This was only a test i was doing for myself to test making a lora on my 12GB gpu. And i think it turned out way better then i would have ever thought it would.
Its not perfect by any means, but i was having so much fun with it, that i decided to upload it.

It removes some of the grain that makes Z so good. And if i make a V2, i'll make sure to include images that have more grain in the dataset.

There is one main keyword(wedgie), but all source images was prompted, so natural language works best, and some times to good. Be careful using "close up", "?????", "panties", "pussy-wedgie" etc if you dont want the image to change quite a bit.

I recommend ~0.5 to make ???????? and 0.75-1.0 for wedgie.

All sample images are raw, without any detailers or "fixes".
Handles 20 step+ without problem.
Personaly i like DDIM + DDIM_uniform. But euler + simple etc works fine to. (Comfy ui)