rav1e

xiph/rav1e

The fastest and safest AV1 encoder.

Latest release

v0.8.1Jun 16, 2025

Last commit

Aug 31, 2026

License

BSD-2-Clause

Stars

4,160

rav1e is an AV1 encoder written in Rust with hand-written assembly, developed by Xiph.org and Mozilla. It is BSD-2-Clause licensed. Its defaults are sensible, it is easy to build, and it is the AV1 encoder behind several Rust and AVIF tools: libavif can use it (avifenc -c rav1e), and the cavif/ravif AVIF encoder is built on it.

Where rav1e stands

The latest stable release is v0.8.1 (June 2025). The weekly pre-release builds stopped in September 2025. Since then the repository has seen mostly maintenance commits (toolchain bumps, lint fixes) rather than encoder work.

For video, most encoders now use SVT-AV1 (or one of its community forks) or aomenc. rav1e is considerably slower than SVT-AV1, and the mendevi encoder dataset found SVT-AV1 better at equal encoding time. rav1e has changed little since, while SVT-AV1 has kept improving. rav1e is still a reasonable choice if you need a pure-Rust library or already use it through libavif or GStreamer. For a new video workflow, start with SVT-AV1.

Install

Prebuilt binaries are on the releases page, and most distributions package rav1e. To build v0.8.1 you need a Rust toolchain and, on x86-64, NASM 2.14.02 or newer:

cargo install --locked rav1e

Or build the tagged release from source:

git clone --branch v0.8.1 https://github.com/xiph/rav1e.git
cd rav1e
cargo build --release

The binary is at target/release/rav1e. rav1e picks its SIMD code paths at runtime, so a normal build already uses your CPU’s AVX2/AVX-512. RUSTFLAGS="-C target-cpu=native" produces a binary that only runs on CPUs like yours.

Encode

rav1e reads Y4M and writes IVF. The input is a positional argument: -i is the minimum keyframe interval, not an input switch, so older commands like rav1e -i input.y4m do not work.

rav1e input.y4m -o output.ivf --quantizer 80 --speed 6

Pipe from FFmpeg to encode anything else. 10-bit output is more efficient and bands less, even from 8-bit sources:

ffmpeg -i input.mkv -pix_fmt yuv420p10le -strict -1 -f yuv4mpegpipe - | \
  rav1e - -o output.ivf --quantizer 80 --speed 6 --photon-noise 8

FFmpeg still needs -strict -1 to write Y4M above 8 bits. Mux the IVF into MKV or MP4 afterwards (for example ffmpeg -i output.ivf -i input.mkv -map 0:v -map 1:a -c copy final.mkv), or use mkvmerge.

For 4K, add tiles. They help both encoding speed and decoding speed:

ffmpeg -i input.mkv -pix_fmt yuv420p10le -strict -1 -f yuv4mpegpipe - | \
  rav1e - -o output.ivf --quantizer 72 --speed 6 --tile-cols 2 --photon-noise 8

Two-pass target bitrate

rav1e input.y4m -o /dev/null --bitrate 3000 --first-pass stats.bin
rav1e input.y4m -o output.ivf --bitrate 3000 --second-pass stats.bin

--bitrate is in kbit/s. The first pass also writes a video file, so send it to /dev/null (otherwise the second pass stops to ask before overwriting). In bitrate mode --quantizer becomes the worst (maximum) quantizer allowed, and --min-quantizer the best. Each pass reads the input, so use a file or regenerate an identical pipe.

