflow backend, you can set its configs with sess_config options — please use tools/python/tf_sess_config.py to get the configs for your system.

dnn_detect 🔍

Does object detection with deep neural networks. detective mode activated.

The filter accepts the following options:

dnn_backend — specify which DNN backend to use for model loading and execution. This option accepts only openvino right now; tensorflow backends will be added later.
model — set the path to the model file specifying network architecture and its parameters. Note that different backends use different file formats.
input — set the input name of the dnn network.
output — set the output name of the dnn network.
confidence — set the confidence threshold (default: 0.5).
labels — set the path to the label file specifying the mapping between label id and name. Each label name is written on one line; tailing spaces and empty lines are skipped. The first line is the name of label id 0 (usually it's 'background'), and the second line is the name of label id 1, etc. The label id is considered as the name if the label file isn't provided.
backend_configs — set the configs to be passed into the backend. To use async execution, set async (default: set). Roll back to sync execution if the backend doesn't support async. 😭

dnn_processing 🦋

Does image processing with deep neural networks. It works together with another filter which converts the pixel format of the Frame to whatever 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 aren't installed into the system path). ✨
- torch — Libtorch backend. To enable this backend you need to build and install the Libtorch for C++ library. Please download the 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 the path to the model file specifying network architecture and its parameters. Note that different backends use different file formats. TensorFlow, OpenVINO and Libtorch backends can load files for only their own 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 the backend. To use async execution, set async (default: set). Roll back to sync execution if the backend doesn't support async.

For the tensorflow backend, you can set its configs with sess_config options — please use tools/python/tf_sess_config.py to get the configs of the TensorFlow backend for your system.

Examples 💖

Remove rain in an rgb24 frame with can.pb (see the 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 the sr filter) for a 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 the 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 the 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`

Draws a colored box right on top of the input image. Simple as that, bestie.

It accepts the following parameters:

x, y — the expressions that say where the top left corner of the box sits. Defaults to 0.
width, w, height, h — the expressions for 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 — the color of the box you're writing. For the general syntax of this option, check the "Color" section in the ffmpeg-utils manual. And listen — if you use the special value invert, the box edge color becomes the same as the video with inverted luma. Cute, right?
thickness, t — the expression that sets the thickness of the box edge. A value of fill gives you a filled box (yes, the whole thing). Default value is 3.

See below for the list of accepted constants.

replace — only applicable if the input has alpha. With value 1, the pixels of the painted box 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's 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, 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. Iconic mutual-reference behavior, honestly.

✨ 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 the same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value. No drama, it just vibes where it is. 🌸

📈 `drawgraph`

Draws a graph using input video metadata. Yes, your metadata goes on a little graph. Main character energy for real.

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. Note the byte order, bestie, don't get it twisted 🙈.

🧚 Example using metadata from `signalstats` filter:

signalstats,drawgraph=lavfi.signalstats.YAVG:min=0:max=255

🧚 Example using metadata from `ebur128` filter:

ebur128=metadata=1,adrawgraph=lavfi.r128.M:min=-120:max=5

🔲 `drawgrid`

Draws a grid on the input image. Grids are lowkey underrated, and this one slays.

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, 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 the image gets framed. Default to 0.
color, c — specify the color of the grid. For the general syntax of this option, check the "Color" section in the ffmpeg-utils manual. If the special value invert is used, the grid color is the same as the video with inverted luma.
thickness, t — the expression 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, 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 energy as drawbox, periodt. 💅

✍️ `drawtext`

Draws a text string — or text from a specified file — on top of a video, using the libfreetype library. This is the one everybody's obsessed with, ngl.

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.

🫶 Heads up bestie: those are build-time configure flags. If your FFmpeg wasn't built with them, this filter (or parts of it) just 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 a 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.

💀 strftime is deprecated, bestie. It still exists, but it's giving legacy vibes.

basetime — set a start time for the count. Value is in microseconds. Only applied in the deprecated strftime expansion mode. To emulate in normal expansion mode use the pts function, supplying the start time (in seconds) as the second argument.
fix_bounds — if it's true, the filter checks your text coords and fixes them so nothing gets clipped off the edge. considerate king behavior 💅
fontcolor — the colour used to draw the text. for the syntax, go peek at the "Color" section in the ffmpeg-utils manual. default is "black".
fontcolor_expr — a string that gets expanded exactly the same way as text to obtain a dynamic fontcolor value. by default this option is empty and just... not processed. but when you set it, it overrides the fontcolor option. so it's lowkey the main character 🎀
font — the font family used for drawing the text. default is Sans.
font file — ok wait, that's fontfile: the actual font file used for drawing text. the path must be included (yes, the whole path, not just the name bestie). this parameter is mandatory if the fontconfig support is disabled.
alpha — draw the text with alpha blending. value can be a number between 0.0 and 1.0. the expression accepts the same variables x, y as well. default value is 1. please see fontcolor_expr (it's giving sibling energy).
fontsize — the font size for drawing text. default is 16.
text_shaping — if set to 1, it attempts to shape the text before drawing it (like reversing the order of right-to-left text and joining Arabic characters). otherwise it just draws the text exactly as given. by default 1 (if supported). ✨
ft_load_flags — the flags used for loading the fonts. these map to the corresponding flags supported by libfreetype, and they're 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

default value is "default". for more info, consult the documentation for the FT_LOAD_* libfreetype flags. ngl, that's a lot of flags, but we respect her 💅

shadowcolor — the colour used for drawing a shadow behind the drawn text. for the syntax, check the "Color" section in the ffmpeg-utils manual. default is "black".
boxw — set the width of the box drawn around the text. by default, boxw is computed automatically to match the text width. 🎁
boxh — set the height of the box drawn around the text. by default, boxh is computed automatically to match the text height.
shadowx, shadowy — the x and y offsets for the text shadow position relative to the position of the text. they can be positive or negative. default for both is "0".
start_number — the starting frame number for the n/frame_num variable. default is "0".
tabsize — the size, in number of spaces, used for rendering a tab. default is 4.
timecode — set the initial timecode representation in "hh:mm:ss[:;.]ff" format. can be used with or without the text parameter. the timecode_rate option must be specified. ⏰
timecode_rate, rate, r — set the timecode frame rate (timecode only). the value gets 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 wraps around at 24 hours. default is 0 (disabled).
text — the text string to be drawn. it must be a sequence of UTF-8 encoded characters. this parameter is mandatory if no file is specified with the textfile parameter.
textfile — a text file containing the 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 text parameter.

if both text and textfile are specified, an error is thrown. pick a lane, babe 😭

text_source — set the text source to side_data_detection_bboxes if you want to use text data in the detection bboxes of side data.

if the text source is set, text and textfile get ignored and it still uses the text data in the detection bboxes of side data. so please don't use this parameter if you're not sure about the text source. > seriously, don't mess with it if you don't know what it does 🫶

reload — the textfile gets reloaded at the specified frame interval. be sure to update textfile atomically, or it may be read partially, or even fail outright. range is 0 to INT_MAX. default is 0.
x, y — the expressions that specify the offsets where text is drawn within the video frame. they're relative to the top/left border of the output image. default value of x and y is "0". see below for the list of accepted constants and functions 👇

🌸 the constants & functions for `x` and `y` 🌸

the parameters for x and y are expressions containing these constants and functions:

dar — input display aspect ratio, same as (w / h) * sar
hsub, vsub — horizontal and vertical chroma subsample values. for example, for the pixel format yuv422p, hsub is 2 and vsub is 1.
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's a positive value, because of 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 one's a negative value, due to the grid's orientation with the Y axis upwards.
max_glyph_h — maximum glyph height, i.e. the maximum height for all the glyphs contained in the rendered text. it's equivalent to ascent - descent.
max_glyph_w — maximum glyph width, i.e. 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 the input frame, starting from 0
rand(min, max) — returns 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 let the x and y expressions refer to each other, so you can for example specify y=x/dar. iconic mutual-reference moment 💫

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. > heads up: this feature is deprecated in favor of "normal" expansion with the gmtime or localtime expansion functions. we're moving on, bestie 💅

if expansion is set to "none", the text is printed verbatim. no notes.

if expansion is set to "normal" (which is the default), this expansion mechanism is used:

the backslash character \, followed by any character, always expands to the second character.

sequences of the form %{...} get 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's up to four levels of escaping 😭 using a text file with the textfile option avoids these problems. huge W for text files honestly.

these are the functions available:

expr, e — the expression evaluation result.

it must take one argument specifying the expression to evaluate, 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 isn't 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 it as a formatted integer.

the first argument is the expression to evaluate, just like for the expr function. the second argument specifies the output format — allowed values are x, X, d and u, 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 is empty.

available metadata can be identified by inspecting entries starting with TAG included within each frame section printed by running ffprobe -show_frames.

string metadata generated in filters leading to the drawtext filter are also available. 🫶

n, frame_num — the frame number, starting from 0.
pict_type — a one character description of the current picture type.
pts — the timestamp of the current frame. it can take up to three arguments.

The first argument is the format of the timestamp, and it defaults to "flt" — that's seconds as a decimal number with microsecond accuracy. "hms" gives you a formatted [-]HH:MM:SS.mmm timestamp with millisecond accuracy. "gmtime" gives you the frame's timestamp formatted as UTC time, and "localtime" gives it to you in the local time zone. 🌍

The second argument is an offset that gets added to the timestamp.

If the format is "hms", you can pass a third argument "24HH" to present the hour part of the formatted timestamp in 24h format (00-23). 🕐

If the format is "localtime" or "gmtime", a third argument may be supplied: a strftime C function format string. If you don't, the default YYYY-MM-DD HH:MM:SS format is what you get. 💖

Commands

Bestie, this filter supports altering parameters via commands:

reinit — alters existing filter parameters. The syntax for the argument is the same as for filter invocation, like:

fontsize=56:fontcolor=green:text='Hello World'

Full filter invocation with sendcmd would look like this:

sendcmd=c='56.0 drawtext reinit fontsize=56\:fontcolor=green\:text=Hello\\ World'

If the entire argument can't be parsed or applied as valid values, then the filter just continues on with its existing parameters. No drama, it just vibes with what it 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). The text is yellow with a red box around it, and 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 aren't necessary if spaces aren't used within the parameter list. Periodt. ✨

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". Ngl, note that the 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). It's giving "I made my own language, no big deal". 👑

The documentation for the language can be found in drawvg - Language Reference. A version of this reference with rendered examples can be found at <https://ayosec.github.io/ffmpeg-drawvg/playground/docs/langref.html>.

Graphics are rendered using <https://cairographics.org/>.

To enable compilation of this filter, you need to configure FFmpeg with --enable-cairo. 💀

Parameters

Either script or file must be set. (Yes, you need one of them. Pick your fighter.)

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. Don't forget this step bestie, or it's gonna be a whole mess. 😭

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. Lowkey iconic. ✨

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. Fr fr, that's the whole vibe.

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. (It tries its best, ok? 🫶)
pal8 — set pal8 output pixel format. This option does not work with codebook length greater than 256. Default is disabled.
use_alpha — include alpha values in the quantization calculation. Allows creating palettized output images (e.g. PNG8) with multiple alpha smooth blending. 💅

📊 entropy

Measure graylevel entropy in histogram of color channels of video frames.

It accepts the following parameters:

mode — can be either normal or diff. Default is normal.

diff mode measures entropy of histogram delta values, absolute differences between neighbour histogram values.

🔍 epx

Apply the EPX magnification filter which is designed for pixel art. Pixel artists, this one's for you. 🎀

It accepts the following option:

n — set the scaling dimension: 2 for "2xEPX", 3 for "3xEPX". Default is 3.

⚖️ eq

Set brightness, contrast, saturation and approximate gamma adjustment.

The filter accepts the following options:

contrast — set the contrast expression. The value must be a float value in range -1000.0 to 1000.0. The default value is "1".
brightness — set the brightness expression. The value must be a float value in range -1.0 to 1.0. The default value is "0".
saturation — the saturation expression. Gotta be a float from 0.0 to 3.0, default is "1". 💅
gamma — the gamma expression. Float in range 0.1 to 10.0, default "1".
gamma_r — gamma expression, but make it red 🍓. Float from 0.1 to 10.0, default "1".
gamma_g — gamma expression for green. Float 0.1 to 10.0, default "1".
gamma_b — gamma expression for blue. Float 0.1 to 10.0, default "1".
gamma_weight — the gamma weight expression. Lowkey the glow-up protector of your highlights: it reduces the effect of a big gamma value on bright image areas so they don't get overamplified into a plain white void 😭. Float from 0.0 to 1.0. 0.0 turns gamma correction alllll the way down, 1.0 leaves it at full strength. Default is "1".
eval — tells the filter when to evaluate the brightness, contrast, saturation and gamma expressions. Takes these values:
- init — evaluate the expressions once, during filter initialization or when a command is processed.
- frame — evaluate the expressions for every incoming frame. ✨

Default value is init.

The expressions accept these 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. > heads up bestie: this one is deprecated, do not use it 💀
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

This filter supports the following commands, no cap:

contrast — set the contrast expression.
brightness — set the brightness expression.
saturation — set the saturation expression.
gamma — set the gamma expression.
gamma_r — set the gamma_r expression.
gamma_g — set gamma_g expression.
gamma_b — set gamma_b expression.
gamma_weight — set gamma_weight expression.

The command accepts the same syntax as the corresponding option. And if the expression you hand it isn't valid? It just chills at its current value 🫶.

`erosion`

Bestie this one literally erodes your video. Apply the erosion effect — it replaces the pixel by the local 3x3 minimum. Iconic, destructive, love her 💅

It accepts the following options:

threshold0, threshold1, threshold2, threshold3 — limit the maximum change for each plane. Default is 65535. If it's 0, that plane stays completely unchanged.
coordinates — a flag saying which pixels to refer to. Default is 255, meaning all eight pixels get used.

The flags map to local 3x3 coordinates like this:

1 2 3
4   5
6 7 8

Commands

This filter supports all the above options as commands. 👑

`estdif`

Deinterlace the input video — and yes, "estdif" stands for "Edge Slope Tracing Deinterlacing Filter", she's got a whole government name 🌸

Spatial only filter that uses the edge slope tracing algorithm to interpolate the missing lines. It accepts the following parameters:

mode — the interlacing mode to adopt. Accepts:
- 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. Accepts:
- 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. Accepts:
- all — deinterlace all frames.
- interlaced — only deinterlace frames marked as interlaced.

Default value is "all".

rslope — the search radius for edge slope tracing. Default value is 1. Allowed range is 1 to 15.
redge — the search radius for best edge matching. Default value is 2. Allowed range is 0 to 15.
ecost — the edge cost for edge matching. Default value is 2. Allowed range is 0 to 50.
mcost — the middle cost for edge matching. Default value is 1. Allowed range is 0 to 50.
dcost — the distance cost for edge matching. Default value is 1. Allowed range is 0 to 50.
interp — the interpolation used. Default is 4-point interpolation. Accepts:
- 2p — two-point interpolation.
- 4p — four-point interpolation.
- 6p — six-point interpolation.

Commands

This filter supports the same commands as options. 💖

`exposure`

Adjust the exposure of the video stream. It's giving camera app slider energy ✨

The filter accepts the following options:

exposure — set the exposure correction in EV. Allowed range is from -3.0 to 3.0 EV. Default value is 0 EV.
black — set the black level correction. Allowed range is from -1.0 to 1.0. Default value is 0.

Commands

This filter supports the same commands as options.

`extractplanes`

Extract color channel components from the input video stream into separate grayscale video streams. One video in, a whole friend group of grayscale streams out 😭💜

The filter accepts the following option:

planes — set the plane(s) to extract.

Available values for planes are:

Choosing planes that aren't available in the input will result in an error. Basically you can't select "r", "g", "b" planes at the same time as "y", "u", "v" planes — pick a lane bestie 🫶

Examples

Extract luma, u and v color channel components from the input video frame into 3 grayscale outputs:

ffmpeg -i video.avi -filter_complex 'extractplanes=y+u+v[y][u][v]' -map '[y]' y.avi -map '[u]' u.avi -map '[v]' v.avi

`fade`

Apply a fade-in/out effect to the input video. Main character entrance AND exit, both handled 🌈

It accepts the following 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 to start applying the fade effect at. Default is 0.
nb_frames, n — how many frames the fade effect lasts. At the end of the fade-in, 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 to start applying the fade effect. 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 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 wins. 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

The command above is equivalent to:

fade=t=in:s=0:n=30

Fade out the last 45 frames of a 200-frame video:

fade=out:155:45
fade=type=out:start_frame=155:nb_frames=45

Fade in the first 25 frames and fade out the last 25 frames of a 1000-frame video:

fade=in:0:25, fade=out:975:25

Make the first 5 frames yellow, then fade in from frame 5-24:

fade=in:5:20:color=yellow

Fade in alpha over 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`

Apply the feedback video filter. This one's a little mind-bender ngl 🔥

This filter passes cropped input frames to the 2nd output. From there you can filter them with other video filters. After the filter receives a frame from the 2nd input, that frame is 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. Sneaky, targeted, we love her 💅

The filter accepts the following 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, 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, 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, size 100x100, with the pixelize filter:

[in][blurin]feedback=x=320:y=240:w=100:h=100[out][blurout];[blurout]pixelize[blurin]

`fftdnoiz`

Denoise frames using 3D FFT (frequency domain filtering). Science girlie behaviour 🧚✨

The filter accepts the following options:

sigma — set the noise sigma constant, which sets the denoising strength. Default value is 1. Allowed range is from 0 to 30. > warning bestie: using a very high sigma with low overlap may give you blocking artifacts 😭
amount — set the amount of denoising. By default all detected noise is reduced. Default value is 1. Allowed range is from 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 from 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 the planes which will be filtered. By default all available planes are filtered, except alpha.

`fftfilt`

Apply arbitrary expressions to samples in the frequency domain — and that's where we pick this up, bestie… 💫

dc_Y — tweaks the dc value (gain, bestie) of the luma plane. Integer, 0 to 1000, default 0. 💛
dc_U — same deal but for the 1st chroma plane. Integer 0 to 1000, default 0.
dc_V — and the 2nd chroma plane, because she deserves one too. Integer 0 to 1000, default 0.
weight_Y — sets the frequency domain weight expression for the luma plane. ✨
weight_U — frequency domain weight expression for the 1st chroma plane.
weight_V — frequency domain weight expression for the 2nd chroma plane.
eval — decides when the expressions get evaluated. Choices are:
- init — only evaluate expressions once, during filter initialization.
- frame — evaluate expressions for every single incoming frame.

Default value is init. Lowkey efficient.

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 input frame, starting from 0.
WS, HS — the size of 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

Extracts a single field from an interlaced image using stride arithmetic, so it doesn't waste CPU time — efficient queen behavior honestly. The output frames are marked as non-interlaced. 🎀

The filter accepts the following options:

type — specifies 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

Makes brand-new frames by copying the top and bottom fields from surrounding frames — and the frame numbers come from a hint file, bestie. It's giving "DIY field assembly" and we're obsessed. 🧚

hint — sets the file containing hints: absolute/relative frame numbers.

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 file can not be out of [N-1,N+1] where N is current frame number for "absolute" mode or out of [-1, 1] range for "relative" mode. First number tells from which frame to pick up top field and second number tells from which frame to pick up bottom field.

If optionally followed by "+" output frame will be marked as interlaced, else if followed by "-" output frame will be marked as progressive, else it will be marked same as input frame. If optionally followed by "t" output frame will use only top field, or in case of "b" it will use only bottom field. If line starts with "#" or ";" that line is skipped. 🫶

mode — can be item "absolute" or "relative" or "pattern". Default is "absolute". The "pattern" mode is same as "relative" mode, except at last entry of file if there are more frames to process than "hint" file is seek back to start. Yes, it loops. Main character energy.

Example of first several lines of "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 — say that five times fast, I dare you. It's meant to reconstruct the progressive frames from a telecined stream. Important thing though: the filter does not drop duplicated frames, so to actually achieve a complete inverse telecine "fieldmatch" needs to be followed by a decimation filter such as decimate in the filtergraph. Don't forget your girl decimate, fr fr. 💀

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. Slay, graceful fallback. ✨

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 later is a light clone of TFM from which "fieldmatch" is based on. While the semantic and usage are very close, some behaviour and options names can differ. So like… same vibe, different font. 👑

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". Copy-paste her, she's iconic.

The filter accepts the following options:

order — specifies 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. 🙈 The stream can lie, bestie.

Default value is auto.

mode — sets the matching mode or strategy to use. pc mode is the safest in the sense that it won't risk creating jerkiness due to duplicate frames when possible, but if there are bad edits or blended fields it 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 p/c/n/u/b meaning section. (Go read it, iykyk.)

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 — marks the main input stream as a pre-processed input, and enables the secondary input stream as the clean source to pick the fields from. See the filter introduction for more details. It is similar to the clip2 feature from VFM/TFM. 🦋

Default value is 0 (disabled).

field — sets 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 — sets 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. Trade-offs, babe. ⚖️

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 defaults to 0.
scthresh — sets the scene change detection threshold as a percentage of maximum change on the luma plane. Good values are in the "[8.0, 14.0]" range. Scene change detection is only relevant in case combmatch=sc. The range for scthresh is "[0.0, 100.0]". 🍓

Default value is 12.0.

combmatch — when combmatch is not none, "fieldmatch" will take into account the combed scores of matches when deciding 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 — forces "fieldmatch" to calculate the combed metrics for certain matches and print them. This setting is known as micout in TFM/VFM vocabulary. Available values are:
- none — no forced calculation.
- pcn — force p/c/n calculations.
- pcnub — force p/c/n/u/b calculations.

Default value is none.

cthresh — the area combing threshold used for combed frame detection. It basically controls how "strong" or "visible" the combing has to be before the filter goes "yep, that's combed." Bigger values = combing must be more visible to get caught, smaller values = combing can be faint and still get detected. Valid settings run from -1 (literally every pixel gets flagged as combed 💀) to 255 (no pixel ever gets flagged as combed). Underneath it's a pixel difference value, and a good range is "[8, 12]".

Default value is 9.

chroma — sets whether chroma gets a vote in the combed-frame decision. Only turn this off if your source has chroma problems (rainbowing, etc.) that are messing with detection while chroma is enabled. Ngl, using chroma=0 is usually more reliable, except for the case where there's chroma-only combing in the source.

Default value is 0.

blockx, blocky — respectively set the x-axis and y-axis size of the window used during combed frame detection. It's about the size of the area where combpel pixels have to be detected as combed for a frame to be declared combed — see the combpel description for more. Possible values are any number that's a power of 2, starting at 4 and going up to 512.

Default value is 16.

combpel — the number of combed pixels inside any of the blocky by blockx size blocks on the frame, for the frame to be detected as combed. So while cthresh controls how "visible" the combing has to be, this one controls how much combing there has to be in any localized area (a window defined by the blockx and blocky settings) on the frame. Minimum value is 0 and maximum is "blocky x blockx" (at which point no frames will ever be detected as combed). In TFM/VFM vocabulary this setting is known as MI 🎀

Default value is 80.

p/c/n/u/b meaning — the lore ✨

p/c/n

We assume the following telecined stream:

Top fields:     1 2 2 3 4
Bottom fields:  1 2 3 4 4

The numbers correspond to the progressive frame the fields relate to. 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

As a result of the field matching, we can see that some frames get duplicated. So to do a complete inverse telecine you gotta rely on a decimation filter after this operation — see for instance the decimate filter, bestie 🫶

The same operation, but now matching from top fields (field=top), looks like this:

Input stream:
                T     1 2 2 3 4   <-- matching reference
                B     1 2 3 4 4

Matches:              c c p p c

Output stream:
                T     1 2 2 3 4
                B     1 2 2 3 4

In these examples you can see what p, c and n mean; basically they refer to the frame and field of the opposite parity:

<p matches the field of the opposite parity in the previous frame>
<c matches the field of the opposite parity in the current frame>
<n matches the field of the opposite parity in the next frame>

u/b

The u and b matching are a lil special, in the sense that they match from the opposite parity flag. In the following examples we assume we're currently matching the 2nd frame (Top:2, bottom:2). According to the match, an x is 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

Transform the field order of the input video. Serving looks, but for fields 💅

It accepts the following parameters:

order — the output field order. Valid values are tff for top field first, or bff for bottom field first.

The default value is tff.

The transformation is done by shifting the picture content up or down by one line, and filling the remaining line with appropriate picture content. This method is consistent with most broadcast field order converters.

If the input video isn't flagged as being interlaced, or it's already flagged as being of the required output field order, then this filter doesn't alter the incoming video at all. It just vibes. 🌸

It's super useful when converting to or from PAL DV material, which is bottom field first.

For example:

ffmpeg -i in.vob -vf "fieldorder=bff" out.dv

🖼️ fillborders

Fill borders of the input video without changing video stream dimensions. Sometimes video has garbage at the four edges and you don't wanna crop the input just to keep the size a multiple of some number. Iconic behaviour, honestly.

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.

It accepts the following values:

smear — fill pixels using outermost pixels
mirror — fill pixels using mirroring (half sample symmetric)
fixed — fill pixels with constant value
reflect — fill pixels using reflecting (whole sample symmetric)
wrap — fill pixels using wrapping
fade — fade pixels to constant value
margins — fill pixels at top and bottom with weighted averages pixels near borders

Default is smear.

color — set color for pixels in fixed or fade mode. Default is black.

Commands

This filter supports the same commands as options. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it's kept at its current value. No drama, no crash 💀

🔍 find_rect

Find a rectangular object in the input video. Detective mode: activated 🕵️‍♀️

The object to search for must be specified as a gray8 image, given with the object option.

For each possible match, a score is computed. If the score reaches the specified threshold, the object is considered found.

If the input video contains multiple instances of the object, the filter will find only one of them. (Rude, but ok.)

When an object is found, the following metadata entries are set in the matching frame:

lavfi.rect.w — width of object
lavfi.rect.h — height of object
lavfi.rect.x — x position of object
lavfi.rect.y — y position of object
lavfi.rect.score — match score of the found object

It accepts the following options:

object — filepath of the object image, needs to be in gray8.
threshold — detection threshold, expressed as a decimal number in the range 0-1.

A threshold value of 0.01 means only exact matches, a threshold of 0.99 means almost everything matches. Choose your fighter ⚔️

Default value is 0.5.

mipmaps — number of mipmaps, default is 3.
xmin, ymin, xmax, ymax — specifies the rectangle in which to search.
discard — discard frames where object is not detected. Default is disabled.

Examples

Cover a rectangular object by the supplied image of a given video using ffmpeg:

ffmpeg -i file.ts -vf find_rect=newref.pgm,cover_rect=cover.jpg:mode=cover new.mkv

Find the position of an object in each frame using ffprobe and write it to a log file:

ffprobe -f lavfi movie=test.mp4,find_rect=object=object.pgm:threshold=0.3 \
  -show_entries frame=pkt_pts_time:frame_tags=lavfi.rect.x,lavfi.rect.y \
  -of csv -o find_rect.csv

🎨 floodfill

Flood area with values of same pixel components with another values. Yes, it's the paint-bucket tool from MS Paint but make it Unix 💖

It accepts the following options:

x — set pixel x coordinate.
y — set pixel y coordinate.
s0 — set source #0 component value.
s1 — set source #1 component value.
s2 — set source #2 component value.
s3 — set source #3 component value.
d0 — set destination #0 component value.
d1 — set destination #1 component value.
d2 — set destination #2 component value.
d3 — set destination #3 component value.

🌈 format

Convert the input video to one of the specified pixel formats. Libavfilter will try to pick one that's suitable as input to the next filter — it's out here doing matchmaking for your filter chain 🫶

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". (Yep, it literally says "color range names" right there in the manpage 🙈 iykyk.)

Examples

Convert the input video to the yuv420p format:

format=pix_fmts=yuv420p

Convert the input video to any of the formats in the list:

format=pix_fmts=yuv420p|yuv444p|yuv410p

🎀 `fps`

Okay so this one is the frame-rate glow-up filter 💅 — it converts your video to a specified constant frame rate by duplicating or dropping frames as necessary. Like, it will literally clone frames or yeet them into the void to hit the number you asked for.

It accepts the following parameters:

fps — the desired output frame rate. Iconic behaviour: it accepts expressions containing the following 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 — assume the first PTS should be the given value, in seconds. This allows for padding/trimming at the start of stream. By default, no assumption is made about the first frame's expected PTS, so no padding or trimming is done. For example, this could be set 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 — action performed when reading the last frame. Possible values are:
- round — use same timestamp rounding method as used for other frames.
- pass — pass through 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`

Pack two different video streams into a stereoscopic video, setting proper metadata on supported codecs. The two views should have the same size and framerate, and processing will stop when the shorter video ends (rude, but fair 😭). Please note that you may conveniently adjust view properties with the scale and fps filters.

It accepts the following parameters:

format — the desired packing format. Supported values are:
- sbs — the views are next to each other (default).
- tab — the views are on top of each other.
- lines — the views are packed by line.
- columns — the views are packed by column.
- frameseq — the views are temporally interleaved.

Some examples:

# Convert left and right views into a frame-sequential video
ffmpeg -i LEFT -i RIGHT -filter_complex framepack=frameseq OUTPUT

# Convert views into a side-by-side video with the same output resolution as the input
ffmpeg -i LEFT -i RIGHT -filter_complex [0:v]scale=w=iw/2[left],[1:v]scale=w=iw/2[right],[left][right]framepack=sbs OUTPUT

⭐ `framerate`

Change the frame rate by interpolating new video output frames from the source frames. Literally manifesting frames that never existed, main character energy ✨.

⚠️ 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 cap, do it in that order.

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. Very "I'm only responding to every fifth text" of it tbh 🙈.

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. It's giving "is this thing even on" 👀.

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. This filter freezes video frames using a frame from the 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.

It accepts the following parameters:

filter_name — the name of the frei0r effect to load. If the environment variable FREI0R_PATH is defined, the frei0r effect is searched for in each of the directories specified by the colon-separated list in FREI0R_PATH. Otherwise, the standard frei0r paths are searched, in this order: HOME/.frei0r-1/lib/, /usr/local/lib/frei0r-1/, /usr/lib/frei0r-1/.
filter_params — a '|'-separated list of parameters to pass to the frei0r effect.

A frei0r effect parameter can be a boolean (its value is either "y" or "n"), a double, a color (specified as R/G/B, where R, G, and B are floating point numbers between 0.0 and 1.0, inclusive) or a color description as specified in the "Color" section in the ffmpeg-utils manual, a position (specified as X/Y, where X and Y are floating point numbers) and/or a string.

The number and types of parameters depend on the loaded effect. If an effect parameter is not specified, the default value is set.

Examples

Apply the distort0r effect, setting the first two double parameters:

frei0r=filter_name=distort0r:filter_params=0.5|0.01

Apply the colordistance effect, taking a color as the first parameter:

frei0r=colordistance:0.2/0.3/0.4
frei0r=colordistance:violet
frei0r=colordistance:0x112233

Apply the perspective effect, specifying the top left and top right image positions:

frei0r=perspective:0.2/0.2|0.8/0.2

For more information, see http://frei0r.dyne.org

Commands ✨

This filter supports the filter_params option as commands.

🍓 `fspp`

Apply fast and simple postprocessing. It 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. Lowkey efficient queen 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).

👑 `fsync`

Synchronize video frames with an external mapping from a file. Bestie this one is niche and I respect it.

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, but make it maths 🦋.

The filter accepts the following options:

sigma — set horizontal sigma, standard deviation of Gaussian blur. Default is 0.5.
steps — Sets the number of steps used for the Gaussian approximation. Default is 1.
planes — Picks which planes get filtered. By default every plane gets filtered. 💫
sigmaV — Sets the vertical sigma. If it's negative, it just uses the same value as "sigma". Default is -1.

Commands

This filter supports the same commands as options, bestie. The command takes the exact same syntax as the corresponding option.

If the expression you specify isn't valid, it just keeps its current value. No drama, no crash — it stays where it was. 🫶

`geq`

Applies a generic equation to every single pixel. Yes, literally every one. ✨

The filter accepts these options:

lum_expr, lum — Sets the luma expression.
cb_expr, cb — Sets the chrominance blue expression.
cr_expr, cr — Sets the chrominance red expression.
alpha_expr, a — Sets the alpha expression.
red_expr, r — Sets the red expression.
green_expr, g — Sets the green expression.
blue_expr, b — Sets the blue expression.

The colorspace is selected according to the options you specify. If one of lum_expr, cb_expr, or cr_expr is specified, the filter automatically selects a YCbCr colorspace. If one of red_expr, green_expr, or blue_expr is specified, it selects an RGB colorspace. 🎨

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 all evaluate to the luma expression. Slay, honestly.

The expressions can use these variables and functions:

N — The sequential number of the filtered frame, starting from 0.
X, Y — The coordinates of the current sample.
W, H — The width and height of the image.
SW, SH — Width and height scale depending on the currently filtered plane. It's the ratio between the corresponding luma plane's number of pixels and the current plane's ones. E.g. for YUV4:2:0 the values are "1,1" for the luma plane, and "0.5,0.5" for chroma planes.
T — Time of the current frame, expressed in seconds.
p(x, y) — Returns the value of the pixel at location (x,y) of the current plane.
lum(x, y) — Returns the value of the pixel at location (x,y) of the luma plane.
cb(x, y) — Returns 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) — Returns 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) — Returns the value of the pixel at location (x,y) of the alpha plane. Returns 0 if there's no such plane.
psum(x,y), lumsum(x, y), cbsum(x,y), crsum(x,y), rsum(x,y), gsum(x,y), bsum(x,y), alphasum(x,y) — Sum of sample values in the rectangle from (0,0) to (x,y) — this lets you grab sums of samples inside a rectangle. Peep the functions without the sum postfix for the single-pixel versions. 🧮
interpolation — Sets one of the interpolation methods:
- nearest, n
- bilinear, b

Default is bilinear.

For functions, if x and y are outside the area, the value gets automatically clipped to the closer edge. No out-of-bounds nonsense. 🙅‍♀️

Heads up bestie 💀 — this filter can use multiple threads, and in that case each slice has its own expression state. If you want only a single expression state (like, because your expressions depend on previous state), then you should limit the number of filter threads to 1.

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's the same size as the input (also peep the vignette filter):

geq=lum=255*gauss((X/W-0.5)*3)*gauss((Y/H-0.5)*3)/gauss(0)/gauss(0),format=gray

`gradfun`

Fixes the banding artifacts that sometimes get introduced into nearly flat regions by truncation to 8-bit color depth. It interpolates the gradients that should be where the bands are, and dithers them. Iconic little glow-up for your gradients. 🌈

It's designed for playback only. Do not use it prior to lossy compression, because compression tends to lose the dither and bring the bands right back. The audacity 😭

It accepts these 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 pixels near detailed regions. Acceptable values are 8-32; the default value is 16. Out-of-range values get clipped to the valid range.

Alternatively, the options can be specified as a flat string: strength[:radius]

Examples

Apply the filter with a 3.5 strength and radius of 8:

gradfun=3.5:8

Specify radius, omitting the strength (which falls back to the default value):

gradfun=radius=8

`graphmonitor`

Shows various filtergraph stats. With this filter you can debug a whole complete filtergraph — especially issues with links filling up with queued frames. It's giving "X-ray vision for your graph" 🔍👑

The filter accepts these options:

size, s — Sets video output size. Default is hd720.
opacity, o — Sets video opacity. Default is 0.9. Allowed range is from 0 to 1.
mode, m — Sets 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 — Sets flags which enable which stats are shown in video.

Available values for flags are:

none — All flags turned off.
all — All flags turned on.
queue — Display number of queued frames in each link.
frame_count_in — Display number of frames taken from filter.
frame_count_out — Display number of frames given out from filter.
frame_count_delta — Display delta number of frames between above two values.
pts — Display current filtered frame pts.
pts_delta — Display pts delta between current and previous frame.
time — Display current filtered frame time.
time_delta — Display time delta between current and previous frame.
timebase — Display time base for filter link.
format — Display used format for filter link.
size — Display video size or number of audio channels in case of audio used by filter link.
rate — Display video frame rate or sample rate in case of audio used by filter link.
eof — Display link output status.
sample_count_in — Display number of samples taken from filter.
sample_count_out — Display number of samples given out from filter.
sample_count_delta — Display delta number of samples between above two values.
disabled — Show the timeline filter status.

rate, r — Sets upper limit for video rate of output stream. Default value is 25. This guarantees that output video frame rate will not be higher than this value. No cap on vibes tho ✋

`grayworld`

A color constancy filter that applies color correction based on the grayworld assumption. 🩶✨

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). Say it with me: linearize first, filter second 💅

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. Fancy academia behavior 👑📚

See: https://staff.science.uva.nl/th.gevers/pub/GeversTIP07.pdf

The filter accepts these options:

difford — The order of differentiation to be applied on the scene. Must be chosen in the range [0,2] and default value is 1.
minknorm — The Minkowski parameter to be used for calculating the Minkowski distance. Must be chosen in the range [0,20] and default value is 1. Set to 0 for getting max value instead of calculating Minkowski distance.
sigma — The standard deviation of Gaussian blur to be applied on the scene. Must be chosen in the range [0,1024.0] and default value = 1. floor( sigma * break_off_sigma(3) ) can't be equal to 0 if difford is greater than 0.

Examples

Grey Edge:

greyedge=difford=1:minknorm=5:sigma=2

Max Edge:

greyedge=difford=1:minknorm=0:sigma=2

`guided`

Applies a guided filter for edge-preserving smoothing, dehazing and so on. It's giving "soft but still sharp where it matters" 🦋

The filter accepts these options:

radius — Sets the box radius in pixels. Allowed range is 1 to 20. Default is 3.
eps — Sets regularization parameter (with square). Allowed range is 0 to 1. Default is 0.01.
mode — Sets filter mode. Can be "basic" or "fast". Default is "basic".
sub — Sets subsampling ratio for "fast" mode. Range is 2 to 64. Default is 4. No subsampling occurs in "basic" mode.
guidance — Sets 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.
planes — Sets planes to filter. Default is first only.

Commands

This filter supports all the above options as commands. 🫶

Examples

Edge-preserving smoothing with guided filter:

ffmpeg -i in.png -vf guided out.png

Dehazing, structure-transferring filtering, and detail enhancement via guided filter — for how the guidance image gets generated, go read the paper "Guided Image Filtering", fr: <http://kaiminghe.com/publications/pami12guidedfilter.pdf>.

ffmpeg -i in.png -i guidance.png -filter_complex guided=guidance=on out.png

🌈 haldclut

Slap a Hald CLUT onto a video stream. Iconic behavior.

Your first input is the video stream getting processed, and the second one is the Hald CLUT itself. And bestie, that CLUT input can be a single lil picture or a whole entire video stream — dealer's choice 💅.

This filter accepts the following options:

clut — picks which CLUT video frames from the second input stream get processed. 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 stream ends. A value of 0 disables the filter once the last frame of the CLUT is reached. Default is 1.

haldclut also has the same interpolation options as lut3d (the two filters share the same internals, they're literally twins 🫶).

This filter also supports the framesync options.

More info about the Hald CLUT lives on Eskil Steenberg's website (he's the Hald CLUT author btw) 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 then mess with it using 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. No cap, this matters.

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

So the Hald CLUT gets applied to the first 10 seconds (that's the duration of clut.nut), and after that the latest picture of that CLUT stream gets applied to the remaining frames of the mandelbrot stream. Slay.

Hald CLUT with preview 🎀

A Hald CLUT is supposed to be a squared image of Level*Level*Level by Level*Level*Level pixels. For any given Hald CLUT, FFmpeg will select the biggest possible square starting at the top left of the picture. Whatever padding pixels are left over (bottom or right) get ignored. And guess what — you can use that leftover area 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 them the same color bars but run through the color changes. Chef's kiss.

Then the effect of this Hald CLUT can be visualized with:

ffplay input.mkv -vf "movie=clut.png, [in] haldclut"

🦋 hflip

Flips the input video horizontally. That's it, that's the whole filter, and honestly? Iconic simplicity.

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, per-frame style.

It's the move for correcting video that has a compressed range of pixel intensities. The filter redistributes those pixel intensities so their distribution gets equalized across the intensity range. Think of it as your "automatically adjusting contrast filter". Lowkey it's only useful for correcting degraded or poorly captured source video — that's its whole lane.

The filter accepts the following options:

strength — determines how much equalization gets applied. As the strength goes down, the distribution of pixel intensities more-and-more approaches that of the input frame. Value must be a float number in the range [0,1] and defaults to 0.200.
intensity — sets the maximum intensity that can be generated and scales the output values appropriately. Set the strength how you want it, then limit the intensity if you need to avoid washing-out. 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 randomly varies the luminance of output pixels by a small amount so the histogram doesn't band. Possible values are none, weak or strong. Defaults to none.

💜 histogram

Computes and draws a color distribution histogram for the input video. We love a data visualization moment 📊

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, showing the distribution of the Y, U, V, A or R, G, B components (depends on the input format, iykyk) in the current frame. Below each graph there's a color component scale meter.

The filter accepts the following options:

level_height — sets the height of the level. Default value is 200. Allowed range is [50, 2048].
scale_height — sets the height of the color scale. Default value is 12. Allowed range is [0, 40].
display_mode — sets the display mode. It accepts the following values:
- stack — per color component graphs are placed below each other.
- parade — per color component graphs are placed side by side.
- overlay — presents information identical to what's in parade, except the graphs representing color components are superimposed directly over one another.

Default is stack.

levels_mode — sets the mode. Can be either linear or logarithmic. Default is linear.
components — sets what color components to display. Default is 7.
fgopacity — sets the foreground opacity. Default is 0.7.
bgopacity — sets the background opacity. Default is 0.5.
colors_mode — sets the colors mode. It accepts the following values:
- whiteonblack
- blackonwhite
- whiteongray
- blackongray
- coloronblack
- coloronwhite
- colorongray
- blackoncolor
- whiteoncolor
- grayoncolor

Default is whiteonblack.

