A SDXL-based LoRA focused on extreme close-up realism of female nipples, using open-license imagery (Wikimedia Commons). Perfect for macro detail, skin texture, and anatomically accurate close-ups. Best used with SDXL photorealistic prompts. Works well at weight 0.6 to 0.9.
Description
FAQ
Comments (8)
unfortunately so many lora's like this remove Asian women and replace them with western women. Even with strong emphasis on Chinese woman and features in a image prompt. it only looks slightly Asian like mixed race. But i turn lora off with same seed and prompt and suddenly i get really authentic looking Asian woman. Lora isn't bad, in my tests the breasts and nipples look amazing. I just wish these lora's weren't always focused only on western women. I know you trained data only on close up of breasts. a couple of other lora's who OP claimed to not include faces had this same problem. I wonder if when they finally release the non-distilled model if lora's will work better without effecting faces as much.
using .6 strength seems to help, it still will create a western looking woman sometimes but also sometimes an actual Asian woman, which without lora will 100% of time make an Asian woman with prompting for it, so not perfect, but usable at least with .6 strength.
Another solution is do not use this lora for the whole image, use it for a Nipple Detailer. Which is basically a Face Detailer setup but change the BBOX detector from a face detector to a nipple detector.
Instead of using this lora when generating an image, use it in a Nipple Detailer. Which a is a face detailer but swap the face detector for a nipple detector, the model for base Z-Image with this lora, and the positive conditioning for a new CLIP-Encoded prompt like "Feminine nipples, extreme closeup of large erect nipples"
That way the overall image is not affected (quality/race/etc) but you can still enjoy improved nipples.
Sorry, I’m new to this. Is there a place where I can read more about how to do this? Or a workflow I can follow?
@TheMadz Look up "face detailer" or "adetailer" which has a lot more information/sample workflows available than nipple detailing. I use ComfyUI so it's easy to just change the settings on the detailer nodes (mainly the loader for the "use this to find which bits of the image to edit" model)
@Nabonidus Thanks for the explanation!
Just to give some feedback: I tried to replicate this idea in ComfyUI (using ZIT, BBOX detector + inpaint instead of face detector), but in practice the base image always changes at least a bit (face, hair or overall look), even with low denoise, tight masks.
I spent around 7 hours trying different workflows (BBOX → mask → inpaint, settings, separate samplers, etc.), but I couldn’t get a true “detail-only” result where the original image stays 100% intact. With ZIT especially, any second diffusion pass seems to introduce some drift.
So I understand the concept and the analogy with face/adetailer, but at least in my tests it didn’t work reliably in ComfyUI for nipples without affecting the original image. Maybe i suck at this hahaha
Still, thanks a lot for the tips and direction... it helped clarify how detailers are supposed to work.
@TheMadz One of the challenges you have is ZIT is not designed for in painting, so it doesn't do a great job at matching the changes to the existing image so you need to be careful with your in-painting settings. I find it works well enough if you increase the mask blur a bit (to soften the edges of the changed area ) and reduce de-noise a bit. There are probably some tricks you can do with color matching as well but I've never looked into that.
I think the full release of Z-Image will do inpainting a lot better, but for now we only have Turbo.