Options that matter

Option	Default	Notes
--quantizer N	100	Constant quantizer, 1–255 (0 aborts: lossless is not implemented). Lower is higher quality. This is the AV1 base_q_idx scale (see below).
--speed N, -s	6	0 (slowest) to 10 (fastest). Faster speeds also shorten the RDO lookahead (40 frames at speeds 0–2, 30 at 3–5, 20 at 6–7, 10 at 8–10).
--photon-noise N	0 (off)	AV1 film-grain synthesis, strength 0–64. The decoder adds the noise, so it costs almost no bitrate and can mask banding. It does not copy the source’s grain. It also lowers full-reference metric scores, so turn it off when comparing encodes with metrics.
--tile-cols N, --tile-rows N, --tiles N	automatic	Tile counts (powers of two), not log2 values. --tiles sets a minimum total and cannot be combined with the other two. More tiles raise thread utilisation and decoder parallelism at a small efficiency cost.
--threads N	all logical CPUs	Thread pool size. A single rav1e encode does not use many cores well without more tiles. On many-core machines, running several chunked encodes (e.g. via Av1an) scales better.
--keyint N, -I / --min-keyint N, -i	240 / 12	Maximum and minimum keyframe distance in frames. Scene-cut keyframes are on by default.
--tune	Psychovisual	Psnr optimises for the metric and looks worse. Leave the default.
--bitrate N, --first-pass, --second-pass	quantizer mode	Bitrate targeting, single or two-pass.
--no-scene-detection	off	For chunked encoders that place keyframes themselves.
--primaries, --transfer, --matrix, --range	unspecified, limited range	Colour signalling written into the stream. It labels the colour; it does not convert it.
--mastering-display, --content-light	unset	HDR10 static metadata.

Converting from aomenc or SVT-AV1

aomenc’s --cq-level and SVT-AV1’s --crf use a 0–63 scale that maps to AV1’s 0–255 quantizer index at roughly ×4. rav1e’s --quantizer is that index directly. A 0–63 value multiplied by 4 (--crf 20 → --quantizer 80) therefore lands on a similar base quantizer. Each encoder distributes bits differently around that base, so file size and quality will not match exactly.

HDR

For HDR10, signal the colour and static metadata from your source:

rav1e input.y4m -o output.ivf --quantizer 80 --speed 6 \
  --primaries BT2020 --transfer SMPTE2084 --matrix BT2020NCL \
  --mastering-display "G(0.17,0.797)B(0.131,0.046)R(0.708,0.292)WP(0.3127,0.329)L(1000,0.0001)" \
  --content-light 1000,400

Replace the mastering-display and content-light values with your source’s. ffprobe -show_frames -read_intervals "%+#1" prints them as fractions, which you need to convert to decimals and cd/m². HDR10+ dynamic metadata is not supported in a release: quietvoid’s pull request #3000, which adds --hdr10plus-json, is still unmerged.

FFmpeg (librav1e)

ffmpeg -i input.mkv -c:v librav1e -qp 80 -speed 6 -c:a libopus output.mkv

-qp maps to --quantizer, and -tiles/-tile-rows/-tile-columns map to the tile options. Other settings go through -rav1e-params key=value:key=value, which only accepts rav1e’s C API keys: speed, quantizer, min_quantizer, bitrate, key_frame_interval, min_key_frame_interval, switch_frame_interval, reservoir_frame_delay, rdo_lookahead_frames, low_latency, no_scene_detection, tune, tiles, tile_rows, tile_cols, threads, still_picture, level and enable_timing_info.

Caution

Photon noise is not one of those keys, so grain synthesis is only available through the rav1e CLI. FFmpeg prints a warning for an unknown -rav1e-params key but encodes anyway.

FFmpeg librav1e private options (n9.0.2)

These are the private options of the librav1e encoder in FFmpeg 9.0.2. Generic FFmpeg options also reach rav1e: -g and -keyint_min set the keyframe interval, -b:v (bit/s) switches to bitrate mode with -qmax (default 255) and -qmin as quantizer limits, -pass 1/-pass 2 run a two-pass encode (needs -b:v), -threads sets the thread pool, and the input’s colour range, primaries, transfer and matrix are signalled automatically. -rav1e-params is applied first, so these and the private options below override the same setting given there.

Private AVOptions (6 options)

Option	Default	Type and range	What it does
-qp	-1	int, -1–255	Constant quantizer, same scale as --quantizer. -1 leaves rav1e’s default of 100. Takes priority over -b:v if both are given.
-speed	-1	int, -1–10	Speed preset, as --speed. -1 leaves rav1e’s default of 6. Setting it resets the speed-derived keys rdo_lookahead_frames and no_scene_detection given in -rav1e-params.
-tiles	0	int, ≥ -1	Minimum total tile count, as --tiles; rav1e picks the rows and columns and ignores -tile-rows/-tile-columns. 0 or -1 leaves tiling to rav1e.
-tile-rows	0	int, ≥ -1	Number of tile rows: a count (power of two), not log2. 0 or -1 leaves it automatic.
-tile-columns	0	int, ≥ -1	Number of tile columns, same rules as -tile-rows. The FFmpeg name is -tile-columns, not -tile-cols.
-rav1e-params	unset	dictionary	key=value pairs separated by :, using rav1e’s C API keys (listed above), not CLI flag names. Booleans must be true or false, and bitrate is in bit/s. Put speed first: it resets lookahead and scene-detection keys. Bad keys only warn.