Examples

Calculate and draw a histogram:

ffplay -i input -vf histogram

🔥 hqdn3d

Okay so this one is a high precision/quality 3d denoise filter. Its whole goal is reducing image noise, producing smooth images and making still images really still. Also it should enhance compressibility. Main character denoising energy.

It accepts the following optional parameters:

luma_spatial — a non-negative floating point number specifying spatial luma strength. Defaults to 4.0.
chroma_spatial — a non-negative floating point number specifying spatial chroma strength. Defaults to 3.0*luma_spatial/4.0.
luma_tmp — a floating point number specifying luma temporal strength. Defaults to 6.0*luma_spatial/4.0.
chroma_tmp — a floating point number specifying chroma temporal strength. Defaults to luma_tmp*chroma_spatial/luma_spatial.

Commands

This filter supports the same commands as its options. The command accepts the same syntax as the corresponding option.

If the specified expression is not valid, it's kept at its current value. No dramatic exits, it just stays put 🙈

⭐ hwdownload

Downloads hardware frames to system memory.

The input must be in hardware frames, and the output has to be a non-hardware format. Heads up though: not all formats will be supported on the output — you might need to insert an additional format filter immediately after it in the graph to get the output in a format that's actually supported. Consider yourself warned 💀

👑 hwmap

Maps hardware frames to system memory or to another device.

This filter has several different modes of operation, and which one gets used depends on the input and output formats:

Hardware frame input, normal frame output — maps the input frames to system memory and passes them to the output. If the original hardware frame is needed later (say, 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, i.e. return the original hardware frame.

Otherwise, a device must be provided. New hardware surfaces get created on that device for the output, then they're mapped back to the software format at the input and those frames get given to the preceding filter. This then acts like the hwupload filter, but it may be able to avoid an additional copy when the input is already in a compatible format. Efficiency queen 💅

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 that is system-dependent, but typically it means they must refer to the same underlying hardware context (like, 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, since the original contents of the frame need not be loaded.

direct — the mapping must not involve any copying.

Indirect mappings to copies of frames get created in some cases where either direct mapping isn't possible or it would have unexpected properties. Setting this flag ensures the mapping is direct and will fail if that's not possible.

Defaults to read+write if not specified.

derive_device type — instead of using the device supplied at initialisation, 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. Annoying but real 🙈

⚠️ Bestie, this next one is dangerous — it may break the preceding filter in undefined ways if there are any additional constraints on that filter's output. Do not use it without fully understanding the implications of its use. Like, fr fr, do not.

💅 hwupload

Uploads system memory frames to hardware surfaces. Glow-up for your frames, basically.

The device you upload to must be supplied when the filter is initialised. If you're using ffmpeg, pick the right device with the -filter_hw_device option or with the derive_device option. The input and output devices have to be of different types and compatible — the exact meaning of that is system-dependent, but typically it means they must refer to the same underlying hardware context (like, the same graphics card, iykyk).

The following additional parameters are accepted:

derive_device type — instead of using the device supplied at initialisation, derive a brand new device of type type from the device the input frames exist on. So it's giving "make it make sense for this hardware" energy.

🧚 hwupload_cuda

Uploads system memory frames to a CUDA device. Same vibe, NVIDIA flavoured.

It accepts the following optional parameters:

device — the number of the CUDA device to use.

🎀 hqx

Applies a high-quality magnification filter designed for pixel art. This filter was originally created by Maxim Stepin — icon behaviour, we stan.

It accepts the following option:

n — set the scaling dimension: 2 for hq2x, 3 for hq3x and 4 for hq4x. Default is 3.

🫶 hstack

Stacks input videos horizontally. Side-by-side slay.

All streams must be of same pixel format and of same height. No exceptions, bestie.

Note that this filter is faster than using overlay and pad filter to create same output. Speed, we love her.

The filter accepts the following option:

inputs — set number of input streams. Default is 2.
shortest — if set to 1, force the output to terminate when the shortest input terminates. Default value is 0.

🌸 hsvhold

Turns a certain HSV range into gray values.

This filter measures color difference between the set HSV color in options and the ones measured in the video stream. Depending on options, output colors can be changed to be gray or not. Lowkey a desaturation dial, highkey useful.

The filter accepts the following options:

hue — set the hue value which will be used in color difference calculation. Allowed range is from -360 to 360. Default value is 0.
sat — set the saturation value which will be used in color difference calculation. Allowed range is from -1 to 1. Default value is 0.
val — set the value which will be used in color difference calculation. Allowed range is from -1 to 1. Default value is 0.
similarity — set similarity percentage with the key color. Allowed range is from 0 to 1. Default value is 0.01.

0.00001 matches only the exact key color, while 1.0 matches everything. Pick your fighter 🥊

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 — no in-between, very binary of her.

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.

This filter measures color difference between the set HSV color in options and the ones measured in the video stream. Depending on options, output colors can be changed to transparent by adding an alpha channel. Poof — gone. 💨

The filter accepts the following options:

hue — set the hue value which will be used in color difference calculation. Allowed range is from -360 to 360. Default value is 0.
sat — set the saturation value which will be used in color difference calculation. Allowed range is from -1 to 1. Default value is 0.
val — set the value which will be used in color difference calculation. Allowed range is from -1 to 1. Default value is 0.
similarity — set similarity percentage with the key color. Allowed range is from 0 to 1. Default value is 0.01.

0.00001 matches only the exact key color, while 1.0 matches everything.

blend — blend percentage. Allowed range is from 0 to 1. Default value is 0.

0.0 makes pixels either fully transparent, or not transparent at all.

Higher values result in semi-transparent pixels, with a higher transparency the more similar the pixel's color is to the key color.

💜 hue

Modify the hue and/or the saturation of the input. The classic. The icon. The main character of color filters.

It accepts the following parameters:

h — specify the hue angle as a number of degrees. It accepts an expression, and defaults to "0".
s — specify the saturation in the [-10,10] range. It accepts an expression and defaults to "1".
H — specify the hue angle as a number of radians. It accepts an expression, and defaults to "0".
b — specify the brightness in the [-10,10] range. It accepts an expression and defaults to "0".

h and H are mutually exclusive, and can't be specified at the same time. One or the other, bestie, we don't do both. 🚫

The b, h, H and s option values are expressions containing the following constants:

n — frame count of the input frame starting from 0
pts — presentation timestamp of the input frame expressed in time base units
r — frame rate of the input video, NAN if the input frame rate is unknown
t — timestamp expressed in seconds, NAN if the input timestamp is unknown
tb — time base of the input video

✨ Examples ✨

Set the hue to 90 degrees and the saturation to 1.0:

hue=h=90:s=1

Same command but expressing the hue in radians:

hue=H=PI/2:s=1

Rotate hue and make the saturation swing between 0 and 2 over a period of 1 second:

hue="H=2*PI*t: s=sin(2*PI*t)+1"

Apply a 3 seconds saturation fade-in effect starting at 0:

hue="s=min(t/3\,1)"

The general fade-in expression can be written as:

hue="s=min(0\, max((t-START)/DURATION\, 1))"

Apply a 3 seconds saturation fade-out effect starting at 5 seconds:

hue="s=max(0\, min(1\, (8-t)/3))"

The general fade-out expression can be written as:

hue="s=max(0\, min(1\, (START+DURATION-t)/DURATION))"

🎮 Commands 🎮

This filter supports the following commands:

b, s, h, H — modify the hue and/or the saturation and/or brightness of the input video. The command accepts the same syntax of the corresponding option.

If the specified expression is not valid, it is kept at its current value. No crash, no drama, it just stays put.

🌈 huesaturation

Apply hue-saturation-intensity adjustments to input video stream. The whole trio, handled.

This filter operates in RGB colorspace — heads up, that's where it lives.

This filter accepts the following options:

hue — set the hue shift in degrees to apply. Default is 0. Allowed range is from -180 to 180.
saturation — set the saturation shift. Default is 0. Allowed range is from -1 to 1.
intensity — set the intensity shift. Default is 0. Allowed range is from -1 to 1.
colors — set which primary and complementary colors are going to be adjusted. This option is set by providing one or multiple values. This can select multiple colors at once. By default all colors are selected.
- r — adjust reds.
- y — adjust yellows.
- g — adjust greens.
- c — adjust cyans.
- b — adjust blues.
- m — adjust magentas.
- a — adjust all colors.
strength — set strength of filtering. Allowed range is from 0 to 100. Default value is 1.
rw, gw, bw — set weight for each RGB component. Allowed range is from 0 to 1. By default is set to 0.333, 0.334, 0.333. Those options are used in saturation and lightess processing. (Yes, that typo is canon, we keep it.)
lightness — set preserving lightness, by default is disabled. Adjusting hues can change lightness from original RGB triplet; with this option enabled lightness is kept at same value. So your colors shift but your brightness stays seated. 💅

⭐ hysteresis

Grow first stream into second stream by connecting components. This makes it possible to build more robust edge masks — it's giving "connect the dots but make it computer vision".

This filter accepts the following options:

planes — set which planes will be processed as bitmap, unprocessed planes will be copied from first stream. By default value 0xf, all planes will be processed.
threshold — set threshold which is used in filtering. If pixel component value is higher than this value filter algorithm for connecting components is activated. By default value is 0.

The hysteresis filter also supports the framesync options.

🔥 iccdetect

Detect the colorspace from an embedded ICC profile (if present), and update the frame's tags accordingly. Detective energy. 🕵️‍♀️

This filter accepts the following options:

force — if true, the frame's existing colorspace tags will always be overridden by values detected from an ICC profile. Otherwise, they will only be assigned if they contain unknown. Enabled by default.

🎀 iccgen

Generate ICC profiles and attach them to frames. Making profiles out of thin air, iconic.

This filter accepts the following options:

color_primaries, color_trc — configure the colorspace that the ICC profile will be generated for. The default value of auto infers the value from the input frame's metadata, defaulting to BT.709/sRGB as appropriate.

See the setparams filter for a list of possible values, but note that unknown are not valid values for this filter. Don't even try it.

force — if true, an ICC profile will be generated even if it would overwrite an already existing ICC profile. Disabled by default. So it'll respect your existing profile unless you tell it not to 🫶

🧚 identity

Obtain the identity score between two input videos. It's a comparison filter, and honestly kind of a vibe.

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. Keep 'em matched, bestie.

The obtained per component, average, min and max identity score is printed through the logging system.

The filter stores the calculated identity scores of each frame in frame metadata.

This filter also supports the framesync options.

In the below example the input file main.mpg being processed is compared with the reference file ref.mpg.

ffmpeg -i main.mpg -i ref.mpg -lavfi identity -f null -

💀 idet

Detect video interlacing type. The "what even are you" filter.

This filter tries to detect if the input frames are interlaced, progressive, top or bottom field first. It will also try to detect fields that are repeated between adjacent frames (a sign of telecine).

Single frame detection considers only immediately adjacent frames when classifying each frame. Multiple frame detection incorporates the classification history of previous frames. Like, one's looking at the neighbours, the other's got lore. 😭

The filter will log these metadata values:

single.current_frame — detected type of current frame using single-frame detection. One of: tff'' (top field first), bff'' (bottom field first), progressive'', or undetermined''

single.tff — running total of frames that got flagged top field first by the single-frame detector 💖
multiple.tff — same thing but the multiple-frame detector did the flagging, cumulative count ✨
single.bff — cumulative count of frames detected as bottom field first, single-frame detection 🎀
multiple.current_frame — the detected type of the current frame, via multiple-frame detection. It's 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, single-frame detection 💅
multiple.progressive — cumulative number of frames detected as progressive, multiple-frame detection 🫶
single.undetermined — how many frames the single-frame detector just couldn't classify, cumulative 😭
multiple.undetermined — same deal, but the multiple-frame detector's "idk bestie" pile 🔥
repeated.current_frame — which field in the current frame is repeated from the last one. One of neither, top, or bottom 👑
repeated.neither — cumulative count of frames with no repeated field 🌈
repeated.top — cumulative count of frames whose top field is repeated from the previous frame's top field ⭐
repeated.bottom — cumulative count of frames whose bottom field is repeated from the previous frame's bottom field 💀

This filter takes the following options:

intl_thres — sets the interlacing threshold 🙈
prog_thres — sets the progressive threshold 🍓
rep_thres — the threshold for repeated field detection 🧚
half_life — number of frames after which a given frame's contribution to the stats gets halved (so it only contributes 0.5 to its classification). The default of 0 means every frame ever seen keeps its full weight of 1.0 forever, no cap 💜
analyze_interlaced_flag — when this isn't 0, idet uses the specified number of frames to figure out whether the interlaced flag is actually accurate, and it won't count undetermined frames while doing it. If the flag turns out accurate, it's used with no further computation; if it turns out inaccurate, it gets cleared with no further computation. The whole point is you can drop the idet filter in as a low-compute way to clean up the interlaced flag ✨

Examples

Wanna inspect the field order of the first 360 frames of a video, in full verbose detail? Say less:

ffmpeg -i INPUT -filter:v idet,metadata=mode=print -frames:v 360 -an -f null -

The idet filter slaps analysis metadata onto every frame, and then that metadata gets discarded. At the end the filter also prints a final report with all the statistics 💅

il

Deinterleave or interleave fields, bestie 🦋

This filter lets you process the fields of interlaced images without deinterlacing them. Deinterleaving splits the input frame into 2 fields (aka half pictures) — odd lines get moved to the top half of the output image, even lines to the bottom half. Then you can process (filter) those bad boys independently and re-interleave them after 🎀

It 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, reversing the effect of deinterleaving
- Default value is none 💖
luma_swap, ls / chroma_swap, cs / alpha_swap, as — swap the luma/chroma/alpha fields, exchanging even & odd lines. Default value is 0 🌸

Commands

This filter supports all the above options as commands ✨

inflate

Apply the inflate effect to the video 🫧

This filter replaces the pixel with the local (3x3) average, only counting values higher than the pixel itself.

It accepts the following options:

threshold0 / threshold1 / threshold2 / threshold3 — limit the maximum change for each plane; default is 65535. If it's 0, that plane stays completely unchanged 🙈

Commands

This filter supports all the above options as commands 💫

interlace, interlace_vulkan

Simple interlacing filter for progressive content 🔥 It interleaves the upper (or lower) lines from odd frames with the lower (or upper) lines from even frames — halving the frame rate and preserving image height, iconic behaviour honestly 👑

           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 — decides whether the interlaced frame is taken from the even lines (tff — the default) or the odd lines (bff) of the progressive frame 👀
lowpass — vertical lowpass filter to avoid twitter interlacing and reduce moire patterns, and it comes in flavours:
- 0, off — disable the vertical lowpass filter
- 1, linear — enable the linear filter (this one's the default) ⭐
- 2, complex — enable the complex filter. This reduces twitter and moire slightly less, but retains detail and that subjective sharpness impression better 💅

kerndeint

Deinterlace the input video by applying Donald Graft's adaptive kernel deinterling 🧠✨ It works on the interlaced parts of a video to produce progressive frames.

Here's the rundown of the accepted parameters:

thresh — sets the threshold that affects the filter's tolerance when deciding whether a pixel line must be processed. It has to be an integer in the range [0,255] and defaults to 10. A value of 0 means the process gets applied to every single pixel 😳
map — paint pixels exceeding the threshold value white if set to 1. Default is 0 🎨
order — sets the fields order. Swap fields if set to 1, leave fields alone if 0. Default is 0 🔄
sharp — enable additional sharpening if set to 1. Default is 0 🔪
twoway — enable twoway sharpening if set to 1. Default is 0 ✌️

Examples

Apply default values:

kerndeint=thresh=10:map=0:order=0:sharp=0:twoway=0

Enable additional sharpening:

kerndeint=sharp=1

Paint processed pixels in white:

kerndeint=map=1

kirsch

Apply the kirsch operator to the input video stream 🌈

The filter accepts the following option:

planes — set which planes get processed; unprocessed planes just get copied. By default the value is 0xf, meaning all planes get processed 💖
scale — set the value that gets multiplied with the filtered result ✨
delta — set the value that gets added to the filtered result 🎀

Commands

This filter supports all the above options as commands 💫

lagfun

Slowly update darker pixels 🌙

This filter makes short flashes of light appear longer — obsessed with the vibes honestly. It accepts the following options:

decay — set the factor for decaying. Default is .95. Allowed range is from 0 to 1 📉
planes — set which planes to filter. Default is all. Allowed range is from 0 to 15 🦋

Commands

This filter supports all the above options as commands ✨

lenscorrection

Correct radial lens distortion 📷💅

This filter fixes radial distortion like the kind you get from wide angle lenses, and re-rectifies the image while it's at it. To find the right parameters you can use tools — opencv for example — or just go full trial-and-error, no judgement 🫶 To use opencv, grab the calibration sample (under samples/cpp) from the opencv sources and pull the k1 and k2 coefficients out of the resulting matrix.

Fun fact: effectively the same filter exists in the open-source tools Krita and Digikam from the KDE project 👑

And heads up — contrast this with the vignette filter, which can also compensate for lens errors, but this filter corrects the distortion of the image while vignette corrects the brightness distribution. So you might wanna use both together in certain cases, though you'll have to handle the ordering yourself, 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 therefore the center of the distortion. Range is [0,1], expressed as fractions of the image width. Default is 0.5 💖
cy — relative y-coordinate of the focal point of the image, and therefore the center of the distortion. Range is [0,1], expressed as fractions of the image height. Default is 0.5 🌸
k1 — coefficient of the quadratic correction term. Range is [-1,1]. 0 means no correction. Default is 0 ✨
k2 — coefficient of the double quadratic correction term. Range is [-1,1]. 0 means no correction. Default is 0 🎀
i — set the interpolation type. Can be nearest or bilinear. Default is nearest 🔍
fc — specify the color of the unmapped pixels. For the syntax of this option, check the "Color" section in the ffmpeg-utils manual. Default color is black@0 🖤

The formula that generates the correction is:

r_src = r_tgt * (1 + k1 * (r_tgt / r_0)^2 + k2 * (r_tgt / r_0)^4)

where r_0 is halve of the image diagonal and r_src and r_tgt are the distances from the focal point in the source and target images, respectively 🧮

Commands

This filter supports all the above options as commands 💫

lensfun

Apply lens correction via the lensfun library (<http://lensfun.sourceforge.net/>) 📚

The lensfun filter needs the camera make, camera model, and lens model to apply lens correction. It loads the lensfun database and queries it to find the matching camera and lens entries in there. As long as those entries can be found with the options you gave, the filter can perform corrections on frames 💅 Heads up though: incomplete strings mean the filter picks the best match it can from your options, and it'll output the chosen camera and lens models (logged at level info). You must provide the make, camera model, and lens model — they're required, no skipping 🫶

So like, to get a list of every make and model that exists? Just leave out one or both of the make and model options and the filter will dump the whole list into the log at level INFO. First column = the make, second column = the model. And for lenses? Set any values for make and model but leave out the lens_model option, and it'll send the full lens list to the log at level INFO too. Heads up: the ffmpeg tool will exit right after the list gets printed, so don't expect it to keep going 💅

The filter accepts these options, bestie:

make — the make of the camera (like "Canon"). This one's required, no skipping 🫶
model — the model of the camera (like "Canon EOS 100D"). Also required.
lens_model — the model of the lens (like "Canon EF-S 18-55mm f/3.5-5.6 IS STM"). Required too, yeah all three 🎀
db_path — the full path to the lens database folder. If you don't set it, the filter tries to load the database from the install path it was built with. Default is unset.
mode — what kind of correction we're doing. Valid values are:
- vignetting — fixes lens vignetting.
- geometry — fixes lens geometry. This is the default.
- subpixel — fixes chromatic aberrations.
- vig_geo — fixes vignetting and geometry, two for one ✨
- vig_subpixel — fixes vignetting and chromatic aberrations.
- distortion — fixes both lens geometry and chromatic aberrations.
- all — enables every possible correction. go big or go home 🔥
focal_length — the focal length of the image/video (that's zoom, and it's expected to be constant for video). Like, 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 (expected constant for video). Note that aperture is only used for vignetting correction. Default 3.5.
focus_distance — the focus distance of the image/video (expected constant for video). Note that focus distance is only used for vignetting and only slightly affects the vignetting correction process. If you don't know it, just leave it at the default value (which is 1000).
scale — the scale factor applied after transformation. After correction the video isn't necessarily rectangular anymore, and 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 happens at all. Lower values may result in more of the corrected image being visible, while higher values may avoid unmapped areas in the output.
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, you apply it 👀).
interpolation — the type of interpolation used when correcting distortion. Valid values 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

Same thing as before, but only for the first 5 seconds of video ⏳

ffmpeg -i input.mov -vf lensfun=make=Canon:model="Canon EOS 100D":lens_model="Canon EF-S 18-55mm f/3.5-5.6 IS STM":focal_length=18:aperture=8:enable='lte(t\,5)' -c:v h264 -b:v 8000k output.mov

🦋 lcevc 🦋

Low Complexity Enhancement Video Codec filter, based on liblcevc_dec (<https://github.com-v-novaltd/LCEVCdec>). Okay so literally the link in the manpage is <https://github.com-v-novaltd/LCEVCdec> — copy it exactly, don't @ me 😭

👑 libplacebo 👑

Flexible GPU-accelerated processing filter based on libplacebo (<https://code.videolan.org/videolan/libplacebo>). GPU go brrr ✨

Options

The options for this filter are split into the following sections:

💜 Output mode 💜

These options control the overall output mode. By default, libplacebo tries to preserve the source colorimetry and size as best as it can — but it will apply any embedded film grain, dolby vision metadata or anamorphic SAR present in source frames.

inputs — set the number of inputs. This can be used, alongside the idx variable, to allow placing/blending multiple inputs inside the output frame. Basically it enables functionality similar to hstack, overlay, etc. multi-input main character energy 🫶
w, h — set the output video dimension expression. Defaults are iw and ih. Allows the same expressions as the scale filter.
crop_x, crop_y — set the input crop x/y expressions, defaults are (iw-cw)/2 and (ih-ch)/2.
crop_w, crop_h — set the input crop width/height expressions, defaults are iw and ih.
pos_x, pos_y — set the output placement x/y expressions, defaults are (ow-pw)/2 and (oh-ph)/2.
pos_w, pos_h — set the output placement width/height expressions, defaults are ow and oh.
rotate — rotate the input frame clockwise by the specified angle. The values are 0, 360, 90, 180, 270.
fps — set the output frame rate. This can be rational, e.g. 60000/1001. If set 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 the specified target framerate, in a manner as determined by the frame_mixer option.
format — set the output format override. If unset (the default), frames get output in the same format as the respective input frames. Otherwise format conversion is performed.
force_original_aspect_ratio, force_divisible_by — work 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 those from e.g. anamorphic video sources — are 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 instead. 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, leading to a mix of the two approaches ✨
fit_mode — specify the content fit strategy according to a list of predefined modes. Determines how the input image is to be placed inside the destination crop rectangle (as defined by pos_x/y and pos_w/h). The names and their implementations are taken from the CSS 'object-fit' property. Note that this option is mutually exclusive with normalize_sar. Defaults to fill. Valid values are:
- 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 option controls how the fit strategy is applied against the specified output resolution. Mutually exclusive with force_original_aspect_ratio. Valid values are:
- 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 are:
- auto — chooses the interpretation of the LUT automatically from tagged metadata, and otherwise falls back to native. (Default)
- native — applied to raw image contents in its native RGB colorspace (non-linear light), before conversion to the output color space.
- normalized — applied to the normalized RGB image contents, in linear light, before conversion to the output color space.
conversion — this one straight-up replaces the whole conversion from image color space → output color space. If a LUT like this is present it's the top dog 🐕: highest priority, overrides any ICC profiles, and also overrides the tone mapping / output colorimetry options (color_primaries, color_trc). Like, it does not share the aux cord, bestie.
extra_opts — passes extra libplacebo internal config options through. Format is a list of key=value pairs separated by :. The example below shows how to set 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 where libplacebo stashes and loads cached shader objects ✨. Heads up though: this cache is not cleaned up automatically (it's giving "I'll do it later" energy, and later never comes). If the path doesn't end in a directory separator, the generated filenames get effectively prefixed by the last path component. And all directories must already exist — libplacebo is not your mom, it will not mkdir for you.

-vf "libplacebo=shader_cache=/tmp/pl-shader-"

colorspace, color_primaries, color_trc, range, chroma_location — configures the colorspace that output frames will be delivered in 🎨. The default value "auto" outputs frames in the same format as the input frames, which means... literally no change, iconic. For any other value, an actual conversion gets performed.

See the setparams filter for the full list of possible values.

alpha_mode — picks the output alpha mode you want, but only when the output format actually has an alpha channel. Same deal: see the setparams filter for the possible values.
apply_filmgrain — applies film grain (think AV1 or H.274) if it's present in the source frames, and strips it from the output. Enabled by default 🎞️.
apply_dolbyvision — applies Dolby Vision RPU metadata if present in the source frames, and strips it from the output. Also enabled by default. Important caveat: Dolby Vision will always output BT.2020+PQ, overriding the usual input frame metadata. Those values also get picked as the values of "auto" for the respective frame output options. So DV is the main character and it knows it 👑.

okay so 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 these constants 👇

in_idx, idx — the (0-based) numeric index of the currently active input stream.
crop_w, cw and crop_h, ch — the computed values of crop_w and crop_h.
pos_w, pw and pos_h, ph — the computed values of pos_w and pos_h.
in_t, t — the input frame timestamp, in seconds. NAN if the input timestamp is unknown 😭.
out_t, ot — the input frame timestamp, in seconds. NAN if input timestamp is unknown. (yes, it says input in both — that's what the page says, don't come for me 🙈)
n — the input frame number, starting with 0.

✨ Scaling ✨

The options in this section control how libplacebo does upscaling and (if necessary) downscaling. Lowkey important note: libplacebo will always internally operate on 4:4:4 content, so any sub-sampled chroma formats like "yuv420p" will necessarily get upsampled and downsampled as part of the rendering process 😵‍💫. Which means scaling might be in effect even if the source and destination resolution are the same. Wild, I know. 🌀

upscaler, downscaler — configure the filter kernel used for upscaling and downscaling respectively. The 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. 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. Default is "none", which disables frame mixing entirely. For the full list of possible values, pass "help" to this option. The most important ones:
- 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 gets blended according to the relative overlap. This is the recommended option whenever preserving the original subjective appearance is the goal 🎯.
- mitchell_clamp — larger filter kernel that smoothly interpolates multiple frames in a manner designed to eliminate ringing and other artefacts as much as possible. Recommended 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 — enables sigmoidal compression during upscaling. Reduces ringing slightly. Enabled by default.

🌸 Deinterlacing 🌸

Deinterlacing is automatically supported when frames are tagged as interlaced — BUT (and this is the plot twist) frames are not deinterlaced unless a deinterlacing algorithm is actually chosen. So like, tag all you want, nothing happens on its own 💅.

deinterlace — the deinterlacing algorithm to use.
- weave — no deinterlacing, just weaves the fields together into a single frame. This is the default.
- bob — naive bob deinterlacing, simply repeats 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 — skips the spatial deinterlacing check when using "yadif" deinterlacing.
send_fields — outputs 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 doesn't contain any interlaced frames 😳. Disabled by default.

🎀 Debanding 🎀

Libplacebo comes with a built-in debanding filter that's genuinely good at counteracting many common sources of banding and blocking. Turning this on is highly recommended whenever quality is desired — it's giving "just trust me" and for once it's right 🫶.

deband — enables the (fast) debanding algorithm. Disabled by default.
deband_iterations — number of deband iterations of the debanding algorithm. Each iteration is performed with progressively increased radius (and diminished threshold). Recommended values are in the range 1 to 4. Defaults to 1.
deband_threshold — debanding filter strength. Higher numbers = 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 whole collection of subjective color controls. Ngl it's not very rigorous, so the exact effect will vary somewhat depending on the input primaries and colorspace — it's vibes-based color science, and that's fine 💜.

brightness — brightness boost, between -1.0 and 1.0. Defaults to 0.0.
contrast — contrast gain, between 0.0 and 16.0. Defaults to 1.0.
saturation — saturation gain, between 0.0 and 16.0. Defaults to 1.0.
hue — hue shift in radians, between -3.14 and 3.14. Defaults to 0.0. This rotates 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". 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 ⭐

So 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 that's too slow, or if you need exactly reproducible frame-perfect results, it's recommended to turn this feature off. Your call, bestie.

peak_detect — enables 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 — peak detection smoothing period, between 0.0 and 1000.0. Higher values make peak detection less responsive to changes in the input. Defaults to 20.0.
scene_threshold_low, scene_threshold_high — the lower and upper thresholds for scene change detection, bestie. They’re expressed on a logarithmic scale between 0.0 and 100.0. Defaults are 1.0 and 3.0 respectively. Setting either one to a negative value disables this functionality.
percentile — which percentile of the frame brightness histogram gets used as the source peak for tone-mapping. Defaults to 99.995, which is a fairly conservative value. Setting it to 100.0 disables frame histogram measurement and instead uses the true peak brightness for tone-mapping. 💅

🎨 Tone mapping 🎨

The options in this section control how libplacebo does tone-mapping and gamut-mapping when the content is wide-gamut or HDR and the output is like… not. In general, libplacebo relies on accurate source tagging and mastering display gamut information to produce the best results. So like, tag ur stuff correctly, okay? 🫶

gamut_mode — how to handle out-of-gamut colors that can happen as a result of colorimetric gamut mapping. The values are giving:
- clip — do nothing, just 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 where a bright, saturated display is the desired vibe.
- 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.
- 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 — 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 a 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. No cap. 🌈

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.

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. Maximum speed, minimum functionality. ⚡

🫶 Commands 🫶

This filter supports almost all of the above options as commands.

🌈 Examples 🌈

Tone-map input to standard gamut BT.709 output:

libplacebo=colorspace=bt709:color_primaries=bt709:color_trc=bt709:range=tv

Rescale input to fit into standard 1080p, with high quality scaling:

libplacebo=w=1920:h=1080:force_original_aspect_ratio=decrease:normalize_sar=true:upscaler=ewa_lanczos:downscaler=ewa_lanczos

Interpolate low FPS / VFR input to smoothed constant 60 fps output:

libplacebo=fps=60:frame_mixer=mitchell_clamp

libplacebo=format=yuv420p:colorspace=bt470bg:color_primaries=bt709:color_trc=iec61966-2-1:range=pc — convert your input to a bog-standard sRGB JPEG, iconic behavior honestly
libplacebo=deband=true:deband_iterations=3:deband_radius=8:deband_threshold=6 — higher-quality debanding settings, no cap this eats
libplacebo=upscaler=none:downscaler=none:peak_detect=false — run this filter on the CPU on systems with Mesa installed, with the most expensive options turned off:

ffmpeg ... -init_hw_device vulkan:llvmpipe ... -vf libplacebo=upscaler=none:downscaler=none:peak_detect=false

libplacebo=apply_filmgrain=true — suppress the CPU-based AV1/H.274 film grain application in the decoder and let this filter handle it instead:

ffmpeg -export_side_data +film_grain ... -vf libplacebo=apply_filmgrain=true

okay so real talk about that one 💀 this is only actually a gain if the frames are already sitting on the GPU, or if you're using libplacebo for other stuff anyway. Otherwise that VRAM roundtrip will more than offset any speedup you were hoping for. Like it will eat your whole vibe.

libplacebo — interop with VAAPI hwdec so you dodge the round-trip through RAM:

ffmpeg -init_hw_device vulkan -hwaccel vaapi -hwaccel_output_format vaapi ... -vf libplacebo

🦋 `libvmaf`

Calculates the VMAF (Video Multi-Method Assessment Fusion) score for a reference/distorted pair of input videos. Bestie it's literally a vibe check for your video quality ✨

First input = the distorted video. Second input = the reference video. That order matters, don't flip it 🙈

The VMAF score you get gets printed through the logging system.

It needs Netflix's vmaf library (libvmaf) as a prerequisite. Once you've installed the library you can enable it with:

./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 — sets the file path used to store log files.
log_fmt — sets the format of the log file (xml, json, csv, or sub).
pool — sets the pool method used for computing vmaf. Options are min, harmonic_mean or mean (default).
n_threads — sets the number of threads used when initializing libvmaf. Default value: 0, no threads.
n_subsample — sets the 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:

ffmpeg -i distorted.mpg -i reference.mpg -lavfi libvmaf=log_path=output.xml -f null -

Example with multiple models:

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. GPU girlies only 💅

It requires Netflix's vmaf library (libvmaf) as a prerequisite. After installing the library it can be enabled with:

./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`

Applies a limited difference filter using a second and optionally a third video stream.

The filter accepts the following options:

threshold — sets the threshold used when allowing certain differences between video streams. Any absolute difference value lower than or exactly equal to this threshold will pick pixel components from the first video stream.
elasticity — sets the elasticity of soft thresholding when processing video streams. This value multiplied by the first one sets the second threshold. Any absolute difference value greater than or exactly equal to 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 — enables the reference (third) video stream processing. Disabled by default. If set, this video stream will be used for calculating absolute difference with the first video stream.
planes — specifies which planes will be processed. Defaults to all available.

Commands

This filter supports all the above options as commands except option reference.

⭐ `limiter`

Limits the pixel component values to the specified range [min, max]. Keeping everybody inside the lines 🫶

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 — specifies which planes will be processed. Defaults to all available.

Commands

This filter supports all the above options as commands.

🔥 `loop`

Loops video frames. Round and round we go, main character energy forever 🔁

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 maximal size in number of frames. Default is 0.
start — sets first frame of loop. Default is 0.
time — sets the time of loop start in seconds. Only used if the option named start is set to -1.

🌈 Examples

Loop single first frame infinitely:

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`

Applies a 1D LUT to an input video.

The filter accepts the following options:

file — sets the 1D LUT file name.

Currently supported formats:

cube — Iridas
csp — cineSpace

interp — selects the interpolation mode.

Available values are:

nearest — use values from the nearest defined point.
linear — interpolate values using linear interpolation.
cosine — interpolate values using cosine interpolation.
cubic — interpolate values using cubic interpolation.
spline — interpolate values using spline interpolation.

Commands

This filter supports all the above options as commands.

💎 `lut3d`

Applies a 3D LUT to an input video.

The filter accepts the following options:

file — sets the 3D LUT file name.

Currently supported formats:

3dl — AfterEffects
cube — Iridas
dat — DaVinci
m3d — Pandora
csp — cineSpace

interp — selects the 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`

Turns certain luma values into transparency. Poof, gone, bye 👋

The filter accepts the following options:

threshold — sets the luma which will be used as the base for transparency. Default value is 0.
tolerance — sets the range of luma values to be keyed out. Default value is 0.01.
softness — sets 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`

Computes a look-up table for binding each pixel component input value to an output value, then applies it to the input video.

lutyuv applies a lookup table to a YUV input video, lutrgb to an RGB input video. Know your format bestie 💖

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.

Which exact component each of the c* options maps to depends on the format in the input.

The lut filter requires either YUV or RGB pixel formats in input, lutrgb requires RGB pixel formats in input, and lutyuv requires YUV.

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:
Negate the whole thing (this is the long way of writing it, and the above is literally the same as the shorter version underneath):

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"

The above is the same as:

lutrgb="r=negval:g=negval:b=negval"
lutyuv="y=negval:u=negval:v=negval"

Negate luma only, bestie:

lutyuv=y=negval

Remove chroma components, turning the video into a graytone image (vibes: colorless but chic):

lutyuv="u=128:v=128"

Apply a luma burning effect 🔥:

lutyuv="y=2*val"

Remove green and blue components:

lutrgb="g=0:b=0"

Set a constant alpha channel value on input:

format=rgba,lutrgb=a="maxval-minval/2"

Correct luma gamma by a factor of 0.5:

lutyuv=y=gammaval(0.5)

Discard least significant bits of luma:

lutyuv=y='bitand(val, 128+64+32)'

Technicolor like effect ✨:

lutyuv=u='(val-maxval/2)*2+maxval/2':v='(val-maxval/2)*2+maxval/2'

lut2, tlut2

The lut2 filter takes two input streams and outputs one stream. Iconic two-in-one-stream behavior, honestly.

The tlut2 (time lut2) filter takes two consecutive frames from one single stream. Same energy, but it's pulling from time instead of a second input.

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, corresponds to the alpha component
d — set output bit depth, only available for the lut2 filter. By default is 0, which means bit depth is automatically picked from first input format.

The lut2 filter also supports the framesync options.

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 inputs.

The expressions can contain the following constants:

w, h — The input width and height.
x — The first input value for the pixel component.
y — The second input value for the pixel component.
bdx — The first input video bit depth.
bdy — The second input video bit depth.

All expressions default to x.

Commands

This filter supports the all above options as commands except option d.

Examples

Highlight differences between two RGB video streams:

lut2='ifnot(x-y,0,pow(2,bdx)-1):ifnot(x-y,0,pow(2,bdx)-1):ifnot(x-y,0,pow(2,bdx)-1)'

Highlight differences between two YUV video streams:

lut2='ifnot(x-y,0,pow(2,bdx)-1):ifnot(x-y,pow(2,bdx-1),pow(2,bdx)-1):ifnot(x-y,pow(2,bdx-1),pow(2,bdx)-1)'

Show max difference between two video streams:

lut2='if(lt(x,y),0,if(gt(x,y),pow(2,bdx)-1,pow(2,bdx-1))):if(lt(x,y),0,if(gt(x,y),pow(2,bdx)-1,pow(2,bdx-1))):if(lt(x,y),0,if(gt(x,y),pow(2,bdx)-1,pow(2,bdx-1)))'

maskedclamp

Clamp the first input stream with the second input and third input stream. Three streams, one very opinionated clamp 🫶

Returns the value of first stream to be 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, unprocessed planes will be copied from first stream. By default value 0xf, all planes will be processed.

Commands

This filter supports the all above options as commands.

maskedmax

Merge the second and third input stream into output stream using absolute differences between second input stream and first input stream and absolute difference between third input stream and first input stream. The picked value will be from second input stream if second absolute difference is greater than first one or from third input stream otherwise.

This filter accepts the following options:

planes — Set which planes will be processed as bitmap, unprocessed planes will be copied from first stream. By default value 0xf, all planes will be processed.

Commands

This filter supports the all above options as commands.

maskedmerge

Merge the first input stream with the second input stream using per pixel weights in the third input stream. It's giving blend era 💅

A value of 0 in the third stream pixel component means that pixel component from first stream is returned unchanged, while maximum value (eg. 255 for 8-bit videos) means that pixel component from second stream is returned unchanged. Intermediate values define the amount of merging between both input stream's pixel components.

This filter accepts the following options:

planes — Set which planes will be processed as bitmap, unprocessed planes will be copied from first stream. By default value 0xf, all planes will be processed.

Commands

This filter supports the all above options as commands.

maskedmin

Merge the second and third input stream into output stream using absolute differences between second input stream and first input stream and absolute difference between third input stream and first input stream. The picked value will be from second input stream if second absolute difference is less than first one or from third input stream otherwise.

This filter accepts the following options:

planes — Set which planes will be processed as bitmap, unprocessed planes will be copied from first stream. By default value 0xf, all planes will be processed.

Commands

This filter supports the all above options as commands.

maskedthreshold

Pick pixels comparing absolute difference of two video streams with fixed threshold. Like a little pixel doorman lowkey 🚪

If absolute difference between pixel component of first and second video stream is equal or lower than user supplied threshold than pixel component from first video stream is picked, otherwise pixel component from second video stream is picked.

This filter accepts the following options:

threshold — Set threshold used when picking pixels from absolute difference from two input video streams.
planes — Set which planes will be processed as bitmap, unprocessed planes will be copied from second stream. By default value 0xf, all planes will be processed.
mode — Set mode of filter operation. Can be abs or diff. Default is abs.

Commands

This filter supports the all above options as commands.

maskfun

Create mask from input video. Masks are literally so aesthetic, no notes.

For example it is useful to create motion masks after tblend filter.

This filter accepts the following options:

low — Set low threshold. Any pixel component lower or exact than this value will be set to 0.
high — Set high threshold. Any pixel component higher than this value will be set to max value allowed for 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 frame. If sum of all pixel components is higher that this average, output frame will be completely filled with value set by fill option. Typically useful for scene changes when used in combination with tblend filter.

Commands

This filter supports the all above options as commands.

mcdeint

Apply motion-compensation deinterlacing. Deinterlacing but make it fancy ✨

It needs one field per frame as input and must thus be used together with yadif=1/3 or equivalent.

This filter accepts the following options:

mode — Set the deinterlacing mode.

It accepts one of the following values:

fast
medium
slow — use iterative motion estimation
extra_slow — like slow, but use multiple reference frames.

Default value is fast.

parity — Set the picture field parity assumed for the input video. It must be one of the following values:
- 0, tff — assume top field first
- 1, bff — assume bottom field first

Default value is bff.

qp — Set 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 median pixel from certain rectangle defined by radius. Median filter said "I'm the main character" and honestly? Correct.

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, value will be picked from horizontal radius option.
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 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.

mergeplanes

Merge color channel components from several video streams. Channel mixing is the group project and everybody's pulling their weight 💖

The filter accepts up to 4 input streams, and merge selected input planes to the output video.

This filter accepts the following options:

mapping — Set input to output plane mapping. Default is 0.

The mappings 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 mappings is similar, Bb describes the mapping for the output stream second plane, Cc describes the mapping for the output stream third plane and Dd describes the mapping for the output stream fourth plane.

format — Set output pixel format. Default is yuva444p.
map0s, map1s, map2s, map3s — Set input to output stream mapping for output Nth plane. Default is 0.
map0p, map1p, map2p, map3p — Set input to output plane mapping for output Nth plane. Default is 0.

