Style and character lora for Yozakura Quartet.
@yozq is trigger qord for the style. "anime screenshot" also does things. But they are not always necessary.
The model knows most of the characters. It can even do multiple characters in single image.
Training data (images, captions) are in optional downloads if anyone is curious. Note that it uses keep_tokens_separator = ";;" .
Illustrious version https://civarchive.com/models/1021790/anime-screencap-style-yozakura-quartet
Pony version https://civarchive.com/models/902479/anime-screencap-style-yozakura-quartet
Description
Trained on Anima v1.0
Some training data improvements.
Trained for 3300 steps, resumed for additional 7200 steps. 2 images per step + gradient accumulation 4 (effectively 8 images per step).