About V2 - I was able to push it a bit feather, without actually loosing fidelity, it still has some problems (fundamental ones which I try to pinpoint how to fix) but its much less overfit, following prompts much better and actually cleaner then the first one. You could probably used overfit one for inpaint, but I doubt its nessesery.

INFO TD:LR

• Model still has a bit of problems

• Using adetailer for face is kinda nessesery, upscaling is not nessesery but advised. I used basic upscale without any fancy stuff.

• I used CFG scale of 5, you probably can use 3-5

• You might need to prompt for body type, since its not stable

• Don't be afraid to put original uniform tags into negative is tag is bleeding

• Outputs might depend on your checkpoint and prompt you use

But at least face and face details are consistent enouth, and arent collapsing to instantly when you prompt for something new, so its still miles better then older versions.

I might comeback to it, when I figure out how to improve training.

USAGE

For preview images, I've used ADetailer on face and hi-rez fix at 0.25 using Remacri.

Besides that really nothing else. You might need to specify poses in your prompt if you want something specific.