Intel To Ring In 2022 With New, Faster AV1 Encoder Release

[FPScholten]

Not sure if this has changed but aomenc was unusably slow when I tried using it. SVT-AV1 is solid, rav1e looks promising although last time I tried it it lost tons of quality on repaint frames.

aomenc's default settings are for best quality, this is like the placebo setting in X265 or preset 1 in SVT_AV1 and tiling is not set so it wil use only 1 tile (the entire frame) with only 1 thread because threading is not enabled by default. I am using libaom-av1 with ffmpeg (both of which I compile myself) with these options:

-cpu-used 4 -row-mt true -threads 8 -crf 35 -tiles 4x2 -aq-mode 3

-cpu-used 4 : similar to presets in SVT_AV1 higher numbers are faster but increase filesize. 8 is fastest, anything between 3 and 6 will do depending on your systems computing power

-row-mt true : this enables tiling and multithreading

-threads 8 : specifies the number of threads to use in tiling mode I have 4 cores with hyperthreading = 8 threads

-crf 35 : single pass encoding ratefactor, 0 for lossless, higher numbers lower quality. (two pass and three pass are not really worth the hassle)

-tiles 4x2 : sets tiling to 4 rows and 2 columns, significant speedup in encoding for a very small quality loss. Do not use more tiles than available threads

aq-mode 3 : best quality setting for regular video, for other stuff like cartoons or images try 1 or 2 as well.

Experiment with these settings to find optimal quality/filesize/encodingtime for your system

[iskra32]

The new 0.8.8 version is good, it fixed a nasty bug, causing it to leak memory during long encodes.
Although it can encode pretty fast, it does not do as good of a job quality wise as the AOM encoder does. I tested a range of settings and compared it with AOM. I found out that encoding a video to a certain size with SvtAV1 always has a lower perceptual quality than with AOM. (both visual as in VMAF score). If you want the quality to be as good as AOM, you end up with a larger file and an encoding time that is about the same as AOM.
However if you do not care about quality/size then SvtAV1 can be good for you with relatively fast encoding leading to relatively big files and/or lower quality.

Not sure if this has changed but aomenc was unusably slow when I tried using it. SVT-AV1 is solid, rav1e looks promising although last time I tried it it lost tons of quality on repaint frames.

[piotrj3]

VMAF is not perfect, but its pretty good at taking PV into account, as it was literally trained on people watching video.

generally VMAF is quite fine to compare diffrent codecs, but no one uses VMAF in process of developing codec. Basicly psychovisual settings are quite limited, and most of changes you make are not on psychovisual plane, but on simple efficiency. So when codec presets are tuned (between speed and quality) those presets are made based on SSIM or PSNR values, this is why codecs by default calcuate them.

For comparison of presets, PSNR or SSIM is perfect. For comparison of diffrent codecs you might take into account VMAF, but SSIM/PSNR should still give you pretty good idea.

[FPScholten]

The new 0.8.8 version is good, it fixed a nasty bug, causing it to leak memory during long encodes.
Although it can encode pretty fast, it does not do as good of a job quality wise as the AOM encoder does. I tested a range of settings and compared it with AOM. I found out that encoding a video to a certain size with SvtAV1 always has a lower perceptual quality than with AOM. (both visual as in VMAF score). If you want the quality to be as good as AOM, you end up with a larger file and an encoding time that is about the same as AOM.
However if you do not care about quality/size then SvtAV1 can be good for you with relatively fast encoding leading to relatively big files and/or lower quality.

[brucethemoose]

Yes, that's what I mean, you can use the built in stats, but they don't really mean anything as far as human perceived quality goes. There's not much value to them given that psycho-visual tuning is what makes a good high quality encoder really good. And the OP was asking about the low setting which are the high quality ones where one would expect PV optimizations to be most important.

VMAF is not perfect, but its pretty good at taking PV into account, as it was literally trained on people watching video.

[Quackdoc]

I never cared about the stats myself, ill just encode a short video that has relevant footage, if it is unacceptable, re-encode it. rinse and repeat until I find an acceptable level, and use that as a preset.

[willmore]

that is not hard to do. Most codecs itself support calculating PSNR/SSIM values. SVT-AV1 supports it, X265 supports it, x264 supports it, vpxenc supports it.

In SVT-AV1 you need to add simple --enable-stat-report 1 and it will calculate PSNR/SSIM. Keep in mind those are values that become worse with good psycho-visual tuning, like good deblocking filters make content look better to user, but makes those values worse.

Yes, that's what I mean, you can use the built in stats, but they don't really mean anything as far as human perceived quality goes. There's not much value to them given that psycho-visual tuning is what makes a good high quality encoder really good. And the OP was asking about the low setting which are the high quality ones where one would expect PV optimizations to be most important.

[piotrj3]

Michael may have chosen not to look into that as length (ratio) is most meaningful when the video quality is determined as well--and that's a very hard thing to do programatically.

that is not hard to do. Most codecs itself support calculating PSNR/SSIM values. SVT-AV1 supports it, X265 supports it, x264 supports it, vpxenc supports it.

In SVT-AV1 you need to add simple --enable-stat-report 1 and it will calculate PSNR/SSIM. Keep in mind those are values that become worse with good psycho-visual tuning, like good deblocking filters make content look better to user, but makes those values worse.

[willmore]

New preset levels and changes in performance of existing levels could possibly be an indication in changes of compression ratios. Were there any or are the outputs still identical for each of the presets?

Michael may have chosen not to look into that as length (ratio) is most meaningful when the video quality is determined as well--and that's a very hard thing to do programatically.

[Quackdoc]

now if only discord supported avif and av1 so we can post things that actually look tolerable instead of paying for nitro. but then again, that would eat into their sales a little too much