Style and character lora for creating anime screencap-like images.
Trained on manually cropped screenshots from Yozakura Quartet.
Training details:
Same dataset and almost same training settings to the pony version.
The images were captioned with wd-vit-large-tagger-v3. Minor manual fixes were applied to the tags.
Training data is available in the file downloads. It includes meta tags so remove those before using.
V2.0
It is now a LoKr. Training duration was increased to 18000 steps. Results in stronger character knowledge and better handling of images with multiple characters.
See also: https://civitai.com/articles/25387/lokr-test-2-how-many-steps
V1.8
Tagged characters in training data, other caption improvements, trained on Illustrious V1.0.
V1.1
Trained from 768 images and more epochs. Trained on Illustrious v0.1
I noticed that I had actually prepared 768 images for the training. Somehow I had accidentally not used them all.
V1
The model was trained from 571 images. Trained on Illustrious v0.1