SwinIR trained on 3000 top-rated AI generated images from derpibooru, with images likely to be upscales already(>2048 in width or height) removed. Images were split into 256x256 patches, 170k patches in total

Should work better than default SwinIR on more anime-2d-looking images

To use,

-download into the SwinIR folder and rename to .pth

-run automatic with --disable-safe-unpickle (I am not sure how to do it any other way and why the default model runs fine without it)