AV1 encoders: AOMenc vs SVT-AV1 vs Rav1e

[foxyshadis]

I've run a few test encodes, but I'm really not sure where the sweet spot of the cream of the crop of encoders lies right now, except that libaom generates the absolute highest quality if encoding time is no barrier (like, 10-100x as long).

This month, SVT is up to 0.75 0.8 and Rav1e is up to 0.2, both claiming major improvements, but no good new quality/size/time comparisons out there so far. Any takers?

[Tommy Carrot]

AWCY has many comparisons, it's just hard to find the relevant ones. If you're interested in the absolute highest quality possible: rav1e vs SVT-AVC vs aomenc (https://beta.arewecompressedyet.com/?job=speed0-master&job=%402019-11-27T00%3A59%3A04.765Z-SVT-AV1-MASTER-NOVEMBER-26-2019&job=Dec18_aomenc_cpu0_consumer_passes2%402019-12-18T07%3A03%3A54.884Z). The SVT-AV1 data is fairly old, and it has improved somewhat since, but there is no newer data there.

For something closer to the normal usage, here (https://beta.arewecompressedyet.com/?job=master-768d7b27ed3e2e5cf2b5b64d9fe5eb116a9c2d02&job=SVT-AV1-enc-mode-3%402019-12-18&job=20_Dec_ref4_cpu4_60frames%402019-12-20T09%3A49%3A08.354Z) is a comparison with faster presets.

Metrics dont tell the whole picture, so here is my experience about the encoders. Rav1e is simply not in the same league as the others, most of the tools are still not implemented, so it cannot reach anywhere near the same quality/efficiency as the others, in my opinion it's still even behind VP9.

SVT-AV1 has improved a lot, with the slowest preset it has actually came quite close to aomenc. But it has some temporal artifacts, so aomenc is still producing better visual quality. It can also be pretty fast with the faster presets, but in my opinion it's not worth going above enc-mode 3, because the quality is getting pretty bad with nasty artifacts.

[Mr_Khyron]

SVT-AV1 0.8 is out
https://github.com/OpenVisualCloud/SVT-AV1/releases/tag/v0.8.0

[foxyshadis]

Metrics dont tell the whole picture, so here is my experience about the encoders. Rav1e is simply not in the same league as the others, most of the tools are still not implemented, so it cannot reach anywhere near the same quality/efficiency as the others, in my opinion it's still even behind VP9.

SVT-AV1 has improved a lot, with the slowest preset it has actually came quite close to aomenc. But it has some temporal artifacts, so aomenc is still producing better visual quality. It can also be pretty fast with the faster presets, but in my opinion it's not worth going above enc-mode 3, because the quality is getting pretty bad with nasty artifacts.

Yeah, I don't necessarily want just a bunch of numbers, but as you noted, most existing comparisons are terribly outdated. Thanks for the link, I'll download a few of those and see how they hold up. The codecs are evolving quite quickly, hopefully fast-but-terrible will turn into fast-but-ok soon.

SVT-AV1 0.8 is out
https://github.com/OpenVisualCloud/SVT-AV1/releases/tag/v0.8.0

Figures, JUST after I set up the toolchain last night to start testing! Nice, though.

[SmilingWolf]

I'm running some tests of my own, SVT-AV1 and libaom should be done by today's evening, rav1e by tomorrow.
Some preliminary numbers seem to highlight around a 6% improvement for SVT-AV1, around the same for rav1e, 0.2% percent improvement for libvpx and 0.1% loss for libaom, this one correlated with big speed gain, with a single clip taking from 38 minutes down to 26 to encode.

Full report, for the first time with start-end (not per-encode) timings, coming tomorrow.

All before-after numbers are related to my previous report from 5 months ago: https://www.reddit.com/r/AV1/comments/cfao4x/codecs_performance_report_5th_and_a_half_edition/