OC catgirl
Description
more detailed after learning more about LoRA generation. Not sure if it requires trigger words, but I tried to set it to minka, and def_outfit.
Comments (4)
These generations are pretty messy.
What training settings did you use, and did the training data look anything like this?
Nope, I don't know much about LoRA training, but the input ones looked better. But I got most of those using a LoRA that looked even worse. I mostly wanted it to force the clothing to be layered correctly.
@grgglegrggle650 Well, I got a settings guide if you ever want to try it. I got some decent results out of it. You may want to set the steps lower if you do not have many images though. Usually the training data should have a variety of styles, poses, backgrounds, and outfits in it.
Also, I do not know about what you want out of the clothing layering, but it should work without a special tag if you know the Danbooru tags to describe each article of clothing.
@Diamond_Letter I have to fight each gen with constant weights and prompt adjustments to get the white clothes and blue clothes in the right areas. I want pink cat ears sticking through a white hood, blue headphones under the hood but in front of both the pink cat ears and the spiral sidelocks, also the hoodie should be a soft furry texture, cropped at mid drift. So many times the crop top and the short cropped hoodie merge, the headphones clip or rest above hood, it forgets the low golden opacity eyeliner, etc. This LoRA still requires me to do prompt work, but it seems to put the colors in the right areas. I'll try to improve it as I learn more. It took me many days to get 15 dif images that were 95% right and I had to do alot of color correction manually. I set the settings to what gemini said to for illustrious, but I'll try the default ones next time.