I wanted to replicate the kind of butts I was able to generate with SDXL checkpoints like Lustify, so I trained a LoRA using a non-AI dataset of mainly face down ass up poses viewed from behind.
v1.0 is experimental; I trained it on 75 manually-tagged images for 5000 steps, but it's a small subset of the 300+ images I intended on using in order to test robustness. The captions were a mix of danbooru-style tags and natural language.
I think it works better for controlnet and img2img, but txt2img can output decent booty depending on how you prompt.
Description
This version is a continuation of v1.0 trained to 10,000 steps with 65 additional images consisting mainly of skirts over the ass and "all fours" pose.
For the additional images I added focus to concepts like "coinslot pussy" for innie pussies and "meaty pussy" that in my use case I prefer putting in the negative prompt to produce nicer pussies. You can also include things like "upskirt", "frilled skirt", and "skirt covering half her ass" along with "anus peek" (which also works well with "asshole behind thong"-type content) if you like skirts.
FAQ
Comments (3)
In v1.1, the butthole/pussy part is SO much blurrier than the rest of the image. Was it trained on very low resolution images? Or maybe it was trained on censored or decensored JAV?
I may have recklessly used the liquify tool in Krita to make the butthole bigger in about 1/4 of the images as a shortcut. Good eye, I hadn't noticed that.
@funkytowns Yeah I noticed that right off the bat myself too. Looks like a complete downgrade from your original 1.0!