Complete native parameter reference

Every option of the rav1e 0.8.1 command-line encoder. Options of the advanced subcommand go after the main options and the advanced keyword (rav1e in.y4m -o out.ivf advanced --save-config enc.toml). Named values such as Psnr, Full or BT2020 are case-insensitive.

Input and output (7 options)

Option	Default	What it does
INPUT	required	Y4M input file, or - for stdin. It is positional: -i is --min-keyint, not an input flag.
--output, -o	required	IVF output file, or - for stdout. A file name must end in .ivf or have no extension.
-y	off	Overwrite an existing output file. Without it rav1e asks for confirmation on stdin, so an encode with piped input exits instead.
--limit, -l	0 (all frames)	Stop after this many frames. --still-picture allows at most 1.
--skip	0	Read and discard this many frames from the start of the input before encoding.
--help, -h	—	Print help; -h gives the short form.
--version, -V	—	Print the version. --version also shows the compiler, assembly level and whether threading and unstable features were built in.

Rate control and speed (9 options)

Option	Default	What it does
--quantizer	100 (255 with --bitrate)	Quantizer index 1–255; lower is higher quality. 0 (lossless) is not implemented and aborts. With --bitrate it is the highest quantizer rate control may use.
--min-quantizer	0	Lowest quantizer (best quality) rate control may use in bitrate mode, 0–255.
--bitrate, -b	unset (quantizer mode)	Target bitrate in kbit/s. Switches from constant quantizer to rate control.
--first-pass	unset	Run the first pass of a two-pass encode and write its statistics to this file. Needs --bitrate.
--second-pass	unset	Run the second pass using statistics from --first-pass. Needs --bitrate and the same input.
--reservoir-frame-delay	min(240, 1.5 × keyint), at least 12	Window, in frames, over which rate control spreads its bit reservoir; minimum 12. A longer window lets the bitrate vary more between scenes.
--speed, -s	6	Preset 0–10; 0 is slowest and best. Slower presets search more partition sizes, transform types and motion vectors. Upstream does not recommend 0 or 10.
--rdo-lookahead-frames	20 at speed 6	Frames analysed ahead for rate–distortion decisions. The preset sets 40 at speeds 0–2, 30 at 3–5, 20 at 6–7 and 10 at 8–10.
--tune	Psychovisual	Psychovisual or Psnr. Psnr optimises for the metric and looks worse.

Keyframes and frame structure (7 options)

Option	Default	What it does
--keyint, -I	240	Maximum frames between keyframes. 0 removes the limit, leaving only scene-cut keyframes.
--min-keyint, -i	12	Minimum frames between scene-cut keyframes; capped at --keyint.
--no-scene-detection	off	Turn off scene-cut keyframes. Use it when a chunking tool such as Av1an already splits at scene cuts; on a full encode it costs quality.
--scd-speed	0 (1 at speed 10)	Scene-detection mode: 0 standard, 1 fast.
--switch-frame-interval, -S	0 (off)	Insert an AV1 switch frame at most every N frames. A player can change to another stream of the same content at a switch frame.
--low-latency	off	Disable frame reordering so no frame waits for later ones, for real-time use. Compresses much less efficiently.
--still-picture	off	Encode only the first frame, with the AV1 still-picture sequence header. The output is still IVF; use an AVIF encoder such as avifenc or cavif for images.

Film grain (2 options)

Option	Default	What it does
--photon-noise	0 (off)	Add synthetic luma grain at strength 0–64 (strength × 100 is the simulated ISO, so 8 ≈ ISO 800). The decoder renders it through AV1 film-grain synthesis; it does not model the source’s grain. Cannot combine with a grain table.
--film-grain-table, --photon-noise-table	unset	Apply grain from a film-grain table file, in the text format aomenc and SVT-AV1 also read. --photon-noise-table is a hidden alias.