Examples

Merge three gray video streams of same width and height into single video stream:

[a0][a1][a2]mergeplanes=0x001020:yuv444p

Merge 1st yuv444p stream and 2nd gray video stream into yuva444p video stream:

[a0][a1]mergeplanes=0x00010210:yuva444p

Swap Y and A plane in yuva444p stream:
format=yuva444p,mergeplanes=0x03010200:yuva444p
- Swap U and V plane in yuv420p stream:

format=yuv420p,mergeplanes=0x000201:yuv420p

Cast a rgb24 clip to yuv444p:

format=rgb24,mergeplanes=0x000102:yuv444p

🌸 mestimate

Bestie, this one estimates and exports motion vectors using block matching algorithms, and then it stashes those motion vectors in frame side data so other filters can come and use them. ✨

This filter accepts the following options:

method — tell it which motion estimation method to use. It takes one of these values:
- 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

Okay so this is the glow-up version of mestimate — it estimates and exports motion vectors too, but using D3D12 hardware-accelerated motion estimation. 🚀 It's leveraging the GPU hardware motion estimation capabilities that live in the DirectX 12 Video APIs, and that's how it scores significant performance improvements over the software-based mestimate filter.

Motion vectors get stored in frame side data so other filters can use them, same deal.

Heads up bestie: this filter requires the input to be in d3d12 hardware pixel format. And it uses quarter-pixel precision for motion vector estimation. No cap.

This filter accepts the following options:

mb_size — Macroblock size. Only 8 and 16 are supported. Default 16.

Example

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

This filter applies the Midway Image Equalization effect using two video streams. Iconic. 💅

So what Midway Image Equalization does is adjust a pair of images so they have the same histogram, all while keeping their dynamics as intact as possible. It's lowkey perfect for stuff like matching exposures from a pair of stereo cameras. 📸

This filter has two inputs and one output, and those inputs must be the same pixel format — but they can be different sizes, slay. The output of the filter is the first input, adjusted with the midway histogram of both inputs.

This filter accepts the following option:

planes — Set which planes to process. Default is 15, which is literally all available planes. ✨

🎀 minterpolate

Convert the video to the specified frame rate using motion interpolation. Periodt.

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, and scene change detection replaces 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. It's giving collab. 💖

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 the number of weights is smaller than the number of frames, the last specified weight will be used for all remaining unset weights.
scale — Specify scale. If it is set, it will be multiplied with the sum of each weight multiplied with pixel values to give the final destination pixel value. By default scale is auto scaled to 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 a custom color filter. Main character energy honestly.

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 all above options as commands. 💅

🧚 morpho

Okay so this filter lets you apply the main morphological grayscale transforms — erode and dilate — with arbitrary structures set in a second input stream.

And bestie, unlike the naive implementation which is much slower with erosion and dilation filters, when speed is critical you should use the morpho filter instead. Not me hyping it up but it's literally the move. 🔥

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. Bye bye redundant frames. 👋

The main use of this filter is for very-low-bitrate encoding (e.g. streaming over dialup modem — yes, we're going back in time 💀), 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 starting to drop 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. Fr fr, that's what it does.

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 that both inputs have the same number of frames, which are compared one by one.

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. ✨

       ffmpeg -i main.mpg -i ref.mpg -lavfi msad -f null -

Like, that's the whole vibe — main character main.mpg gets compared against the reference girlie ref.mpg, no cap. ✨

multiply

Multiply the pixel values of the first video stream by the pixel values of the second video stream. Math but make it cinema. 🔥

She accepts these options:

scale — sets the scale that gets applied to the second video stream. Default is 1. Allowed range goes from 0 to 9.
offset — sets the offset applied to the second video stream. Default is 0.5. Allowed range is -1 to 1.
planes — specify which planes from the input video stream get processed. By default, all planes are processed, she's not picking favorites.

Commands

This filter supports the same commands as options. 💅

negate

Negate (aka invert) the input video. Yes it's giving photo-negative, and yes it ate.

It accepts this option:

components — set which components to negate.

Available values for components are:

y
u
v
a
r
g
b
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 your frames using the Non-Local Means algorithm. Basically she looks around for pixels with similar contexts and adjusts each pixel accordingly — very "you're not alone bestie" of her. 🌸

That context similarity is defined by comparing their surrounding patches of size pxp. Patches get searched in an area of rxr around the pixel.

Heads up: the research area defines centers for patches, which means some patches will be made of pixels outside that research area. Not a bug, just how she rolls.

The filter accepts the following options.

s — set denoising strength. Default is 1.0. Must be in range [1.0, 30.0].
p — set patch size. Default is 7. Must be an odd number in range [0, 99].
pc — same as p but for the 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 the chroma planes. The default value is 0 and means automatic.

nnedi

Deinterlace video using neural network edge directed interpolation. Neural net but for your interlaced mess, slay. 🧠✨

This filter accepts the following options:

weights — mandatory option, without the binary file the filter literally cannot 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 the 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 the size of the 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 get 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 gets 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 to not use any of the specified pixel formats for the input to the next filter. It's giving "girl, literally anything but those." 💀

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 sprinkled on purpose, we love to see it. 🎀

The filter accepts the following 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) — she's academic AND cute. 📚💖

For each channel of each frame, the filter computes the input range and maps it linearly to the user-specified output range. The output range defaults to the full dynamic range from pure black to pure white.

Temporal smoothing can be used on the input range to reduce flickering (rapid changes in brightness) caused when small dark or bright objects enter or leave the scene. This is similar to the auto-exposure (automatic gain control) on a video camera, and, like a video camera, it may cause a period of over- or under-exposure of the video. 😭

The R,G,B channels can be normalized independently, which may cause some color shifting, or linked together as a single channel, which prevents color shifting. Linked normalization preserves hue. Independent normalization does not, so it can be used to remove some color casts. Independent and linked normalization can be combined in any ratio.

The normalize filter accepts the following options:

blackpt, whitept — colors which define the output range. The minimum input value is mapped to the blackpt. The maximum input value is mapped to the whitept. The defaults are black and white respectively. Specifying white for blackpt and black for whitept will give color-inverted, normalized video. Shades of grey can be used to reduce the dynamic range (contrast). Specifying saturated colors here can create some interesting effects. 🌈
smoothing — the number of previous frames to use for temporal smoothing. The input range of each channel is smoothed using a rolling average over the current frame and the smoothing previous frames. The default is 0 (no temporal smoothing).
independence — controls the ratio of independent (color shifting) channel normalization to linked (color preserving) normalization. 0.0 is fully linked, 1.0 is fully independent. Defaults to 1.0 (fully independent).
strength — overall strength of the filter. 1.0 is full strength. 0.0 is a rather expensive no-op. Defaults to 1.0 (full strength).

Commands

This filter supports 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. She does nothing and she does it flawlessly. 👑

ocio

OpenColorIO library filter. 💅

This filter allows you to do color management using the OpenColorIO library. See https://opencolorio.org/ for more details. To enable compilation of this filter, you need to configure FFmpeg with --enable-libopencolorio.

It accepts the following options:

config — by default the filter will use the OCIO config defined by the OCIO environment variable, but this parameter allows you to explicitly specify its location. If you are getting started, you can use config=ocio://studio-config-v1.0.0_aces-v1.3_ocio-v2.1 which specifies one of the built in defaults. ✨
input — set the input colorspace.
output — set the output colorspace.
display — set the display colorspace, used in combination with view.
view — set the view colorspace, used in combination with display.
inverse — when used in combination with display and view, this inverts the transform, so going from a display/view to the "input colorspace". 🦋
filetransform — lets you point at an external file-transform to use instead of the OCIO config file. Super handy when you just wanna apply one single transform without dragging a whole OCIO config along. ✨
format — lets you set the output pix_fmt of the OCIO filter. Heads up bestie, this has to be an RGB colorspace, so you're genuinely limited to rgb24, rgba, rgb48, rgba48, gbrp10, gbrp12, gbrpf32le, gbrapf32le. For most encoding we'd recommend rgb48. 💅
context_params — lets you specify extra context parameters for the OCIO filter. It's a list of key=value pairs, separated by colons. Easy.

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 — again, this assumes the OCIO file is defined with the OCIO environment variable. Note that you'll need to wrap the argument in quotes so the spaces get interpreted correctly, no cap:

input=ACEScg:display=sRGB - Display:view=ACES 1.0 - SDR Video:format=rgb48

Same vibe as above, but using the OCIO file studio-config-v1.0.0_aces-v1.3_ocio-v2.1_ns.ocio instead of the OCIO environment variable:

config=studio-config-v1.0.0_aces-v1.3_ocio-v2.1_ns.ocio:input=ACEScg:display=sRGB - Display:view=ACES 1.0 - SDR Video:format=rgb48

If you're converting to YCrCb you still wanna set the color matrix for the conversion — this one is a genuinely good example of combining the two. 🔥

ffmpeg -y -i SOURCEFRAMES.%05d.exr -c:v libx265 -vf "ocio=input=ACEScg:output=ACEScct:format=rgb48,scale=in_color_matrix=bt709:out_color_matrix=bt709,format=yuv444p10"  OUTPUTFILE.mov

ocr

Optical Character Recognition — yes, literally reading text out of your frames. 👑

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. It also exports the confidence of recognized words as the frame metadata lavfi.ocr.confidence. Cute, right? 🫶

ocv

Apply a video transform using libopencv.

To enable this filter, install the libopencv library and headers and configure FFmpeg with --enable-libopencv.

It accepts the following parameters:

filter_name — the name of the libopencv filter to apply.
filter_params — the parameters to pass to the libopencv filter. If not specified, the default values are assumed.

Refer to the official libopencv documentation for more precise info: <http://docs.opencv.org/master/modules/imgproc/doc/filtering.html>

Several libopencv filters are supported — see the following subsections. 🦋

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 of columns and rows of the read file are assumed instead. 🤯

The default value for struct_el is "3x3+0x0/rect".

nb_iterations specifies the number of times the transform is applied to the image, and defaults to 1.

Some examples:

# Use the default values
ocv=dilate

# Dilate using a structuring element with a 5x5 cross, iterating two times
ocv=filter_name=dilate:filter_params=5x5+2x2/cross|2

# Read the shape from the file diamond.shape, iterating two times.
# The file diamond.shape may contain a pattern of characters like this
#   *
#  ***
# *****
#  ***
#   *
# The specified columns and rows are ignored
# but the anchor point coordinates are not
ocv=dilate:0x0+2x2/custom=diamond.shape|2

erode

Erode an image by using a specific structuring element. It corresponds to the libopencv function cvErode.

It accepts the parameters: struct_el:nb_iterations, with the same syntax and semantics as the dilate filter. (Yes, the separator is a colon here — do not come for me, that's just how it is. 💀)

smooth

Smooth the input video.

The filter takes the following parameters: type|param1|param2|param3|param4.

type is the type of smooth filter to apply, and must be one of these values: "blur", "blur_no_scale", "median", "gaussian", or "bilateral". The default value is "gaussian".

The meaning of param1, param2, param3, and param4 depends on the smooth type. param1 and param2 accept integer positive values or 0. param3 and param4 accept floating point values.

The default value for param1 is 3. The default value for the other parameters is 0.

These parameters correspond to the parameters assigned to the libopencv function cvSmooth.

oscilloscope

2D Video Oscilloscope. 📈

Useful to measure spatial impulse, step responses, chroma delays, etc.

It accepts the following parameters:

x — set scope center x position.
y — set scope center y position.
s — set scope size, relative to frame diagonal.
t — set scope tilt/rotation.
o — set trace opacity.
tx — set trace center x position.
ty — set trace center y position.
tw — set trace width, relative to width of frame.
th — set trace height, relative to height of frame.
c — set which components to trace. By default it traces 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. Literal main character energy — one video is the moment, the other is the accessory. 💅

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 and main_h, H — the main input width and height.
overlay_w, w and 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. (Like, genuinely don't. It's giving "why is this still here." 💀)
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:

…and here's the code for the logo-drop, bestie 💅

ffmpeg -i input -i logo -filter_complex 'overlay=10:main_h-overlay_h-10' output

Insert 2 different transparent PNG logos (second logo on bottom right corner) using the ffmpeg tool: — yes, two logos, two overlays, one command, iconic:

ffmpeg -i input -i logo1 -i logo2 -filter_complex 'overlay=x=10:y=H-h-10,overlay=x=W-w-10:y=H-h-10' output

Add a transparent color layer on top of the main video; WxH must specify the size of the main input to the overlay filter — so like, WxH is literally the size of your main input, no improvising:

color=color=red@.3:size=WxH [over]; [in][over] overlay [out]

Play an original video and a filtered version (here with the deshake filter) side by side using the ffplay tool — split screen moment fr fr:

ffplay input.avi -vf 'split[a][b]; [a]pad=iw*2:ih[src]; [b]deshake[filt]; [src][filt]overlay=w'

The above command is the same as:

ffplay input.avi -vf 'split[b], pad=iw*2[src], [b]deshake, [src]overlay=w'

Make a sliding overlay appearing from the left to the right top part of the screen starting since time 2 — entrance? served. ✨

overlay=x='if(gte(t,2), -w+(t-2)*20, NAN)':y=0

Compose output by putting two input videos side to side — the split-screen collage of your dreams:

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 — censoring era 🫣:

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 — overlay-ception, and honestly? ate:

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 — it's giving "smooth out the noise but keep the vibes" energy.

The filter accepts the following options:

depth — sets the depth. Larger depth values will denoise lower frequency components more, but slow down filtering. Must be an int in the range 8-16, default is 8. (So yeah, more denoise = more waiting, lowkey the tradeoff of all time.)
luma_strength, ls — sets the luma strength. Must be a double value in the range 0-1000, default is 1.0.
chroma_strength, cs — sets the chroma strength. Must be a double value in the range 0-1000, default is 1.0.

🖼️ pad

Adds paddings to the input image, and places the original input at the provided x, y coordinates. It's the frame around the picture, bestie — the matte, the border, the whole presentation. 💖

It accepts the following parameters:

width, w
height, h — Specify an expression for the size of the output image with the paddings added. If the value for width or height is 0, the corresponding input size is used for the output.

The width expression can reference the value set by the height expression, and vice versa. (They're gossip besties, they know each other's business.)

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. (Not me auto-centering your whole life 🫶)

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 — violet era, obsessed:

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 the color mood board for your entire video, and yes it's doing the absolute most. ✨

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 the leftovers are just chilling in black, no drama.)
reserve_transparent — Create a palette of 255 colors maximum and reserve the last one for transparency. Reserving the transparency color is useful for GIF optimization. If not set, the maximum of colors in the palette will be 256. You probably want to disable this option for a standalone image. Set by default.
transparency_color — Set the color that will be used as background for transparency.
stats_mode — Set statistics mode. It accepts the following values:
- full — Compute full frame histograms.
- diff — Compute histograms only for the part that differs from previous frame. This might be relevant to give more importance to the moving part of your input if the background is static.
- single — Compute new histogram for each frame.

Default value is full.

The filter also exports the frame metadata "lavfi.color_quant_ratio" ("nb_color_in / nb_color_out") which you can use to evaluate the degree of color quantization of the palette. This information is also visible at info logging level. 👀

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. Basically palettegen's other half — the one that actually does the coloring-in. 💅

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.
- floyd_steinberg — Floyd and Steingberg dithering (error diffusion)
- sierra2 — Frankie Sierra dithering v2 (error diffusion)
- sierra2_4a — Frankie Sierra dithering v2 "Lite" (error diffusion)
- sierra3 — Frankie Sierra dithering v3 (error diffusion)
- burkes — Burkes dithering (error diffusion)
- atkinson — Atkinson dithering by Bill Atkinson at Apple Computer (error diffusion)
- none — Disable dithering.

Default is sierra2_4a.

bayer_scale — When bayer dithering is selected, this option defines the scale of the pattern (how much the crosshatch pattern is visible). A low value means more visible pattern for less banding, and higher value means less visible pattern at the cost of more banding.

The option must be an integer value in the range [0,5]. Default is 2.

diff_mode — If set, define the zone to process.
- rectangle — Only the changing rectangle will be reprocessed. This is similar to GIF cropping/offsetting compression mechanism. This option can be useful for speed if only a part of the image is changing, and has use cases such as limiting the scope of the error diffusal dither to the rectangle that bounds the moving scene (it leads to more deterministic output if the scene doesn't change much, and as a result less moving noise and better GIF compression).

Default is none.

new — Take new palette for each output frame.
alpha_threshold — Sets the alpha threshold for transparency. Alpha values above this threshold will be treated as completely opaque, and values below this threshold will be treated as completely transparent.

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 cursed angle? We fix. Un-cursed. 💜)

okay so here's exactly what it takes, bestie:

x0, y0, x1, y1, x2, y2, x3, y3 — set the coordinate expressions for the top left, top right, bottom left and bottom right corners. Default is "0:0:W:0:0:H:W:H", and with that the perspective just... stays the same, no drama. If the sense option is "source", then the points you gave get sent to the corners of the destination. If sense is "destination", then the corners of the source get sent to the coordinates you specified. 🦋

Your expressions can use these variables:

W, H — the width and height of the video frame.
in — input frame count.
on — output frame count.

interpolation — set the interpolation used for the perspective correction. Accepted values:
- linear
- cubic

Default value is linear. 💅

sense — set how the coordinate options are interpreted. Accepted values:
- 0, source — send the point in the source specified by the given coordinates to the corners of the destination.
- 1, destination — send the corners of the source to the point in the destination specified by the given coordinates.

Default value is source. ✨

eval — set when the expressions for the coordinates x0,y0,...x3,y3 get evaluated. Accepted values:
- init — only evaluate expressions once during filter initialization, or when a command is processed.
- frame — evaluate expressions for each incoming frame.

Default value is init.

🌸 `phase` 🌸

This one delays interlaced video by one field time so the field order gets flipped. It's giving "fixer upper".

The intended use is fixing PAL movies that got captured with the opposite field order to the film-to-video transfer. 🎀

A description of the accepted parameters follows:

mode — set the phase mode. Accepted 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 have something to refer to in the docs — but if you actually select it, the filter will faithfully do nothing. no cap, it's a professional no-op. 😭
- a — capture field order determined automatically by field flags, transfer opposite. The filter picks between t and b modes on a frame-by-frame basis 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 on a frame-by-frame basis by analyzing the images and selecting the alternative that 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 can help users with epilepsy. genuinely such a kind filter, we stan. 💖

It accepts the following options:

frames, f — set how many frames to use when filtering. Default is 30.
threshold, t — set the detection threshold factor. Default is 1. Lower is stricter.
skip — set how many pixels to skip when sampling frames. Default is 1. Allowed range is from 1 to 1024.
bypass — leave frames unchanged. Default is disabled.

⭐ `pixdesctest` ⭐

Pixel format descriptor test filter, mainly useful for internal testing. The output video should be equal to the input video — like, that's the whole vibe check.

For example:

format=monow, pixdesctest

can be used to test the monowhite pixel format descriptor definition.

🎀 `pixelize` 🎀

Apply pixelization to video stream. artsy. blocky. iconic.

The filter accepts the following options:

width, w and height, h — set the block dimensions that will be 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` 🦋

Display sample values of the color channels. Mainly useful for checking color and levels. Minimum supported resolution is 640x480 — so no tiny frames, she needs room to breathe.

The filters accept the following 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` 💜

Apply 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 the following options:

qp — force a constant quantization parameter. It accepts an integer in range 0 to 63. If not set, the filter will use the QP from the video stream (if available).
mode — set thresholding mode. Available modes are:
- hard — set hard thresholding.
- soft — set soft thresholding (better de-ringing effect, but likely blurrier).
- medium — set medium thresholding (good results, default). 🌈

💅 `premultiply` 💅

Apply alpha premultiply effect to the input video stream, using the first plane of the second stream as alpha.

Both streams must have the same dimensions and the same pixel format. no exceptions, bestie.

The filter accepts the following option:

planes — set which planes will be processed, unprocessed planes will be copied. By default value 0xf, all planes will be processed.
inplace — do not require a 2nd input for processing, instead use the alpha plane from the input stream.

✨ `premultiply_dynamic` ✨

Dynamically premultiply or unpremultiply 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. she's adaptable like that. 🧚

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.

🔥 `prewitt` 🔥

Apply the prewitt operator to the input video stream. edge detection main character energy.

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 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` 🌈

Alter frame colors in video with pseudocolors. it's giving aesthetic overhaul, honestly.

This filter accepts the following 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 is 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 the following 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` 🍓

Obtain the average, maximum and minimum PSNR (Peak Signal to Noise Ratio) between two input videos. quality-checking queen.

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 PSNR.

Both video inputs 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 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 is 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 — decides whether the max value gets output to the stats log. Default value is 0. Requires stats_version >= 2. If this is set and stats_version < 2, the filter will return an error.

bestie, this filter also supports the framesync options. 💅

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.

If a stats_version greater than 1 is specified, a header line precedes the list of per-frame-pair stats, with key value pairs following the frame format with the following 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 follows, bestie:

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.

Another example with different containers:

ffmpeg -i main.mpg -i ref.mkv -lavfi  "[0:v]settb=AVTB,setpts=PTS-STARTPTS[main];[1:v]settb=AVTB,setpts=PTS-STARTPTS[ref];[main][ref]psnr" -f null -

🎀 pullup 🎀

Pulldown reversal (inverse telecine) filter, capable of handling mixed hard-telecine, 24000/1001 fps progressive, and 30000/1001 fps progressive content. Iconic, honestly.

The pullup filter is designed to take advantage of future context in making its decisions. This filter is stateless in the sense that it does not lock onto a pattern to follow, but it instead looks forward to the following fields in order to identify matches and rebuild progressive frames. future context queen ✨

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.

This 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. realtime on slow machines, we love efficiency 🫶

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). short and powerful 💅

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. QR code era ✨

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 overlaied. (yes, that spelling is in the original, love that for us) 🙈

This filter 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 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 versa.

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. set set, iconic 💀

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 a higher correction level, the encoding size will increase but the code will be more robust to corruption. Lower level is L.
- L
- M
- Q
- H
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, 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, 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.

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.

expr_formatted, ef — evaluates an expression and spits it back out as a formatted string 💅 the first argument is the expression (same deal as the expr function), the second is the output format, and the allowed values are x, X, d and u — all treated exactly like in the "printf" function. Third parameter is optional and sets how many positions the output takes up, so you can pad with zeros on the left 🎀
gmtime — the time the filter is running, expressed in UTC ⏰ it can take an argument: a "strftime" C function format string. That format string is extended to support the variable %[1-6]N, which prints fractions of the second with an optionally specified number of digits ✨
localtime — same energy as gmtime but in the local time zone 🌍 takes an argument too, a "strftime" C function format string, extended to support %[1-6]N for fractions of the second with optionally specified digits 🕰️
metadata — frame metadata, and it takes one or two arguments 🏷️ The first argument is mandatory and is the metadata key. The second is optional and is the default value used when that metadata key isn't found or is empty. Wanna know what metadata you actually have? Go inspect the entries starting with TAG inside each frame section printed by running "ffprobe -show_frames" 👀 String metadata generated in filters that lead into the qrencode filter is available too 💖
rand(min, max) — returns a random number between min and max 🎲

✨ Examples ✨

Generate a QR code encoding the specified text with the default size, overlaid in the top left corner of the input video, with the default size:

qrencode=text=www.ffmpeg.org

Same as below, but select blue on pink colors:

qrencode=text=www.ffmpeg.org:bc=pink@0.5:fc=blue

Place the QR code in the bottom right corner of the input video:

qrencode=text=www.ffmpeg.org:x=W-Q:y=H-Q

Generate a QR code with width of 200 pixels and padding, making the padded width 4/3 of the QR code width:

qrencode=text=www.ffmpeg.org:q=200:Q=4/3*q

Generate a QR code with padded width of 200 pixels and padding, making the QR code width 3/4 of the padded width:

qrencode=text=www.ffmpeg.org:Q=200:q=3/4*Q

Make the QR code a fraction of the input video width:

qrencode=text=www.ffmpeg.org:q=W/5

Generate a QR code encoding the frame number:

qrencode=text=%{n}

Generate a QR code encoding the GMT timestamp:

qrencode=text=%{gmtime}

Generate a QR code encoding the timestamp expressed as a float:

qrencode=text=%{pts}

🌈 quirc 🌈

Identifies and decodes a QR code using the libquirc library (check it out at <https://github.com/dlbeer/quirc/>), and prints the identified QR codes' positions and payload as metadata 👑

To even compile this filter, you have to configure FFmpeg with "--enable-libquirc" — no flag, no quirc, that's the whole vibe 🔧

For every QR code found in the input video, some metadata entries get added with the prefix lavfi.quirc.N, where N is the index starting from 0 that's associated with that QR code 🦋

Description of each metadata value, coming right up:

lavfi.quirc.count — the number of QR codes found, and it is not set if none were found 🙈
lavfi.quirc.N.corner.M.x
lavfi.quirc.N.coreer.M.y — the x/y positions of the four corners of the square containing the QR code, where M is the index of the corner starting from 0 💅 (yes, coreer is literally how it's spelled in the page — i'm not touching it, don't come for me 💀)
lavfi.quirc.N.payload — the payload of the QR code 📦

🎲 random 🎲

Flushes video frames from the internal frame cache into a random order. Not a single frame gets discarded, bestie — nothing is lost 🫶 Inspired by the frei0r nervous filter.

frames — sets the size in number of frames of the internal cache, in range from 2 to 512. Default is 30.
seed — sets the seed for the random number generator, and it must be an integer between 0 and "UINT32_MAX". If it's not specified, or if it's explicitly set to less than 0, the filter will try to use a good random seed on a best effort basis ✨

📺 readeia608 📺

Reads closed captioning (EIA-608) info 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). Each metadata value explained:

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. If there's a parity error, the filter outputs 0x00 for that character instead. Default is false 🙈
lp — lowpass lines prior to further processing. Default is enabled ✅

🎀 Commands 🎀

This filter supports all of the 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 ⏱️

Reads vertical interval timecode (VITC) info from the top lines of a video frame 🎬

The filter adds the frame metadata key "lavfi.readvitc.tc_str" holding the timecode value — but only if a valid timecode has actually been detected. There's also another metadata key, "lavfi.readvitc.found", set to 0/1 depending on whether timecode data was 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 gets 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 to 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 to 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 🗺️

Remaps pixels using the 2nd input video stream as Xmap and the 3rd as Ymap 🌀

The destination pixel at position (X, Y) gets picked from the source position (x, y), where x = Xmap(X, Y) and y = Ymap(X, Y). If the mapping values land out of range, a zero value for the pixel is used for that destination pixel 🕳️

The Xmap and Ymap input video streams must be the same dimensions. The output video stream will have Xmap/Ymap video stream dimensions. Xmap and Ymap input video streams are 16bit depth, single channel 📐

format — specify the pixel format of the 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 ✨

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. Here's what each mode does, buckle up bestie:

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 neighbour 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 neighbour pixels are the closest.
14 — bob mode, interpolates bottom field from the line where the neighbour 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 🎭

Suppresses a TV station logo, using an image file to figure out which pixels make up 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's recommended to take a screen capture of a black frame with the logo visible, and then use a threshold filter followed by the erode filter once or twice 📸

If you need to, you can also patch little splotches by hand. But bestie, listen: if any logo pixels end up not covered, the filter quality drops hard. Marking too many pixels as part of the logo doesn't hurt as much, but it means more blurring is needed to cover the image up, which destroys more information than necessary — and all those extra pixels will slow things down on a large logo. So like… be reasonable 💅

repeatfields

This one peeks at the repeat_field flag in the Video ES headers and hard-repeats fields based on whatever value it finds. Respect the header's authority, basically 👑

reverse

Reverses a video clip. That's it, that's the whole vibe. ⏪

⚠️ Warning: this filter needs enough memory to buffer the entire clip, so trimming first is strongly suggested. Do not just yeet a two-hour video at it, babe.

Examples

Take the first 5 seconds of a clip and reverse it:

trim=end=5,reverse

rgbashift

Shifts the R/G/B/A pixels horizontally and/or vertically. Little nudge left, little nudge up, whatever you're feeling 💫

The filter takes these options:

rh — how much to shift red horizontally.
rv — how much to shift red vertically.
gh — how much to shift green horizontally.
gv — how much to shift green vertically.
bh — how much to shift blue horizontally.
bv — how much to shift blue vertically.
ah — how much to shift alpha horizontally.
av — how much to shift alpha vertically.
edge — set the edge mode. Can be smear, default, or warp.

Commands

This filter supports all the above options as commands. Every single one, slay 🫶

roberts

Applies the roberts cross operator to the input video stream. Edge detection era ✨

The filter takes this option:

planes — sets which planes get processed; unprocessed planes just get copied along. Default value is 0xf, which means all planes are processed.
scale — sets a value that gets multiplied with the filtered result.
delta — sets a value that gets added to the filtered result.

Commands

This filter supports all the above options as commands too 💖

rotate

Rotates the video by an arbitrary angle, expressed in radians. Fancy, angular, main character energy 🔄

The filter takes these options:

Here's the description of the optional parameters, mwah:

angle, a — sets an expression for the angle by which to rotate the input video clockwise, expressed as a number of radians. A negative value gives you a counter-clockwise rotation. Default is "0".

This expression is evaluated for each frame.

out_w, ow — sets the output width expression. Default value is "iw". This expression is evaluated just once, during configuration.
out_h, oh — sets the output height expression. Default value is "ih". This expression is evaluated just once, during configuration.
bilinear — turns bilinear interpolation on if set to 1, and a value of 0 disables it. Default value is 1.
fillcolor, c — sets the color used to fill the output area that the rotated image doesn't cover. For the general syntax of this option, go check the Color section in the ffmpeg-utils manual. If you pick the special value "none", then no background gets printed at all (super useful when the background is never shown, iykyk).

Default value is "black".

The expressions for the angle and the output size can contain these 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's 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 and in_h, ih — the input video width and height.
out_w, ow and 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 to completely contain the input video rotated by a radians.

These are only available when computing the out_w and out_h expressions. Periodt.

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 oscillate with a period of T seconds and an amplitude of A radians:

rotate=A*sin(2*PI/T*t)

Rotate the video, with the output size 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, and 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 these commands:

a, angle — sets the angle expression. The command accepts the same syntax as the corresponding option.

If the expression you give it isn't valid, it just stays at its current value. No drama 💅

sab

Applies Shape Adaptive Blur. Soft and smart about it 🌸

The filter takes these options:

luma_radius, lr — sets the luma blur filter strength. Must be a value in the range 0.1-4.0, default value is 1.0. A greater value gives you a more blurred image, and slower processing.
luma_pre_filter_radius, lpfr — sets the luma pre-filter radius. Must be a value in the 0.1-2.0 range, default value is 1.0.
luma_strength, ls — sets the 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 — sets the chroma blur filter strength. Must be a value in the range -0.9-4.0. A greater value gives you a more blurred image, and slower processing.
chroma_pre_filter_radius, cpfr — sets the chroma pre-filter radius. Must be a value in the -0.9-2.0 range.
chroma_strength, cs — sets the chroma maximum difference between pixels to still be considered. Must be a value in the -0.9-100.0 range.

Every chroma option value, if you don't explicitly specify it, gets set to the corresponding luma option value. Copying its bestie's homework fr fr 🎀

scale

Scales (resizes) the input video, using the libswscale library. The classic glow-up filter 👑

The scale filter forces the output display aspect ratio to be the same as the input's, 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 converts the input to the requested format. Helpful king.

Options

The filter accepts the following options, plus any of the options supported by the libswscale scaler, plus any of the framesync options.

📖 See the ffmpeg-scaler manual for the complete list of scaler options. Go read it, bestie, it's giving reference material.

width, w and 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 uses a value that maintains the aspect ratio of the input image, calculated from the other specified dimension. After that, however, it makes sure the calculated dimension is divisible by n and adjusts the value if necessary.

If both values are -n with n >= 1, the behavior is identical to both values being set to 0 as detailed above.

See below for the list of accepted constants to use in the dimension expression.

eval — specifies when to evaluate the width and height expressions. It accepts these values:
- init — only evaluate expressions once during filter initialization, or when a command is processed.
- frame — evaluate expressions for each incoming frame.

Default value is init.

interl — sets the interlacing mode. It accepts these 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 — sets the 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 — sets 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 — sets the ICC rendering intent to use when transforming between different color spaces. It accepts these 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 reproduces 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, it stretches the input color volume directly onto the output color volume, in a non-linear fashion that preserves the original signal appearance as much as possible. This intent is recommended for signal content evaluation, since it will not lead to any clipping. It's 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 — sets 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 the in/output YCbCr color space type.

This lets you override whatever got autodetected, and it also lets you force a specific value that gets used for the output and the encoder. If you don't specify it, the color space type just 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 the in/output YCbCr sample range. 🎚️

Same deal: override the autodetected value, or force a specific one for the output and encoder. If you don't specify it, the range depends on the pixel format. Possible values:

auto/unknown — choose automatically.
jpeg/full/pc — set full range (0-255 in case of 8-bit luma).
mpeg/limited/tv — set "MPEG" range (16-235 in case of 8-bit luma).

in_chroma_loc, out_chroma_loc — set the in/output chroma sample location. If you don't specify it, center-sited chroma is used by default. Possible values: 🎯

auto, unknown
left
center
topleft
top
bottomleft
bottom

in_primaries, out_primaries — set the in/output RGB primaries. 🌈

Override the autodetected value, or force a specific one for the output and encoder, bestie. 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 the in/output transfer response curve (TRC). 📈

Override the autodetected value, or force a specific one 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 — lets the output video width or height decrease or increase when it needs to, so the original aspect ratio stays intact. 🫶 Possible values:

disable — scale the video exactly as specified and turn this feature off.
decrease — the output video dimensions get automatically decreased if needed.
increase — the output video dimensions get automatically increased if needed.

One genuinely iconic use case: when you know a specific device's maximum allowed resolution, you can use this to cap the output video at that, while keeping the aspect ratio. Say device A allows 1280x720 playback and your video is 1920x800. Use this option (set it to decrease) and pass 1280x720 on the command line, and the output becomes 1280x533. ✨

Heads up, this is a different thing than specifying -1 for w or h — you still need to specify the output resolution for this option to actually work. No cap, fr.

force_divisible_by — makes sure both output dimensions, width and height, are divisible by the integer you give it, when used together with force_original_aspect_ratio. Works similar to using -n in the w and h options. 🔢

This option respects whatever value you set for force_original_aspect_ratio, bumping the resolution up or down accordingly. The video's aspect ratio may be slightly modified — just so you know.

It's lowkey handy if you need a video to fit within (or exceed) a defined resolution via force_original_aspect_ratio, but your encoder has opinions about width or height divisibility.

reset_sar — enabling this leads to the output SAR being reset to 1. On top of that, if the user asks for proportional scaling through the width or height expressions — like w=-4:h=360 or w=iw/2:h=-1 — or by enabling force_original_aspect_ratio, then the input DAR gets 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, and they can contain these constants:

in_w, in_h — the input width and height.

iw, ih — these are the same as in_w and in_h.

out_w, out_h — the output (scaled) width and height.

ow, oh — these are the same as out_w and out_h.

a — the same as iw / ih.

sar — input sample aspect ratio.

dar — the input display aspect ratio. Calculated from (iw / ih) * sar.

hsub, vsub — horizontal and vertical input chroma subsample values. For example, for the pixel format yuv422p, hsub is 2 and vsub is 1.

ohsub, ovsub — horizontal and vertical output chroma subsample values. For example, for the pixel format yuv422p, hsub is 2 and vsub is 1.

n — the (sequential) number of the input frame, starting from 0. Only available with eval=frame.

t — the presentation timestamp of the input frame, expressed as a number of seconds. Only available with eval=frame.

pos — the position (byte offset) of the frame in the input stream, or NaN if this info is unavailable and/or meaningless (like, in the 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, 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's kept at its current value. No drama. 😌

`scharr`

Apply the scharr operator to the input video stream. 🌀

The filter accepts the following option:

planes — set which planes will be processed; unprocessed planes get copied. By default the value is 0xf, so all planes are processed.
scale — set the value that gets multiplied with the filtered result.
delta — set the value that gets added to the filtered result.

Commands

This filter supports all the above options as commands. 💅

`scroll`

Scroll the input video horizontally and/or vertically at a 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 changes. 🔍

This filter sets frame metadata with mafd between frame, the scene score, and forwards the frame on to the next filter, so they can use that metadata to detect scene changes or whatever else they're into. ✨

In addition, this filter logs a message and sets frame metadata when it detects a scene change by threshold.

lavfi.scd.mafd — metadata key set with mafd for every single frame, no cap 🫶
lavfi.scd.score — metadata key set with the scene change score for every frame, so you can actually detect a scene change
lavfi.scd.time — metadata key set with the current filtered frame time that detected a scene change with the 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 itself 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. Enable it if you literally only want a snapshot of the scene change frames 💅

🎀 selectivecolor 🎀

Adjust 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 — meaning how saturated it already is. Iconic.

This filter is similar to the Adobe Photoshop Selective Color tool. Photoshop girlies, this one's for you ✨

The filter accepts the following options:

correction_method — selects the color correction method. Available values are:
- absolute — specified adjustments are applied "as-is" (added/subtracted to the original pixel component value).
- relative — 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. Bestie, you can straight-up import your presets, we're so back 🦋

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 areas, 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 components fields, producing a new half height clip with twice the frame rate and twice the frame count. Half the height, double the drama, periodt.

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. Trust. 🙈

💜 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 — like if something scales, or if another setdar or a setsar filter gets applied. So don't come for me if a later filter undoes your work 😭

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. Same vibe, same warning 💅

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 🦋

Force field for the output video frame. Main character energy for your interlacing.

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). It's giving "just setting the vibe, not touching the pixels."

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 ⭐

Force frame parameter 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 before: no pixel touching, just labels 🏷️

field_mode — available values are:
- auto — keep the same field property (default).
- bff — mark the frame as bottom-field-first.
- tff — mark the frame as top-field-first.
- prog — mark the frame as progressive.
range — available values are:
- auto — keep the same color range property (default).
- unspecified, unknown — mark the frame as unspecified color range.
- limited, tv, mpeg — mark the frame as limited range.
- full, pc, jpeg — mark the frame as full range.
color_primaries — set the color primaries. Available values are:
- auto — keep the same color primaries property (default).
- bt709
- unknown
- bt470m
- bt470bg
- smpte170m
- smpte240m
- film
- bt2020
- smpte428
- smpte431
- smpte432
- jedec-p22
color_trc — set the color transfer. Available values are:
- auto — keep the same color trc property (default).
- bt709
- unknown
- bt470m
- bt470bg
- smpte170m
- smpte240m
- linear
- log100
- log316
- iec61966-2-4
- bt1361e
- iec61966-2-1
- bt2020-10
- bt2020-12
- smpte2084
- smpte428
- arib-std-b67
colorspace — set the colorspace. Available values are:
- auto — keep the same colorspace property (default).
- gbr
- bt709
- unknown
- fcc
- bt470bg
- smpte170m
- smpte240m
- ycgco
- bt2020nc
- bt2020c
- smpte2085
- chroma-derived-nc
- chroma-derived-c
- ictcp
chroma_location — set the chroma sample location. Available values are:
- auto — keep the same chroma location (default).
- unspecified, unknown
- left
- center
- topleft
- top
- bottomleft
- bottom
alpha_mode — set the alpha moda. (Yes, the manpage really says "moda" and honestly? We're keeping her 💀💅) Available values are:
- auto — keep the same alpha mode (default).
- unspecified, unknown
- premultiplied
- straight

🔥 shear 🔥

Apply shear transform to input video. Slanting your whole video, why not.

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 the all above options as commands. Yup, all of them, live, at runtime, go off 👑

🌈 showinfo 🌈

Show a line containing various information for each input video frame. The input video is not modified — she's just reading, not touching, we love a respectful icon.

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, periodt. For the actual syntax of this option, go peek at the "Video size" section in the ffmpeg-utils manual, bestie, it's all there 💅
i — the type of interlaced mode: "P" for "progressive", "T" for top field first, "B" for bottom field first. Three letters, three vibes ✨
iskey — this is 1 if the frame is a key frame, 0 otherwise. Main character frame or background extra, no in between 🎀
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 lowkey go refer to the documentation of the "AVPictureType" enum and of the "av_get_picture_type_char" function defined in libavutil/avutil.h, if you wanna go deeper 🦋
checksum — the Adler-32 checksum (printed in hexadecimal) of all the planes of the input frame. Whole-frame fingerprint energy 🔥
plane_checksum — the Adler-32 checksum (printed in hexadecimal) of each plane of the input frame, expressed in the form "[c0 c1 c2 c3]" 🫶
mean — the mean value of pixels in each plane of the input frame, expressed in the form "[mean0 mean1 mean2 mean3]" 🌸
stdev — the standard deviation of pixel values in each plane of the input frame, expressed in the form "[stdev0 stdev1 stdev2 stdev3]" ⭐

💖 showpalette 💖

Displays the 256 colors palette of each frame. Heads up: this filter is only relevant for pal8 pixel format frames — like, it will simply not serve for other formats, ngl 🙈

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). Cute little squares, one per color 🎨

🦋 shuffleframes 🦋

Reorder and/or duplicate and/or drop video frames. Chaos, but make it intentional 💅

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. The number of indexes also sets the maximal value that each index may have. The '-1' index has special meaning and that is to drop the frame 💀

The first frame has the index 0. The default is to keep the input unchanged.

✨ Examples ✨

Swap second and third frame of every three frames of the input:

ffmpeg -i INPUT -vf "shuffleframes=0 2 1" OUTPUT

Swap 10th and 1st frame of every ten frames of the input:

ffmpeg -i INPUT -vf "shuffleframes=9 1 2 3 4 5 6 7 8 0" OUTPUT

🎀 shufflepixels 🎀

Reorder pixels in video frames. Yes, pixels. Yes, it's unhinged. Yes, we love it 🔥

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 the size is used, and in case of vertical shuffle mode only the height part of the size is used. iykyk 🫶
seed, s — set random seed used with shuffling pixels. Mainly useful to set so you're able to reverse the filtering process and get the original input back. For example, to reverse forward shuffle you need to use the same parameters and the exact same seed, and to set direction to inverse. Screenshot the seed, bestie, or it's gone forever 😭

🌈 shuffleplanes 🌈

Reorder and/or duplicate video planes. Plane remix, no cap 👑

It accepts the following parameters:

map0 — the index of the input plane to be used as the first output plane.
map1 — the index of the input plane to be used as the second output plane.
map2 — the index of the input plane to be used as the third output plane.
map3 — the index of the input plane to be used as the fourth output plane.

