I was getting tired of waiting for a model like this to exist, which i guess it wasn't going to happen since there's not enough pics to find, so i edited a lot of images to make this.
The illustrious version is better looking and can make headless girls with much lower strength.
It has very strong chibi-like cuteness on generations. You might need to reduce strength to 0.6 and put chibi, solid oval eyes, on the negative prompt.
Also, this model is sometimes able to make SWBs when using the bottomless tag, which is unintended but cool.
Description
Same dataset trained on Illustrious.
Trained on 10 epochs, 543 steps and 7 repeats.
FAQ
Comments (8)
Does anyone know how to make this model not have a mouth? In other words, just some eyes on the crotch?
using no_mouth seems to work, but it removes all mouths. The only way i know to only take out the crotch mouth is manually drawing over the mouth and inpainting.
@fominhavi Thank you for the response! I'm trying to make a LoRa that basically puts eyes on the crotch of another character, but I'm failing miserably.
@Ranma3333 Do you have image editing skills? The way i made this lora was by copying and pasting the faces into the crotch area on a lot of images. You could try that, just know that it would only work with one character.
@fominhavi Gotcha. You've got more skill than I do in making LoRa's, did you want to take a look at my model?
@Ranma3333 Okay, i'll see what i can do with it. It might take some time, since i may run out of buzz.
@Ranma3333 I think i found the problem. You're only using the trigger word when training. That means that everything in the image gets linked to the trigger, which is really bad for a concept lora. A concept lora needs many other tags, so when you prompt those, the model knows they are to be used with the trigger.
What you should do is use the auto label button on the dataset screen. My configuration is 20 max tags, 0,3 Min Threshold.
Also, you could put better prompts on the sample image prompts space. Putting something you would prompt like "1girl, nude, short hair" would work.
So basically: You will have to train another lora with proper dataset tagging.
@fominhavi Ah, okay. That makes sense. I'll give it a try!