Even tho Pony's models are good with nudes, they still struggle making vulvas. What they do feels too weird/unrealistic/unnatural, so I wanted to finetune them. Using LoRAs didn't work good because either the model lose quality inferencing other stuff, or changed the overall style of the model (because most of the training data are real photos), or didn't catch how to make vulvas good enough.

So I finetuned an entire unet for it. Using a mix of a huge image collection of close up vulvas with manual tags on the style of Pony's models, inpainted images and original images from the model so it doesnt lose its quality or style. I picked Solus25D's model because I liked its style more than others, but I plan to use this training method with others. Also with Flux and Hunyuan video in the future.