2023-04-24, 20:49
https://blogs.nvidia.com/blog/2023/02/28...esolution/
This was initially announced only for Chrome and Edge webbrowsers but someone built a fork of MADVR that could take advantage of it and I did some testing using MPC-HC and this build:
https://github.com/emoose/VideoRenderer/releases
It also seems that since then VLC made it available in Beta.
I thought this might be a terrible thing, one of those 'It tries and it ties too hard mushing everything up' but I've been frankly impressed with it's restraint? But maybe that's because it seems to be trained for a very specific task: Modern HD video scaled down to lower bitrates and resolution for streaming, that needs to be scaled back up.
Feeding clean but highly compressed video seems to benefit the most which is kinda a 'duh'. It does some nice work on compression artifacts, that chunkiness you see in reds, and such. I tried very clean, DVD Remuxed mid2000's animation into it and it mostly just did a better scale to 4K than you'd see with bilinear or other typical scaling methods without going all 'AI Destructive'. Contrasting that, with similar animation that was low bitrate 480p stream copies and it did a nice job on the compression artifacts. I also tried SD 90's television (Sourced from DVD without re-encoding of course), with grain, bounciness, composite dotcrawl and the like and frankly it seemed to do nothing at all other than minor scaling improvements and improvements on compression artifacts. It seemingly looks for some specific traits and scales better but doesn't suffer from 'Overkill' as I'd once expected. When you go outside of what the model was likely trained for, it seems to hold back and not do much but it does some neat things on bitrate staved encodes.
This would be pretty neat to see as an internally supported scaling option in the Windows build of Kodi.
This was initially announced only for Chrome and Edge webbrowsers but someone built a fork of MADVR that could take advantage of it and I did some testing using MPC-HC and this build:
https://github.com/emoose/VideoRenderer/releases
It also seems that since then VLC made it available in Beta.
I thought this might be a terrible thing, one of those 'It tries and it ties too hard mushing everything up' but I've been frankly impressed with it's restraint? But maybe that's because it seems to be trained for a very specific task: Modern HD video scaled down to lower bitrates and resolution for streaming, that needs to be scaled back up.
Feeding clean but highly compressed video seems to benefit the most which is kinda a 'duh'. It does some nice work on compression artifacts, that chunkiness you see in reds, and such. I tried very clean, DVD Remuxed mid2000's animation into it and it mostly just did a better scale to 4K than you'd see with bilinear or other typical scaling methods without going all 'AI Destructive'. Contrasting that, with similar animation that was low bitrate 480p stream copies and it did a nice job on the compression artifacts. I also tried SD 90's television (Sourced from DVD without re-encoding of course), with grain, bounciness, composite dotcrawl and the like and frankly it seemed to do nothing at all other than minor scaling improvements and improvements on compression artifacts. It seemingly looks for some specific traits and scales better but doesn't suffer from 'Overkill' as I'd once expected. When you go outside of what the model was likely trained for, it seems to hold back and not do much but it does some neat things on bitrate staved encodes.
This would be pretty neat to see as an internally supported scaling option in the Windows build of Kodi.