The first plane has the index 0. The default is to keep the input unchanged.

✨ Examples ✨

Swap the second and third planes of the input:

ffmpeg -i INPUT -vf shuffleplanes=0:2:1:3 OUTPUT

👑 signalstats 👑

Evaluate various visual metrics that assist in determining issues associated with the digitization of analog video media. Think of it as a vibe check for your analog captures 💅

By default the filter will log these metadata values:

YMIN — display the minimal Y value contained within the input frame. Expressed in range of [0-255].
YLOW — display the Y value at the 10% percentile within the input frame. Expressed in range of [0-255].
YAVG — display the average Y value within the input frame. Expressed in range of [0-255].
YHIGH — display the Y value at the 90% percentile within the input frame. Expressed in range of [0-255].
YMAX — display the maximum Y value contained within the input frame. Expressed in range of [0-255].
UMIN — display the minimal U value contained within the input frame. Expressed in range of [0-255].
ULOW — display the U value at the 10% percentile within the input frame. Expressed in range of [0-255].
UAVG — display the average U value within the input frame. Expressed in range of [0-255].
UHIGH — display the U value at the 90% percentile within the input frame. Expressed in range of [0-255].
UMAX — display the maximum U value contained within the input frame. Expressed in range of [0-255].
VMIN — display the minimal V value contained within the input frame. Expressed in range of [0-255].
VLOW — display the V value at the 10% percentile within the input frame. Expressed in range of [0-255].
VAVG — display the average V value within the input frame. Expressed in range of [0-255].
VHIGH — display the V value at the 90% percentile within the input frame. Expressed in range of [0-255].
VMAX — display the maximum V value contained within the input frame. Expressed in range of [0-255].
SATMIN — display the minimal saturation value contained within the input frame. Expressed in range of [0-~181.02].
SATLOW — display the saturation value at the 10% percentile within the input frame. Expressed in range of [0-~181.02].
SATAVG — display the average saturation value within the input frame. Expressed in range of [0-~181.02].
SATHIGH — display the saturation value at the 90% percentile within the input frame. Expressed in range of [0-~181.02].
SATMAX — display the maximum saturation value contained within the input frame. Expressed in range of [0-~181.02].
HUEMED — display the median value for hue within the input frame. Expressed in range of [0-360].
HUEAVG — display the average value for hue within the input frame. Expressed in range of [0-360].
YDIF — display the average of sample value difference between all values of the Y plane in the current frame and corresponding values of the previous input frame. Expressed in range of [0-255].
UDIF — display the average of sample value difference between all values of the U plane in the current frame and corresponding values of the previous input frame. Expressed in range of [0-255].
VDIF — display the average of sample value difference between all values of the V plane in the current frame and corresponding values of the previous input frame. Expressed in range of [0-255].
YBITDEPTH — display bit depth of Y plane in current frame. Expressed in range of [0-16].
UBITDEPTH — display bit depth of U plane in current frame. Expressed in range of [0-16].
VBITDEPTH — display bit depth of V plane in current frame. Expressed in range of [0-16].

The filter accepts the following options:

stat, out — stat specifies 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. Bestie, this filter can handle more than one input — and when it does, 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 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) — that bit gets swapped out for the input number. And if you don't specify a filename at all? No output gets written, periodt. That's literally the default. 💅
format — picks the output format. Your options are:
- binary — use the specified binary representation (this is the default, bestie).
- xml — use the specified xml representation.
th_d — sets the threshold for detecting one word as similar. Value's gotta be an integer greater than zero. Default is 9000.
th_dc — sets the threshold for detecting all words as similar. Integer greater than zero, please. Default is 60000.
th_xh — sets the threshold for detecting frames as similar. Integer greater than zero. Default is 116.
th_di — sets the minimum length of a sequence, in frames, to actually recognize it as a matching sequence. Gotta be a non-negative integer. Default is 0.
th_it — sets the minimum relation that matching frames to all frames must have. Gotta be a double between 0 and 1. Default is 0.5.

🌸 Examples 🌸

To calculate the signature of an input video and store it in signature.bin:

ffmpeg -i input.mkv -vf signature=filename=signature.bin -map 0:v -f null -

To detect whether two videos match and store the signatures in XML format in signature0.xml and signature1.xml:

ffmpeg -i input1.mkv -i input2.mkv -filter_complex "[0:v][1:v] signature=nb_inputs=2:detectmode=full:format=xml:filename=signature%d.xml" -map :v -f null -

siti 🎀

Calculates Spatial Information (SI) and Temporal Information (TI) scores for a video, as defined in ITU-T Rec. P.910 (11/21): Subjective video quality assessment methods for multimedia applications. The PDF lives at <https://www.itu.int/rec/T-REC-P.910-202111-S/en>. ⭐

Okay so real talk: this is a legacy implementation that corresponds to a superseded recommendation. If you want 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 this one option:

print_summary — if set to 1, Summary statistics get printed to the console. Default is 0.

🦋 Examples 🦋

To calculate SI/TI metrics and print summary:

ffmpeg -i input.mp4 -vf siti=print_summary=1 -f null -

smartblur ✨

Blurs the input video without messing with the outlines. Iconic behaviour honestly. 🫶

It accepts these options:

luma_radius, lr — sets the luma radius. Gotta be a float in the range [0.1,5.0] that specifies the variance of the gaussian filter used to blur the image (slower if larger). Default is 1.0.
luma_strength, ls — sets the luma strength. Float in the range [-1.0,1.0] that configures the blurring. A value in [0.0,1.0] blurs the image, whereas a value in [-1.0,0.0] sharpens it. Default is 1.0.
luma_threshold, lt — sets the luma threshold, used as a coefficient to decide whether a pixel should be blurred or not. Integer in the range [-30,30]. A value of 0 filters the whole image, a value in [0,30] filters flat areas, and a value in [-30,0] filters edges. Default is 0.
chroma_radius, cr — sets the chroma radius. Float in the range [0.1,5.0] specifying the variance of the gaussian filter used to blur (slower if larger). Default is luma_radius.
chroma_strength, cs — sets the chroma strength. Float in the range [-1.0,1.0] configuring the blurring. [0.0,1.0] blurs, [-1.0,0.0] sharpens. Default is luma_strength.
chroma_threshold, ct — sets the chroma threshold, the coefficient for deciding whether a pixel gets blurred. Integer in the range [-30,30]. 0 filters the whole image, [0,30] filters flat areas, [-30,0] filters edges. Default is luma_threshold.
alpha_radius, ar — sets the alpha radius. Float in the range [0.1,5.0] specifying the variance of the gaussian filter used to blur (slower if larger). Default is luma_radius.
alpha_strength, as — sets the alpha strength. Float in the range [-1.0,1.0] configuring the blurring. [0.0,1.0] blurs, [-1.0,0.0] sharpens. Default is luma_strength.
alpha_threshold, at — sets the alpha threshold, the coefficient for deciding whether a pixel gets blurred. Integer in the range [-30,30]. 0 filters the whole image, [0,30] filters flat areas, [-30,0] filters edges. Default is luma_threshold.

And here's the little rule that ties it all together, bestie: 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. Edges? Detected. Slay.

The filter accepts these options:

planes — sets which planes get processed; the unprocessed ones are copied as-is. Default value is 0xf, meaning 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 all the above options as commands. iykyk.

spp 💖

Applies a simple postprocessing filter that compresses and decompresses the image at several shifts (or — when the quality level is 6 — all of them) and averages the results.

The filter accepts these options:

quality — sets the quality. This option defines the number of levels for averaging. It takes an integer in the range 0-6. If set to 0, the filter has no effect, zero, nada. A value of 6 means the higher quality. And bestie, for each increment of that value, the speed drops by a factor of approximately 2. Default 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:
- 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.

Ngl, using this option may cause flicker, since the B-Frames often have a larger QP. Default is 0 (not enabled). 🙈

🌈 Commands 🌈

This filter supports these commands:

quality, level — sets the quality level. The value "max" can be used to set the maximum level, currently 6.

sr 👑

Scales the input by applying one of the super-resolution methods based on convolutional neural networks. Supported models:

Super-Resolution Convolutional Neural Network model (SRCNN). See <https://arxiv.org/abs/1501.00092>.
Efficient Sub-Pixel Convolutional Neural Network model (ESPCN). See <https://arxiv.org/abs/1609.05158>.

Training scripts as well as scripts for model file (.pb) saving can be found at <https://github.com/XueweiMeng/sr/tree/sr_dnn_native>. Original repository is at <https://github.com/HighVoltageRocknRoll/sr.git>. ⭐

The filter accepts these options:

dnn_backend — specifies which DNN backend to use for model loading and execution. Accepts these values:
- tensorflow — TensorFlow backend. To enable it 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 backend can load files for only its format.
scale_factor — sets the scale factor for the SRCNN model. Allowed values are 2, 3 and 4. Default is 2. This factor is necessary for the SRCNN model, because it accepts input upscaled using bicubic upscaling with proper scale factor.

If you want the full functionality (like async execution), please use the dnn_processing filter instead. 💅

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 these 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.

okay so real talk bestie: both video inputs must have the same resolution and pixel format for this filter to work correctly. It also assumes both inputs have the same number of frames, and those frames get compared one by one. 🫶

The filter stores the calculated SSIM of each frame.

The description of the accepted parameters follows.

stats_file, f — if specified, the filter will use the named file to save the SSIM of each individual frame. When filename equals -, the data is sent to standard output.

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.

✨ Examples

For example:

movie=ref_movie.mpg, setpts=PTS-STARTPTS [main];
[main][ref] ssim="stats_file=stats.log" [out]

On this example the input file being processed is compared with the reference file ref_movie.mpg. The SSIM of each individual frame is stored in stats.log.

Another example with both psnr and ssim at same time:

ffmpeg -i main.mpg -i ref.mpg -lavfi  "ssim;[0:v][1:v]psnr" -f null -

Another example with different containers:

ffmpeg -i main.mpg -i ref.mkv -lavfi  "[0:v]settb=AVTB,setpts=PTS-STARTPTS[main];[1:v]settb=AVTB,setpts=PTS-STARTPTS[ref];[main][ref]ssim" -f null -

💫 stereo3d

Convert between different stereoscopic image formats.

The filters accept the following options:

in — set stereoscopic image format of input.

Available values for input image formats are:

sbsl — side by side parallel (left eye left, right eye right)
sbsr — side by side crosseye (right eye left, left eye right)
sbs2l — side by side parallel with half width resolution (left eye left, right eye right)
sbs2r — side by side crosseye with half width resolution (right eye left, left eye right)
abl, tbl — above-below (left eye above, right eye below)
abr, tbr — above-below (right eye above, left eye below)
ab2l, tb2l — above-below with half height resolution (left eye above, right eye below)
ab2r, tb2r — above-below with half height resolution (right eye above, left eye below)
al — alternating frames (left eye first, right eye second)
ar — alternating frames (right eye first, left eye second)
irl — interleaved rows (left eye has top row, right eye starts on next row)
irr — interleaved rows (right eye has top row, left eye starts on next row)
icl — interleaved columns, left eye first
icr — interleaved columns, right eye first

Default value is sbsl.

out — set stereoscopic image format of output.
- sbsl — side by side parallel (left eye left, right eye right)
- sbsr — side by side crosseye (right eye left, left eye right)
- sbs2l — side by side parallel with half width resolution (left eye left, right eye right)
- sbs2r — side by side crosseye with half width resolution (right eye left, left eye right)
- abl, tbl — above-below (left eye above, right eye below)
- abr, tbr — above-below (right eye above, left eye below)
- ab2l, tb2l — above-below with half height resolution (left eye above, right eye below)
- ab2r, tb2r — above-below with half height resolution (right eye above, left eye below)
- al — alternating frames (left eye first, right eye second)
- ar — alternating frames (right eye first, left eye second)
- irl — interleaved rows (left eye has top row, right eye starts on next row)
- irr — interleaved rows (right eye has top row, left eye starts on next row)
- arbg — anaglyph red/blue gray (red filter on left eye, blue filter on right eye)
- argg — anaglyph red/green gray (red filter on left eye, green filter on right eye)
- arcg — anaglyph red/cyan gray (red filter on left eye, cyan filter on right eye)
- arch — anaglyph red/cyan half colored (red filter on left eye, cyan filter on right eye)
- arcc — anaglyph red/cyan color (red filter on left eye, cyan filter on right eye)
- arcd — anaglyph red/cyan color optimized with the least squares projection of dubois (red filter on left eye, cyan filter on right eye)
- agmg — anaglyph green/magenta gray (green filter on left eye, magenta filter on right eye)
- agmh — anaglyph green/magenta half colored (green filter on left eye, magenta filter on right eye)
- agmc — anaglyph green/magenta colored (green filter on left eye, magenta filter on right eye)
- agmd — anaglyph green/magenta color optimized with the least squares projection of dubois (green filter on left eye, magenta filter on right eye)
- aybg — anaglyph yellow/blue gray (yellow filter on left eye, blue filter on right eye)
- aybh — anaglyph yellow/blue half colored (yellow filter on left eye, blue filter on right eye)
- aybc — anaglyph yellow/blue colored (yellow filter on left eye, blue filter on right eye)
- aybd — anaglyph yellow/blue color optimized with the least squares projection of dubois (yellow filter on left eye, blue filter on right eye)
- ml — mono output (left eye only)
- mr — mono output (right eye only)
- chl — checkerboard, left eye first
- chr — checkerboard, right eye first
- icl — interleaved columns, left eye first
- icr — interleaved columns, right eye first
- hdmi — HDMI frame pack

Default value is arcd.

🌈 Examples

Convert input video from side by side parallel to anaglyph yellow/blue dubois:

stereo3d=sbsl:aybd

Convert input video from above below (left eye above, right eye below) to side by side crosseye.

stereo3d=abl:sbsr

🎛️ streamselect, astreamselect

Select video or audio streams.

The filter accepts the following options:

inputs — set number of inputs. Default is 2.
map — set input indexes to remap to outputs.

⌨️ Commands

The streamselect and astreamselect filter supports the following commands:

map — set input indexes to remap to outputs.

✨ Examples

Select first 5 seconds 1st stream and rest of time 2nd stream:

sendcmd='5.0 streamselect map 1',streamselect=inputs=2:map=0

Same as above, but for audio:

asendcmd='5.0 astreamselect map 1',astreamselect=inputs=2:map=0

💬 subtitles

Draw subtitles on top of input video using the libass library.

To enable compilation of this filter you need to configure FFmpeg with --enable-libass. This filter also requires a build with libavcodec and libavformat to convert the passed subtitles file to ASS (Advanced Substation Alpha) subtitles format.

The filter accepts the following options:

filename, f — set the filename of the subtitle file to read. It must be specified.
original_size — specify the size of the original video, the video for which the ASS file was composed. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Due to a misdesign in ASS aspect ratio arithmetic, this is necessary to correctly scale the fonts if the aspect ratio has been changed.
fontsdir — set a directory path containing fonts that can be used by the filter. These fonts will be used in addition to whatever the font provider uses.
alpha — process alpha channel, by default alpha channel is untouched.
charenc — set subtitles input character encoding. subtitles filter only. Only useful if not UTF-8.
stream_index, si — set subtitles stream index. subtitles filter only.
force_style — override default style or script info parameters of the subtitles. It accepts a string containing ASS style format KEY=VALUE couples separated by ,.
wrap_unicode — break lines according to the Unicode Line Breaking Algorithm. Availability requires at least libass release 0.17.0 (or LIBASS_VERSION 0x01600010), and libass must have been built with libunibreak.

The option is enabled by default except for native ASS.

If the first key is not specified, it is assumed that the first value specifies the filename.

For example, to render the file sub.srt on top of the input video, use the command:

subtitles=sub.srt

which is equivalent to:

subtitles=filename=sub.srt

To render the default subtitles stream from file video.mkv, use:

subtitles=video.mkv

To render the second subtitles stream from that file, use:

subtitles=video.mkv:si=1

To make the subtitles stream from sub.srt appear in 80% transparent blue "DejaVu Serif", use:

subtitles=sub.srt:force_style='Fontname=DejaVu Serif,PrimaryColour=&HCCFF0000'

🧚 super2xsai

Scale the input by 2x and smooth using the Super2xSaI (Scale and Interpolate) pixel art scaling algorithm.

Useful for enlarging pixel art images without reducing sharpness. obsessed 💖

🔄 swaprect

Swap two rectangular objects in video.

This filter accepts the following options:

w — set object width.
h — set object height.
x1 — set 1st rect x coordinate.
y1 — set 1st rect y coordinate.
x2 — set 2nd rect x coordinate.
y2 — set 2nd rect y coordinate.

All expressions are evaluated once for each frame.

The all options are expressions containing the following constants:

w, h — the input width and height.
a — same as w / h.
sar — input sample aspect ratio.
dar — input display aspect ratio, it is the same as (w / h) * sar.
n — the number of the input frame, starting from 0.
t — the timestamp expressed in seconds. It's NAN if the input timestamp is unknown.
pos — the position in the file of the input frame, NAN if unknown; deprecated, do not use.

⌨️ Commands

This filter supports the all above options as commands.

`swapuv`

Swaps the U and V planes. That's it. That's the whole filter. 💅

`tblend`

Blends successive video frames together. 🌀

See blend for the deets — iykyk.

`telecine`

Applies the telecine process to your video. ✨

This filter accepts the following options:

first_field — which field goes first:
- top, t — top field first 🥇
- bottom, b — bottom field first. The default value is "top".
pattern — a string of numbers representing the pulldown pattern you wanna apply. The default value is 23.

Some typical patterns, since you're gonna ask:

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. ⏳📊

Unlike the histogram video filter, which only shows the histogram of a single input frame at a certain time, this filter also shows past histograms, for however many frames the width option says.

The computed histogram is a representation of the color component distribution in an image. Fr fr.

The filter accepts the following options:

width, w — sets the width of a single color component output. Default value is 0. A value of 0 means the width will be picked from the input video. This also sets the number of passed histograms to keep. Allowed range is [0, 8192]. 📏
display_mode, d — sets the display mode. It accepts the following values:
- stack — per color component graphs are placed below each other.
- parade — per color component graphs are placed side by side.
- overlay — presents information identical to that in the "parade" mode, 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 for slide are:

frame — draw a new frame when the right border is reached.
replace — replace old columns with new ones.
scroll — scroll from right to left.
rscroll — scroll from left to right.
picture — draw a single picture.

Default is "replace".

`threshold`

Applies a threshold effect to the video stream. 🔲

This filter needs four video streams to do its thresholding thing. The first stream is the stream we're filtering. The second stream holds threshold values, the third stream holds min values, and the last, fourth stream holds max values.

The filter accepts the following option:

planes — sets which planes will be processed; unprocessed planes will be copied. By default the value is 0xf, meaning all planes get processed.

So like, for example: if the first stream's pixel component value is less than the threshold value of that pixel component from the 2nd (threshold) stream, then the third stream's value gets picked — otherwise the fourth stream's pixel component value gets picked.

Using the color source filter, one can perform various types of thresholding:

Commands

This filter supports all options as commands. 💬

Examples

Binary threshold, using gray color as threshold:

ffmpeg -i 320x240.avi -f lavfi -i color=gray -f lavfi -i color=black -f lavfi -i color=white -lavfi threshold output.avi

Inverted binary threshold, using gray color as threshold:

ffmpeg -i 320x240.avi -f lavfi -i color=gray -f lavfi -i color=white -f lavfi -i color=black -lavfi threshold output.avi

Truncate binary threshold, using gray color as threshold:

ffmpeg -i 320x240.avi -f lavfi -i color=gray -i 320x240.avi -f lavfi -i color=gray -lavfi threshold output.avi

Threshold to zero, using gray color as threshold:

ffmpeg -i 320x240.avi -f lavfi -i color=gray -f lavfi -i color=white -i 320x240.avi -lavfi threshold output.avi

Inverted threshold to zero, using gray color as threshold:

ffmpeg -i 320x240.avi -f lavfi -i color=gray -i 320x240.avi -f lavfi -i color=white -lavfi threshold output.avi

`thumbnail`

Picks the most representative frame out of a given sequence of consecutive frames. 🖼️✨ Main character detection, basically.

The filter accepts the following 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".

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. 😭 Be nice to your RAM.

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. 🧩

The untile filter can do the reverse, bestie.

The filter accepts the following options:

layout — sets the grid size in the form COLUMNSxROWS. Range goes 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 (like 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'll get your 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. Don't skip it, seriously. 💀

Display 5 pictures in an area of 3x2 frames, with 7 pixels between them, and 2 pixels of initial margin, using mixed flat and named options:

tile=3x2:nb_frames=5:padding=7:margin=2

`tiltandshift`

Applies a tilt-and-shift effect. 🌀🌄

So like… what even happens when you invert time and space? 🤔

Normally a video is composed of several frames that represent a different instant of time, and shows a scene that evolves in the space captured by the frame. This filter is the antipode of that concept, taking inspiration from tilt and shift photography.

A filtered frame contains the whole timeline of events composing the sequence, and this is obtained by placing a slice of pixels from each frame into a single one. However — since there are no infinite-width frames — this is done up to the width of the input frame, and a video is recomposed by shifting away one column for each subsequent frame. In order to map space to time, the filter tilts each input frame as well, so that motion is preserved. This is accomplished by progressively selecting a different column from each input frame.

The end result is a sort of inverted parallax, so that far-away objects move much faster than the ones in the front. The ideal conditions for this video effect are when there's either very little motion and the background is static, or when there's a lot of motion and a very wide depth of field (e.g. wide panorama, while moving on a train). 🚆✨

The filter accepts the following parameters:

tilt — tilt video while shifting (default). When unset, the video will be sliding a static image, composed of the first column of each frame.
start — what to do at the start of filtering (see below).
end — what to do at the end of filtering (see below).
hold — how many columns should pass through before start of filtering.
pad — how many columns should be inserted before end of filtering.

Normally the filter shifts and tilts from the very first frame, and stops when the last one is received. However, before filtering starts, normal video may be preserved, so that the effect is slowly shifted in its place. Similarly, the last video frame may be reconstructed at the end. Alternatively it's possible to just start and end with black.

none — filtering starts immediately and ends when the last frame is received.
frame — the first frames or the very last frame are kept intact during processing.
black — black is padded at the beginning or at the end of filtering.

`tinterlace`

Performs various types of temporal field interlacing. 🔀

Frames are counted starting from 1, so the first input frame is considered odd. (Iconic numbering choice honestly.)

The filter accepts the following 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

…and that last mode finishes off with this exact output:

11111                           33333
22222                           44444
11111                           33333
22222                           44444
11111                           33333
22222                           44444
11111                           33333
22222                           44444

okay now the rest of the modes, buckle up bestie 💖

`drop_even`, `1`

Only outputs odd frames — the even frames get dropped, and you get a frame with unchanged height but at half frame rate. slay, but make it half.

         ------> 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 mirror twin 🪞 — only outputs even frames, the odd ones get yeeted, 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
                        22222                           44444
                        22222                           44444
                        22222                           44444

`pad`, `3`

Expands each frame to full height, but the alternate lines get padded with black. Result: a frame with double height at the same input frame rate. it's giving letterbox but line-by-line 🖤

         ------> 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`

Interleaves the upper field from odd frames with the lower field from even frames → unchanged height, half frame rate. the way it just ✨merges✨

         ------> time
        Input:
        Frame 1         Frame 2         Frame 3         Frame 4

        11111<-         22222           33333<-         44444
        11111           22222<-         33333           44444<-
        11111<-         22222           33333<-         44444
        11111           22222<-         33333           44444<-

        Output:
        11111                           33333
        22222                           44444
        11111                           33333
        22222                           44444

`interleave_bottom`, `5`

The other way round, bestie: 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 frame rate, height unchanged 👑 Frames get inserted, each one containing the second temporal field from the previous input frame plus the first temporal field from the next input frame. This mode relies on the top_field_first flag. Useful for interlaced video displays with no field synchronisation — iykyk.

         ------> 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`

Moves odd frames into the upper field, even into the lower field → a double height frame at the same frame rate. absolute unit 🫶

         ------> 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 are still accepted for backward compatibility reasons. So like… they work, but don't be that girl.

Default mode is "merge".

`flags`

Specifies flags that influence the filter process 🚩

The available value for flags is:

low_pass_filter, vlpf — enables linear vertical low-pass filtering in the filter. Vertical low-pass filtering is required 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. (iconic behaviour honestly)
complex_filter, cvlpf — enables complex vertical low-pass filtering. This will slightly less reduce interlace 'twitter' and Moire patterning, but better retains detail and the subjective sharpness impression ✨
bypass_il — bypasses already interlaced frames and only adjusts the frame rate.

💀 heads up: vertical low-pass filtering and bypassing already interlaced frames can only be enabled for mode interleave_top and interleave_bottom. no exceptions bestie.

🧚 tmedian 🧚

Picks median pixels from several successive input video frames. it's giving "majority vote but make it pixels" 🗳️

The filter accepts the following options:

radius — sets the radius of the median filter. Default is 1. Allowed range is from 1 to 127.
planes — sets which planes to filter. Default value is 15, and with that all planes get processed.
percentile — sets the median percentile. Default value is 0.5. That default 0.5 will always pick median values, while 0 picks minimum values and 1 picks maximum values.

Commands

This filter supports all above options as commands, excluding option radius. (sorry radius, you're not command-able 💅)

🦋 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 maintaining their dynamics as much as possible. It's useful for e.g. matching exposures from a video frames sequence — no cap, this one actually eats 🔥

This filter accepts the following option:

radius — sets filtering radius. Default is 5. Allowed range is from 1 to 127.
sigma — sets filtering sigma. Default is 0.5. This controls the strength of filtering. Setting this option to 0 effectively does nothing (so don't bother, bestie 🙈).
planes — sets which planes to process. Default is 15, which is all available planes.

🌈 tmix 🌈

Mixes successive video frames 🌀 like a smoothie but for pixels.

A description of the accepted options follows.

frames — the number of successive frames to mix. If unspecified, it defaults to 3.
weights — specifies the weight of each input video frame. Each weight is separated by space. If the number of weights is smaller than the number of frames, the last specified weight will be used for all the remaining unset weights.
scale — specifies the scale. If it's set, it gets multiplied with the sum of each weight multiplied with pixel values, giving the final destination pixel value. By default scale is auto scaled to the sum of weights.
planes — sets which planes to filter. Default is all. Allowed range is from 0 to 15.

Examples

Average 7 successive frames:

tmix=frames=7:weights="1 1 1 1 1 1 1"

Apply simple temporal convolution:

tmix=frames=3:weights="-1 3 -1"

Similar as above but only showing temporal differences:

tmix=frames=3:weights="-1 2 -1":scale=1

Commands

This filter supports the following commands:

weights
scale
planes

Syntax is the same as the option with the same name 💖

🌸 `tonemap` — tone map colors from different dynamic ranges

okay so this one is genuinely iconic for HDR girlies: it tone maps colors from different dynamic ranges 💅

Real talk though — this filter expects data in single precision floating point, because it literally needs to operate on (and can output) out-of-range values. So you need another filter, like zscale, to convert the resulting frame back into a usable format. It's giving relay race 🏃‍♀️

Also important: the tonemapping algorithms implemented only work on linear light, so your input data should be linearized beforehand (and possibly correctly tagged). No skipping steps 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. 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. (she said "not today" 💀)
- 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. iykyk 🫶

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 breather for your video stream. 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 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.

Numerical values are deprecated, and should be dropped in favor of symbolic constants. (she's retired, use the names 💅)

passthrough — Do not apply the transposition if the input geometry matches the one specified by the specified value. It accepts the following values:
- none — Always apply transposition.
- portrait — Preserve portrait geometry (when height >= width).
- landscape — Preserve landscape geometry (when width >= height).

Default value is "none".

For example to rotate by 90 degrees clockwise and preserve portrait layout:

transpose=dir=1:passthrough=portrait

The command above can also be specified as:

transpose=1:portrait

✨ `trim` — trim the input so that the output contains one continuous subpart of the input

It 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.

Note that the first two sets of the start/end options and the duration option look at the frame timestamp, while the _frame variants simply count the frames that pass through the filter. Also note that this filter does not modify the timestamps. If you wish for the output timestamps to start at zero, insert a setpts filter after the trim filter.

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. (fr fr, she's an overachiever 😭)

The defaults are such that all the input is kept. So it is possible to set e.g. just the end values to keep everything before the specified time.

Examples:

Drop everything except the second minute of input:

ffmpeg -i INPUT -vf trim=60:120

Keep only the first second:

ffmpeg -i INPUT -vf trim=duration=1

💜 `unpremultiply` — apply alpha unpremultiply effect to input video stream using first plane of second stream as alpha

Both streams must have same dimensions and same pixel format. (matching outfits only 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.

inplace — Do not require 2nd input for processing, instead use alpha plane from input stream. (solo mode, main character energy 👑)

🔥 `unsharp` — sharpen or blur the input video

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, while positive values sharpen it — and a value of 0 disables the effect entirely, no cap.

Default value is 0.0. 🫶

alpha_msize_x, ax — sets the alpha matrix horizontal size. Gotta be an odd integer between 3 and 23, bestie. Default value is 5. ✨
alpha_msize_y, ay — sets the alpha matrix vertical size. Also an odd integer between 3 and 23. Default value is 5. 💅
alpha_amount, aa — sets the alpha effect strength. Must be a floating point number, and the reasonable values live between -1.5 and 1.5.

Negative values blur the input video, positive values sharpen it, and 0 means no effect at all.

Default value is 0.0.

All the parameters are optional, and they default to the equivalent of the string '5:5:1.0:5:5:0.0'. 🎀

Examples

Apply a strong luma sharpen effect — iconic behaviour fr:

unsharp=luma_msize_x=7:luma_msize_y=7:luma_amount=2.5

Apply a strong blur of both luma and chroma parameters — the way this one goes blurry on purpose:

unsharp=7:7:-2:7:7:-2

untile

Decomposes a video that's made of tiled images into the individual images. Like, un-tiling it. 🧩

The frame rate of the output video is the frame rate of the input video multiplied by the number of tiles — yes really, it speeds up like that.

This filter does the exact 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, go check the "Video size" section in the ffmpeg-utils manual. 📐

Examples

Produce a 1-second video from a still image file made of 25 frames stacked vertically, like an analogic film reel — vintage film-girl energy, obsessed:

ffmpeg -r 1 -i image.jpg -vf untile=1x25 movie.mkv

uspp

Applies the ultra slow/simple postprocessing filter, which compresses and decompresses the image at several shifts (or — in the case of quality level 8 — all of them) and averages the results. 💫

The way this differs from the behaviour 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. iykyk.

Heads up bestie: this filter is not available in ffmpeg versions between 5.0 and 6.0. 😭

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 you set it to 0, the filter just has no effect. A value of 8 means the higher quality. And for each increment of that value, the speed drops by a factor of approximately 2. Default value is 3. 😮‍💨
qp — force a constant quantization parameter. If it's not set, the filter will use the QP from the video stream (if available).
codec — use the specified codec instead of snow.

v360

Converts 360 videos between all the various formats. This one is a whole personality, buckle up. 🌈

The filter accepts the following options:

input
output — set the format of the input/output video.

Available formats:

e, equirect — equirectangular projection.
c3x2, c6x1, c1x6 — cubemap with 3x2/6x1/1x6 layout.

Format specific options:

in_pad
out_pad — set the padding proportion for the input/output cubemap. Values in decimals.

Example values:

0 — no padding.
0.01 — 1% of the face is padding. For example, with 1920x1280 resolution the face size would be 640x640 and the 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 it's greater than zero, it overrides the other padding options.

in_forder
out_forder — set the 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 the 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. Like, normal-normal.

Format specific options:

h_fov
v_fov
d_fov — set the output horizontal/vertical/diagonal field of view. Values in degrees.

If the diagonal field of view is set, it overrides the horizontal and vertical field of view. 👑

ih_fov
iv_fov
id_fov — set the input horizontal/vertical/diagonal field of view. Values in degrees.

If the diagonal field of view is set, it overrides the horizontal and vertical field of view.

dfisheye — dual fisheye.

Format specific options:

h_fov
v_fov
d_fov — set the output horizontal/vertical/diagonal field of view. Values in degrees.

If the diagonal field of view is set, it overrides the horizontal and vertical field of view.

ih_fov
iv_fov
id_fov — set the input horizontal/vertical/diagonal field of view. Values in degrees.

If the diagonal field of view is set, it overrides the horizontal and vertical field of view.

barrel, fb, barrelsplit — Facebook's 360 formats. 🦋
sg — stereographic format.

Format specific options:

h_fov
v_fov
d_fov — set the output horizontal/vertical/diagonal field of view. Values in degrees.

If the diagonal field of view is set, it overrides the horizontal and vertical field of view.

ih_fov
iv_fov
id_fov — set the input horizontal/vertical/diagonal field of view. Values in degrees.

If the diagonal field of view is set, it overrides the horizontal and vertical field of view.

mercator — Mercator format.
ball — ball format, gives significant distortion toward the back.
hammer — Hammer-Aitoff map projection format.
sinusoidal — sinusoidal map projection format.
fisheye — fisheye projection.

Format specific options:

h_fov
v_fov
d_fov — set the output horizontal/vertical/diagonal field of view. Values in degrees.

If the diagonal field of view is set, it overrides the horizontal and vertical field of view.

ih_fov
iv_fov
id_fov — set the input horizontal/vertical/diagonal field of view. Values in degrees.

If the diagonal field of view is set, it overrides the horizontal and vertical field of view.

pannini — Pannini projection.

Format specific options:

h_fov — set the output pannini parameter.
ih_fov — set the input pannini parameter.

cylindrical — cylindrical projection. 💜

Format specific options:

h_fov
v_fov
d_fov — set the output horizontal/vertical/diagonal field of view. Values in degrees.

If the diagonal field of view is set, it overrides the horizontal and vertical field of view.

ih_fov
iv_fov
id_fov — set the input horizontal/vertical/diagonal field of view. Values in degrees.

If the diagonal field of view is set, it overrides the horizontal and vertical field of view.

perspective — perspective projection. (output only)

Format specific options:

v_fov — set perspective parameter.

tetrahedron — tetrahedron projection.
tsp — truncated square pyramid projection.
he, hequirect — half equirectangular projection.
equisolid — equisolid format.

Format specific options:

h_fov
v_fov
d_fov — set the output horizontal/vertical/diagonal field of view. Values in degrees.

If the diagonal field of view is set, it overrides the horizontal and vertical field of view.

ih_fov
iv_fov
id_fov — set the input horizontal/vertical/diagonal field of view. Values in degrees.

If the diagonal field of view is set, it overrides the horizontal and vertical field of view.

og — orthographic format.

Format specific options:

h_fov
v_fov
d_fov — set the output horizontal/vertical/diagonal field of view. Values in degrees.

If the diagonal field of view is set, it overrides the horizontal and vertical field of view.

ih_fov
iv_fov
id_fov — set the input horizontal/vertical/diagonal field of view. Values in degrees.

If the diagonal field of view is set, it overrides the horizontal and vertical field of view.

octahedron — octahedron projection.
cylindricalea — Cylindrical Equal Area projection. ✨

interp — sets the interpolation method. Note: more complex interpolation methods require much more memory to run. You've been warned, bestie. 💀

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 — set the output video resolution. The default resolution depends on formats.
in_stereo
out_stereo — set the input/output stereo format:
- 2d — 2D mono
- sbs — side by side
- tb — top bottom

Default value is 2d for both input and output format.

yaw
pitch
roll — set rotation for the output video. Values are in degrees.
rorder — set 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.

h_flip
v_flip
d_flip — flip the output video horizontally (swaps left-right) / vertically (swaps up-down) / in-depth (swaps back-forward). Boolean values.
ih_flip
iv_flip — set if the input video is flipped horizontally / vertically. Boolean values.
in_trans — set if the input video is transposed. Boolean value, by default disabled.
out_trans — set if the output video needs to be transposed. Boolean value, by default disabled.
h_offset
v_offset — set output horizontal / vertical off-axis offset. Default is set to 0. Allowed range is from -1 to 1.
alpha_mask — build a mask in the alpha plane for all unmapped pixels by marking them fully transparent. Boolean value, by default disabled.
reset_rot — reset the rotation of the output video. Boolean value, by default disabled.

Examples 🎀

Convert equirectangular video to cubemap with 3x2 layout and 1% padding using bicubic interpolation:

ffmpeg -i input.mkv -vf v360=e:c3x2:cubic:out_pad=0.01 output.mkv

Extract back view of Equi-Angular Cubemap:

ffmpeg -i input.mkv -vf v360=eac:flat:yaw=180 output.mkv

Convert transposed and horizontally flipped Equi-Angular Cubemap in side-by-side stereo format to equirectangular top-bottom stereo format:

v360=eac:equirect:in_stereo=sbs:in_trans=1:ih_flip=1:out_stereo=tb

Commands 💅

This filter supports a subset of the above options as commands.

`vaguedenoiser`

Bestie, this one applies a wavelet based denoiser. ✨

So what it does: it transforms each frame from the video input into the wavelet domain, using Cohen-Daubechies-Feauveau 9/7. Then it applies some filtering to the obtained coefficients. Then it does an inverse wavelet transform after. Because wavelet properties are lowkey iconic, it should give you a nice smoothed result, and reduced noise, without blurring picture features. No cap.

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 looks 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 — intermediary between (more) soft and (less) hard thresholding.

Default is garrote.

nsteps — number of times the wavelet will decompose the picture. The picture can't be decomposed beyond a particular point (typically, 8 for a 640x480 frame — as 2^9 = 512 > 480). Valid values are integers between 1 and 32. Default value is 6.
percent — partial of full denoising (limited coefficients shrinking), from 0 to 100. Default value is 85.
planes — a list of the planes to process. By default all planes are processed.
type — the threshold type the filter will use. It accepts the following values:
- universal — threshold used is same for all decompositions.
- bayes — threshold used depends also on each decomposition coefficients.

Default is universal.

`varblur`

Apply a variable blur filter by using the 2nd video stream to set the blur radius. Heads up: the 2nd stream must have the same dimensions. 🫶

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.

`vectorscope`

Display 2 color component values in the two dimensional graph (which is called a vectorscope). Honestly? So satisfying to look at. 🌈

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 the 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.
- 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.
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.
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
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.

`vidstabdetect`

Analyze video stabilization/deshaking. Perform pass 1 of 2, see vidstabtransform for pass 2. 🎬

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. Don't skip this bestie, it will not build otherwise. 💀

This filter accepts the following options:

result — set the path to the file used to write the transforms information. Default value is transforms.trf.
shakiness — set how shaky the video is and how quick the camera is. It accepts an integer in the range 1-10, a value of 1 means little shakiness, a value of 10 means strong shakiness. Default value is 5.
accuracy — set the accuracy of the detection process. It must be a value in the range 1-15. A value of 1 means low accuracy, a value of 15 means high accuracy. Default value is 15.
stepsize — set stepsize of the search process. The region around 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.

show — show fields and transforms in the resulting frames. It accepts an integer in the range 0-2. Default value is 0, which disables any visualization.
fileformat — format for the transforms data file to be written. Acceptable values are:
- ascii — human-readable plain text
- binary — binary format, roughly 40% smaller than ascii. (default)

Examples 🦋

Use default values:

vidstabdetect

Analyze strongly shaky movie and put the results in file mytransforms.trf:

vidstabdetect=shakiness=10:accuracy=15:result="mytransforms.trf"

Visualize the result of internal transformations in the resulting video:

vidstabdetect=show=1

Analyze a video with medium shakiness using ffmpeg:

ffmpeg -i input -vf vidstabdetect=shakiness=5:show=1 dummy.avi

🎀 vidstabtransform

Video stabilization/deshaking: pass 2 of 2, see vidstabdetect for pass 1.

So basically this one reads a file full of transform info for each frame and then applies/compensates them. Pair it with the vidstabdetect filter and you can deshake your videos, bestie. See also <http://public.hronopik.de/vid.stab>. And listen — it is important to also use the unsharp filter, see below. Not optional-optional, like actually important.

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

Options

input — sets the path to the file we read the transforms from. Default value is transforms.trf.
smoothing — sets the number of frames (value*2 + 1) used for lowpass filtering the camera movements. Default value is 10.

So for example a number of 10 means 21 frames get used (10 in the past and 10 in the future) to smoothen the motion in the video. A larger value gives you a smoother video, but it limits the acceleration of the camera (pan/tilt movements). 0 is that special case where a static camera is simulated.

optalgo — sets the camera path optimization algorithm.

Accepted values are:

gauss — gaussian kernel low-pass filter on camera motion (default)
avg — averaging on transformations

maxshift — sets the maximal number of pixels to translate frames. Default value is -1, meaning no limit.
maxangle — sets the maximal angle in radians (degree*PI/180) to rotate frames. Default value is -1, meaning no limit.
crop — specifies how to deal with borders that may be visible due to movement compensation.

Available values are:

keep — keep image information from previous frame (default)
black — fill the border black

invert — invert transforms if set to 1. Default value is 0.
relative — consider transforms as relative to previous frame if set to 1, absolute if set to 0. Default value is 0.
zoom — sets percentage to zoom. A positive value gives you a zoom-in effect, a negative value a zoom-out effect. Default value is 0 (no zoom).
optzoom — sets optimal zooming to avoid borders.

Accepted values are:

0 — disabled
1 — optimal static zoom value is determined (only very strong movements will lead to visible borders) (default)
2 — optimal adaptive zoom value is determined (no borders will be visible), see zoomspeed

Note that the value you give at zoom gets added to the one calculated here. Math!

zoomspeed — sets percent to zoom maximally each frame (enabled when optzoom is set to 2). Range is from 0 to 5, default value is 0.25.
interpol — specifies the 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.

