hevc-aq adaptive quantization

[ShortKatz]

I was wondering if somebody already started to experiment with the new --hevc-aq option?
https://bitbucket.org/multicoreware/x265/commits/3cd0b5ed0b91bcb3d5d6cfa1395cb502fc6d01ca

Until now I was using --aq-mode=3 because of the bias to dark scenes. Does --hevc-aq also bias dark scenes? And what would be useful values for --qp-adaptation-range for film or animation?

[AzraelNewtype]

https://forum.doom9.org/showpost.php?p=1861433&postcount=6579

[ShortKatz]

Thanks. Interesting results. In my test hevc-aq created some bad blocking artifacts on moving objects.

[benwaggoner]

It'd be nice to see some screenshots and sample clips. Going just off the metrics may not be helpful, since AQ is a psychovisual optimization, so drops in PSNR and SSIM can happen even when visual quality improves.

[zub35]

On this test sample (3 scenes of 60 frames) in 2pass mode, the final bitrate does not match the specified. Up to +25% from the given.
https://yadi.sk/i/OYxgXdaLo99zxw
In the case of a single scene, the final bitrate is as specified.
This only occurs in hevc-aq mode.

[benwaggoner]

On this test sample (3 scenes of 60 frames) in 2pass mode, the final bitrate does not match the specified. Up to +25% from the given.
https://yadi.sk/i/OYxgXdaLo99zxw
In the case of a single scene, the final bitrate is as specified.
This only occurs in hevc-aq mode.

Yeah, I saw some rate control issues as well. MCW confirmed that that HEVC-AQ is an experimental mode, ala AQ-motion, so this kind of experimentation is appropriate right now. I’d suggest opening an issue in Bitbucket with a repro for the problem you’ve seen. I’ve also seen some VBV violations with complex sources.


Sent from my iPad using Tapatalk