>>43567271Lora itself is bad because it's just a quick test, and images were not matching perfectly, but idea behind it is interesting.
It's trained on pairs of game screenshots taken at the same locations: one with lowest graphics settings, another one with highest settings and raytracing. First used as regularization image, second - as training data.
It will be interesting to try this approach for bad hands - good hands pairs too, using blender/ue and some scripting to generate different hand positions.
image:
https://litter.catbox.moe/98qif3.pnglora:
https://litter.catbox.moe/4mm4e7.safetensors 