[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]
2023-11: Warosu is now out of extended maintenance.

/vr/ - Retro Games

Search:


View post   

>> No.8777248 [View]
File: 545 KB, 2880x2114, mpv-shot (2022-04-03 23_09_41) 01.jpg [View same] [iqdb] [saucenao] [google]
8777248

>>8777189
I've recently been doing a lot of hobby AI upscaling too, so I was interested in doing a test upscale with some footage I successfully de-interlaced, but little else.
https://streamable.com/8154s5
https://files.catbox.moe/zeyr8o.mkv

The first problem was bad input, as mentioned before the source has other artefacts that need correcting like ringing or dotcrawl (the de-interlacing corrects that to an extent, but doesn't eliminate it).
This means that later AI upscaling will upscale the artefacts instead of removing them, unless the AI model is designed to do that.
The de-interlacing process is also not correct, the animation as far as I can tell is 24fps that's been telecined to 29.97fps. My de-interlaced output has doubled the frame rate to 59.94fps
For the sake of processing (video encoding and later AI processing), this means I'm more than doubling the number of stored frames.

The second problem is the AI model I used wasn't appropriate.
I used the default RealESRGAN model that is sort of suitable for real-world material, but is prone to producing blotches of light and dark vertical stripes on flat color painted cels.

tldr; AI upscaling takes work, fix your sources and use the appropriate model.
A lot of upscaled stuff you find on YouTube will do poor jobs like these, because the initial impression is still pretty good before scrutiny.

Navigation
View posts[+24][+48][+96]