/vr/ - Retro Games

Currently training on the vanilla DIV2K dataset. I want to see how my results differ from the stock models, which could give me a clue as to where I'm going wrong with my own models.

I'd like to give FF7 backgrounds a try, but there would need to be a way of automatically compositing and decompositing backgrounds, as FF7 backgrounds are divided into transparent layers and you can't just take out the transparency without ruining the upscaling.