Works best at 0.95 weight.
Put 'nipples, blurry, censored' in the negative prompt
Description
FAQ
Comments (4)
This is great! Is there any chance you can train it for base sdxl?
Maybe, I'm currently having an issue where the model will generate blurry circles on the tits which is a dataset issue.
@EmilyMars I started trying to make a lora like this before I found yours, and I'm still playing with it; but no matter how diverse I think I've made the images or how well I've captioned the images, my loras still end up really locking down the composition of any images I make with them.
If you don't mind my asking, how did you train yours? What kinds of images did you have in your dataset? How did you make them? How specific were your tags?
@Rholenor I train my pony models using the default on-site training, just with clip skip two. I use anime images if I can and fanart when training character loras so that the lora isn't restricted to one style. I also remove watermarks/text from the images so the computer doesn't learn that, my tagging is danbooru tags, and I tag each part of an outfit or appearance ie "black hair, brown shirt, white jacket" ect.