Open hohMiyazawa opened 3 years ago
10bit encoding from 8bit sources gives some coding gain, but nothing impressive.
Yeah. I remeber it being the way anime groups squezing debanding into 8 bit source without overbloating. ALbeit ithere is some paper that claims that 10 bit (at least on h264 and vp9) just improves efficiency without further manipulations
Resizing an 8bit source can also increase the effective bit depth slightly.
But taking advantage of this likely requires rewriting the video scaling portion of the script. By default, ffmpeg uses the same pixel format for the input and output of the scaling.
The default behaviour is right now to do "10bit when 10bit input". That's at least better than what it does for libvpx
That's at least better than what it does for libvpx
Dunno about that. VP9 10bit non-HDR playback in browsers is really untested grounds, unlike AV1 10bit
What I mean is that mpv-webm never made 10bit encodes before, since libvpx automatically used yuv420p (even if the input was yuv420p10le) with the settings in the script. libaom on the other hand automatically preserves the pixel format.
So before, you never got 10bit encodes, and now with this AV1 patch you get 10bit encodes when they matter the most.
But I agree that I have no idea how 10bit non-HDR behaves. Probably not worth implementing.
mpv-webm never made 10bit encodes before
In case of using x264 through mpv-webm it does encode 10 bit source into 10 bit clip (and it is really annoying behavior when web sharing is primary use). But for libvpx hacknyed implementation of 10 bit on HDR and 10bit AV1 playback is a standard for hardware playback it does not matter
As you can see, I have never made h264 encodes for the web :)
Is that something that should be changed too? 10bit h264 sounds actively bad for a quick and convenient clip script. The use cases for 10bit h264 aren't overlapping with this at all.
Is that something that should be changed too?
That would be really neat since copy-pasta "--ovcopts-add=format=yuv420p" do not work
Just to document this. The sane options would be:
10bit encoding from 8bit sources gives some coding gain, but nothing impressive.