Also use the tripod option of vidstabdetect. Teamwork 💅

debug — increase log verbosity if set to 1. Also the detected global motions get 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. Always. 🫶

Zoom in a bit more and load transform data from a given file:

vidstabtransform=zoom=5:input="mytransforms.trf"

Smoothen the video even more:

vidstabtransform=smoothing=30

🌸 vflip

Flips the input video vertically. That's it, that's the whole vibe. Upside-down era.

For example, to vertically flip a video with ffmpeg:

ffmpeg -i in.avi -vf "vflip" out.avi

🦋 vfrdet

Detects 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 ones with constant delta pts. And if there were frames with variable delta, then it will also show the min, max and average delta encountered. Receipts! 💀

🍓 vibrance

Boost or alter saturation. Iconic little color filter, ngl.

The filter accepts the following options:

intensity — sets the strength of the boost if it's a positive value, or the strength of the alter if it's a negative value. Default is 0. Allowed range is from -2 to 2.
rbal — sets the red balance. Default is 1. Allowed range is from -10 to 10.
gbal — sets the green balance. Default is 1. Allowed range is from -10 to 10.
bbal — sets the blue balance. Default is 1. Allowed range is from -10 to 10.
rlum — sets the red luma coefficient.
glum — sets the green luma coefficient.
blum — sets 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 of the above options as commands. Yes bestie, runtime control, we love to see it. 👑

🔥 vif

Obtains the average VIF (Visual Information Fidelity) between two input videos.

This filter takes two input videos. Two! It's a two-input filter, okay?

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. No skipping ahead, no falling behind. 🙈

The obtained average VIF score gets 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

Makes or reverses a natural vignetting effect. Aesthetic af.

The filter accepts the following options:

angle, a — set lens angle expression as a number of radians.

The value gets 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

Obtains the average VMAF motion score of a video. It's one of the component metrics of VMAF. Little metric, big job.

The obtained average motion score gets 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 the AMD Advanced Media Framework library for hardware acceleration. GPU go brrr. ⚡ 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, the HDR one 💜
smpte428 — SMPTE428
arib-std-b67 — ARIB_STD_B67

🎨 `primaries`

Tell it what the output color primaries are, bestie — this is the "what do pure red/green/blue actually look like" of it all. Accepted values:

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 — the classic glow-up 💅

vpp_amf=-2:720:format=yuv420p

Upscale to 4K and change color profile to bt2020 — main character energy, huge canvas 🦋

vpp_amf=4096:2160:color_profile=bt2020

🥞 `vstack`

Stacks your input videos vertically, one on top of the other, like a little video sandwich 🥪

All streams must be of same pixel format and of same width — no cap, that's a hard requirement.

Fun fact: this filter is faster than using overlay and pad filter to create the same output. Slay, efficiency.

It accepts the following options:

inputs — Set number of input streams. Default is 2.
shortest — If set to 1, force the output to terminate when the shortest input terminates. Default value is 0.

🎞️ `w3fdif`

Deinterlace the input video — and yes, "w3fdif" stands for "Weston 3 Field Deinterlacing Filter", iykyk.

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. Literally BBC-approved deinterlacing, iconic 👑

This filter uses field-dominance information in frame to decide which of each pair of fields to place first in the output. If it gets it wrong use setfield filter before "w3fdif" filter — noted, bestie 🫶

There are two sets of filter coefficients, so called "simple" and "complex". Which set of filter coefficients is used can be set by passing an optional parameter:

filter — Set the interlacing filter coefficients. Accepts one of the following values:
- simple — Simple filter coefficient set.
- complex — More-complex filter coefficient set.

Default value is complex.

mode — The interlacing mode to adopt. It accepts one of the following values:
- frame — Output one frame for each frame.
- field — Output one frame for each field.

The default value is "field".

parity — The picture field parity assumed for the input interlaced video. It accepts one of the following values:
- tff — Assume the top field is first.
- bff — Assume the bottom field is first.
- auto — Enable automatic detection of field parity.

The default value is "auto". If the interlacing is unknown or the decoder does not export this information, top field first will be assumed.

deint — Specify which frames to deinterlace. Accepts one of the following values:
- all — Deinterlace all frames,
- interlaced — Only deinterlace frames marked as interlaced.

Default value is all.

Commands

This filter supports same commands as options — cute and consistent 💖

📈 `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 accepts the following options:

mode, m — Can be either "row", or "column". Default is "column". In row mode, the graph on the left side represents color component value 0 and the right side represents value = 255. In column mode, the top side represents color component value = 0 and bottom side represents value = 255.
intensity, i — Set intensity. Smaller values are useful to find out how many values of the same luminance are distributed across input rows/columns. Default value is 0.04. Allowed range is [0, 1].
mirror, r — Set mirroring mode. 0 means unmirrored, 1 means mirrored. In mirrored mode, higher values will be represented on the left side for "row" mode and at the top for "column" mode. Default is 1 (mirrored).
display, d — Set display mode. It accepts the following values:
- overlay — Presents information identical to that in the "parade", except that the graphs representing color components are superimposed directly over one another. This display mode makes it easier to spot relative differences or similarities in overlapping areas of the color components that are supposed to be identical, such as neutral whites, grays, or blacks.
- stack — Display separate graph for the color components side by side in "row" mode or one below the other in "column" mode.
- parade — Display separate graph for the color components side by side in "column" mode or one below the other in "row" mode. Using this display mode makes it easy to spot color casts in the highlights and shadows of an image, by comparing the contours of the top and the bottom graphs of each waveform. Since whites, grays, and blacks are characterized by exactly equal amounts of red, green, and blue, neutral areas of the picture should display three waveforms of roughly equal width/height. If not, the correction is easy to perform by making level adjustments the three waveforms.

Default is "stack".

components, c — Set which color components to display. Default is 1, which means only luma or red color component if input is in RGB colorspace. If is set for example to 7 it will display all 3 (if) available color components.
envelope, e
- none — No envelope, this is default.
- instant — Instant envelope, minimum and maximum values presented in graph will be easily visible even with small "step" value.
- peak — Hold minimum and maximum values presented in graph across time. This way you can still spot out of range values without constantly looking at waveforms.
- peak+instant — Peak and instant envelope combined together.
filter, f
- lowpass — No filtering, this is default.
- flat — Luma and chroma combined together.
- aflat — Similar as above, but shows difference between blue and red chroma.
- xflat — Similar as above, but use different colors.
- yflat — Similar as above, but again with different colors.
- chroma — Displays only chroma.
- color — Displays actual color value on waveform.
- acolor — Similar as above, but with luma showing frequency of chroma values.
graticule, g — Set which graticule to display.
- none — Do not display graticule.
- green — Display green graticule showing legal broadcast ranges.
- orange — Display orange graticule showing legal broadcast ranges.
- invert — Display invert graticule showing legal broadcast ranges.
opacity, o — Set graticule opacity.
flags, fl — Set graticule flags.
- numbers — Draw numbers above lines. By default enabled.
- dots — Draw dots instead of lines.
scale, s — Set scale used for displaying graticule.
- digital
- millivolts
- ire

Default is digital.

bgopacity, b — Set background opacity.
tint0, t0
tint1, t1

Set tint for output. Only used with lowpass filter and when display is not overlay and input pixel formats are not RGB.

fitmode, fm — Set sample aspect ratio of video output frames. Can be used to configure waveform so it is not stretched too much in one of directions.
- none — Set sample aspect ration to 1/1.
- size — Set sample aspect ratio to match input size of video

Default is none.

input — Set input formats for filter to pick from. Can be all, for selecting from all available formats, or first, for selecting first available format. Default is first.

🪡 `weave`, `doubleweave`

The "weave" takes a field-based video input and join each two sequential fields into single frame, producing a new double height clip with half the frame rate and half the frame count. It's giving loom energy 🧶

The "doubleweave" works same as "weave" but without halving frame rate and frame count.

It accepts the following option:

first_field — Set first field. Available values are:
- top, t — Set the frame as top-field-first.
- bottom, b — Set the frame as bottom-field-first.

🌈 Examples

Interlace video using select and separatefields filter:

separatefields,select=eq(mod(n,4),0)+eq(mod(n,4),3),weave

🔍 `xbr`

Apply the xBR high-quality magnification filter which is designed for pixel art. It follows a set of edge-detection rules, see <https://forums.libretro.com/t/xbr-algorithm-tutorial/123> — pixel art girlies stay winning 🎮✨

It accepts the following option:

n — Set the scaling dimension: 2 for "2xBR", 3 for "3xBR" and 4 for "4xBR". Default is 3.

🧮 `xcorrelate`

Apply normalized cross-correlation between first and second input video stream.

Second input video stream dimensions must be lower than first input video stream — no exceptions, bestie, that's the vibe check ⚠️

The filter accepts the following options:

planes — Set which planes to process.
secondary — Set which secondary video frames will be processed from second input video stream, can be first or all. Default is all.

The "xcorrelate" filter also supports the framesync options — free real estate, we love to see it 🫶

🎬 `xfade`

Apply cross fade from one input video stream to another input video stream. The cross fade is applied for specified duration — smooth transition hours 🌸

Both inputs must be constant frame-rate and have the same resolution, pixel format, frame rate and timebase. All matching, periodt.

The filter accepts the following options:

transition — Set one of available transition effects:

okay so the transition names you can pick from are literally a whole menu — no cap, here’s the full lineup:

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 — she’s the default for a reason 💅

duration — sets cross fade duration in seconds. Range is 0 to 60 seconds. Default duration is 1 second.
offset — sets cross fade start relative to first input stream in seconds. Default offset is 0.
expr — sets expression for custom transition effect.

The expressions can use these variables and functions, bestie:

X, Y — the coordinates of the current sample.
W, H — the width and height of the image.
P — progress of 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 — it’s giving group project, but make it median ✨

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, then 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. iconic metric behavior 🔥

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. All streams must be of same pixel format. pixel format matching is mandatory, bestie — no mixing and matching 🫶

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

No cap: if the inputs are different sizes, you might get gaps or overlaps. Just so you know 💅

`yadif`

Deinterlaces your video, bestie — and yeah, yadif stands for "yet another deinterlacing filter" 😭 the devs really said "we've done this before" and did it again, and honestly? Iconic.

It takes these parameters:

mode — the interlacing mode to adopt. It accepts 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. It accepts 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. (So it just assumes, and we love a confident filter 🫶)

deint — specify which frames to deinterlace. Accepts one of these values:
- 0, all — deinterlace all frames.
- 1, interlaced — only deinterlace frames marked as interlaced.

The default value is all.

`yaepblur`

Blur filter that preserves edges — yaepblur means "yet another edge preserving blur filter" 💀 these names, I swear. 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, actual academic citation, we stan a well-read filter 🎀

It accepts these 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 the same commands as options. Slay ✨

`zoompan`

Applies the Zoom & Pan effect. Main character energy, literally moving the camera around your frame 🔥

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 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. (iykyk, NAN means "no thoughts head empty" here 🙈)
out_time, time, ot — the output timestamp expressed in seconds.
x, y — last calculated x and y position from the x and y expression for the current input frame.
px, py — x and y of the last output frame of the previous input frame, or 0 when there wasn't such a frame yet (first input frame).
zoom — last calculated zoom from the z expression for the current input frame.
pzoom — last calculated zoom of the last output frame of the previous input frame.
duration — number of output frames for the current input frame. Calculated from the d expression for each input frame.
pduration — number of output frames created for the 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 the same time to some spot near the center of the 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 the center of the 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 the center of the 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`

Scales (resizes) the input video using the z.lib library: <https://github.com/sekrit-twc/zimg>. To enable compilation of this filter, you need to configure FFmpeg with --enable-libzimg 🛠️

The zscale filter forces the output display aspect ratio to be the same as the input, by changing the output sample aspect ratio. Sneaky but correct 💅

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. (No crash, no drama, it just stays where it is 🧚)

🦋 CUDA VIDEO FILTERS 🦋

To enable CUDA and/or NPP filters, please refer to the configuration guidelines for CUDA and for CUDA NPP filters. Lowkey a whole setup quest, but we move 🫶

Running CUDA filters requires you to initialize a hardware device and to pass that device to all filters in any filter graph. Non-negotiable, bestie.

-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 a 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

So like, CUDA filters live and breathe exclusively on GPU memory, which means frame data sometimes has to get uploaded (hwupload) to hardware surfaces tied to the right CUDA device before processing, and then downloaded (hwdownload) back to regular memory afterward if you need it there. Whether hwupload or hwdownload is actually necessary is giving "depends on your workflow" energy:

If the input frames are already in GPU memory (like, say, when you're using -hwaccel cuda or -hwaccel_output_format cuda), you do not need to explicitly use hwupload — the data is already chilling in the appropriate memory space. Slay. 💅
If the input frames are in CPU memory (software-decoded frames, or frames that got processed by CPU-based filters), then yeah, you do need hwupload to transfer that data over to GPU memory so CUDA processing can happen.
If the output of the CUDA filters needs more processing by software-based filters, or needs to be saved in a format the GPU-based encoders don't support, then hwdownload is required to haul the data back to CPU memory.

Heads up bestie 🫶 — hwupload uploads data to a surface with the same layout as the software frame, so you may need to drop a format filter immediately before hwupload to make sure the input is in the correct format. Same vibe in reverse: hwdownload may not support all output formats, so an extra format filter might need to be slipped in immediately after hwdownload in the filter graph to keep everything compatible.

CUDA

Here's the rundown on the currently available Nvidia CUDA video filters, bestie ✨

Prerequisites:

Install Nvidia CUDA Toolkit

Note: if FFmpeg detects the Nvidia CUDA Toolkit during configuration, it will enable CUDA filters automatically, no extra flags needed. But if you want to explicitly enable them, use these options:

Configure FFmpeg with --enable-cuda-nvcc --enable-nonfree.
Configure FFmpeg with --enable-cuda-llvm. Bonus requirement: the llvm lib must be installed.

bilateral_cuda

CUDA accelerated bilateral filter — an edge preserving filter, and it's mathematically accurate thanks to that GPU acceleration, no cap 🔥 For the best output quality, use one to one chroma subsampling, i.e. yuv444p format.

The filter accepts the following options:

sigmaS — Sets the sigma of the gaussian function used to calculate the spatial weight, also called sigma space. Allowed range is 0.1 to 512. Default is 0.1.
sigmaR — Sets the sigma of the gaussian function used to calculate the color range weight, also called sigma color. Allowed range is 0.1 to 512. Default is 0.1.
window_size — Sets the window size of the bilateral function to determine how many neighbours to loop on. If the number you enter is even, one will be added automatically (sneaky lil auto-correct 💀). 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

Deinterlaces the input video using the bwdif algorithm, but implemented in CUDA so it can run as part of a GPU accelerated pipeline with nvdec and/or nvenc. Iconic teamwork honestly 👑

It accepts the following parameters:

mode — The interlacing mode to adopt. 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 the following values:
- 0, tff — Assume the top field is first.
- 1, bff — Assume the bottom field is first.
- -1, auto — Enable automatic detection of field parity.

The default value is auto. If the interlacing is unknown or the decoder doesn't export this information, top field first will be assumed.

deint — Specifies which frames to deinterlace. Accepts one of the following values:
- 0, all — Deinterlace all frames.
- 1, interlaced — Only deinterlace frames marked as interlaced.

The default value is all.

chromakey_cuda

CUDA accelerated YUV colorspace color/chroma keying 🌈

This filter works like the normal chromakey filter but operates on CUDA frames. For more details and parameters see chromakey.

Examples

Make all the green pixels in the input video transparent and use it as an overlay for another video:

./ffmpeg \
    -hwaccel cuda -hwaccel_output_format cuda -i input_green.mp4  \
    -hwaccel cuda -hwaccel_output_format cuda -i base_video.mp4 \
    -init_hw_device cuda \
    -filter_complex \
    " \
        [0:v]chromakey_cuda=0x25302D:0.1:0.12:1[overlay_video]; \
        [1:v]scale_cuda=format=yuv420p[base]; \
        [base][overlay_video]overlay_cuda" \
    -an -sn -c:v h264_nvenc -cq 20 output.mp4

Process two software sources, explicitly uploading the frames:

./ffmpeg -init_hw_device cuda=cuda -filter_hw_device cuda \
    -f lavfi -i color=size=800x600:color=white,format=yuv420p \
    -f lavfi -i yuvtestsrc=size=200x200,format=yuv420p \
    -filter_complex \
    " \
        [0]hwupload[under]; \
        [1]hwupload,chromakey_cuda=green:0.1:0.12[over]; \
        [under][over]overlay_cuda" \
    -c:v hevc_nvenc -cq 18 -preset slow output.mp4

colorspace_cuda

CUDA accelerated implementation of the colorspace filter.

Real talk: it is by no means feature complete compared to the software colorspace filter — at the current time it only supports color range conversion between jpeg/full and mpeg/limited range. That's it, that's the whole menu 😭

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. Classic main character move ✨

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, and the second input gets overlaid on top of it.

It accepts the following parameters:

x, y — Set expressions for the x and y coordinates of the overlaid video on the main video.

They can contain the following parameters:

main_w, W — The main input width and height.
main_h, H — The main input width and height.
overlay_w, w — The overlay input width and height.
overlay_h, h — The overlay input width and height.
x, y — The computed values for x and y. They are evaluated for each new frame.
n — The ordinal index of the main input frame, starting from 0.
pos — The byte offset position in the file of the main input frame, NAN if unknown. Deprecated, do not use. 🚫
t — The timestamp of the main input frame, expressed in seconds, NAN if unknown.

Default value is 0 for both expressions.

eval — Set when the expressions for x and y are evaluated. It accepts the following values:
- init — Evaluate expressions once during filter initialization or when a command is processed.
- frame — Evaluate expressions for each incoming frame

Default value is frame.

eof_action — See framesync.
shortest — See framesync.
repeatlast — See framesync.

This filter also supports the framesync options.

pad_cuda

Add paddings to an input video stream using CUDA 🎀

This filter is the CUDA-accelerated version of the pad filter. It accepts the same options and expressions and provides the same core functionality. For a detailed description of the 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. (sorry babes 💀)

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 input to 720p, keeping the aspect ratio and making sure the output is yuv420p. Obsessed with how clean this one is:

scale_cuda=-2:720:format=yuv420p

Upscale to 4K using the nearest neighbour algorithm. Big pixel energy, no smoothing, no thoughts:

scale_cuda=4096:2160:interp_algo=nearest

Don't convert or scale anything at all — just copy every input frame into a freshly allocated one. Sounds pointless, but it's actually a slay: it's useful when your filter + encode chain would otherwise exhaust the decoder's frame pool 💀

scale_cuda=passthrough=0

thumbnail_cuda

Picks the most representative frame out of a whole sequence of consecutive frames, and it does it with CUDA so it's fast. Basically the filter is scrolling through your timeline going "this one. this is the moment." 👑

It accepts these options:

n — sets how many frames go in each batch to analyze. In a set of n frames, the filter picks one of them, then moves on to the next batch of n frames and keeps doing that till the end. Default is 100.

Fr fr, since this filter keeps track of the whole frames sequence, a bigger n value means higher memory usage. So like... don't go wild with a huge value, bestie 😭

Example

Thumbnails get extracted from every n=150-frame batch, one per batch. Then the chosen frames get 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 the input video using the yadif algorithm — but implemented in CUDA, so it can be part of a fully GPU-accelerated pipeline with nvdec and/or nvenc. It's giving teamwork, it's giving efficiency 🫶

It accepts the following parameters:

mode — the interlacing mode to adopt. Takes one of these values:
- 0, send_frame — outputs one frame for each frame.
- 1, send_field — outputs 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 — assumes the top field is first.
- 1, bff — assumes the bottom field is first.
- -1, auto — enables automatic detection of field parity.

The default value is auto. And heads up: if the interlacing is unknown or the decoder doesn't export this info, top field first will be assumed.

deint — specifies which frames to deinterlace. Accepts one of these values:
- 0, all — deinterlaces all frames.
- 1, interlaced — only deinterlaces frames marked as interlaced.

The default value is all.

🎀 CUDA NPP 🎀

Okay so below here is a description of the currently available NVIDIA Performance Primitives (libnpp) video 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 exactly the same way as it does for the scale filter. Same vibes, GPU edition 💅

These 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 the following:
- nn — nearest neighbour.
- linear
- cubic
- cubic2p_bspline — 2-parameter cubic (B=1, C=0)
- cubic2p_catmullrom — 2-parameter cubic (B=0, C=1/2)
- cubic2p_b05c03 — 2-parameter cubic (B=1/2, C=3/10)
- super — supersampling
- lanczos
force_original_aspect_ratio — lets you decrease or increase the output video width or height when necessary to keep the original aspect ratio. Possible values:
- disable — scales the video as specified and disables 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 instance of this option: 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 on the command line makes the output 1280x533. Math ate, honestly 🔥

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.

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. 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. Warning though: the video's aspect ratio may be slightly modified 😬

It's handy if you need a video to fit within or exceed a defined resolution using force_original_aspect_ratio but you also have encoder restrictions on width or height divisibility.

reset_sar — works the same as the identical scale filter option.
eval — specifies when to evaluate the 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.

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 information is unavailable and/or meaningless (for example in 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. It's literally the main character scaling itself to match its fave 🫶

See the scale_npp filter for the 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 these 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 information is unavailable and/or meaningless (for example in 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 ✨

These additional options are accepted:

border_type — the type of sampling to be used at frame borders. One of the following:
- 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, which shares mostly the same options. Flip era 💅

It accepts the following parameters:

dir — specifies 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 — doesn't apply the transposition if the input geometry matches the one specified by the given value. 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 rise 🦋

To enable compilation of these filters you need to configure FFmpeg with --enable-opencl.

Running OpenCL filters means you've gotta initialize a hardware device and pass that device to all filters in any filter graph. It's giving "no device left behind" 🙈

-init_hw_device opencl[=name][:device[,key=value...]] — initialises a new hardware device of type opencl called name, using the given device parameters.
-filter_hw_device name — passes 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 on the second platform and running the avgblur_opencl filter with default parameters on it:

-init_hw_device opencl=gpu:1.0 -filter_hw_device gpu -i INPUT -vf "hwupload, avgblur_opencl, hwdownload" OUTPUT

okay so real talk bestie — OpenCL filters literally cannot reach frame data while it's just chilling in normal memory 😭 so every single frame has to be uploaded first with hwupload onto hardware surfaces that are connected to the appropriate device, used, and then downloaded back into normal memory with hwdownload.

⚠️ heads up: hwupload uploads to a surface with the same layout as the software frame, so you may need to add a format filter immediately before it to get the input into the right format 💅 and hwdownload does not support all formats on the output — it may be necessary to insert an additional format filter immediately following it in the graph to get the output into a supported format. no cap, plan your graph accordingly 🫶

🎀 avgblur_opencl

Apply that average blur filter, bestie ✨

The filter accepts the following options:

sizeX — sets the horizontal radius size. Range is [1, 1024] and the default value is 1.
planes — picks which planes get filtered. Default value is 0xf, which means all planes are processed 💅
sizeY — sets the vertical radius size. Range is [1, 1024] and the default value is 0. If it's zero, the sizeX value gets used instead.

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

Applies a boxblur algorithm to the input video 🦋

It accepts the following parameters:

luma_radius, lr
luma_power, lp
chroma_radius, cr
chroma_power, cp
alpha_radius, ar
alpha_power, ap

A description of the accepted options follows 👇

`luma_radius`, `lr` / `chroma_radius`, `cr` / `alpha_radius`, `ar`

Set an expression for the box radius in pixels used for blurring the corresponding input plane.

The radius value must be a non-negative number, and must not be greater than the value of the expression min(w,h)/2 for the luma and alpha planes, and of min(cw,ch)/2 for the chroma planes 👑

Default value for luma_radius is 2. If not specified, chroma_radius and alpha_radius default to the corresponding value set for luma_radius.

The expressions can contain the following constants:

w, h — the input width and height in pixels.
cw, ch — the input chroma image width and height in pixels.
hsub, vsub — the horizontal and vertical chroma subsample values. For example, for the pixel format yuv422p, hsub is 2 and vsub is 1.

`luma_power`, `lp` / `chroma_power`, `cp` / `alpha_power`, `ap`

Specify how many times the boxblur filter is applied to the corresponding plane.

Default value for luma_power is 2. If not specified, chroma_power and alpha_power default to the corresponding value set for luma_power.

A value of 0 will disable the effect 🙈

Examples

Apply boxblur filter, setting each pixel of the output to the average value of box-radiuses luma_radius, chroma_radius, alpha_radius for each plane respectively. The filter will apply luma_power, chroma_power, alpha_power times onto the corresponding plane. For pixels on the edges of the image, the radius does not extend beyond the image boundaries, and so out-of-range coordinates are not used in the calculations.

Apply a boxblur filter with the luma, chroma, and alpha radius set to 2 and luma, chroma, and alpha power set to 3. The filter will run 3 times with box-radius set to 2 for every plane of the image.

-i INPUT -vf "hwupload, boxblur_opencl=luma_radius=2:luma_power=3, hwdownload" OUTPUT
-i INPUT -vf "hwupload, boxblur_opencl=2:3, hwdownload" OUTPUT

Apply a boxblur filter with luma radius set to 2, luma_power to 1, chroma_radius to 4, chroma_power to 5, alpha_radius to 3 and alpha_power to 7.

For the luma plane, a 2x2 box radius will be run once.

For the chroma plane, a 4x4 box radius will be run 5 times.

For the alpha plane, a 3x3 box radius will be run 7 times.

-i INPUT -vf "hwupload, boxblur_opencl=2:1:4:5:3:7, hwdownload" OUTPUT

🌈 colorkey_opencl

RGB colorspace color keying 💚 literally making colors vanish.

The filter accepts the following options:

color — the color which will be replaced with transparency.
similarity — similarity percentage with the key color. 0.01 matches only the exact key color, while 1.0 matches everything 👀
blend — blend percentage. 0.0 makes pixels either fully transparent, or not transparent at all. Higher values result in semi-transparent pixels, with a higher transparency the more similar the pixels color is to the key color ✨

Examples

Make every semi-green pixel in the input transparent with some slight blending:

-i INPUT -vf "hwupload, colorkey_opencl=green:0.3:0.1, hwdownload" OUTPUT

💜 convolution_opencl

Apply convolution of 3x3, 5x5, 7x7 matrix 🔥 math but make it fashion.

The filter accepts the following options:

0m, 1m, 2m, 3m — set matrix for each plane. Matrix is sequence of 9, 25 or 49 signed numbers. Default value for each plane is 0 0 0 0 1 0 0 0 0.
0rdiv, 1rdiv, 2rdiv, 3rdiv — set multiplier for calculated value for each plane. If unset or 0, it will be sum of all matrix elements. The option value must be a float number greater or equal to 0.0. Default value is 1.0.
0bias, 1bias, 2bias, 3bias — set bias for each plane. This value is added to the result of the multiplication. Useful for making the overall image brighter or darker. The option value must be a float number greater or equal to 0.0. Default value is 0.0.

Examples

Apply sharpen:

-i INPUT -vf "hwupload, convolution_opencl=0 -1 0 -1 5 -1 0 -1 0:0 -1 0 -1 5 -1 0 -1 0:0 -1 0 -1 5 -1 0 -1 0:0 -1 0 -1 5 -1 0 -1 0, hwdownload" OUTPUT

Apply blur:

-i INPUT -vf "hwupload, convolution_opencl=1 1 1 1 1 1 1 1 1:1 1 1 1 1 1 1 1 1:1 1 1 1 1 1 1 1 1:1 1 1 1 1 1 1 1 1:1/9:1/9:1/9:1/9, hwdownload" OUTPUT

Apply edge enhance:

-i INPUT -vf "hwupload, convolution_opencl=0 0 0 -1 1 0 0 0 0:0 0 0 -1 1 0 0 0 0:0 0 0 -1 1 0 0 0 0:0 0 0 -1 1 0 0 0 0:5:1:1:1:0:128:128:128, hwdownload" OUTPUT

Apply edge detect:

-i INPUT -vf "hwupload, convolution_opencl=0 1 0 1 -4 1 0 1 0:0 1 0 1 -4 1 0 1 0:0 1 0 1 -4 1 0 1 0:0 1 0 1 -4 1 0 1 0:5:5:5:1:0:128:128:128, hwdownload" OUTPUT

Apply laplacian edge detector which includes diagonals:

-i INPUT -vf "hwupload, convolution_opencl=1 1 1 1 -8 1 1 1 1:1 1 1 1 -8 1 1 1 1:1 1 1 1 -8 1 1 1 1:1 1 1 1 -8 1 1 1 1:5:5:5:1:0:128:128:0, hwdownload" OUTPUT

Apply emboss:

-i INPUT -vf "hwupload, convolution_opencl=-2 -1 0 -1 1 1 0 1 2:-2 -1 0 -1 1 1 0 1 2:-2 -1 0 -1 1 1 0 1 2:-2 -1 0 -1 1 1 0 1 2, hwdownload" OUTPUT

🌸 erosion_opencl

Apply erosion effect to the video — this filter replaces the pixel by the local(3x3) minimum 🕳️

It accepts the following options:

threshold0, threshold1, threshold2, threshold3 — limit the maximum change for each plane. Range is [0, 65535] and default value is 65535. If 0, plane will remain unchanged.
coordinates — flag which specifies the pixel to refer to. Range is [0, 255] and default value is 255, i.e. all eight pixels are used.

Flags to local 3x3 coordinates region centered on x:

Example

Apply erosion filter with threshold0 set to 30, threshold1 set 40, threshold2 set to 50 and coordinates set to 231, setting each pixel of the output to the local minimum between pixels: 1, 2, 3, 6, 7, 8 of the 3x3 region centered on it in the input. If the difference between input pixel and local minimum is more then threshold of the corresponding plane, output pixel will be set to input pixel - threshold of corresponding plane.

-i INPUT -vf "hwupload, erosion_opencl=30:40:50:coordinates=231, hwdownload" OUTPUT

🦋 deshake_opencl

Feature-point based video stabilization filter 📹✨ bye bye shaky cam.

The filter accepts the following options:

tripod — simulates a tripod by preventing any camera movement whatsoever from the original frame. Defaults to 0.
debug — whether or not additional debug info should be displayed, both in the processed output and in the console.

Note that in order to see console debug output you will also need to pass -v verbose to ffmpeg.

Viewing point matches in the output video is only supported for RGB input.

Defaults to 0.

adaptive_crop — whether or not to do a tiny bit of cropping at the borders to cut down on the amount of mirrored pixels. Defaults to 1.
refine_features — whether or not feature points should be refined at a sub-pixel level. This can be turned off for a slight performance gain at the cost of precision. Defaults to 1.
smooth_strength — the strength of the smoothing applied to the camera path from 0.0 to 1.0. 1.0 is the maximum smoothing strength while values less than that result in less smoothing. 0.0 causes the filter to adaptively choose a smoothing strength on a per-frame basis. Defaults to 0.0.
smooth_window_multiplier — controls the size of the smoothing window (the number of frames buffered to determine motion information from). The size of the smoothing window is determined by multiplying the framerate of the video by this number. Acceptable values range from 0.1 to 10.0. Larger values increase the amount of motion data available for determining how to smooth the camera path, potentially improving smoothness, but also increase latency and memory usage. Defaults to 2.0.

Examples

Stabilize a video with a fixed, medium smoothing strength:

-i INPUT -vf "hwupload, deshake_opencl=smooth_strength=0.5, hwdownload" OUTPUT

Stabilize a video with debugging (both in console and in rendered video):

-i INPUT -filter_complex "[0:v]format=rgba, hwupload, deshake_opencl=debug=1, hwdownload, format=rgba, format=yuv420p" -v verbose OUTPUT

💀 dilation_opencl

Apply dilation effect to the video 🌟 this one replaces the pixel by the local(3x3) maximum.

It accepts the following options:

threshold0 — limits the max change for plane 0. Range: [0, 65535], default 65535. If it’s 0, that plane stays completely unchanged. 💅
threshold1 — same vibe for plane 1: range [0, 65535], default 65535; if 0, plane stays unchanged.
threshold2 — plane 2’s max-change cap. Range [0, 65535], default 65535; 0 means no change.
threshold3 — plane 3’s max-change cap. Range [0, 65535], default 65535; 0 means no change.
coordinates — flag that says which pixel to refer to. Range is [0, 255], default is 255, i.e. all eight pixels are used. 🎀

Flags map to a local 3x3 coordinate region centered on x:

🦋 Example

Apply dilation filter with threshold0 set to 30, threshold1 set 40, threshold2 set to 50 and coordinates set to 231, setting each pixel of the output to the local maximum between pixels: 1, 2, 3, 6, 7, 8 of the 3x3 region centered on it in the input. If the difference between input pixel and local maximum is more then threshold of the corresponding plane, output pixel will be set to input pixel + threshold of corresponding plane. ✨

-i INPUT -vf "hwupload, dilation_opencl=30:40:50:coordinates=231, hwdownload" OUTPUT

💖 `nlmeans_opencl`

Non-local Means denoise filter through OpenCL — this filter accepts the same options as nlmeans. No cap, she’s just borrowing the fit. 👑

🌸 `overlay_opencl`

Overlay one video on top of another. It takes two inputs and has one output. The first input is the "main" video on which the second input is overlaid. This filter requires same memory layout for all the inputs, so format conversion may be needed. Lowkey important, bestie. 🫶

The filter accepts the following options:

x — set the x coordinate of the overlaid video on the main video. Default value is 0.
y — set the y coordinate of the overlaid video on the main video. Default value is 0.

🌈 Examples

Overlay an image LOGO at the top-left corner of the INPUT video. Both inputs are yuv420p format.

-i INPUT -i LOGO -filter_complex "[0:v]hwupload[a], [1:v]format=yuv420p, hwupload[b], [a][b]overlay_opencl, hwdownload" OUTPUT

The inputs have same memory layout for color channels, the overlay has additional alpha plane, like INPUT is yuv420p, and the LOGO is yuva420p.

-i INPUT -i LOGO -filter_complex "[0:v]hwupload[a], [1:v]format=yuva420p, hwupload[b], [a][b]overlay_opencl, hwdownload" OUTPUT

🧚 `pad_opencl`

Add paddings to the input image, and place the original input at the provided x, y coordinates. It’s giving “make room, main character incoming.” 💅

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 is used for the output. The width expression can reference the value set by the height expression, and vice versa. The default value of width and height is 0.
x, y — specify the offsets to place the input image at within the padded area, with respect to the top/left border of the output image. The x expression can reference the value set by the y expression, and vice versa. The default value of x and y is 0. If x or y evaluate to a negative number, they’ll be changed so the input image is centered on the padded area. 🎀
color — specify the color of the padded area. For the syntax of this option, check the "Color" section in the ffmpeg-utils manual.
aspect — pad to an aspect instead to a resolution.

The value for the width, height, x, and y options are expressions containing the following constants:

in_w, in_h — the input video width and height.
iw, ih — these are the same as in_w and in_h.
out_w, out_h — the output width and height (the size of the padded area), as specified by the width and height expressions.
ow, oh — these are the same as out_w and out_h.
x, y — the x and y offsets as specified by the x and y expressions, or NAN if not yet specified.
a — same as iw / ih.
sar — input sample aspect ratio.
dar — input display aspect ratio, it is the same as (iw / ih) * sar.

🔥 `prewitt_opencl`

Apply the Prewitt operator (https://en.wikipedia.org/wiki/Prewitt_operator) to input video stream. Edge-detection slay, no cap. ⭐

The filter accepts the following option:

planes — set which planes to filter. Default value is 0xf, by which all planes are processed.
scale — set value which will be multiplied with filtered result. Range is [0.0, 65535] and default value is 1.0.
delta — set value which will be added to filtered result. Range is [-65535, 65535] and default value is 0.0.

🍓 Example

Apply the Prewitt operator with scale set to 2 and delta set to 10.

-i INPUT -vf "hwupload, prewitt_opencl=scale=2:delta=10, hwdownload" OUTPUT

👑 `program_opencl`

Filter video using an OpenCL program. Custom kernel realness. 💜

source — OpenCL program source file.
kernel — kernel name in program.
inputs — number of inputs to the filter. Defaults to 1.
size, s — size of output frames. Defaults to the same as the first input.

The program_opencl filter also supports the framesync options.

The program source file must contain a kernel function with the given name, which will be run once for each plane of the output. Each run on a plane gets enqueued as a separate 2D global NDRange with one work-item for each pixel to be generated. The global ID offset for each work-item is therefore the coordinates of a pixel in the destination image. ✨

The kernel function needs to take the following arguments:

Destination image, __write_only image2d_t. This image will become the output; the kernel should write all of it.
Frame index, unsigned int. This is a counter starting from zero and increasing by one for each frame.
Source images, __read_only image2d_t. These are the most recent images on each input. The kernel may read from them to generate the output, but they can’t be written to.

🧚 Example programs

Copy the input to the output (output must be the same size as the input).

__kernel void copy(__write_only image2d_t destination,
                   unsigned int index,
                   __read_only  image2d_t source)
{
    const sampler_t sampler = CLK_NORMALIZED_COORDS_FALSE;

    int2 location = (int2)(get_global_id(0), get_global_id(1));

    float4 value = read_imagef(source, sampler, location);

    write_imagef(destination, location, value);
}

Apply a simple transformation, rotating the input by an amount increasing with the index counter. Pixel values are linearly interpolated by the sampler, and the output need not have the same dimensions as the input.

__kernel void rotate_image(__write_only image2d_t dst,
                           unsigned int index,
                           __read_only  image2d_t src)
{
    const sampler_t sampler = (CLK_NORMALIZED_COORDS_FALSE |
                               CLK_FILTER_LINEAR);

    float angle = (float)index / 100.0f;

    float2 dst_dim = convert_float2(get_image_dim(dst));
    float2 src_dim = convert_float2(get_image_dim(src));

    float2 dst_cen = dst_dim / 2.0f;
    float2 src_cen = src_dim / 2.0f;

    int2   dst_loc = (int2)(get_global_id(0), get_global_id(1));

    float2 dst_pos = convert_float2(dst_loc) - dst_cen;
    float2 src_pos = {
        cos(angle) * dst_pos.x - sin(angle) * dst_pos.y,
        sin(angle) * dst_pos.x + cos(angle) * dst_pos.y
    };
    src_pos = src_pos * src_dim / dst_dim;

    float2 src_loc = src_pos + src_cen;

    if (src_loc.x < 0.0f      || src_loc.y < 0.0f ||
        src_loc.x > src_dim.x || src_loc.y > src_dim.y)
        write_imagef(dst, dst_loc, 0.5f);
    else
        write_imagef(dst, dst_loc, read_imagef(src, sampler, src_loc));
}

Blend two inputs together, with the amount of each input used varying with the index counter.

__kernel void blend_images(__write_only image2d_t dst,
                           unsigned int index,
                           __read_only  image2d_t src1,
                           __read_only  image2d_t src2)
{
    const sampler_t sampler = (CLK_NORMALIZED_COORDS_FALSE |
                               CLK_FILTER_LINEAR);

    float blend = (cos((float)index / 50.0f) + 1.0f) / 2.0f;

    int2  dst_loc = (int2)(get_global_id(0), get_global_id(1));
    int2 src1_loc = dst_loc * get_image_dim(src1) / get_image_dim(dst);
    int2 src2_loc = dst_loc * get_image_dim(src2) / get_image_dim(dst);

    float4 val1 = read_imagef(src1, sampler, src1_loc);
    float4 val2 = read_imagef(src2, sampler, src2_loc);

    write_imagef(dst, dst_loc, val1 * blend + val2 * (1.0f - blend));
}

💀 `remap_opencl`

Remap pixels using 2nd: Xmap and 3rd: Ymap input video stream. It’s giving “pixel teleportation.” 🦋

Destination pixel at position (X, Y) will be picked from source (x, y) position where x = Xmap(X, Y) and y = Ymap(X, Y). If mapping values are out of range, zero value for pixel will be used for destination pixel.

Xmap and Ymap input video streams must be of same dimensions. Output video stream will have Xmap/Ymap video stream dimensions. Xmap and Ymap input video streams are 32bit float pixel format, single channel. 🫶

interp — specify interpolation used for remapping of pixels. Allowed values are "near" and "linear". Default value is "linear".
fill — specify the color of the unmapped pixels. For the syntax of this option, check the "Color" section in the ffmpeg-utils manual. Default color is "black".

🌟 `roberts_opencl`

Apply the Roberts cross operator (https://en.wikipedia.org/wiki/Roberts_cross) to input video stream. Another icon behavior. 💅

The filter accepts the following option:

planes — set which planes to filter. Default value is 0xf, by which all planes are processed.
scale — set value which will be multiplied with filtered result. Range is [0.0, 65535] and default value is 1.0.
delta — set value which will be added to filtered result. Range is [-65535, 65535] and default value is 0.0.

🎀 Example

Apply the Roberts cross operator with scale set to 2 and delta set to 10.

-i INPUT -vf "hwupload, roberts_opencl=scale=2:delta=10, hwdownload" OUTPUT

💅 sobel_opencl

Okay so this one literally applies the Sobel operator (<https://en.wikipedia.org/wiki/Sobel_operator>) to your input video stream. Edge detection core, iconic, no notes.

The filter accepts the following option:

planes — sets which planes get filtered. Default value is 0xf, which means all the planes get processed. Slay.
scale — sets the value that gets multiplied with the filtered result. Range is [0.0, 65535] and default value is 1.0.
delta — sets the value that gets added to the filtered result. Range is [-65535, 65535] and default value is 0.0.

🎀 Example

Apply 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 performs HDR(PQ/HLG) to SDR conversion with tone-mapping. It's giving "make your HDR footage actually watchable on a normal screen" energy 💖

It accepts the following parameters:

tonemap — specify the tone-mapping operator to be used. Same as the tonemap option in tonemap.
param — tune the tone mapping algorithm. Same as the param option in tonemap.
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 value is 0.5, and the algorithm here is a little different from the cpu version tonemap currently. A setting of 0.0 disables this option.

threshold — the tonemapping algorithm parameters is fine-tuned per each scene. And a threshold is used to detect whether the scene has changed or not. If the distance between the current frame average brightness and the current running average exceeds a threshold value, we would re-calculate scene average and peak brightness. The default value 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.

