[anima-p2] (experimental)
NSFW focused, female focused. Pretty pleased with this one. Still lots of artifacts with hands that require rerolls, but that may just be the preview model. I trained this with a bunch of photos too, so you can try realistic if you want, but I was not impressed with the results, so I recommend sticking with illustration.
[Z-Image] (experimental)
Warning: This LORA is NSFW focused. The dataset has alot of nudity and sex. Unfortunately, it's not really enough data to move the needle to make Z-Image good at that. Hopefully we will get a good checkpoint soon.
[QWEN]
Warning: This LORA is NSFW focused. You can use it for other stuff, but your characters may get nipples.
Goal was to get a little bit of the pony/chroma styling I like into qwen. It's not fantastic, but good enough for the moment.
Recommendation: I highly recommend writing descriptive prose prompt first, then adding a few tags at the end.
Settings: Model weight seems best 0.6-0.9
Sampler: I really liked the res_multistep/simple combo with the 8-step lightning lora
Most of the sample images I just looked for highly rated qwen images that week and put them through. I didn't think to track the images I used, so if you notice one is yours, feel free to comment.
Known Limitations:
-Dataset highly human female focused
-I apparently did poorly at body type variety, so too skinny and lots of ribcages. Maybe that's your thing, but I'm not thrilled with it.
-Nipples show up even when clothed (use lora model weight 0.6-0.7 helps some)
-Genitals understanding and detail still pretty poor, couldn't correct that very well.
-Character interactions are poor
-Faces seemed to converge a little too heavily
-Doesn't seem to play nice with the other NSFW loras, at least in my tests.
-I guess maybe too many untagged body freckles in the dataset. Because, damn, this lora likes body freckles. Feature not a bug?
Description
This can be used with Z-Image Base or with Z-Image Turbo. Most of the images I posted were with ZIT. ZIB does much better with the illustrations because ZIT is finetuned for realism. Still lots of artifacts, unfortunately. The images I posted are very cherry-picked.