Mastodawn

The following works (for a json file stored in D:\mypath\):

ffmpeg -i H:\nu\main.mkv -i H:\nu\ref.mkv -lavfi libvmaf=model='path=D\\\:/mypath/vmaf_v0.6.1.json' -f null -

https://www.reddit.com/r/ffmpeg/comments/11fnnou/new_syntax_libvmaf_model_the_model_path_option_is/ #ffmpeg #vmaf

Okay, so we''ve avoided the Custom Keyboard rabbithole, unfortunately only because I'm already neck deep into a far more annoying one;

Compression video quality, and video quality testing.

So I have a couple video files, and I'd like to get them down to far more sane sizes while not losing lots of quality. Sounds reasonable?

However; If you've ever used #ffmpeg, #handbrake, #vmaf, etc. you know that there are only ~1,827,199 different CLI options that can be paired together with different results.

Thankfully; Tools like #Netflix's #VMAF help, however (on my system) they only process at ~110fps. That sounds like a lot until you realize a 45minute show might have 65,000+ frames. That's a slog for anything lengthy.

But hey! I found out that both ffmpeg and vmaf have #CUDA support and can hit 1400fps on 4k video; Fuck yeah! LET'S GO!

oh, but wait. They're non-free libraries from #Nvidia, so distributed builds of vmaf/ffmpeg can't have them enabled. The end user has to compile them.

To boot; Some of them won't compile if they can't find a compatible en/decoder. So they have to touch the gpu. Which isn't available in virtualbox.

SO much headache, but the delta between a 10 minute slog at 110fps to sub-1 minute, for seasons of video means it's worth it, I GUESSSssss

Gianni Rosato Apr 9, 2023

To everyone still using VMAF: please don't, use SSIMU2 instead.

https://nitter.kitsuna.net/jonsneyers/status/1573371624132419585

#vmaf #av1 #video

Jon Sneyers (@jonsneyers)

I don't want to diss Netflix, but VMAF is a pretty bad metric, at least for still images. It thinks the image on the left looks quite a bit better than the image on the right.

Nitter