VMAF calculation issue with recent FFmpeg git

[veikk0]

I'm having an issue with extremely low (~1) VMAF scores when using an FFmpeg built from recent git code. The score is always this low with any CRF, and ab-av1 keeps going lower and lower until the size of max-encoded-percent is reached and it exits.

I've tracked the issue down to the following FFmpeg commit: FFmpeg/FFmpeg@c94e9d0

Calculating VMAF normally using FFmpeg still works fine.

[veikk0]

Looking through the ab-av1 repository, I can't really find a reason why VMAF calculation is done with pipes instead of a single FFmpeg command.

If the reason is pixel format conversion, then the format filter is the way to go. I believe -pix_fmt is an alias for the format filter anyway and more of a legacy option, but it's more convenient when typing a command for a simple encode.

[alexheretic]

Yep pixel format & vfilter. It may be worth trying to crunch it into a single call. I'd need to regression test such a change.

[alexheretic]

#177 changes vmaf to a single ffmpeg call. In my testing this produces identical results to previous code though much simpler and easier to reproduce in a manual ffmpeg call.

[veikk0]

Nice! I'll do some before-after testing too and report back.

[veikk0]

#WorksForMe