Threading and tiles (5 options)

Option	Default	What it does
--threads	0 (logical CPUs)	Maximum worker threads. One encode needs several tiles to keep many threads busy.
--tile-cols	0 (automatic)	Number of tile columns: a count (power of two, up to 64), not log2. 0 uses the fewest the frame size allows, and rav1e adjusts other values to fit the frame.
--tile-rows	0 (automatic)	Number of tile rows, same rules as --tile-cols.
--tiles	unset	Minimum total tile count; rav1e picks the rows and columns. Cannot be combined with --tile-rows or --tile-cols.
--slots	0 (off)	Encode up to N GOPs in parallel; values above 1 enable it. Requires a build with the unstable feature, and only the rav1e-ch binary (also needing channel-api) uses it.

Colour and HDR metadata (6 options)

These options only label the stream. They do not convert pixels.

Option	Default	What it does
--range	Limited	Limited or Full. rav1e does not read the range from the Y4M header, so set Full yourself for full-range input.
--primaries	Unspecified	BT709, BT470M, BT470BG, BT601, SMPTE240, GenericFilm, BT2020, XYZ, SMPTE431, SMPTE432, EBU3213 or Unspecified.
--transfer	Unspecified	BT709, BT470M, BT470BG, BT601, SMPTE240, Linear, Log100, Log100Sqrt10, IEC61966, BT1361, SRGB, BT2020_10Bit, BT2020_12Bit, SMPTE2084 (PQ), SMPTE428, HLG or Unspecified.
--matrix	Unspecified	Identity, BT709, FCC, BT470BG, BT601, SMPTE240, YCgCo, BT2020NCL, BT2020CL, SMPTE2085, ChromatNCL, ChromatCL, ICtCp or Unspecified.
--mastering-display	unset	HDR mastering display as G(x,y)B(x,y)R(x,y)WP(x,y)L(max,min): CIE xy chromaticities as decimals 0–1, luminance in cd/m².
--content-light	unset	MaxCLL and MaxFALL in cd/m², as max_cll,max_fall (for example 1000,400). 0,0 writes nothing.

Frame rate and level (3 options)

Option	Default	What it does
--frame-rate	from Y4M header	Override the output frame rate’s numerator; needs --time-scale (--frame-rate 24000 --time-scale 1001 is 23.976 fps). Changes timestamps only.
--time-scale	0 (unset)	Frame-rate denominator for --frame-rate; use 1 for whole-number rates. Ignored without --frame-rate.
--level	auto	AV1 level to signal and enforce: 2.0, 2.1, 3.0, 3.1, 4.0, 4.1, 5.0–5.3 or 6.0–6.3. auto and unconstrained both write level 31 (no level). The encode fails if the frame size exceeds the level.

Statistics and debugging (7 options)

Option	Default	What it does
--verbose, -v	off	Print statistics for every frame. Cannot combine with --quiet.
--quiet, -q	off	Print no progress or summary.
--benchmark	off	Print a timing summary at the end.
--psnr	off	Report PSNR of the reconstruction against the input. Cannot combine with --metrics.
--metrics	off	Report PSNR, PSNR-HVS, SSIM, MS-SSIM and CIEDE2000 against the input. Slows the encode.
--reconstruction, -r	unset	Also write the encoder’s reconstructed frames (what a decoder outputs) to this Y4M file.
--high-bitdepth	off	Use the high-bit-depth code path for 8-bit input too. For debugging.

Advanced subcommand (4 options)

Option	Default	What it does
advanced --completion, advanced -c	unset	Print a shell completion script for bash, elvish, fish, powershell or zsh; zsh panics in 0.8.1. An input and -o are still required (rav1e - -o - advanced -c bash).
advanced --save-config, advanced -s	unset	Write the encoder configuration to a TOML file. Requires a build with the serialize feature, which default builds lack.
advanced --load-config, advanced -l	unset	Load the encoder configuration from a TOML file made by --save-config; encoder options on the command line are then ignored. Requires serialize.
advanced --help, advanced -h	—	Print help for the subcommand.