AV1 with SVT-AV1

SVT-AV1 is the AV1 encoder hosted by AOMedia. It is fast, scales well across CPU cores, and gives excellent compression at its slower presets. The current release is v4.2.0. The easiest way to use it is through FFmpeg’s libsvtav1 encoder.

ffmpeg -i input.mkv -c:v libsvtav1 -preset 4 -crf 30 -pix_fmt yuv420p10le \
  -svtav1-params tune=0:enable-variance-boost=1:keyint=10s -c:a copy output.mkv

That is a sound starting point for a high-quality 1080p encode. The audio is copied unchanged; the audio and subtitles guide covers re-encoding it. Settings that are not FFmpeg options go in -svtav1-params as colon-separated key=value pairs, using the same names as the standalone encoder without the leading dashes. The sections below explain each choice and the options worth adding.

Preset

Presets run from 0 (slowest, most efficient) to 13 (fastest); negative values are for debugging. The default is 8.

Presets	Typical use
2–3	Maximum efficiency when encode time does not matter much. Much slower than 4.
4–6	The usual range for quality encodes at home: a good balance of efficiency and time.
7–9	Faster encodes with a noticeable efficiency loss.
10–13	Real-time and automation use. The encoder warns that visual artifacts may occur at these presets.

Outside real-time mode (rtc=1), presets 12 and 13 run as 11.

Use the lowest preset you can afford. In CRF mode SVT-AV1’s multithreading does not change the output, so CPU count affects speed but not quality.

CRF

CRF ranges from 1 to 70; the default is 35. Lower is better quality and larger. Upstream suggests starting around 30 for 1080p and uses 25 in its high-efficiency VOD example. 2D animation usually holds up well at higher values than live action. FFmpeg’s own -crf option takes integers up to 63 (-crf 27.5 becomes 28); set -svtav1-params crf=… for quarter steps or higher values.

SVT-AV1 CRF values are not comparable with x264 or x265 values. The wider range means a given quality sits at a higher number than in those encoders.

Tune

The default is tune=1, which optimizes for PSNR. tune=0 (visual quality) is designed to look better to people, usually with more retained sharpness, at the cost of PSNR scores. Use tune=0 for anything meant to be watched. The other tunes (SSIM, MS-SSIM, VMAF, and IQ for still images) target metrics or images, not video viewing.

10-bit output

Use -pix_fmt yuv420p10le even for 8-bit sources. Encoding at 10 bits uses less rounding inside the encoder and reduces banding in gradients and dark areas, for little extra size and little extra encoding time. AV1’s Main profile includes 10-bit 4:2:0, so every AV1 decoder, hardware included, can play it. Without -pix_fmt, SVT-AV1 keeps the source bit depth.

Film grain

Grain and sensor noise are random, so they take a lot of bits to encode faithfully. AV1 can describe grain with a small set of parameters and let the decoder add synthetic grain on top of a cleaner picture. film-grain=N turns this on; N is the strength (0–50).

Source	Starting film-grain
Live action with normal grain	8
Heavy grain or noise	10–15
Hand-drawn 2D animation	4–6
Clean or CGI sources with no grain	Leave it off

film-grain-denoise controls what happens to the source grain:

• film-grain-denoise=0 (the default since v2.0.0) keeps the source as it is and adds synthetic grain on top. This preserves the most detail, but the encoder still spends bits on the real grain, and too high a level stacks extra noise on the picture.
• film-grain-denoise=1 denoises the source before encoding and lets the synthetic grain replace it. This saves the most bits, but the denoiser can also remove fine detail such as small particles or skin texture.

ffmpeg -i input.mkv -c:v libsvtav1 -preset 4 -crf 30 -pix_fmt yuv420p10le \
  -svtav1-params tune=0:enable-variance-boost=1:keyint=10s:film-grain=8 -c:a copy output.mkv

Check the result at normal speed, since grain is a moving texture. Grain synthesis is meant for presets 6 and below; the encoder warns about the compute cost above that. Synthetic grain lowers full-reference metric scores even when it looks right, because it does not match the source grain pixel for pixel.

Keyframe interval

By default SVT-AV1 inserts a keyframe about every 5 seconds (160 frames at 24 fps). Keyframes are large, so a longer interval improves efficiency at the cost of slower seeking. For local playback, 5–10 seconds is common, which is why the starting command sets keyint=10s. FFmpeg’s -g sets the same thing in frames (-g 240 at 24 fps) but has no seconds form. Video-on-demand streaming typically uses keyframes about once per second.

SVT-AV1 never places keyframes at scene changes; scd=1 is ignored with a warning. For scene-aligned keyframes, encode in scene chunks with Av1an.

Other useful options

Option	Default	What it does
enable-variance-boost=1	0	Gives more bits to low-contrast areas (skin, clouds, dark and foggy scenes) that otherwise blur or band.
variance-boost-strength	2	1–4. Upstream recommends 2 for most content, 1 for flat animation and 3 for films mixing very dark and bright scenes.
ac-bias	0.0	Energy-preserving psychovisual bias, similar to x264/x265 psy-rd. Upstream suggests 1.0–1.5 for texture retention.
fast-decode=1 or 2	0	Tunes the encode so it decodes faster on weak devices; 2 is the stronger level.

Using SvtAv1EncApp directly

The standalone encoder accepts every option, including some not exposed through FFmpeg. Feed it Y4M from FFmpeg:

ffmpeg -i input.mkv -pix_fmt yuv420p10le -strict -1 -f yuv4mpegpipe - | \
  SvtAv1EncApp -i - --preset 4 --crf 30 --tune 0 --enable-variance-boost 1 --keyint 10s -b video.ivf

FFmpeg needs -strict -1 to write 10-bit Y4M. The output is a bare IVF video stream. Mux it with the audio, subtitles and chapters from the source:

mkvmerge -o output.mkv video.ivf --no-video input.mkv

Y4M does not carry color metadata. For HDR sources, pass it on the command line, for example --color-primaries bt2020 --transfer-characteristics smpte2084 --matrix-coefficients bt2020-ncl plus --mastering-display and --content-light with the source’s values. FFmpeg’s libsvtav1 passes the color tags and static HDR metadata through automatically. See preserve color and HDR, and the SVT-AV1-HDR fork for HDR10+ and Dolby Vision work.

Chunked encoding with Av1an

Even with good threading, one SVT-AV1 process does not keep a large CPU fully busy at slow presets. Av1an splits the video at scene changes, encodes the chunks in parallel, and can pick a CRF per scene to hit a quality target. It is the usual tool for long, high-quality AV1 encodes.

The SVT-AV1 encoder page covers the full option set.