LoRA for Kizumonagatari style, from Monogatari Series by Nisio Isin. Trigger word: KizuStyle. But LoRa works without it. Although the results may vary. This is my first more or less normal attempt to train this LoRa for style. Choose different weights for different models and LoRA's.
Description
A more stable version. I mainly used a weight of 0.7-0.8. But it's better to choose weights for different loras and models. I've tried to train this style so many times. So, i'm not really sure which versions are better. XD I don't think I'll go back to this style again..
FAQ
Comments (1)
If you need more images here is the full collection
A tip would be to remove any character tags that it might suggest since you don't want their names burnt into the dataset and turn every "1girl, brown hair" into hanekawa as example. The series has too many short characters so I'd also suggest reducing their images so that their features don't get burnt in, and instead increase backgrounds so it picks up the shading, compositions and the color usage while the characters retain more flexibility. Training steps for illustrious styles around 2500-5000 for an authentic look.