🦋 Example

Convert HDR(PQ/HLG) video to bt2020-transfer-characteristic p010 format using 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 the input video. Your one-stop glow-up (or soft-focus) shop, bestie 💅

It accepts the following parameters:

luma_msize_x, lx — set the luma matrix horizontal size. Range is [1, 23] and default value is 5.
luma_msize_y, ly — set the luma matrix vertical size. Range is [1, 23] and default value is 5.
luma_amount, la — set the luma effect strength. Range is [-10, 10] and default value is 1.0.

Negative values will blur the input video, while positive values will sharpen it, a value of zero will disable the effect.

chroma_msize_x, cx — set the chroma matrix horizontal size. Range is [1, 23] and default value is 5.
chroma_msize_y, cy — set the chroma matrix vertical size. Range is [1, 23] and default value is 5.
chroma_amount, ca — set the chroma effect strength. Range is [-10, 10] and default value is 0.0.

Negative values will blur the input video, while positive values will sharpen it, a value of zero will disable the effect.

All parameters are optional and default to the equivalent of the string '5:5:1.0:5:5:0.0'. So if you just raw-dog it with no options, that's the vibe you're getting 🫶

🌸 Examples

Apply 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 with custom transition effect by using OpenCL. Two videos walk in, one video walks out — main character energy fr fr 👑

It accepts the following options:

transition — set one of possible transition effects.
- custom — select custom transition effect, the actual transition description will be picked from source and kernel options.
- fade
- wipeleft
- wiperight
- wipeup
- wipedown
- slideleft
- slideright
- slideup
- slidedown

Default transition is fade.

source — OpenCL program source file for custom transition.
kernel — set name of kernel to use for custom transition from program source file.
duration — set duration of video transition.
offset — set time of start of transition relative to first video.

The kernel function needs to take the following arguments:

Destination image, __write_only image2d_t.

This image will become the output; the kernel should write all of it.

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 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);
}

✨ VAAPI VIDEO FILTERS ✨

Okay bestie, so VAAPI Video filters are usually used with a VAAPI decoder and a VAAPI encoder. Below is the description of VAAPI video filters. 💅

To enable compilation of these filters you need to configure FFmpeg with --enable-vaapi.

To use vaapi filters, you need to setup the vaapi device correctly. For more information, please read https://trac.ffmpeg.org/wiki/Hardware/VAAPI.

💖 overlay_vaapi

Overlay one video on the 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. Main character energy fr fr.

The filter accepts the following options:

x, y — Set expressions for the x and y coordinates of the overlaid video on the main video. Default value is 0 for both expressions.
w, h — Set expressions for the width and height of the overlaid video on the main video. Default values are overlay_iw for w and overlay_ih*w/overlay_iw for h.

The expressions can contain the following parameters:

main_w, W — The main input width.
main_h, H — The main input height.
overlay_iw — The overlay input width.
overlay_ih — The overlay input height.
overlay_w, w — The overlay output width.
overlay_h, h — The overlay output height.
overlay_x, x — Position of the overlay layer inside of main.
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.
eof_action — See framesync.
shortest — See framesync.
repeatlast — See framesync.

This filter also supports the framesync options.

✨ Examples ✨

Top-left LOGO overlay: 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

Offset overlay at (200, 100): 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. We love a format with range.

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 sets 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 ✨

HDR10 to p010: Convert HDR(HDR10) video to bt2020-transfer-characteristic p010 format.

tonemap_vaapi=format=p010:t=bt2020-10

HDR to HDR: Convert HDR video to HDR video.

tonemap_vaapi=display=7500\ 3000|34000\ 16000|13250\ 34500|15635\ 16450|500\ 10000000

🦋 hstack_vaapi

Stack input videos horizontally.

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. Side-by-side slay.

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.

This is the VA-API variant of the vstack filter. Each input stream may have different width, and this filter will scale down/up each input stream while keeping the original aspect.

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 VA-API variant of the xstack filter. Each input stream may have different size, and this filter will scale down/up each input stream to the given output size, or the size of the first input stream. Custom layout = main character energy.

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 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 is used for the output. The width expression can reference the value set by the height expression, and vice versa. The default value of width and height is 0.
x, y — Specify the offsets to place the input image at within the padded area, with respect to the top/left border of the output image. The x expression can reference the value set by the y expression, and vice versa. The default value of x and y is 0. If x or y evaluate to a negative number, they'll be changed so the input image is centered on the padded area.
color — Specify the color of the padded area. For the syntax of this option, check the Color section in the ffmpeg-utils manual.
aspect — Pad to an aspect instead to a resolution.

The value for the width, height, x, and y options are expressions containing the following constants:

in_w, in_h — The input video width and height.
iw, ih — These are the same as in_w and in_h.
out_w, out_h — The output width and height (the size of the padded area), as specified by the width and height expressions.
ow, oh — These are the same as out_w and out_h.
x, y — The x and y offsets as specified by the x and y expressions, or NAN if not yet specified.
a — same as iw / ih.
sar — input sample aspect ratio.
dar — input display aspect ratio, it is the same as (iw / ih) * sar.

💜 drawbox_vaapi

Draw a colored box on the input image. Arts and crafts era.

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 and 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 — The input height.
in_w, iw — The input width.
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 ✨

Black border box: Draw a black box around the edge of the input image.

drawbox

Red box with 50% opacity: 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

Pink filled box: Fill the box with pink color.

drawbox=x=10:y=10:w=100:h=100:color=pink@0.5:t=fill

Red 2.40:1 mask: Draw a 2-pixel red 2.40:1 mask.

drawbox=x=-t:y=0.5*(ih-iw/2.4)-t:w=iw+t*2:h=iw/2.4+t*2:t=2:c=red

💻 VIDEOTOOLBOX VIDEO FILTERS 💻

Okay bestie, so this is the little crew of VideoToolbox video filters we've got rn — here's the rundown on each one 🎀

The VideoToolbox filter depends on the VideoToolbox framework, and it gets auto-detected when you're building ffmpeg for an Apple platform like macOS. If autodetect is disabled for some reason, just add --enable-videotoolbox to configure and you're golden ✨

🦋 scale_vt

This one scales and converts the color parameters using VTPixelTransferSession — two jobs, one filter, we love an efficient queen 👑

The filter accepts the following options:

w
h — Set the output video dimension expression. Default value is the input dimension. (So if you don't say anything, it just keeps whatever size you gave it, no drama 💅)
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, look at her go — HDR in, SDR out, half the size, audio copied straight through, hvc1 tag slapped on at the end. Iconic behavior fr fr 🔥

🧚 transpose_vt

This one transposes rows with columns in the input video, and optionally flips it too. For more in-depth examples, go peek at the transpose video filter — it shares mostly the same options, so you already kinda know her 🫶

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) — yeah she's the default, don't come for her 💀
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 the passthrough option is the "respect the vibe of the original" switch — set it and she'll skip the whole transpose if your geometry already matches. Very considerate, very demure ✨

🔥 VULKAN VIDEO FILTERS 🔥

okay bestie, buckle up — here's the rundown on all the Vulkan video filters we currently ship. GPU go brrr 💜

So to even compile these filters you gotta configure FFmpeg with --enable-vulkan and either --enable-libglslang or --enable-libshaderc. Both, like, pick one of the last two, but you need the vulkan one regardless. No cap. ✨

And running Vulkan filters means you have to initialize a hardware device and then pass that device to all the filters in any filter graph. Every single one. The whole squad. 🫶

-init_hw_device vulkan[=name][:device[,key=value...]] — Initialises a shiny new hardware device of type vulkan called name, using the device parameters and options you give it in key=value. Here's what's supported:
- debug — Switches validation layers on if set to 1.
- linear_images — Allocates linear images. Nope, does not apply to decoding.
- disable_multiplane — Disables multiplane images. Also does not apply to decoding.
- avoid_host_import — Avoids using dynamic host memory imports, in favor of a regular memcpy() into a previously mapped buffer.
-filter_hw_device name — Passes the hardware device called name to all filters in any filter graph.

For way more detailed info, go peek at <https://www.ffmpeg.org/ffmpeg.html#Advanced-Video-options> 👑

💖 Example of choosing the first device and running nlmeans_vulkan with default parameters on it:

-init_hw_device vulkan=vk:0 -filter_hw_device vk -i INPUT -vf "hwupload,nlmeans_vulkan,hwdownload" OUTPUT

Here's the tea though: Vulkan filters cannot 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. Important detail: hwupload will upload to a frame with the same layout as the software frame, so you may need to add a format filter immediately before it to get your input into the right format. And hwdownload does not support all formats on the output — usually you gotta insert an additional format filter immediately following in the graph to get the output into a supported format. Annoying but iconic, that's just how it is 💅

✨ `avgblur_vulkan`

Applies an average blur filter, implemented on the GPU using Vulkan.

It accepts the following options:

sizeX — Sets the horizontal radius size. Range is [1, 32], default value is 3.
sizeY — Sets the vertical radius size. Range is [1, 32], default value is 3.
planes — Sets which planes to filter. Default value is 0xf, which means all planes get processed. 🌀

🦋 `blend_vulkan`

Blends two Vulkan frames into each other. Literally mixing vibes.

The blend filter takes two input streams and outputs one stream — first input is the "top" layer and second input is the "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 — Sets the blend mode for a specific pixel component, or for all pixel components in the case of all_mode. Default value is "normal".

Available values for component modes are:

normal
multiply

🌸 `bwdif_vulkan`

A deinterlacer using bwdif, the "Bob Weaver Deinterlacing Filter" algorithm, implemented on the GPU using Vulkan. Look at that name, so main character 💫

It accepts the following parameters:

mode — The interlacing mode to adopt. It accepts one of these:
- 0, send_frame — Output one frame for each frame.
- 1, send_field — Output one frame for each field.

Default value is "send_field".

parity — The picture field parity assumed for the input interlaced video. Accepts one of:
- 0, tff — Assume the top field is first.
- 1, bff — Assume the bottom field is first.
- -1, auto — Enables automatic detection of field parity.

Default value is "auto". Heads up: if the interlacing is unknown or the decoder doesn't export this information, top field first will be assumed.

deint — Specifies which frames to deinterlace. Accepts one of:
- 0, all — Deinterlace all frames.
- 1, interlaced — Only deinterlace frames marked as interlaced.

Default value is "all".

🎀 `chromaber_vulkan`

Applies an effect that emulates chromatic aberration. Works best with RGB inputs, but honestly gives a similar vibe with YCbCr inputs too.

dist_x — Horizontal displacement multiplier. Each chroma pixel's position gets multiplied by this amount, starting from the center of the image. Default is 0.
dist_y — Same energy, but this sets the vertical displacement multiplier. Default is 0.

🌈 `color_vulkan`

A video source that creates a Vulkan frame of a solid color. Super useful for benchmarking, or overlaying. 💯

It accepts the following parameters:

color — The color to use. Either a name, or a hexadecimal value. Default value is "black".
size — The size of the output frame. Default value is "1920x1080".
rate — The framerate to output at. Default value is 60 frames per second.
duration — The video duration. Default value is -0.000001.
sar — The video signal aspect ratio. Default value is "1/1".
format — The pixel format of the output Vulkan frames. Default value is "yuv444p".
out_range — Sets the output YCbCr sample range.

This lets you override the autodetected value, and also lets you 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).

🧚 `vflip_vulkan`

Flips an image vertically. Upside down bestie 🙃

🧚 `hflip_vulkan`

Flips an image horizontally. Mirror mirror 🪞

🧚 `flip_vulkan`

Flips an image along both the vertical and horizontal axis. Double flip, double slay.

💜 `gblur_vulkan`

Applies a Gaussian blur filter on Vulkan frames. Soft focus aesthetic ✨

The filter accepts the following options:

sigma — Sets horizontal sigma, standard deviation of Gaussian blur. Default is 0.5.
sigmaV — Sets vertical sigma; if negative it will be the same as sigma. Default is -1.
planes — Sets which planes to filter. By default all planes are filtered.
size — Sets the kernel size along the horizontal axis. Default is 19.
sizeV — Sets the kernel size along the vertical axis. Default is 0, which sets it to use the same value as size.

👑 `nlmeans_vulkan`

Denoises frames using the Non-Local Means algorithm, implemented on the GPU using Vulkan. It supports more pixel formats than nlmeans or nlmeans_opencl, including alpha channel support. Obsessed with the range honestly 🔥

The filter accepts the following options:

s — Sets denoising strength for all components. Default is 1.0. Must be in range [0.0, 100.0].
p — Sets patch size for all planes. Default is 7. Must be an odd number in range [0, 99].
r — Sets research size. Default is 15. Must be an odd number in range [0, 99].
t — Sets parallelism. Default is 8. Must be a number in the range [1, 64]. Larger values will use more VRAM but may not actually result in greater speed — the optimal value is hardware and input dependent, so like, don't just crank it 😭
s0
s1
s2
s3 — Sets denoising strength for a specific component. Default is 1.0, equal to s. Must be in range [0.0, 100.0]. Setting 0.0 disables denoising in that component.
p0
p1
p2
p3 — Sets patch size for a specific component. Default is 7, equal to p. Must be an odd number in range [0, 99].

🍓 `overlay_vulkan`

Overlays one video on top of another.

It takes two inputs and has one output. The first input is the "main" video, and the second input is overlaid on top of it. Lowkey important: this filter requires all inputs to use the same pixel format. So, format conversion may be needed. Plan ahead bestie 🫶

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.

🌟 `transpose_vulkan`

Transposes rows with columns in the input video, and optionally flips it. For more in-depth examples check out the transpose video filter, which shares mostly the same options. iykyk 💫

It accepts the following parameters:

dir — Specifies 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 — Does 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).

✨ QSV VIDEO FILTERS ✨

okay so these are the QSV filters that are currently available, bestie 💅

to even compile these filters you gotta configure FFmpeg with --enable-libmfx or --enable-libvpl. like, both exist, pick your fighter 🫶

and to actually use QSV filters you need to set up the QSV device correctly first. for more info pls go read <https://trac.ffmpeg.org/wiki/Hardware/QuickSync> — it's giving required reading, ngl 📚

🎀 hstack_qsv

Stacks input videos horizontally. literally side-by-side vibes ✨

This is the QSV variant of the hstack filter, and the way it works: each input stream may have a different height, and this filter will scale down/up each input stream while keeping the original aspect. no squishing, we respect the aspect ratio 👑

It accepts the following options:

inputs — see hstack. iykyk
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

Stacks input videos vertically. stacking upward, main character energy going down the screen 💜

This is the QSV 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_qsv

Stack video inputs into custom layout. yes custom, you're the layout designer now 💅

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. like so:

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. ok so these are where your video comes from, the origin story if you will 🧚

💖 buffer

Buffer video frames, and make them available to the filter chain. it's literally the bridge ok 🫶

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

fr fr don't use that flat string version, it's giving legacy 😭

⭐ cellauto

Create a pattern generated by an elementary cellular automaton. yes it's math, yes it's gorgeous, no I will not elaborate 💅✨

The initial state of the cellular automaton can be defined through the filename and pattern options. If such options are not specified an initial state is created randomly. chaos as a treat 🎲

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. apple girlies rise 🍎

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. no CPU round-trip, we stay on the GPU, slay 💅

🎀 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. iconic, honestly 👑

💀 ddagrab

Captures the Windows Desktop via Desktop Duplication API. windows besties this one's for you 🪟

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.

literally cannot stress this enough — hardware frames only. want software processing? hwdownload first or you're getting nothing 😭

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. no cap, your custom game cursor is coming along whether you like it or not 🙈
framerate — the max framerate your desktop gets captured at. The gap between frames will never be smaller than the inverse of this value, so it's a ceiling, not a promise 💅. When dup_frames is true (which it is by default) and your desktop isn't updating fast enough, the filter just duplicates the last frame to keep things moving. Heads up tho: there's no background buffering happening, so if the filter doesn't get polled often enough, the actual inter-frame interval can end up way bigger than you'd expect 😭.

Defaults to 30 FPS.

video_size — pick the size of the captured video.

Defaults to the full size of the screen.

Cropped from the bottom/right if it's smaller than the screen size.

offset_x — horizontal offset of the captured video.
offset_y — vertical offset of the captured video.
output_fmt — the filter output format you want. Defaults to 8 Bit BGRA.

It accepts these values:

auto — passes all supported output formats to DDA and just returns whatever DDA decides to use.
8bit, bgra — 8 Bit formats always work, and DDA will convert to them if it has to.
10bit, x2bgr10 — filter initialization will fail if you request 10 bit but it's not available.

dup_frames — when this is set to true (the default), the filter duplicates frames whenever the desktop hasn't updated, so it can keep approximately constant target framerate 🎯. When it's set to false, the filter waits for the desktop to update instead (inter-frame intervals may vary a lot in that 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. This one also demonstrates downloading the frame and encoding with libx264. Explicit output format specification is required in this case:

ffmpeg -filter_complex ddagrab=output_idx=1:framerate=60,hwdownload,format=bgra -c:v libx264 -crf 18 output.mp4

If you want to capture only a subsection of the desktop, that's doable 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 the Windows.Graphics.Capture API ✨.

This source gives you low overhead capture of application windows or entire monitors. The filter outputs hardware frames in d3d11 format; use hwdownload,format= if you need system memory frames.

Which window gets captured can be picked 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 the expressions you provide to be selected. The first matching window gets picked, in whatever order Windows hands them back to us 🙃.

Explicit handles (hwnd, hmonitor) override pattern or index based selection. If you give neither handles nor a monitor index, the first window matching the regular expressions you provided is captured.

This source does NOT hold a stable FPS. It returns frames at whatever rate the compositor provides them, only capped by max_framerate. If you need a stable rate, you gotta add an fps filter to drop/duplicate frames as needed 👑

If the capture source disappears mid-capture (window closed, monitor disconnected), the filter will return EOF.

This source accepts the following options:

window_title — ECMAScript regular expression matched against the window title. Supports a PCRE style (?i) prefix for case-insensitive matching.
window_class — same as window_title, but matched against the window class name.
window_exe — same as window_title, but matched against the executable file name of the window's process.
monitor_idx — zero-based index of the monitor to capture.

Can also be set to window to capture the monitor the selected window is displayed on at filter initialization time.

hwnd — explicit native window handle (HWND).
hmonitor — explicit native monitor handle (HMONITOR).
capture_cursor — capture the mouse cursor. Enabled by default.
capture_border — capture the full area of the window, including its window 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 you pass a negative number, the width gets 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 you pass a negative number, the height gets rounded down to the next multiple of that number.

See resize_mode.

crop_left — crop this many pixels from the left side of the captured frames.
crop_top — crop this many pixels from the left side of the captured frames.
crop_right — crop this many pixels from the left side of the captured frames.
crop_bottom — crop this many pixels from the left side of the captured frames.

Not me noticing all four of those say "left side" in the original docs 💀 that's literally how it's written, I'm not touching it.

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. The remaining area is filled with black.
scale_mode — the 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 🌈. It's giving aesthetic.

size, s — set frame size. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Default value is 640x480.
rate, r — set frame rate, expressed as number of frames per second. Default value is 25.
c0, c1, c2, c3, c4, c5, c6, c7 — set 8 colors. The default 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 manual really does list y0 twice — I'm preserving the icon's typos 💅.)
nb_colors, n — set number of colors to use at once. Allowed range is from 2 to 8. Default value is 2.
seed — set seed for picking gradient line points.
duration, d — set the duration of the sourced video. See the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax.

If not specified, or the expressed duration is negative, the video is supposed to be generated forever ♾️.

speed — set speed of gradients rotation.
type, t — set type of gradients. Available values are:
- linear
- radial
- circular
- spiral
- square

Default type is linear.

Commands

This source supports the some above options as commands.

mandelbrot

Generate a Mandelbrot set fractal, and progressively zoom towards the point specified with start_x and start_y 🌀. Infinite zoom, main character energy.

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 the 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.

Those default coordinates are lowkey iconic — that's the classic seahorse valley zoom spot. But like, the manual just says they're the defaults, so that's all I'm saying 👀

🎀 `mptestsrc` 🎀

Okay bestie, this one generates a whole little buffet of test patterns — literally the ones the MPlayer test filter makes. The size of the generated video is fixed at 512x512 (no, you can't change it, don't try 💅). This source is highkey useful for testing encoding features, iykyk.

It accepts these options:

rate, r — the frame rate of the sourced video, i.e. how many frames get generated per second. Gotta 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".
duration, d — how long the sourced video lasts. Peep the Time duration section in the ffmpeg-utils(1) manual for the accepted syntax. If you don't specify it, or the duration you express is negative, the video is just supposed to be generated forever 😭 (main character energy: it never stops).
test, t — 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 cycles through the list of all tests. Iconic, honestly.

Some examples:

mptestsrc=t=dc_luma

…will generate a "dc_luma" test pattern. Slay.

💫 `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.

It takes these 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. May be a string of the form num/den or a frame rate abbreviation.
filter_name — the name of the frei0r source to load. Wanna know more about 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. This source is based on a generalization of John Conway's life game — yes, that life game, the OG cellular automaton 🧚.

So the sourced input represents a life grid, and every pixel is a cell that can be in one of two possible states: alive or dead. Every cell interacts with its eight neighbours, which are the cells horizontally, vertically, or diagonally adjacent to it.

At each interaction the grid evolves according to the adopted rule, which specifies the number of neighbor alive cells that will make a cell stay alive or be born. And the rule option is literally how you specify which rule to adopt. Obsessed.

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 delimits the end of each row. If you don't specify this option, 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 ranging from 0 to 1, defaults to 1/PHI. It gets ignored when a file is specified.
random_seed, seed — set the seed for filling the initial random grid; must be an integer between 0 and UINT32_MAX. If it's not specified, or if you explicitly set it to -1, the filter will try to use a good random seed on a best effort basis.
rule — set the life rule. You can specify a rule with a code of the kind SNS/BNB, where NS and NB are sequences of numbers in the range 0-8. NS specifies the number of alive neighbor cells which make a live cell stay alive, and NB the number of alive neighbor cells which make a dead cell become alive (i.e. to "born"). s and b can be used in place of S and B, respectively. Alternatively, a rule can be specified by an 18-bits integer. The 9 high order bits encode the next cell state if it is alive for each number of neighbor alive cells, and the low order bits specify the rule for "borning" new cells. Higher order bits encode for a higher number of neighbor cells. For example the number 6153 = (12<<9)+9 specifies a stay alive rule of 12 and a born rule of 9, which corresponds to S23/B03. Default value is S23/B3, which is the original Conway's game of life rule — it'll 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 defaults to the same size as the input file. If size is set, it must contain the size specified in the input file, and the initial grid defined in that file gets centered in the larger resulting area. If a filename is not specified, the size value defaults to 320x240 (used for a randomly generated initial grid).
stitch — if set to 1, stitch the left and right grid edges together, and the top and bottom edges too. Defaults to 1.
mold — 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, bestie. Perlin noise is a kind of noise with local continuity in space — so you can use it 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 that case the persistence option specifies the ratio of the amplitude with respect to the previous component. More octave components let you specify more high frequency details in the generated noise (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. 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. 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 🫶.

It accepts these 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.

Ngl bestie, some of these constants are literally just not there for the source — like the x or t or ¸n ones — because they only make sense when you're evaluating the expression for each frame rather than at initialization time. So don't come for me when they're missing, okay? 💅

rate, r — sets the frame rate of the sourced video, aka how many frames get generated per second. Gotta be a string shaped like frame_rate_num/frame_rate_den, or a plain integer, or a floating point number, or a valid video frame rate abbreviation. Default is "25". ✨
case_sensitive, cs — tells libqrencode to do case sensitive encoding. Enabled by default, icon behavior. You can disable it to reduce the QR encoding size, so if you're size-obsessed (same), that's your move.
level, l — specifies 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. Either "none" or "normal" (that one's the default). Peep the qrencode Text expansion section for the deets 🫶
text, textfile — define the text to be rendered. If neither is specified, no QR gets encoded at all (you just get an empty colored frame, kind of a vibe honestly). If expansion is enabled, the text is treated as a text template using the qrencode expansion mechanism. Again: qrencode Text expansion section, go read it 💖
background_color, bc and foreground_color, fc — set the QR code and background color. foreground_color defaults to "black", background_color defaults to "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 this is a whole lil squad of sources, so let me introduce them one by one, no cap 👑

The allrgb source returns frames of size 4096x4096 of all rgb colors. Every single one. Iconic.

The allyuv source returns frames of size 4096x4096 of all yuv colors. Same energy, different color space.

The color source provides an uniformly colored input. Simple. Clean. Aesthetic.

The colorchart source provides a colors checker chart.

The colorspectrum source provides a color spectrum input. 🌈

The haldclutsrc source provides an identity Hald CLUT. Also peep the haldclut filter, iykyk.

The nullsrc source returns unprocessed video frames. It's mainly useful for analysis / debugging tools, or as the source for filters which ignore the input data. Lowkey the main character of "I'm just here so the graph works."

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, something's wrong and it's not me 💀

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. Mainly intended for testing purposes. Fr fr.

The testsrc2 source is similar to testsrc, but supports more pixel formats instead of just rgb24. That means you can use it as an input for other tests without needing a format conversion. Glow up behavior ✨

The yuvtestsrc source generates a YUV test pattern. You should see a y, cb and cr stripe from top to bottom.

The sources accept the following parameters:

level — specifies 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 — specifies 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 — specifies 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. Heads up: this option is not available with the allrgb, allyuv, and haldclutsrc filters.
rate, r — specifies the frame rate of the sourced video, as the number of frames generated per second. 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 it's not specified, or 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 isn't a multiple of the frame duration, it gets rounded up. Math is math 🧮
sar — sets the sample aspect ratio of the sourced video.
alpha — specifies the alpha (opacity) of the background, only available in the testsrc2 source. The value must be between 0 (fully transparent) and 255 (fully opaque, and that's 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 these:
- black
- white
- all
patch_size — sets the patch size of a 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 as the corresponding color option.

🌸 openclsrc

Generate video using an OpenCL program. We love a GPU moment, honestly 🫶

source — OpenCL program source file.
kernel — kernel name in program.
size, s — size of frames to generate. This one must be set. No skipping.
format — pixel format to use for the generated frames. This must be set too, bestie.
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.

__kernel void sierpinski_carpet(__write_only image2d_t dst,
                                unsigned int index)
{
    int2 loc = (int2)(get_global_id(0), get_global_id(1));

    float4 value = 0.0f;
    int x = loc.x + index;
    int y = loc.y + index;
    while (x > 0 || y > 0) {
        if (x % 3 == 1 && y % 3 == 1) {
            value = 1.0f;
            break;
        }
        x /= 3;
        y /= 3;
    }

    write_imagef(dst, loc, value);
}

🦋 sierpinski

Generate a Sierpinski carpet/triangle fractal, and randomly pan around. Fractals are so main character, I can't even 😭

This source accepts the following options:

size, s — set frame size. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Default value is 640x480.
rate, r — set frame rate, expressed as number of frames per second. Default value is "25".
seed — set seed which is used for random panning.
jump — set max jump for single pan destination. Allowed range is from 1 to 10000.
type — set fractal type, can be default "carpet" or "triangle".

💜 zoneplate

Generate a zoneplate test video pattern. The name alone is a vibe check.

This source accepts the following options:

size, s — set frame size. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Default value is 320x240.
rate, r — Set the frame rate as frames per second. Default is "25".
duration, d — Set how long the sourced video lasts. The accepted syntax lives in the Time duration section of the ffmpeg-utils(1) manual. If you don’t set it, or you give it a negative duration, the video is generated forever, no cap.
sar — Set the sample aspect ratio of the sourced video.
precision — Set precision in bits for the sine lookup table. Default is 10. Allowed range is from 4 to 16.
xo — Set the horizontal axis offset for the output signal. Default is 0.
yo — Set the vertical axis offset for the output signal. Default is 0.
to — Set the time axis offset for the output signal. Default is 0.
k0 — Set the 0-order constant added to the signal phase. Default is 0.
kx — Set the 1-order phase factor multiplier for the horizontal axis. Default is 0.
ky — Set the 1-order phase factor multiplier for the vertical axis. Default is 0.
kt — Set the 1-order phase factor multiplier for the time axis. Default is 0.
kxt, kyt, kxy — Set phase factor multipliers for combinations of spatial and temporal axes. Default is 0.
kx2 — Set the 2-order phase factor multiplier for the horizontal axis. Default is 0.
ky2 — Set the 2-order phase factor multiplier for the vertical axis. Default is 0.
kt2 — Set the 2-order phase factor multiplier for the time axis. Default is 0.
ku — Set the constant added to the final phase to produce the chroma-blue component of the signal. Default is 0.
kv — Set the constant added to the final phase to produce the chroma-red component of the signal. Default is 0.

💬 Commands 💬

This source supports some of the options above as commands too.

🎀 Examples 🎀

Generate horizontal color sine sweep:

zoneplate=ku=512:kv=0:kt2=0:kx2=256:s=wvga:xo=-426:kt=11

Generate vertical color sine sweep:

zoneplate=ku=512:kv=0:kt2=0:ky2=156:s=wvga:yo=-240:kt=11

Generate circular zone-plate:

zoneplate=ku=512:kv=100:kt2=0:ky2=256:kx2=556:s=wvga:yo=0:kt=11

🎥 VIDEO SINKS 🎥

Here’s the current lineup of video sinks.

💾 buffersink

Buffer video frames and make them available to the end of the filter graph. This sink is mainly for programmatic use, especially through the interface defined in libavfilter/buffersink.h or the options system.

It accepts a pointer to an AVBufferSinkContext structure, which defines the incoming buffers’ formats, passed as the opaque parameter to avfilter_init_filter for initialization.

🫥 nullsink

Null video sink: it does absolutely nothing with the input video. It’s mainly useful as a template and for analysis/debugging tools.

🌈 MULTIMEDIA FILTERS 🌈

Here’s the description of the multimedia filters currently available.

🌀 a3dscope

Convert input audio to a 3D scope video output.

This filter accepts these options:

rate, r — Set the frame rate as frames per second. Default is "25".
size, s — Set the video size for the output. For the syntax, check the Video size section in the ffmpeg-utils manual. Default is "hd720".
fov — Set the camera field of view. Default is 90 degrees. Allowed range is from 40 to 150.
roll — Set the camera roll.
pitch — Set the camera pitch.
yaw — Set the camera yaw.
xzoom — Set the camera zoom on the X-axis.
yzoom — Set the camera zoom on the Y-axis.
zzoom — Set the camera zoom on the Z-axis.
xpos — Set the camera position on the X-axis.
ypos — Set the camera position on the Y-axis.
zpos — Set the camera position on the Z-axis.
length — Set the length of displayed audio waves in number of frames.

💬 Commands 💬

The filter supports some of the options above as commands.

🎚️ abitscope

Convert input audio to a video output, showing the audio bit scope.

This filter accepts these options:

rate, r — Set the frame rate as frames per second. Default is "25".
size, s — Set the video size for the output. For the syntax, check the Video size section in the ffmpeg-utils manual. Default is "1024x256".
colors — Set a list of colors separated by a space or by '|', used to draw channels. Unrecognized or missing colors get replaced with white.
mode, m — Set the output mode. Can be "bars" or "trace". Default is "bars".

📈 adrawgraph

Draw a graph using input audio metadata.

See drawgraph.

👀 agraphmonitor

See graphmonitor.

📊 ahistogram

Convert input audio to a video output, showing the volume histogram.

This filter accepts these options:

dmode — Set how the histogram is calculated. It accepts:
- single — use a single histogram for all channels.
- separate — use a separate histogram for each channel.

Default is "single".

rate, r — Set the frame rate as frames per second. Default is "25".
size, s — Set the video size for the output. For the syntax, check the Video size section in the ffmpeg-utils manual. Default is "hd720".
scale — Set the display scale. It accepts:
- log — logarithmic.
- sqrt — square root.
- cbrt — cubic root.
- lin — linear.
- rlog — reverse logarithmic.

Default is "log".

ascale — Set the amplitude scale. It accepts:
- log — logarithmic.
- lin — linear.

Default is "log".

acount — Set how many frames to accumulate in the histogram. Default is 1. Setting it to -1 accumulates all frames.
rheight — Set the histogram ratio of window height.
slide — Set the sonogram sliding. It accepts:
- replace — replace old rows with new ones.
- scroll — scroll from top to bottom.

Default is "replace".

hmode — Set the histogram mode. It accepts:
- abs — use absolute values of samples.
- sign — use untouched values of samples.

Default is "abs".

🌗 aphasemeter

Measures the phase of input audio, exported as metadata lavfi.aphasemeter.phase, representing 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.

Audio gets rematrixed to stereo if it has a different channel layout. Phase value is in range [-1, 1], where -1 means left and right channels are completely out of phase, and 1 means the channels are in phase.

This filter accepts these options, all related to its video output:

rate, r — Set the output frame rate. Default is 25.
size, s — Set the video size for the output. For the syntax, check the Video size section in the ffmpeg-utils manual. Default is "800x400".
rc, gc, bc — Set the red, green, and blue contrast. Default values are 2, 7, and 1. Allowed range is [0, 255].
mpc — Set the color used for drawing median phase. If the color is "none", which is the default, no median phase value gets drawn.
video — Enable 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.

The filter accepts these options for this detection:

phasing — Enable mono and out-of-phase detection. Default is disabled.
tolerance, t — Set the phase tolerance for mono detection, in amplitude ratio. Default is 0. Allowed range is [0, 1].
angle, a — Set the angle threshold for out-of-phase detection, in degrees. Default is 170. Allowed range is [90, 180].
duration, d — Set 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

Convert input audio to a video output, representing the audio vector scope.

This filter measures the difference between channels of a stereo audio stream. A monaural signal made of identical left and right signals gives you a straight vertical line. Any stereo separation shows up as a deviation from that line, creating a Lissajous figure. If a straight horizontal line appears — or a deviation that’s horizontal — that means the left and right channels are out of phase.

This filter accepts these options:

mode, m — Set 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 — Set the video size for the output. For the syntax, check the Video size section in the ffmpeg-utils manual. Default is "400x400".
rate, r — Set the output frame rate. Default is 25.
rc, gc, bc, ac — Set the red, green, blue, and alpha contrast. Default values are 40, 160, 80, and 255. Allowed range is [0, 255].
rf, gf, bf, af — Set the red, green, blue, and alpha fade. Default values are 15, 10, 5, and 5. Allowed range is [0, 255].
zoom — Set the zoom factor. Default is 1. Allowed range is [0, 10]. Values lower than 1 auto-adjust the zoom factor to the maximum possible value.
draw — Set 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 — Set the amplitude scale of audio samples. Available values are:
- lin — linear.
- sqrt — square root.
- cbrt — cubic root.
- log — logarithmic.
swap — Swap the left channel axis with the right channel axis.
mirror — Mirror the axis.
- none — no mirror.
x — mirror the x axis only, that's it 💅
y — mirror the y axis only, one-axis queen ✨
xy — mirror both axes at once, go fully symmetrical bestie 🦋

🌸 Examples 🌸

okay here's the full flex, a complete example using ffplay:

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 of the above options as commands — except size and rate, those two are not command-able, no cap 😭

🎯 bench, abench

Benchmark part of a filtergraph. Lowkey it's a stopwatch for your filter chain ⏱️✨

The filter accepts the following options:

action — start or stop a timer. Iconic little on/off switch 🫶

Available values are:

start — grab the current time, stash it as frame metadata (using the key lavfi.bench.start_time), and forward the frame on to the next filter 💖
stop — grab the current time and fetch the lavfi.bench.start_time metadata from the input frame's metadata to get the time difference. Then the time difference, average, maximum and minimum time (respectively t, avg, max and min) get printed. The timestamps are expressed in seconds ⭐

🦋 Examples 🦋

benchmarking the selectivecolor filter, look at her go:

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 "glue stick but for media" 💀💖

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. Periodt, no mixing and matching 👑

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 🙈 (yes bestie, it's literally called unsafe, act accordingly)

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, don't be chaotic about it 🌈

Related streams do not always have exactly the same duration, for various reasons including codec frame size or sloppy authoring. For that reason, related synchronized streams (e.g. a video and its audio track) should be concatenated at once. The concat filter will use the duration of the longest stream in each segment (except the last one), and if necessary pad shorter audio streams with silence 🤫

For this filter to work correctly, all segments must start at timestamp 0. That's the vibe check it will not pass if you skip it 😭

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. Yeah, scale is your bestie here 💅

Different frame rates are acceptable but will result in variable frame rate at output; be sure to configure the output file to handle it. Don't say nobody warned you 🫶

✨ Examples ✨

Concatenate an opening, an episode and an ending, all in bilingual version (video in stream 0, audio in streams 1 and 2):

ffmpeg -i opening.mkv -i episode.mkv -i ending.mkv -filter_complex \
  '[0:0] [0:1] [0:2] [1:0] [1:1] [1:2] [2:0] [2:1] [2:2]
   concat=n=3:v=1:a=2 [v] [a1] [a2]' \
  -map '[v]' -map '[a1]' -map '[a2]' output.mkv

Concatenate two parts, handling audio and video separately, using the (a)movie sources, and adjusting the resolution:

movie=part1.mp4, scale=512:288 [v1] ; amovie=part1.mp4 [a1] ;
movie=part2.mp4, scale=512:288 [v2] ; amovie=part2.mp4 [a2] ;
[v1] [v2] concat [outv] ; [a1] [a2] concat=v=0:a=1 [outa]

Note that a desync will happen at the stitch if the audio and video streams do not have exactly the same duration in the first file. Like, actual desync. Not a drill 💀

🎀 Commands 🎀

This filter supports the following commands:

next — close the current segment and step to the next one 👑

🌟 ebur128

EBU R128 scanner filter. This filter takes an audio stream and analyzes its loudness level. By default, it logs a message at a frequency of 10Hz with the Momentary loudness (identified by M), Short-term loudness (S), Integrated loudness (I) and Loudness Range (LRA). She's a whole loudness scientist 🔬💖

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. FYI so you don't get caught out 🫶

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.
metadata — set metadata injection. If set to 1, the audio input will be segmented into 100ms output frames, each of them containing various loudness information in metadata. All the metadata keys are prefixed with lavfi.r128..

Default is 0.

framelog — force the frame logging level.

Available values are:

quiet — logging disabled 🤫
info — information logging level
verbose — verbose logging level

By default, the logging level is set to info. If the video or the metadata options are set, it switches to verbose.

peak — set peak mode(s).

Available modes can be cumulated (the option is a flag type). Possible values are:

none — disable any peak mode (default).
sample — enable sample-peak mode.

Simple peak mode looking for the higher sample value. It logs a message for sample-peak (identified by SPK).

true — enable true-peak mode.

If enabled, the peak lookup is done on an over-sampled version of the input stream for better peak accuracy. It logs a message for true-peak (identified by TPK) and true-peak per frame (identified by FTPK). This mode requires a build with libswresample.

dualmono — treat mono input files as "dual mono". If a mono file is intended for playback on a stereo system, its EBU R128 measurement will be perceptually incorrect. If set to true, this option will compensate for this effect. Multi-channel input files are not affected by this option.
panlaw — set a specific pan law to be used for the measurement of dual mono files. This parameter is optional, and has a default value of -3.01dB.
target — set a specific target level (in LUFS) used as relative zero in the visualization. This parameter is optional and has a default value of -23LUFS as specified by EBU R128. However, material published online may prefer a level of -16LUFS (e.g. for use with podcasts or video platforms).
gauge — set the value displayed by the gauge. Valid values are momentary and shortterm. By default the momentary value will be used, but in certain scenarios it may be more useful to observe the short term value instead (e.g. live mixing).
scale — sets the display scale for the loudness. Valid parameters are absolute (in LUFS) or relative (LU) relative to the target. This only affects the video output, not the summary or continuous log output.
integrated — read-only exported value for measured integrated loudness, in LUFS.
range — read-only exported value for measured loudness range, in LU.
lra_low — read-only exported value for measured LRA low, in LUFS.
lra_high — read-only exported value for measured LRA high, in LUFS.
sample_peak — read-only exported value for measured sample peak, in dBFS.
true_peak — read-only exported value for measured true peak, in dBFS.

💜 Examples 💜

Real-time graph using ffplay, with a EBU scale meter +18:

ffplay -f lavfi -i "amovie=input.mp3,ebur128=video=1:meter=18 [out0][out1]"

Run an analysis with ffmpeg:

ffmpeg -nostats -i input.mp3 -filter_complex ebur128 -f null -

🦋 interleave, ainterleave

Temporally interleave frames from several inputs. Like shuffling cards but the cards are frames, and the dealer is very particular 💅

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, no vibes-based timestamps 🕰️

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. It's a whole group project — everyone has to show up 😭

For example consider the case when one input is a select filter which always drops input frames. The interleave filter will keep reading from that input, but it will never be able to send new frames to output until the input sends an end-of-stream signal.

