>>2932004If you still have suggestions I'll gladly listen, but in the meanwhile I think I got somewhere with the model. Maybe.
It will fuck up horrendously on about 10% of the outputs, but the majority are much better than before, about comparable with EPS.
I'm still testing what samplers work for image gen and IMG2IMG/inpainting. SMEA works much better than before, but it still loses detail notcieably.
So I went back to Euler A, and switched to the basic Align your Step instead of the GITS one.
Picrel is an IMG2IMG at 0.6 denoise. 0.6 is fairly high, intentional to see which ones fuck up and which ones do well.
Main issue is all my tests in IMG2IMG so far noticeably brighten up the image.