Also, depending on inputs synchronization, the filters will drop frames in case one input receives more frames than the other ones, and the queue is already filled. Sorry bestie, some frames are getting left on read 💀

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 — the duration of the first input. that's it, that's the whole vibe 💅

🌸 Examples

interleave frames from different streams using ffmpeg, bestie:

ffmpeg -i bambi.avi -i pr0n.mkv -filter_complex "[0:v][1:v] interleave" out.avi

and if you want a flickering blur effect (yes it's as chaotic as it sounds 😭):

select='if(gt(random(0), 0.2), 1, 2)':n=2 [tmp], boxblur=2:2, [tmp] interleave

💫 `latency`, `alatency`

these two literally just measure filtering latency ✨

they report the previous filter's filtering latency — so the delay in number of audio samples for audio filters, or number of video frames for video filters 🎀

and when the input stream ends, the filter reports the min and max measured latency for the previous running filter in the filtergraph. cute little report card fr fr 📊

🔍 `metadata`, `ametadata`

manipulate frame metadata. yes, the little invisible tags riding along on your frames 🏷️

this filter accepts the following options:

options

mode — sets the mode of operation of the filter. can be one of these:
- select — if both value and key are set, it selects frames that have such metadata. if only key is set, it selects every frame that has such key in its metadata. lowkey the picky one 🫶
- add — add new metadata key and value. if the key is already there, do nothing. polite king behavior 👑
- modify — modify the value of a key that's already present.
- delete — if value is set, delete only the keys that have such value. otherwise, delete the key. and if key isn't set at all, delete all metadata values in the frame. scorched earth 💀
- print — print the key and its value if metadata was found. if key isn't set, print all metadata values available in the frame. show and tell 🌈
key — set the key used with all modes. must be set for all modes except print and delete. no cap, it's required.
value — set the metadata value that will be used. this option is mandatory for modify and add mode. don't forget it bestie 😤
function — which function to use when comparing the metadata value and value. can be one of the following:
- same_str — values are interpreted as strings, returns true if the metadata value is the same as value.
- starts_with — values are interpreted as strings, returns true if the metadata value starts with the value option string.
- less — values are interpreted as floats, returns true if the metadata value is less than value.
- equal — values are interpreted as floats, returns true if value is equal to the metadata value.
- greater — values are interpreted as floats, returns true if the metadata value is greater than value.
- expr — values are interpreted as floats, returns true if the expression from the expr option evaluates to true.
- ends_with — values are interpreted as strings, returns true if the metadata value ends with the value option string.
expr — set the expression used when function is set to expr. it's evaluated through the eval API and can contain these 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 gets written to the named file. instead of a plain filename you can specify any writable url. the filename - is shorthand for standard output. and if the file option isn't set, output is written to the log with the AV_LOG_INFO loglevel 📝
direct — reduces buffering in print mode when output is written to a URL set using file. smooth brain move for streaming output 🦋

🌸 Examples

print all metadata values for frames with key lavfi.signalstats.YDIF with values between 0 and 1:

signalstats,metadata=print:key=lavfi.signalstats.YDIF:value=0:function=expr:expr='between(VALUE1,0,1)'

print silencedetect output to the file metadata.txt:

silencedetect,ametadata=mode=print:file=metadata.txt

direct all metadata to a pipe with file descriptor 4:

metadata=mode=print:file='pipe\:4'

🔐 `perms`, `aperms`

set read/write permissions for the output frames. yes we're doing access control on frames now, it's giving tiny bouncer energy 🚪

these filters are mainly aimed at developers to test the direct path in the following filter in the filtergraph. so like, if that's not you, no stress 💖

the filters accept the following options:

mode — select the permissions mode. accepts these values:
- none — do nothing. this is the default.
- ro — set all the output frames read-only.
- rw — set all the output frames directly writable.
- toggle — make the frame read-only if it's writable, and writable if it's read-only. chaotic flip-flop, iconic 🌗
- random — set each output frame read-only or writable randomly.
seed — set the seed for the random mode, must be an integer 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. lowkey a gotcha, highkey important 🫠

⏱️ `realtime`, `arealtime`

slow down filtering to match real time approximately. vibe check: it's the "chill, we have time" filter 🧘‍♀️

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 — the time limit for the pauses. any pause longer than that is 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 result in faster-than-realtime processing, smaller values slow processing down. the limit is automatically adapted accordingly. default is 1.0.

🫶 heads up bestie: a processing speed faster than what's possible without these filters cannot be achieved. physics said no 😭

Commands

both filters support all the above options as commands. periodt 💅

✂️ `segment`, `asegment`

split a single input stream into multiple streams. it literally does the opposite of the concat filters 🎬

segment works on video frames, asegment works on audio samples.

this filter accepts the following options:

timestamps — timestamps of the 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. bookends, slay ✨
frames, samples — exact frame/sample count to split the segments.

in all cases, prefixing each segment with + makes it relative to the previous segment. tiny detail, big difference 🎀

🌸 Examples

split the input audio stream into three output audio streams — starting at the start of the input audio stream and storing that in the 1st output audio stream, then following at the 60th second and storing that in the 2nd output audio stream, and last after the 150th second of the input audio stream stored in the 3rd output audio stream:

asegment=timestamps="60|150"

👑 `select`, `aselect`

select frames to pass in output. the main character filter fr fr 🌟

this filter accepts the following options:

expr, e — set 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, honestly 📐

outputs, n — set the number of outputs. which output the selected frame goes to is based on the result of the evaluation. default value is 1.

the expression can contain the following constants:

n — the (sequential) number of the filtered frame, starting from 0.
selected_n — the (sequential) number of the selected frame, starting from 0.
prev_selected_n — the sequential number of the last selected frame. it's NAN if undefined.
TB — the timebase of the input timestamps.
pts — the PTS (Presentation TimeStamp) of the filtered frame, expressed in TB units. it's NAN if undefined.
t — the PTS of the filtered frame, expressed in seconds. it's NAN if undefined.
prev_pts — the PTS of the previously filtered frame. it's NAN if undefined.
prev_selected_pts — the PTS of the last previously filtered frame. it's NAN if undefined.
prev_selected_t — the PTS of the last previously selected frame, expressed in seconds. it's NAN if undefined.
start_pts — the first PTS in the stream which is not NAN. it remains NAN if not found.
start_t — the first PTS, in seconds, in the stream which is not NAN. it remains NAN if not found.
pict_type (video only) — the type of the filtered frame. it can assume one of these values: I, P, B, S, SI, SP, BI.
interlace_type (video only) — the frame interlace type. it can assume one of these values:
- PROGRESSIVE — the frame is progressive (not interlaced).
- TOPFIRST — the frame is top-field-first.
- BOTTOMFIRST — the frame is bottom-field-first.
consumed_sample_n (audio only) — the number of selected samples before the current frame.
samples_n (audio only) — the number of samples in the current frame.
sample_rate (audio only) — the input sample rate.
key — this is 1 if the filtered frame is a key-frame, 0 otherwise. 🔑
pos — the position in the file of the filtered frame, -1 if the information is not available (e.g. for synthetic video). deprecated, do not use 🚫
scene (video only) — a 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). dramatic scene change detected 🎭
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 are not exactly contained in the selected interval.

this works by comparing the frame pts against the lavf.concat.start_time and the lavf.concat.duration packet metadata values which are also present in the decoded frames. iykyk 🧚

concatdec_select — this one's -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. Otherwise it's 0, and it's NaN if the start_time metadata is missing 😭

So basically, bestie: an input frame gets selected if its pts lands inside the interval the concat demuxer set up. 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) — the view ID for multi-view video 👀

The default value of the select expression is "1".

🌸 Examples 🌸

Select all frames in input:

select

Which, like, is literally the same as:

select=1

Skip all frames (ghosting everyone, iconic behaviour):

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

Ngl, 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. Like passing notes in class, but the notes are legal filter commands ✨

These filters read commands that then get sent to other filters in the filtergraph.

sendcmd must be inserted between two video filters, and asendcmd must be inserted between two audio filters — but other than that, they act the exact same way. Twins fr fr 👑

The specification of commands can be provided 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 each one holds a list of commands to run 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.

A frame's current time counts as inside the interval if it's in [START, END), meaning the time is greater or equal to START and lesser than END — half-open intervals, we love to see it 💅

COMMANDS is a sequence of one or more command specifications separated by ,, all relating to that interval. The syntax of a command specification is:

[<FLAGS>] <TARGET> <COMMAND> <ARG>

FLAGS is optional and says which type of events relating to the time interval actually enable sending the command. It has to 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 isn't 👋
expr — the command ARG is interpreted as an expression, and the result of that expression gets passed as ARG.

The expression is evaluated through the eval API and can contain 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 isn't specified, a default value of [enter] is assumed. No cap 🧢

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 argument 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 you can use them 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 allows you to 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. Time is a social construct anyway ⏳💫

setpts works on video frames, asetpts works on audio frames.

This filter accepts the following options:

expr — the expression which is evaluated for each frame to construct its timestamp.
strip_fps (video only) — boolean option that determines if the original framerate and frame duration metadata is unset. If you set it 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. Slay, equal opportunity icon behaviour 👑

🧚 `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 — it only sets the corresponding property, which affects how the frame is treated by the filters that come after it 👀✨

The filter accepts the following options:

range — bestie, this one picks the color range. Available values are:
- auto — keep the same color range property.
- unspecified, unknown — set the color range as unspecified.
- limited, tv, mpeg — set the color range as limited.
- full, pc, jpeg — set the color range as full.

🎛️ `settb`, `asettb`

This sets the timebase used for the output frames’ timestamps. It’s mainly useful for testing timebase configuration. It accepts these parameters:

expr, tb — the expression that gets evaluated into the output timebase.

The value for tb is an arithmetic expression representing a rational. That expression can contain the constants "AVTB" (the default timebase), "intb" (the input timebase), and "sr" (the sample rate, audio only). Default value is "intb".

Examples:

Set the timebase to 1/25:

settb=expr=1/25

Set the timebase to 1/10:

settb=expr=0.1

Set the timebase to 1001/1000:

settb=1+0.001

Set the timebase to 2*intb:

settb=2*intb

Set the default timebase value:

settb=AVTB

🌈 `showcqt` 🌈

Okay so this one 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 transform itself is not really constant Q — instead the Q factor is actually variable/clamped. It gives you a musical tone scale from E0 to D#10, which is honestly iconic. 💅

The filter accepts the following options:

size, s — specify the video size for the output. It must be even. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Default value is "1920x1080".
fps, rate, r — set the output frame rate. Default value is 25.
bar_h — set the bargraph height. It must be even. Default value is -1, which computes the bargraph height automatically.
axis_h — set the axis height. It must be even. Default value is -1, which computes the axis height automatically.
sono_h — set the sonogram height. It must be even. Default value is -1, which computes the sonogram height automatically.
fullhd — set the fullhd resolution. This option is deprecated — use size, s instead. Default value is 1.
sono_v, volume — specify the sonogram volume expression. It can contain 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 — specify 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 — specify the sonogram gamma. Lower gamma makes the spectrum more contrast, higher gamma makes the spectrum have more range. Default value is 3. Acceptable range is "[1, 7]".
bar_g, gamma2 — specify the bargraph gamma. Default value is 1. Acceptable range is "[1, 7]".
bar_t — specify the bargraph transparency level. Lower value makes the bargraph sharper. Default value is 1. Acceptable range is "[0, 1]".
timeclamp, tc — specify the transform timeclamp. At low frequency, there is a trade-off between accuracy in the time domain and accuracy in the frequency domain. If timeclamp is lower, events in the time domain are represented more accurately (such as a fast bass drum); otherwise, events in the frequency domain are represented more accurately (such as a bass guitar). Acceptable range is "[0.002, 1]". Default value is 0.17.
attack — set attack time in seconds. The default is 0 (disabled). Otherwise, it limits future samples by applying asymmetric windowing in the time domain, useful when low latency is required. Accepted range is "[0, 1]".
basefreq — specify the transform base frequency. Default value is 20.01523126408007475, which is a frequency 50 cents below E0. Acceptable range is "[10, 100000]".
endfreq — specify the transform end frequency. Default value is 20495.59681441799654, which is a frequency 50 cents above D#10. Acceptable range is "[10, 100000]".
coeffclamp — this option is deprecated and ignored.
tlength — specify the transform length in the time domain. Use this option 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 — specify the transform count for every video frame. Default value is 6. Acceptable range is "[1, 30]".
fcount — specify the transform count for every single pixel. Default value is 0, which makes it computed automatically. Acceptable range is "[0, 10]".
fontfile — specify a font file for use with freetype to draw the axis. If not specified, use the embedded font. Note that drawing with a font file or embedded font is not implemented with custom basefreq and endfreq — use the axisfile option instead.
font — specify a fontconfig pattern. This has lower priority than fontfile. The ":" in the pattern may be replaced by "|" to avoid unnecessary escaping.
fontcolor — specify a font color expression. This is an arithmetic expression that should return 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: E0(16), C1(24), C2(36), A4(69).
r(x), g(x), b(x) — red, green, and blue value of intensity x.

Default value is "st(0, (midi(f)-59.5)/12); st(1, if(between(ld(0),0,1), 0.5-0.5*cos(2*PI*ld(0)), 0)); r(1-ld(1)) + b(ld(1))".

axisfile — specify 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 is set to 0, drawing to the axis is disabled, ignoring the fontfile and axisfile options. Default value is 1.
csp — set colorspace. The accepted values are:
- unspecified — unspecified (default).
- bt709 — BT.709.
- fcc — FCC.
- bt470bg — BT.470BG or BT.601-6 625.
- smpte170m — SMPTE-170M or BT.601-6 525.
- smpte240m — SMPTE-240M.
- bt2020ncl — BT.2020 with non-constant luminance.
cscheme — set the spectrogram color scheme. This is a list of floating point values with format "left_r|left_g|left_b|right_r|right_g|right_b". The default is "1|0.5|0|0|0.5|1".

Examples:

Playing audio while showing the spectrum:

ffplay -f lavfi 'amovie=a.mp3, asplit [a][out1]; [a] showcqt [out0]'

Same as above, but with frame rate 30 fps:

ffplay -f lavfi 'amovie=a.mp3, asplit [a][out1]; [a] showcqt=fps=30:count=5 [out0]'

Playing at 1280x720:

ffplay -f lavfi 'amovie=a.mp3, asplit [a][out1]; [a] showcqt=s=1280x720:count=4 [out0]'

Disable sonogram display:

sono_h=0

A1 and its harmonics: A1, A2, (near)E3, A3:

ffplay -f lavfi 'aevalsrc=0.1*sin(2*PI*55*t)+0.1*sin(4*PI*55*t)+0.1*sin(6*PI*55*t)+0.1*sin(8*PI*55*t),
                                  asplit[a][out1]; [a] showcqt [out0]'

Same as above, but with more accuracy in 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` 🌀

This one converts input audio into a video output representing the frequency spectrum using Continuous Wavelet Transform and Morlet wavelet. The filter accepts the following options:

size, s — specify the video size for the output. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Default value is "640x512".
rate, r — set the output frame rate. Default value is 25.
scale — set the frequency scale used. Allowed values are:
- linear
- log
- bark
- mel
- erbs
- sqrt
- cbrt
- qdrt
- fm

Default value is "linear".

iscale — set the intensity scale used. Allowed values are:
- linear
- log
- sqrt
- cbrt
- qdrt

Default value is "log".

min — set the minimum frequency that will be used in output. Default is 20 Hz.
max — set the maximum frequency that will be used in output. Default is 20000 Hz. The real frequency upper limit depends on the input audio’s sample rate, and such will be enforced on this value when it is set to a value greater than the Nyquist frequency.
imin — set the minimum intensity that will be used in output.
imax — set the maximum intensity that will be used in output.
logb — sets the logarithmic basis for brightness strength when mapping calculated magnitude values onto pixel values. Allowed range is 0 to 1, default is 0.0001 ✨
deviation — sets the frequency deviation. Lower values than 1 are more frequency oriented, higher values than 1 are more time oriented. Allowed range is 0 to 10, default is 1 🌀
pps — sets the number of pixel output per each second in one row. Allowed range is 1 to 1024, default 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 🌈
- 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 🌸

Okay so this one converts input audio into a video output representing the audio power spectrum — audio amplitude is on the Y-axis while frequency is on the X-axis 📈 literally the whole vibe in one picture.

The filter accepts the following options:

size, s — specify size of video. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Default is "1024x512" 🖼️
rate, r — set video rate. Default is 25 ⏱️
mode — set display mode. This sets how each frequency bin will be represented. It accepts the following values:
- line
- bar
- dot

Default is "bar".

ascale — set amplitude scale. It accepts the following values:
- lin — linear scale 📏
- sqrt — square root scale 🌱
- cbrt — cubic root scale 🧊
- log — logarithmic scale 🌲

Default is "log".

fscale — set frequency scale. It accepts the following values:
- lin — linear scale 📏
- log — logarithmic scale 🌲
- rlog — reverse logarithmic scale 🔄

Default is "lin".

win_size — set window size. Allowed range is from 16 to 65536. Default is 2048 🪟
win_func — set windowing function. It accepts the following values: rect, bartlett, hanning, hamming, blackman, welch, flattop, bharris, bnuttall, bhann, sine, nuttall, lanczos, gauss, tukey, dolph, cauchy, parzen, poisson, bohman, kaiser. Default is "hanning" 🧚
overlap — set window overlap. In range [0, 1]. Default is 1, which means optimal overlap for the selected window function will be picked 🫶
averaging — set time averaging. Setting this to 0 will display current maximal peaks. Default is 1, which means time averaging is disabled ⏳
colors — specify list of colors separated by space or by '|' which will be used to draw channel frequencies. Unrecognized or missing colors will be replaced by white color 🤍
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 🦋

This one converts stereo input audio into a video output, representing the spatial relationship between two channels — the whole left-right lore, visualized 💜

The filter accepts the following options:

size, s — specify the video size for the output. For the syntax of this option, check the "Video size" section in the ffmpeg-utils manual. Default value is "512x512" 🖼️
win_size — set window size. Allowed range is from 1024 to 65536. Default size is 4096 🪟
win_func — set window function. It accepts the following values: rect, bartlett, hann, hanning, hamming, blackman, welch, flattop, bharris, bnuttall, bhann, sine, nuttall, lanczos, gauss, tukey, dolph, cauchy, parzen, poisson, bohman, kaiser. Default value is "hann" ✨
rate, r — set output framerate ⏱️

🔥 showspectrum 🔥

Converts input audio into a video output representing the audio frequency spectrum — no cap, this is the one you've definitely seen in every music visualizer ever 👑

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: rect, bartlett, hann, hanning, hamming, blackman, welch, flattop, bharris, bnuttall, bhann, sine, nuttall, lanczos, gauss, tukey, dolph, cauchy, parzen, poisson, kaiser. 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 the specific window function currently used 🫶
gain — set scale gain for calculating intensity color values. Default value is 1 🔊
data — set which data to display. Can be "magnitude", default, or "phase", or unwrapped phase: "uphase" 🌙
rotation — set color rotation, must be in [-1.0, 1.0] range. Default value is 0 💫
start — set start frequency from which to display spectrogram. Default is 0 🎚️
stop — set stop frequency to which to display spectrogram. Default is 0 🎚️
fps — set upper frame rate limit. Default is "auto", unlimited 🚀
legend — draw time and frequency axes and legends. Default is disabled 📝
drange — set dynamic range used to calculate intensity color values. Default is 120 dBFS. Allowed range is from 10 to 200 📏
limit — set upper limit of input audio samples volume in dBFS. Default is 0 dBFS. Allowed range is from -100 to 100 🔉
opacity — set opacity strength when using pixel format output with alpha component 👻

The usage is very similar to the showwaves filter; see the examples in that section 👀

✨ Examples ✨

Large window with logarithmic color scaling:

showspectrum=s=1280x480:scale=log

Complete example for a colored and sliding spectrum per channel using ffplay:

ffplay -f lavfi 'amovie=input.mp3, asplit [a][out1];
             [a] showspectrum=mode=separate:color=intensity:slide=1:scale=cbrt [out0]'

🎀 showspectrumpic 🎀

Converts input audio into a single video frame, representing the audio frequency spectrum — one frame, one gorgeous spectrum, no motion 🖼️✨

The filter accepts the following options:

okay bestie we're still in the thick of it — carrying straight on with the back half of the showspectrumpic options, no recap, no yapping, let's go ✨

size, s — sets the video size for the output. For the syntax of this option, go check the "Video size" section in the ffmpeg-utils manual. Default value is 4096x2048. 💖
mode — picks the display mode, aka how the channels get laid out. 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 — sets the 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 intensity. 🌈 (the color scheme roster is genuinely giving aesthetic mood board, ngl)

scale — the 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 log.

💀 okay so the page literally tags sqrt with "default" up there and then says the default is log right after. I'm not fixing it, I'm just reporting it, bestie — the manpage said what it said.

fscale — sets the frequency scale. It accepts the following values:
- lin — linear
- log — logarithmic

Default value is lin.

saturation — sets the saturation modifier for displayed colors. Negative values give you an alternative color scheme, and 0 is no saturation at all. Saturation must be in the [-10.0, 10.0] range. Default value is 1. 🎀
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 value is hann. 🦋

orientation — sets the orientation of the time vs frequency axis. Can be vertical or horizontal. Default is vertical.
gain — sets the scale gain used for calculating intensity color values. Default value is 1.
legend — draws the time and frequency axes plus legends. Default is enabled. 👑
rotation — sets color rotation, must be in the [-1.0, 1.0] range. Default value is 0.
start — sets the start frequency from which to display the spectrogram. Default is 0.
stop — sets the stop frequency to which to display the spectrogram. Default is 0.
drange — sets the dynamic range used to calculate intensity color values. Default is 120 dBFS. Allowed range is from 10 to 200.
limit — sets the upper limit of input audio samples volume in dBFS. Default is 0 dBFS. Allowed range is from -100 to 100.
opacity — sets the opacity strength when you're using a pixel format output that has an alpha component. 🧚

Examples 🌸

Extract an audio spectrogram of a whole audio track into a 1024x1024 picture using ffmpeg:

ffmpeg -i audio.flac -lavfi showspectrumpic=s=1024x1024 spectrogram.png

showvolume 💜

Turns input audio volume into a video output. The filter accepts the following options:

rate, r — sets the video rate.
b — sets the border width, allowed range is [0, 5]. Default is 1.
w — sets the channel width, allowed range is [80, 8192]. Default is 400.
h — sets the channel height, allowed range is [1, 900]. Default is 20.
f — sets the fade, allowed range is [0, 1]. Default is 0.95.
c — sets the volume color expression. The expression can use the following 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. Default is enabled.
v — if set, displays volume values. Default is enabled.
o — sets orientation, can be horizontal: h or vertical: v, default is h.
s — sets step size, allowed range is [0, 5]. Default is 0, which means step is disabled.
p — sets background opacity, allowed range is [0, 1]. Default is 0.
m — sets metering mode, can be peak: p or rms: r, default is p.
ds — sets display scale, can be linear: lin or log: log, default is lin.
dm — in second. If set to > 0., display a line for the max level in the previous seconds. Default is disabled: 0.
dmc — the color of the max line. Use when dm option is set to > 0. Default is: orange. 🍓

showwaves 💅

Converts input audio to a video output that represents the samples waves. The filter accepts the following options:

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 600x240.
mode — sets the display mode. Available values are:
- 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 value is point.

n — sets the number of samples which are printed on the same column. A larger value will decrease the frame rate. Must be a positive integer. 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 option n. Default value is 25.
split_channels — sets whether channels should be drawn separately or overlap. Default value is 0.
colors — sets colors separated by '|' which are going to be used for drawing of each channel.
scale — sets the amplitude scale. Available values are:
- 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 are:
- scale — scale pixel values for each drawn sample.
- full — draw every sample directly.

Default value is scale. 🔥

Examples 🎀

Output the input file audio and the corresponding video representation at the same time:

amovie=a.mp3,asplit[out0],showwaves[out1]

Create a synthetic signal and show it with showwaves, forcing a frame rate of 30 frames per second:

aevalsrc=sin(1*2*PI*t)*sin(880*2*PI*t):cos(2*PI*200*t),asplit[out0],showwaves=r=30[out1]

showwavespic 🌟

Converts input audio to a single video frame that represents the samples waves. One frame, iconic behavior. The filter accepts the following options:

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 600x240.
split_channels — sets whether channels should be drawn separately or overlap. Default value is 0.
colors — sets colors separated by '|' which are going to be used for drawing of each channel.
scale — sets the amplitude scale. Available values are:
- lin — linear.
- log — logarithmic.
- sqrt — square root.
- cbrt — cubic root.

Default is linear.

draw — sets the draw mode. Available values are:
- scale — scale pixel values for each drawn sample.
- full — draw every sample directly.

Default value is scale.

filter — sets the filter mode. Available values are:
- average — use average samples values for each drawn sample.
- peak — use peak samples values for each drawn sample.

Default value 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. Very much the "block or keep" energy. This filter accepts the following options:

mode — sets the mode of operation of the filter. Can be one of the following:
- 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

(yes, all twenty-nine of them, I counted so you don't have to 😭)

spectrumsynth 🔮

Synthesizes audio from two input video spectrums — first input stream represents magnitude across time, 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 reversing the whole art project back into sound, we love a glow-up in both directions 🫶

still more to come, bestie — don't close the tab 💜

Okay so this filter is mainly here to reverse processed showspectrum filter outputs, but bestie it can also synthesize sound from other spectrograms too. Lowkey in that case the results are gonna be poor if the phase data is not available, because then the phase data has to be recreated, usually just from random noise. For best results use gray-only output ("channel" color mode in showspectrum filter) and "log" scale for magnitude video and "lin" scale for phase video. To produce phase, for the 2nd video, use "data" option. Input videos should generally use "fullframe" slide mode because that saves resources needed for decoding video.

The filter accepts the following options:

sample_rate — set the sample rate of the output audio; the sample rate of the audio that the spectrum was generated from may differ, no cap.
channels — set the number of channels represented in the input video spectrums.
scale — set the scale that was used when generating the magnitude input spectrum. Can be "lin" or "log". Default is "log".
slide — set the slide that was used when generating the inputs 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 input videos. Can be "vertical" or "horizontal". Default is "vertical".

Examples

First create magnitude and phase videos from audio, assuming the audio is stereo with 44100 sample rate, then resynthesize the 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. Bestie, it's giving copy-paste energy but make it filtergraph.

"asplit" works with audio input, and "split" works with video.

The filter accepts a single parameter which specifies the number of outputs. If unspecified, it defaults to 2.

Examples

Create two separate outputs from the same input:

[in] split [out0][out1]

To create 3 or more outputs, you need to specify the number of outputs, like in:

[in] asplit=3 [out0][out1][out2]

Create two separate outputs from the same input, one cropped and one padded:

[in] split [splitout1][splitout2];
[splitout1] crop=100:100:0:0    [cropout];
[splitout2] pad=200:200:100:100 [padout];

Create 5 copies of the input audio with ffmpeg:

ffmpeg -i INPUT -filter_complex asplit=5 OUTPUT

💅 zmq, azmq 💅

Receive commands sent through a libzmq client, and forward them to filters in the filtergraph. Literally the group-chat relay of the filter world.

"zmq" and "azmq" work as pass-through filters. "zmq" must be inserted between two video filters, and "azmq" between two audio filters. Both are capable of sending messages to any filter type.

To enable these filters you need to install the libzmq library and headers and configure FFmpeg with --enable-libzmq.

For more information about libzmq see: http://www.zeromq.org/

The "zmq" and "azmq" filters work as a libzmq server, which receives messages sent through a network interface defined by the bind_address (or the abbreviation "b") option. Default value of this option is tcp://localhost:5555. You may want to alter this value to your needs, but do not forget to escape any ':' signs (see filtergraph escaping).

The received message must be in the form:

<TARGET> <COMMAND> [<ARG>]

TARGET specifies the target of the command, usually the name of the filter class or a specific filter instance name. The default filter instance name uses the pattern Parsed_<filter_name>_<index>, but you can override this by using the filter_name@id syntax (see Filtergraph syntax).

COMMAND specifies the name of the command for the target filter.

ARG is optional and specifies the optional argument list for the given COMMAND.

Upon reception, the message is processed and the corresponding command is injected into the filtergraph. Depending on the result, the filter will send a reply to the client, adopting the format:

<ERROR_CODE> <ERROR_REASON>
<MESSAGE>

MESSAGE is optional.

Examples

Look at tools/zmqsend for an example of a zmq client which can be used to send commands processed by these filters.

Consider the following filtergraph generated by ffplay. In this example the last overlay filter has an instance name. All other filters will have default instance names.

ffplay -dumpgraph 1 -f lavfi "
color=s=100x100:c=red  [l];
color=s=100x100:c=blue [r];
nullsrc=s=200x100, zmq [bg];
[bg][l]   overlay     [bg+l];
[bg+l][r] overlay@my=x=100 "

To change the color of the left side of the video, the following command can be used:

echo Parsed_color_0 c yellow | tools/zmqsend

To change the right side:

echo Parsed_color_1 c pink | tools/zmqsend

To change the position of the right side:

echo overlay@my x 150 | tools/zmqsend

💖 MULTIMEDIA SOURCES 💖

okay bestie so here's the rundown of all the multimedia sources we've got rn:

🎀 amovie

literally the same energy as the movie source, except it picks an audio stream by default. that's the whole vibe. iykyk ✨

🦋 avsynctest

generates an Audio/Video Sync Test, no cap. the stream it makes periodically flashes a video frame and drops a beep in the audio — super useful for checking out A/V sync issues. it's giving diagnostic icon 💅

it accepts these options:

size, s — set the output video size. default is "hd720".
framerate, fr — set the output video frame rate. default is 30.
samplerate, sr — set the output audio sample rate. default is 44100.
amplitude, a — set the output audio beep amplitude. default is 0.7.
period, p — set the output audio beep period in seconds. default is 3.
delay, dl — set the output video flash delay in number of frames. default is 0.
cycle, c — enable cycling of video delays; by default it's disabled.
duration, d — set stream output duration. by default duration is unlimited.
fg, bg, ag — set foreground/background/additional color.

Commands

this source supports some of the above options as commands too 🫶

🌸 movie

reads audio and/or video stream(s) out of a movie container. main character behavior honestly.

it accepts these parameters:

filename — the name of the resource to read. and like, 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 is 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 give several streams, separated by +. the source will then have as many outputs, in the same order. 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 auto-selected. default value is -1. deprecated btw. if the filter is called amovie, it'll select audio instead of video.
loop — specifies how many times to read the stream in sequence. if the value is 0, the stream will be looped infinitely. default value is 1.

heads up bestie: when the movie is looped the source timestamps are not changed, so it'll generate non monotonically increasing timestamps. no cap, that's a real thing.

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 given 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 lets you overlay a second video on top of the main input of a filtergraph, like this graph right here:

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 these commands:

seek — perform seek using av_seek_frame. the syntax is:

seek stream_index|timestamp|flags

stream_index: if stream_index is -1, a default stream is selected, and timestamp is automatically converted from AV_TIME_BASE units to the stream specific time_base.
timestamp: timestamp in AVStream.time_base units or, if no stream is specified, in AV_TIME_BASE units.
flags: flags which select direction and seeking mode.

get_duration — get movie duration in AV_TIME_BASE units.

👑 EXTERNAL LIBRARIES 👑

FFmpeg can be hooked up with a whole bunch 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. that's the rule 💅

⭐ Alliance for Open Media (AOM)

FFmpeg can make use of the AOM library for AV1 decoding and encoding. slay.

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). iconic.

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 happens 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, amdgru-pro version 19.20+ and the amf-amdgpu-pro package are required (amdgru-pro contains it, but doesn't install it automatically). fr fr.

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, the 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.

okay so 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's 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. periodt. 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>. lowkey obsessed with fingerprinting ngl.

💜 codec2

FFmpeg can make use of the codec2 library for codec2 decoding and encoding. there's 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.

⚠️ okay important: 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), so you must upgrade FFmpeg's license to GPL in order to use it.

🦋 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 bestie, FFmpeg can lowkey use the Game Music Emu library to read audio out of all those supported video game music file formats — yes, the vibes of your fave retro soundtrack, decoded. To turn it on, pass --enable-libgme to configure. Go peep <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. Iconic. But there's a catch, bestie: to use QSV, FFmpeg must be linked against the libmfx dispatcher, which is the thing that loads the actual decoding libraries.

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. No .pc files, no party. 🫶

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. slay.

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. that's the MP3 encoder of all time fr fr.

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. noted, bestie — keep that license clean.

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 — like, it's giving "the next main character image format." 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. vintage tracker music nation rise 🙈

OpenCORE, VisualOn, and Fraunhofer libraries 🔊

Spun off Google Android sources, OpenCore, VisualOn and Fraunhofer libraries provide encoders for a number of audio codecs.

⚠️ okay listen up because this part is important. OpenCORE and VisualOn libraries are under the Apache License 2.0 (see <http://www.apache.org/licenses/LICENSE-2.0> for details), which is incompatible to the LGPL version 2.1 and GPL version 2. You have to upgrade FFmpeg's license to LGPL version 3 (or if you have enabled GPL components, GPL version 3) by passing --enable-version3 to configure in order to use it.

⚠️ and ALSO: the license of the Fraunhofer AAC library is incompatible with the GPL. Therefore, for GPL builds, you have to pass --enable-nonfree to configure in order to use it. To the best of our knowledge, it is compatible with the LGPL. no cap, read the fine print bestie.

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. (both flags, together, no skipping 💅)

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 — like, 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>). so like, know what you're getting into, k? 😭

OpenJPEG 🌈

FFmpeg can use the OpenJPEG libraries for decoding/encoding J2K videos. Go to <http://www.openjpeg.org/> to get the libraries and follow the installation instructions. To enable using OpenJPEG in FFmpeg, pass --enable-libopenjpeg to ./configure.

rav1e 🦀

FFmpeg can make use of rav1e (Rust AV1 Encoder) via its C bindings to encode videos. Go to <https://github.com/xiph/rav1e/> and follow the instructions to build the C library. To enable using rav1e in FFmpeg, pass --enable-librav1e to ./configure. rust girlies stay winning ✨

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. like:

-f vapoursynth -i yourscript.vpy

don't forget it or it simply won't work, bestie 💀

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. (yes, the little text pages on old TV — iconic behaviour.)

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 FULL list, no tea left unspilled. 👑

File Formats 🗂️

FFmpeg supports the following file formats through the libavformat library:

the columns go like this, mkay:

Name  :  Encoding @tab Decoding @tab Comments

3dostr — decode: X
4xm — decode: X — 4X Technologies format, used in some games.
8088flex TMV — decode: X
AAX — decode: X — Audible Enhanced Audio format, used in audiobooks.
AA — decode: X — Audible Format 2, 3, and 4, used in audiobooks.
ACT Voice — decode: X — contains G.729 audio
Adobe Filmstrip — encode: X · decode: X
Audio IFF (AIFF) — encode: X · decode: X
American Laser Games MM — decode: X — Multimedia format used in games like Mad Dog McCree.
3GPP AMR — demux X, mux X 💖
Amazing Studio Packed Animation File — mux X only; multimedia format used in game Heart Of Darkness. 🎮
Apple HTTP Live Streaming — mux X only. 🍎
Artworx Data Format — mux X only. 🎨
Interplay ACM — mux X only; audio only format used in some Interplay games. 🎧
ADP — mux X only; audio format used on the Nintendo Gamecube. 🎮
AFC — mux X only; audio format used on the Nintendo Gamecube. 🎮
ADS/SS2 — mux X only; audio format used on the PS2. 🎮
APNG — demux X, mux X ✨
ASF — demux X, mux X; Advanced / Active Streaming Format. 🦋
AST — demux X, mux X; audio format used on the Nintendo Wii. 🎮
AVI — demux X, mux X 🎀
AviSynth — mux X only. 🧚
AVR — mux X only; audio format used on Mac. 🍎
AVS — mux X only; multimedia format used by the Creature Shock game. 👾
Beam Software SIFF — mux X only; audio and video format used in some games by Beam Software. 🎮
Bethesda Softworks VID — mux X only; used in some games from Bethesda Softworks. 🎮
Binary text — mux X only. 🔤
Bink — mux X only; multimedia format used by many games. 🎮
Bink Audio — mux X only; audio only multimedia format used by some games. 🎧
Bitmap Brothers JV — mux X only; used in Z and Z95 games. 🕹️
BRP — mux X only; Argonaut Games format. 🎮
Brute Force & Ignorance — mux X only; used in the game Flash Traffic: City of Angels. 🚓
BFSTM — mux X only; audio format used on the Nintendo WiiU (based on BRSTM). 🎮
BRSTM — mux X only; audio format used on the Nintendo Wii. 🎮
BW64 — mux X only; Broadcast Wave 64bit. 🌊
BWF — demux X, mux X 💖
codec2 (raw) — demux X, mux X; must be given -mode format option to decode correctly. ⚠️
codec2 (.c2 files) — demux X, mux X; contains header with version and mode info, simplifying playback. 🎧
CRI ADX — demux X, mux X; audio-only format used in console video games. 🎮
CRI AIX — mux X only. 🎮
CRI HCA — mux X only; audio-only format used in console video games. 🎮
Discworld II BMV — mux X only. 🪄
Interplay C93 — mux X only; used in the game Cyberia from Interplay. 🎮
Delphine Software International CIN — mux X only; multimedia format used by Delphine Software games. 🎮
Digital Speech Standard (DSS) — mux X only. 🎙️
CD+G — mux X only; video format used by CD+G karaoke disks. 🎤
Phantom Cine — mux X only. 🎥
Commodore CDXL — mux X only; Amiga CD video format. 💿
Core Audio Format — demux X, mux X; Apple Core Audio Format. 🍎
CRC testing format — demux X only. 🧪
Creative Voice — demux X, mux X; created for the Sound Blaster Pro. 🔊
CRYO APC — mux X only; audio format used in some games by CRYO Interactive Entertainment. 🎮
D-Cinema audio — demux X, mux X 🎬
Deluxe Paint Animation — mux X only. 🎨
DCSTR — mux X only. 🎮
DFA — mux X only; this format is used in Chronomaster game. ⏳
DirectDraw Surface — mux X only. 🖼️
DSD Stream File (DSF) — mux X only. 🎧
DV video — demux X, mux X 📹
DXA — mux X 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 X only. 🎮
Electronic Arts Multimedia — mux X only; used in various EA games; files have extensions like WVE and UV2. 🎮
Ensoniq Paris Audio File — mux X only. 🎹
FFM (FFserver live feed) — demux X, mux X 📡
Flash (SWF) — demux X, mux X ⚡
Flash 9 (AVM2) — demux X, mux X; only embedded audio is decoded. 🔊
FLI/FLC/FLX animation — mux X only; .fli/.flc files. 🎞️
Flash Video (FLV) — demux X, mux X; Macromedia Flash video files. 📼
framecrc testing format — demux X only. 🧪
FunCom ISS — mux X only; audio format used in various games from FunCom like The Longest Journey. 🎮
G.723.1 — demux X, mux X ☎️
G.726 — mux X only; Both left- and right-justified. ⚖️
G.729 BIT — demux X, mux X ☎️
G.729 raw — mux X only. ☎️
GENH — mux X only; audio format for various games. 🎮
GIF Animation — demux X, mux X 🌈
GXF — demux X, mux X; General eXchange Format SMPTE 360M, used by Thomson Grass Valley playout servers. 📺
HNM — mux X only; only version 4 supported, used in some games from Cryo Interactive. 🎮
iCEDraw File — mux X only. 🧊
ICO — demux X, mux X; Microsoft Windows ICO. 🪟
id Quake II CIN video — mux X only. 🎮
id RoQ — demux X, mux X; used in Quake III, Jedi Knight 2 and other computer games. 🕹️
IEC61937 encapsulation — demux X, mux X 🔌
IFF — mux X only; Interchange File Format. 📦
IFV — mux X only; a format used by some old CCTV DVRs. 📹
iLBC — demux X, mux X 🎧
Interplay MVE — mux X only; format used in various Interplay computer games. 🎮
Iterated Systems ClearVideo — mux X only; I-frames only. 🖼️
IV8 — mux X only; a format generated by IndigoVision 8000 video server. 📹
IVF (On2) — demux X, mux X; a format used by libvpx. 🎥
Internet Video Recording — mux X only. 🌐
IRCAM — demux X, mux X 🎧
LAF — mux X only; Limitless Audio Format. 🎧
LATM — demux X, mux X 🎧
LMLM4 — mux X only; used by Linux Media Labs MPEG-4 PCI boards. 🐧
LOAS — mux X only; contains LATM multiplexed AAC audio. 🔊
LRC — demux X, mux X 📝
LVF — mux X only. 🎮
LXF — mux X only; VR native stream format, used by Leitch/Harris' video servers. 📺
Magic Lantern Video (MLV) — mux X only. 📷
Matroska — demux X, mux X 🎀
Matroska audio — demux X only. 🎧
FFmpeg metadata — demux X, mux X; metadata in text format. 🏷️
MAXIS XA — mux X only; used in Sim City 3000; file extension .xa. 🏙️
MCA — mux X only; used in some games from Capcom; file extension .mca. 🎮
MD Studio — mux X only. 🎮
Metal Gear Solid: The Twin Snakes — mux X only. 🐍
Megalux Frame — mux X only; used by Megalux Ultimate Paint. 🎨
MobiClip MODS — mux X only. 📱
MobiClip MOFLEX — mux X only. 📱
Mobotix .mxg — mux X only. 📹
Monkey's Audio — mux X only. 🐒
Motion Pixels MVI — mux X only. 🎞️
MOV/QuickTime/MP4 — demux X, mux X; 3GP, 3GP2, PSP, iPod variants supported. 🍎
MP2 — demux X, mux X 🎧
MP3 — demux X, mux X 🎧
MPEG-1 System — demux X, mux X; muxed audio and video, VCD format supported. 📀
MPEG-PS (program stream) — demux X, mux X; also known as VOB file, SVCD and DVD format supported. 📀
MPEG-TS (transport stream) — demux X, mux X; also known as DVB Transport Stream. 📡
MPEG-4 — demux X, mux X; MPEG-4 is a variant of QuickTime. 🎬
MSF — mux X only; audio format used on the PS3. 🎮
Mirillis FIC video — mux X only; no cursor rendering. 🖱️
MIDI Sample Dump Standard — mux X only. 🎹
MIME multipart JPEG — demux X only. 🖼️
MSN TCP webcam — mux X only; used by MSN Messenger webcam streams. 📹
MTV — mux X only. 📺
Musepack — mux X only. 🎧
Musepack SV8 — mux X only. 🎧
Material eXchange Format (MXF) — demux X, mux X; SMPTE 377M, used by D-Cinema, broadcast industry. 🎬
Material eXchange Format (MXF), D-10 Mapping — demux X, mux X; SMPTE 386M, D-10/IMX Mapping. 🎬
NC camera feed — mux X only; NC (AVIP NC4600) camera streams. 📹
NIST SPeech HEader REsources — mux X only. 🗣️
Computerized Speech Lab NSP — mux X only. 🗣️
NTT TwinVQ (VQF) — mux X only; Nippon Telegraph and Telephone Corporation TwinVQ. ☎️
Nullsoft Streaming Video — mux X only. 📺
NuppelVideo — mux X only. 🎞️
NUT — demux X, mux X; NUT Open Container Format. 🥜
Ogg — demux X, mux X 🎧
Playstation Portable PMP — mux X only. 🎮
Portable Voice Format — mux X only. 🎙️
RK Audio (RKA) — mux X only. 🎧
TechnoTrend PVA — mux X only; used by TechnoTrend DVB PCI boards. 📺
QCP — decode only, bestie (encode column is blank, decode is X) 💅
raw ADTS (AAC) — both ways, encode X + decode X ✨
raw AC-3 — both ways, encode X + decode X 🔥
raw AMR-NB — decode only, encode blank / decode X 🎀
raw AMR-WB — decode only, encode blank / decode X 🎀
raw APAC — decode only, encode blank / decode X 🦋
raw APV — both ways, encode X + decode X ✨
raw aptX — both ways, encode X + decode X 🎧
raw aptX HD — both ways, encode X + decode X 🎧
raw Bonk — decode only, encode blank / decode X 💀
raw Chinese AVS video — both ways, encode X + decode X 🌸
raw DFPWM — both ways, encode X + decode X ✨
raw Dirac — both ways, encode X + decode X 🌀
raw DNxHD — both ways, encode X + decode X 🎬
raw DTS — both ways, encode X + decode X 🔥
raw DTS-HD — decode only, encode blank / decode X 🎧
raw E-AC-3 — both ways, encode X + decode X 🔥
raw EVC — both ways, encode X + decode X ✨
raw FLAC — both ways, encode X + decode X 🎶
raw G.728 — decode only, encode blank / decode X 📞
raw GSM — decode only, encode blank / decode X 📞
raw H.261 — both ways, encode X + decode X 🎥
raw H.263 — both ways, encode X + decode X 🎥
raw H.264 — both ways, encode X + decode X 👑
raw HEVC — both ways, encode X + decode X 👑
raw Ingenient MJPEG — decode only, encode blank / decode X 🖼️
raw MJPEG — both ways, encode X + decode X 🖼️
raw MLP — decode only, encode blank / decode X 🎧
raw MPEG — decode only, encode blank / decode X 📼
raw MPEG-1 — decode only, encode blank / decode X 📼
raw MPEG-2 — decode only, encode blank / decode X 📼
raw MPEG-4 — both ways, encode X + decode X 📼
raw NULL — encode only, encode X / decode column blank 💀
raw video — both ways, encode X + decode X 🎞️
raw id RoQ — encode only, encode X / decode column blank 🎮
raw OBU — both ways, encode X + decode X ✨
raw OSQ — decode only, encode blank / decode X 🎵
raw SBC — both ways, encode X + decode X 🎧
raw Shorten — decode only, encode blank / decode X 🩳
raw TAK — decode only, encode blank / decode X 🎶
raw TrueHD — both ways, encode X + decode X 🔊
raw VC-1 — both ways, encode X + decode X 🎬
raw PCM A-law — both ways, encode X + decode X 🔉
raw PCM mu-law — both ways, encode X + decode X 🔉
raw PCM Archimedes VIDC — both ways, encode X + decode X 🏛️
raw PCM signed 8 bit — both ways, encode X + decode X 🎚️
raw PCM signed 16 bit big-endian — both ways, encode X + decode X 🎚️
raw PCM signed 16 bit little-endian — both ways, encode X + decode X 🎚️
raw PCM signed 24 bit big-endian — both ways, encode X + decode X 🎚️
raw PCM signed 24 bit little-endian — both ways, encode X + decode X 🎚️
raw PCM signed 32 bit big-endian — both ways, encode X + decode X 🎚️
raw PCM signed 32 bit little-endian — both ways, encode X + decode X 🎚️
raw PCM signed 64 bit big-endian — both ways, encode X + decode X 🎚️
raw PCM signed 64 bit little-endian — both ways, encode X + decode X 🎚️
raw PCM unsigned 8 bit — both ways, encode X + decode X 🎚️
raw PCM unsigned 16 bit big-endian — both ways, encode X + decode X 🎚️
raw PCM unsigned 16 bit little-endian — both ways, encode X + decode X 🎚️
raw PCM unsigned 24 bit big-endian — both ways, encode X + decode X 🎚️
raw PCM unsigned 24 bit little-endian — both ways, encode X + decode X 🎚️
raw PCM unsigned 32 bit big-endian — both ways, encode X + decode X 🎚️
raw PCM unsigned 32 bit little-endian — both ways, encode X + decode X 🎚️
raw PCM 16.8 floating point little-endian — decode only, encode blank / decode X 🌊
raw PCM 24.0 floating point little-endian — decode only, encode blank / decode X 🌊
raw PCM floating-point 32 bit big-endian — both ways, encode X + decode X 🌊
raw PCM floating-point 32 bit little-endian — both ways, encode X + decode X 🌊
raw PCM floating-point 64 bit big-endian — both ways, encode X + decode X 🌊
raw PCM floating-point 64 bit little-endian — both ways, encode X + decode X 🌊
RDT — decode only, encode blank / decode X 📡
REDCODE R3D — decode only, encode blank / decode X 🎥 — 💬 File format used by RED Digital cameras, contains JPEG 2000 frames and PCM audio.
RealMedia — both ways, encode X + decode X 📺
Redirector — decode only, encode blank / decode X 🔀
RedSpark — decode only, encode blank / decode X ⚡
Renderware TeXture Dictionary — decode only, encode blank / decode X 🧱
Resolume DXV — both ways, encode X + decode X 🎛️ — 💬 Encoding is only supported for the DXT1 (Normal Quality, No Alpha) texture format.
RF64 — decode only, encode blank / decode X 📦
RL2 — decode only, encode blank / decode X 🎮 — 💬 Audio and video format used in some games by Entertainment Software Partners.
RPL/ARMovie — decode only, encode blank / decode X 🎞️
Lego Mindstorms RSO — both ways, encode X + decode X 🤖
RSD — decode only, encode blank / decode X 🎵
RTMP — both ways, encode X + decode X 📡 — 💬 Output is performed by publishing stream to RTMP server
RTP — both ways, encode X + decode X 📡
RTSP — both ways, encode X + decode X 🎥
Sample Dump eXchange — decode only, encode blank / decode X 🎛️
SAP — both ways, encode X + decode X 📡
SBG — decode only, encode blank / decode X 🎮
SDNS — decode only, encode blank / decode X 🔉
SDP — decode only, encode blank / decode X 📄
SER — decode only, encode blank / decode X 🔭
Digital Pictures SGA — decode only, encode blank / decode X 🖼️
Sega FILM/CPK — both ways, encode X + decode X 🎮 — 💬 Used in many Sega Saturn console games.
Silicon Graphics Movie — decode only, encode blank / decode X 🎬
Sierra SOL — decode only, encode blank / decode X 🕹️ — 💬 .sol files used in Sierra Online games.
Sierra VMD — decode only, encode blank / decode X 💿 — 💬 Used in Sierra CD-ROM games.
Smacker — decode only, encode blank / decode X 🎮 — 💬 Multimedia format used by many games.
SMJPEG — both ways, encode X + decode X 🎞️ — 💬 Used in certain Loki game ports.
SMPTE 337M encapsulation — decode only, encode blank / decode X 📻
Smush — decode only, encode blank / decode X 🎮 — 💬 Multimedia format used in some LucasArts games.
Sony OpenMG (OMA) — both ways, encode X + decode X 🎧 — 💬 Audio format used in Sony Sonic Stage and Sony Vegas.
Sony PlayStation STR — decode only, encode blank / decode X 🎮
Sony Wave64 (W64) — both ways, encode X + decode X 🌊
SoX native format — both ways, encode X + decode X 🎚️
SUN AU format — both ways, encode X + decode X ☀️
SUP raw PGS subtitles — both ways, encode X + decode X 💬
SVAG — decode only, encode blank / decode X 🎮 — 💬 Audio format used in Konami PS2 games.
TDSC — decode only, encode blank / decode X 📷
Text files — decode only, encode blank / decode X 📝
THP — decode only, encode blank / decode X 🎮 — 💬 Used on the Nintendo GameCube.
Tiertex Limited SEQ — decode only, encode blank / decode X 💾 — 💬 Tiertex .seq files used in the DOS CD-ROM version of the game Flashback.
True Audio — both ways, encode X + decode X 🎶
VAG — decode only, encode blank / decode X 🎮 — 💬 Audio format used in many Sony PS2 games.
VC-1 test bitstream — both ways, encode X + decode X 🧪
Vidvox Hap — both ways, encode X + decode X 🎬
Vivo — decode only, encode blank / decode X 🌐
VPK — decode only, encode blank / decode X 🎮 — 💬 Audio format used in Sony PS games.
Marble WADY — decode only, encode blank / decode X 🎲
WAV — both ways, encode X + decode X 🌊
Waveform Archiver — decode only, encode blank / decode X 📼
WavPack — both ways, encode X + decode X 🎶
WebM — both ways, encode X + decode X 🌐
Windows Televison (WTV) — both ways, encode X + decode X 📺
Wing Commander III movie — decode only, encode blank / decode X 🚀 — 💬 Multimedia format used in Origin's Wing Commander III computer game.
Westwood Studios audio — both ways, encode X + decode X 🎧 — 💬 Multimedia format used in Westwood Studios games.
Westwood Studios VQA — both ways, encode X + decode X 🎬 — 💬 Multimedia format used in Westwood Studios games.
Wideband Single-bit Data (WSD) — decode only, encode blank / decode X 📻
WVE — decode only, encode blank / decode X 🎵
Konami XMD — decode only, encode blank / decode X 🎮
XMV — decode only, encode blank / decode X 🎮 — 💬 Microsoft video container used in Xbox games.
XVAG — decode only, encode blank / decode X 🎮 — 💬 Audio format used on the PS3.
xWMA — decode only, encode blank / decode X 🎧 — 💬 Microsoft audio container used by XAudio 2.
eXtended BINary text (XBIN) — both ways, encode X + decode X 📝
YUV4MPEG pipe — both ways, encode X + decode X 🎞️
Psygnosis YOP — decode only, encode blank / decode X 🎮

Okay so the whole vibe of those two little columns: X means that the feature in that column — encoding / decoding — is supported. That's it, that's the whole legend 🫶

🖼️ Image Formats 🖼️

FFmpeg can read and write images for each frame of a video sequence, which is honestly iconic of it 💅 The following image formats are supported:

.Y.U.V — both ways, encode X + decode X 🌈 — 💬 one raw file per component
Alias PIX — both ways, encode X + decode X 🎨 — 💬 Alias/Wavefront PIX image format
animated GIF — both ways, encode X + decode X 🌀
APNG — both ways, encode X + decode X ✨ — 💬 Animated Portable Network Graphics
BMP — both ways, encode X + decode X 🖼️ — 💬 Microsoft BMP image
BRender PIX — decode only, encode blank / decode X 🎮 — 💬 Argonaut BRender 3D engine image format.
CRI — decode only, encode blank / decode X 🎥 — 💬 Cintel RAW
DPX — both ways, encode X + decode X 🎞️ — 💬 Digital Picture Exchange
EXR — decode only, encode blank / decode X 🌟 — 💬 OpenEXR
FITS — both ways, encode X + decode X 🔭 — 💬 Flexible Image Transport System
HDR — both ways, encode X + decode X ☀️ — 💬 Radiance HDR RGBE Image format
IMG — decode only, encode blank / decode X 💾 — 💬 GEM Raster image
JPEG — both ways, encode X + decode X 📷 — 💬 Progressive JPEG is not supported.
JPEG 2000 — both ways, encode X + decode X 🖼️
JPEG-LS — both ways, encode X + decode X 🖼️
JPEG-XS — E + E, so both ways via the external-library route 🌸 — 💬 Supported through external library libsvtjpegxs
LJPEG — encode only, encode X / decode column blank 💅 — 💬 Lossless JPEG
Media 100 — decode only, encode blank / decode X 🎬
MSP — decode only, encode blank / decode X 🎨 — 💬 Microsoft Paint image
PAM — both ways, encode X + decode X 🖌️ — 💬 PAM is a PNM extension with alpha support.
PBM — both ways, encode X + decode X ⚫ — 💬 Portable BitMap image
PCD — decode only, encode blank / decode X 💿 — 💬 PhotoCD
PCX — both ways, encode X + decode X 🖍️ — 💬 PC Paintbrush
PFM — both ways, encode X + decode X 🌊 — 💬 Portable FloatMap image
PGM — both ways, encode X + decode X ⬜ — 💬 Portable GrayMap image
PGMYUV — both ways, encode X + decode X 🎚️ — 💬 PGM with U and V components in YUV 4:2:0
PGX — decode only, encode blank / decode X 📄 — 💬 PGX file decoder
PHM — both ways, encode X + decode X 🌊 — 💬 Portable HalfFloatMap image
PIC — decode only, encode blank / decode X 🎨 — 💬 Pictor/PC Paint
PNG — both ways, encode X + decode X 🦋 — 💬 Portable Network Graphics image
PPM — both ways, encode X + decode X 🎨 — 💬 Portable PixelMap image
PSD — decode only, encode blank / decode X 🖌️ — 💬 Photoshop
PTX — decode only, encode blank / decode X 🎮 — 💬 V.Flash PTX format
QOI — both ways, encode X + decode X 💖 — 💬 Quite OK Image format
SGI — both ways, encode X + decode X 🌈 — 💬 SGI RGB image format
Sun Rasterfile — both ways, encode X + decode X ☀️ — 💬 Sun RAS image format
TIFF — both ways, encode X + decode X 📷 — 💬 YUV, JPEG and some extension is not supported yet.
Truevision Targa — both ways, encode X + decode X 🎨 — 💬 Targa (.TGA) image format
VBN — both ways, encode X + decode X 📺 — 💬 Vizrt Binary Image format
WBMP — both ways, encode X + decode X 📱 — 💬 Wireless Application Protocol Bitmap image format
WebP — encode E / decode X 🦋 — 💬 WebP image format, encoding supported through external library libwebp
XBM — both ways, encode X + decode X 🖼️ — 💬 X BitMap image format
XFace — both ways, encode X + decode X 🙂 — 💬 X-Face image format
XPM — decode only, encode blank / decode X 🎨 — 💬 X PixMap image format
XWD — both ways, encode X + decode X 🪟 — 💬 X Window Dump image format

Same legend as before, bestie: X means that the feature in that column — encoding / decoding — is supported. 💅

And E means that support is provided through an external library. So if you see an E, that's a "go install the library first" moment, not a "ffmpeg lied to me" moment 🫶

🎬 Video Codecs 🎬

Name : Encoding @tab Decoding @tab Comments — same deal, here come the codecs:

4X Movie — decode only, encode blank / decode X 🎮 — 💬 Used in certain computer games.
8088flex TMV — enc: — · dec: X
A64 multicolor — enc: X · dec: — — creates video suitable to be played on a commodore 64 (multicolor mode) 🕹️
Amazing Studio PAF Video — enc: — · dec: X
American Laser Games MM — enc: — · dec: X — used in games like Mad Dog McCree
Amuse Graphics Movie — enc: — · dec: X
AMV Video — enc: X · dec: X — used in Chinese MP3 players ✨
ANSI/ASCII art — enc: — · dec: X
Apple Intermediate Codec — enc: — · dec: X
Apple MJPEG-B — enc: — · dec: X
Apple Pixlet — enc: — · dec: X
Apple ProRes — enc: X · dec: X — fourcc: apch,apcn,apcs,apco,ap4h,ap4x
Apple QuickDraw — enc: — · dec: X — fourcc: qdrw
APV — enc: — · dec: X
Argonaut Video — enc: — · dec: X — used in some Argonaut games
Asus v1 — enc: X · dec: X — fourcc: ASV1
Asus v2 — enc: X · dec: X — fourcc: ASV2
ATI VCR1 — enc: — · dec: X — fourcc: VCR1
ATI VCR2 — enc: — · dec: X — fourcc: VCR2
Auravision Aura — enc: — · dec: X
Auravision Aura 2 — enc: — · dec: X
Autodesk Animator Flic video — enc: — · dec: X
Autodesk RLE — enc: — · dec: X — fourcc: AASC
AV1 — enc: E · dec: E — supported through external libraries libaom, libdav1d, librav1e and libsvtav1 🔥
Avid 1:1 10-bit RGB Packer — enc: X · dec: X — fourcc: AVrp
AVS (Audio Video Standard) video — enc: — · dec: X — video encoding used by the Creature Shock game
AVS2-P2/IEEE1857.4 — enc: E · dec: E — supported through external libraries libxavs2 and libdavs2
AVS3-P2/IEEE1857.10 — enc: — · dec: E — supported through external library libuavs3d
AYUV — enc: X · dec: X — Microsoft uncompressed packed 4:4:4:4
Beam Software VB — enc: — · dec: X
Bethesda VID video — enc: — · dec: X — used in some games from Bethesda Softworks
Bink Video — enc: — · dec: X
BitJazz SheerVideo — enc: — · dec: X
Bitmap Brothers JV video — enc: — · dec: X
y41p Brooktree uncompressed 4:1:1 12-bit — enc: X · dec: X
Brooktree ProSumer Video — enc: — · dec: X — fourcc: BT20
Brute Force & Ignorance — enc: — · dec: X — used in the game Flash Traffic: City of Angels
C93 video — enc: — · dec: X — codec used in Cyberia game
CamStudio — enc: — · dec: X — fourcc: CSCD
CD+G — enc: — · dec: X — video codec for CD+G karaoke disks 🎤
CDXL — enc: — · dec: X — Amiga CD video codec
Chinese AVS video — enc: E · dec: X — AVS1-P2, JiZhun profile, encoding through external library libxavs
Delphine Software International CIN video — enc: — · dec: X — codec used in Delphine Software International games
Discworld II BMV Video — enc: — · dec: X
CineForm HD — enc: X · dec: X
Canopus HQ — enc: — · dec: X
Canopus HQA — enc: — · dec: X
Canopus HQX — enc: — · dec: X
Canopus Lossless Codec — enc: — · dec: X
CDToons — enc: — · dec: X — codec used in various Broderbund games
Cinepak — enc: — · dec: X
Cirrus Logic AccuPak — enc: X · dec: X — fourcc: CLJR
CPiA Video Format — enc: — · dec: X
Creative YUV (CYUV) — enc: — · dec: X
DFA — enc: — · dec: X — codec used in Chronomaster game
Dirac — enc: E · dec: X — supported through the native vc2 (Dirac Pro) encoder
Deluxe Paint Animation — enc: — · dec: X
DNxHD — enc: X · dec: X — aka SMPTE VC3
Duck TrueMotion 1.0 — enc: — · dec: X — fourcc: DUCK
Duck TrueMotion 2.0 — enc: — · dec: X — fourcc: TM20
Duck TrueMotion 2.0 RT — enc: — · dec: X — fourcc: TR20
DV (Digital Video) — enc: X · dec: X
Dxtory capture format — enc: — · dec: X
Feeble Files/ScummVM DXA — enc: — · dec: X — codec originally used in Feeble Files game
Electronic Arts CMV video — enc: — · dec: X — used in NHL 95 game
Electronic Arts Madcow video — enc: — · dec: X
Electronic Arts TGV video — enc: — · dec: X
Electronic Arts TGQ video — enc: — · dec: X
Electronic Arts TQI video — enc: — · dec: X
Escape 124 — enc: — · dec: X
Escape 130 — enc: — · dec: X
EVC / MPEG-5 Part 1 — enc: E · dec: E — encoding and decoding supported through external libraries libxeve and libxevd
FFmpeg video codec #1 — enc: X · dec: X — lossless codec (fourcc: FFV1)
Flash Screen Video v1 — enc: X · dec: X — fourcc: FSV1
Flash Screen Video v2 — enc: X · dec: X
Flash Video (FLV) — enc: X · dec: X — Sorenson H.263 used in Flash
FM Screen Capture Codec — enc: — · dec: X
Forward Uncompressed — enc: — · dec: X
Fraps — enc: — · dec: X
Go2Meeting — enc: — · dec: X — fourcc: G2M2, G2M3
Go2Webinar — enc: — · dec: X — fourcc: G2M4
Gremlin Digital Video — enc: — · dec: X
H.261 — enc: X · dec: X
H.263 / H.263-1996 — enc: X · dec: X
H.263+ / H.263-1998 / H.263 version 2 — enc: X · dec: X
H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 — enc: E · dec: X — encoding supported through external library libx264 and OpenH264 👑
HEVC — enc: X · dec: X — encoding supported through external library libx265 and libkvazaar
HNM version 4 — enc: — · dec: X
HuffYUV — enc: X · dec: X
HuffYUV FFmpeg variant — enc: X · dec: X
IBM Ultimotion — enc: — · dec: X — fourcc: ULTI
id Cinematic video — enc: — · dec: X — used in Quake II
id RoQ video — enc: X · dec: X — used in Quake III, Jedi Knight 2, other computer games 🎮
IFF ILBM — enc: — · dec: X — IFF interleaved bitmap
IFF ByteRun1 — enc: — · dec: X — IFF run length encoded bitmap
Infinity IMM4 — enc: — · dec: X
Intel H.263 — enc: — · dec: X
Intel Indeo 2 — enc: — · dec: X
Intel Indeo 3 — enc: — · dec: X
Intel Indeo 4 — enc: — · dec: X
Intel Indeo 5 — enc: — · dec: X
Interplay C93 — enc: — · dec: X — used in the game Cyberia from Interplay
Interplay MVE video — enc: — · dec: X — used in Interplay .MVE files
J2K — enc: X · dec: X
Karl Morton's video codec — enc: — · dec: X — codec used in Worms games 🪱
Kega Game Video (KGV1) — enc: — · dec: X — Kega emulator screen capture codec
Lagarith — enc: — · dec: X
LCEVC / MPEG-5 LCEVC / MPEG-5 Part 2 — enc: — · dec: E — decoding supported through external library liblcevc-dec
LCL (LossLess Codec Library) MSZH — enc: — · dec: X
LCL (LossLess Codec Library) ZLIB — enc: E · dec: E
LEAD MCMP — enc: — · dec: X
LOCO — enc: — · dec: X
LucasArts SANM/Smush — enc: — · dec: X — used in LucasArts games / SMUSH animations
lossless MJPEG — enc: X · dec: X
MagicYUV Video — enc: X · dec: X
Mandsoft Screen Capture Codec — enc: — · dec: X
Microsoft ATC Screen — enc: — · dec: X — also known as Microsoft Screen 3
Microsoft Expression Encoder Screen — enc: — · dec: X — also known as Microsoft Titanium Screen 2
Microsoft RLE — enc: X · dec: X
Microsoft Screen 1 — enc: — · dec: X — also known as Windows Media Video V7 Screen
Microsoft Screen 2 — enc: — · dec: X — also known as Windows Media Video V9 Screen
Microsoft Video 1 — enc: — · dec: X
Mimic — enc: — · dec: X — used in MSN Messenger Webcam streams 💻
Miro VideoXL — enc: — · dec: X — fourcc: VIXL
MJPEG (Motion JPEG) — enc: X · dec: X
Mobotix MxPEG video — enc: — · dec: X
Motion Pixels video — enc: — · dec: X
MPEG-1 video — enc: X · dec: X
MPEG-2 video — enc: X · dec: X
MPEG-4 part 2 — enc: X · dec: X — libxvidcore can be used alternatively for encoding
MPEG-4 part 2 Microsoft variant version 1 — enc: — · dec: X
MPEG-4 part 2 Microsoft variant version 2 — enc: X · dec: X
MPEG-4 part 2 Microsoft variant version 3 — enc: X · dec: X
Newtek SpeedHQ — enc: X · dec: X
Nintendo Gamecube THP video — enc: — · dec: X
NotchLC — enc: — · dec: X
NuppelVideo/RTjpeg — enc: — · dec: X — video encoding used in NuppelVideo files
On2 VP3 — enc: — · dec: X — still experimental ⚠️
On2 VP4 — enc: — · dec: X — fourcc: VP40
On2 VP5 — enc: — · dec: X — fourcc: VP50
On2 VP6 — enc: — · dec: X — fourcc: VP60,VP61,VP62
On2 VP7 — enc: — · dec: X — fourcc: VP70,VP71
VP8 — enc: E · dec: X — fourcc: VP80, encoding supported through external library libvpx
VP9 — enc: E · dec: X — encoding supported through external library libvpx
Pinnacle TARGA CineWave YUV16 — enc: — · dec: X — fourcc: Y216
Q-team QPEG — enc: — · dec: X — fourccs: QPEG, Q1.0, Q1.1
QuickTime 8BPS video — enc: — · dec: X
QuickTime Animation (RLE) video — enc: X · dec: X — fourcc: 'rle '
QuickTime Graphics (SMC) — enc: X · dec: X — fourcc: 'smc '
QuickTime video (RPZA) — enc: X · dec: X — fourcc: rpza
R10K AJA Kona 10-bit RGB Codec — enc: X · dec: X
R210 Quicktime Uncompressed RGB 10-bit — enc: X · dec: X
Raw Video — enc: X · dec: X
RealVideo 1.0 — enc: X · dec: X
RealVideo 2.0 — enc: X · dec: X
RealVideo 3.0 — enc: — · dec: X — still far from ideal 😬
RealVideo 4.0 — enc: — · dec: X
RealVideo 6.0 — enc: — · dec: X
Renderware TXD (TeXture Dictionary) — enc: — · dec: X — texture dictionaries used by the Renderware Engine
RivaTuner Video — enc: — · dec: X — fourcc: 'RTV1'
RL2 video — enc: — · dec: X — used in some games by Entertainment Software Partners
ScreenPressor — dec: X
Screenpresso — dec: X
Screen Recorder Gold Codec — dec: X
Sierra VMD video — dec: X
- the video codec living inside Sierra VMD files 🎀
Silicon Graphics Motion Video Compressor 1 (MVC1) — dec: X
Silicon Graphics Motion Video Compressor 2 (MVC2) — dec: X
Silicon Graphics RLE 8-bit video — dec: X
Smacker video — dec: X
- video encoding used in Smacker ✨
SMPTE VC-1 — dec: X
Snow — enc: X · dec: X
- experimental wavelet codec (fourcc: SNOW) 🧚
Sony PlayStation MDEC (Motion DECoder) — dec: X
Sorenson Vector Quantizer 1 — enc: X · dec: X
- fourcc: SVQ1
Sorenson Vector Quantizer 3 — dec: X
- fourcc: SVQ3
Sunplus JPEG (SP5X) — dec: X
- fourcc: SP5X
TechSmith Screen Capture Codec — dec: X
- fourcc: TSCC
TechSmith Screen Capture Codec 2 — dec: X
- fourcc: TSC2
Theora — enc: E · dec: X
- encoding supported through external library libtheora
Tiertex Limited SEQ video — dec: X
- codec used in the DOS CD-ROM FlashBack game 🌸
Ut Video — enc: X · dec: X
v210 QuickTime uncompressed 4:2:2 10-bit — enc: X · dec: X
v308 QuickTime uncompressed 4:4:4 — enc: X · dec: X
v408 QuickTime uncompressed 4:4:4:4 — enc: X · dec: X
v410 QuickTime uncompressed 4:4:4 10-bit — enc: X · dec: X
VBLE Lossless Codec — dec: X
vMix Video — dec: X
- fourcc: 'VMX1'
VMware Screen Codec / VMware Video — dec: X
- the codec used in videos captured by VMware 👑
Westwood Studios VQA (Vector Quantized Animation) video — dec: X
Windows Media Image — dec: X
Windows Media Video 7 — enc: X · dec: X
Windows Media Video 8 — enc: X · dec: X
Windows Media Video 9 — dec: X
- not completely working — bestie, that's literally a warning label 💀
Wing Commander III / Xan — dec: X
- used in Wing Commander III .MVE files
Wing Commander IV / Xan — dec: X
- used in Wing Commander IV
Winnov WNV1 — dec: X
WMV7 — enc: X · dec: X
YAMAHA SMAF — enc: X · dec: X
Psygnosis YOP Video — dec: X
yuv4 — enc: X · dec: X
- libquicktime uncompressed packed 4:2:0
ZeroCodec Lossless Video — dec: X
ZLIB — enc: X · dec: X
- part of LCL, encoder is experimental 🙈
Zip Motion Blocks Video — enc: X · dec: X
- encoder works only in PAL8 ⭐

Okay so the legend, because this table has lore:

X means that the feature in that column (encoding / decoding) is supported 💖
E means that support is provided through an external library

🎀 Audio Codecs 🎀

The columns are: Name · Encoding · Decoding · Comments 🦋 (same deal as before, X = supported, E = external library).

8SVX exponential — dec: X
8SVX fibonacci — dec: X
AAC — enc: EX · dec: X
- encoding supported through internal encoder and external library libfdk-aac 🔥
AAC+ — enc: E · dec: IX
- encoding supported through external library libfdk-aac
AC-3 — enc: IX · dec: IX
ACELP.KELVIN — dec: X
ADPCM 4X Movie — dec: X
ADPCM Yamaha AICA — dec: X
ADPCM AmuseGraphics Movie — dec: X
ADPCM Argonaut Games — enc: X · dec: X
ADPCM CDROM XA — dec: X
ADPCM Creative Technology — dec: X
- 16 -> 4, 8 -> 4, 8 -> 3, 8 -> 2
ADPCM Electronic Arts — dec: X
- used in various EA titles 🎮
ADPCM Electronic Arts Maxis CDROM XS — dec: X
- used in Sim City 3000
ADPCM Electronic Arts R1 — dec: X
ADPCM Electronic Arts R2 — dec: X
ADPCM Electronic Arts R3 — dec: X
ADPCM Electronic Arts XAS — dec: X
ADPCM G.722 — enc: X · dec: X
ADPCM G.726 — enc: X · dec: X
ADPCM IMA Acorn Replay — dec: X
ADPCM IMA AMV — enc: X · dec: X
- used in AMV files
ADPCM IMA Cunning Developments — dec: X
ADPCM IMA Electronic Arts EACS — dec: X
ADPCM IMA Electronic Arts SEAD — dec: X
ADPCM IMA Funcom — dec: X
ADPCM IMA High Voltage Software ALP — enc: X · dec: X
ADPCM IMA Mobiclip MOFLEX — dec: X
ADPCM IMA QuickTime — enc: X · dec: X
ADPCM IMA Simon & Schuster Interactive — enc: X · dec: X
ADPCM IMA Ubisoft APM — enc: X · dec: X
ADPCM IMA Loki SDL MJPEG — dec: X
ADPCM IMA WAV — enc: X · dec: X
ADPCM IMA Westwood — dec: X
ADPCM ISS IMA — dec: X
- used in FunCom games
ADPCM IMA Dialogic — dec: X
ADPCM IMA Duck DK3 — dec: X
- used in some Sega Saturn console games 🕹️
ADPCM IMA Duck DK4 — dec: X
- used in some Sega Saturn console games
ADPCM IMA Radical — dec: X
ADPCM IMA Xbox — dec: X
ADPCM Microsoft — enc: X · dec: X
ADPCM MS IMA — enc: X · dec: X
ADPCM Nintendo Gamecube AFC — dec: X
ADPCM Nintendo Gamecube DTK — dec: X
ADPCM Nintendo THP — dec: X
ADPCM Playstation — dec: X
ADPCM QT IMA — enc: X · dec: X
ADPCM Sanyo — dec: X
ADPCM SEGA CRI ADX — enc: X · dec: X
- used in Sega Dreamcast games
ADPCM Shockwave Flash — enc: X · dec: X
ADPCM Sound Blaster Pro 2-bit — dec: X
ADPCM Sound Blaster Pro 2.6-bit — dec: X
ADPCM Sound Blaster Pro 4-bit — dec: X
ADPCM VIMA — dec: X
- used in LucasArts SMUSH animations
ADPCM Konami XMD — dec: X
ADPCM Westwood Studios IMA — enc: X · dec: X
- used in Westwood Studios games like Command and Conquer 💅
ADPCM Yamaha — enc: X · dec: X
ADPCM Zork — dec: X
AMR-NB — enc: E · dec: X
- encoding supported through external library libopencore-amrnb
AMR-WB — enc: E · dec: X
- encoding supported through external library libvo-amrwbenc
Amazing Studio PAF Audio — dec: X
Apple lossless audio — enc: X · dec: X
- QuickTime fourcc 'alac' 🍎
aptX — enc: X · dec: X
- used in Bluetooth A2DP
aptX HD — enc: X · dec: X
- used in Bluetooth A2DP
ATRAC1 — dec: X
ATRAC3 — dec: X
ATRAC3+ — dec: X
ATRAC9 — dec: X
Bink Audio — dec: X
- used in Bink and Smacker files in many games
Bonk audio — dec: X
CELT — dec: E
- decoding supported through external library libcelt
codec2 — enc: E · dec: E
- en/decoding supported through external library libcodec2
CRI HCA — dec: X
Delphine Software International CIN audio — dec: X
- codec used in Delphine Software International games
DFPWM — enc: X · dec: X
Digital Speech Standard - Standard Play mode (DSS SP) — dec: X
Discworld II BMV Audio — dec: X
COOK — dec: X
- all versions except 5.1 are supported 👀
DCA (DTS Coherent Acoustics) — enc: X · dec: X
- supported extensions: XCh, XXCH, X96, XBR, XLL, LBR (partially)
Dolby E — dec: X
DPCM Cuberoot-Delta-Exact — dec: X
- used in a few games
DPCM Gremlin — dec: X
DPCM id RoQ — enc: X · dec: X
- used in Quake III, Jedi Knight 2 and other computer games 🔥
DPCM Marble WADY — dec: X
DPCM Interplay — dec: X
- used in various Interplay computer games
DPCM Squareroot-Delta-Exact — dec: X
- used in various games
DPCM Sierra Online — dec: X
- used in Sierra Online game audio files
DPCM Sol — dec: X
DPCM Xan — dec: X
- used in Origin's Wing Commander IV AVI files
DPCM Xilam DERF — dec: X
DSD (Direct Stream Digital), least significant bit first — dec: X
DSD (Direct Stream Digital), most significant bit first — dec: X
DSD (Direct Stream Digital), least significant bit first, planar — dec: X
DSD (Direct Stream Digital), most significant bit first, planar — dec: X
DSP Group TrueSpeech — dec: X
DST (Direct Stream Transfer) — dec: X
DV audio — dec: X
Enhanced AC-3 — enc: X · dec: X
EVRC (Enhanced Variable Rate Codec) — dec: X
FLAC (Free Lossless Audio Codec) — enc: X · dec: IX 🦋
FTR Voice — dec: X
G.723.1 — enc: X · dec: X
G.728 — dec: X
G.729 — dec: X
GSM — enc: E · dec: X
- encoding supported through external library libgsm
GSM Microsoft variant — enc: E · dec: X
- encoding supported through external library libgsm
IAC (Indeo Audio Coder) — dec: X
iLBC (Internet Low Bitrate Codec) — enc: E · dec: EX
- encoding and decoding supported through external library libilbc
IMC (Intel Music Coder) — dec: X
Interplay ACM — dec: X
LC3 — enc: E · dec: E
- supported through external library liblc3
MACE (Macintosh Audio Compression/Expansion) 6:1 — dec: X
Marian's A-pac audio — dec: X
MI-SC4 (Micronas SC-4 Audio) — dec: X
MLP (Meridian Lossless Packing) — enc: X · dec: X
- used in DVD-Audio discs 💿
Monkey's Audio — dec: X
MP1 (MPEG audio layer 1) — dec: IX
MP2 (MPEG audio layer 2) — enc: IX · dec: IX
- encoding is also supported through external library TwoLAME
MP3 (MPEG audio layer 3) — enc: E · dec: IX
- encoding supported through external library LAME; ADU MP3 and MP3onMP4 are also supported 🎧
MPEG-4 Audio Lossless Coding (ALS) — dec: X
MPEG-H 3D Audio — dec: E
- decoding supported through external library libmpeghdec

…and we're sliding straight back into the audio codec roll-call, no intro needed 💅. Reading order for each line: encoding then decoding, and X means it's a yes.

MobiClip FastAudio — decode X only. Encoder said no 🚫
Musepack SV7 — decode X only
Musepack SV8 — decode X only
Nellymoser Asao — encode X, decode X. Both ways, slay ✨
On2 AVC (Audio for Video Codec) — decode X only
Opus — encode E, decode X. Note on the page: encoding supported through external library libopus 🫶
OSQ (Original Sound Quality) — decode X only
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 X only. So specific, love that for it 🎀
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 X only
QCELP / PureVoice — decode X only
QDesign Music Codec 1 — decode X only
QDesign Music Codec 2 — decode X only. Page note: There are still some distortions. 💀 (real of them to admit it)
RealAudio 1.0 (14.4K) — encode X, decode X. Page note: Real 14400 bit/s codec
RealAudio 2.0 (28.8K) — decode X only. Page note: Real 28800 bit/s codec
RealAudio 3.0 (dnet) — encode IX, decode X. Page note: Real low bitrate AC-3 codec
RealAudio Lossless — decode X only
RealAudio SIPR / ACELP.NET — decode X only
RK Audio (RKA) — decode X only
SBC (low-complexity subband codec) — encode X, decode X. Page note: Used in Bluetooth A2DP 🦋
Shorten — decode X only
Sierra VMD audio — decode X only. Page note: Used in Sierra VMD files.
Smacker audio — decode X only
SMPTE 302M AES3 audio — encode X, decode X
Sonic — encode X, decode X. Page note: experimental codec ⚗️
Sonic lossless — encode X, decode X. Page note: experimental codec
Speex — encode E, decode EX. Page note: supported through external library libspeex
TAK (Tom's lossless Audio Kompressor) — decode X only
True Audio (TTA) — encode X, decode X
TrueHD — encode X, decode X. Page note: Used in HD-DVD and Blu-Ray discs.
TwinVQ (VQF flavor) — decode X only
VIMA — decode X only. Page note: Used in LucasArts SMUSH animations.
ViewQuest VQC — decode X only
Vorbis — encode E, decode X. Page note: A native but very primitive encoder exists. 🙈
Voxware MetaSound — decode X only
Waveform Archiver — decode X only
WavPack — encode X, decode X
Westwood Audio (SND1) — decode X only
Windows Media Audio 1 — encode X, decode X
Windows Media Audio 2 — encode X, decode X
Windows Media Audio Lossless — decode X only
Windows Media Audio Pro — decode X only
Windows Media Audio Voice — decode X only
Xbox Media Audio 1 — decode X only
Xbox Media Audio 2 — decode X only

Now the legend, because the letters are doing a lot of heavy lifting here:

X means that the feature in that column (encoding / decoding) is supported.
E means that support is provided through an external library.
I means that an integer-only version is available, too (ensures high performance on systems without hardware floating point support).

💬 Subtitle Formats 💬

Columns go Muxing, Demuxing, Encoding, Decoding — in that order, every line, no cap.

3GPP Timed Text — encode X, decode X (muxing and demuxing stay blank)
AQTitle — demux X, decode X
DVB — mux X, demux X, encode X, decode X. Full house, icon behavior 👑
DVB teletext — demux X, decode E
DVD — mux X, demux X, encode X, decode X
JACOsub — mux X, demux X, decode X
MicroDVD — mux X, demux X, decode X
MPL2 — demux X, decode X
MPsub (MPlayer) — demux X, decode X
PGS — decode X only
PJS (Phoenix) — demux X, decode X
RealText — demux X, decode X
SAMI — demux X, decode X
Spruce format (STL) — demux X, decode X
SSA/ASS — mux X, demux X, encode X, decode X
SubRip (SRT) — mux X, demux X, encode X, decode X
SubViewer v1 — demux X, decode X
SubViewer — demux X, decode X
TED Talks captions — demux X, decode X
TTML — mux X, encode X (demuxing and decoding stay blank)
VobSub (IDX+SUB) — demux X, decode X
VPlayer — demux X, decode X
WebVTT — mux X, demux X, encode X, decode X
XSUB — encode X, decode X

Tiny legend again, because consistency:

X means that the feature is supported.
E means that support is provided through an external library.

🌐 Network Protocols 🌐

Name : Support. That's the whole vibe.

AMQP — E
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
SCTP — X
SFTP — E
TCP — X
TLS — X
UDP — X
ZMQ — E

And the legend, one more time:

X means that the protocol is supported.
E means that support is provided through an external library.

🎛️ Input/Output Devices 🎛️

Two columns here: Input and Output. If a cell's empty in the original, it just doesn't get a X, bestie.

ALSA — input X, output X
BKTR — input X (no output)
caca — output X (no input)
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

Legend:

X means that input/output is supported.

⏱️ Timecode ⏱️

Read and Write, and honestly it's X's across the board, we love a clean sweep 🔥

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

👀 SEE ALSO 👀

Go read your sisters, bestie, they're all iconic:

💖 AUTHORS 💖

The FFmpeg developers. That's the credit line, periodt 🫶

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 a wrap, babe 🌈 — the footer is the source/version-plus-date line that closes out FFPROBE-ALL(1), and you just read the whole thing, every last X. Genuinely obsessed with you rn 💅✨

📦 where this page comes from

alpine 3.24: ffmpeg-doc 8.1.3-r0 · GPL-2.0-or-later AND LGPL-2.1-or-later
nixos 26.05: ffmpeg 8.1.3
nixos 26.05: ffmpeg-headless 8.1.3
💖 girlified by deepseek-ai/DeepSeek-V4.1-Flash-1017 on 2026-11-06: a language model's rewrite of the original, under its terms