User LoRA
Community DiT
User LoRA
@zzzdcatl
My attempt to train a style of characters in ZZZ "dream comes true at last" event. There weren't much samples for train, just 6-8 ig, after augmentation I could achieve 16 samples but removed some because they seemed to be repetitive and could cause more overfit. It doesn't mean it doesn't overfit now, but still shows good results. If this work will go well I will try to expand the dataset and update the lora.It was trained on NAI EPS 1.1 but works well with NAI vpred and Illustrious checkpoints (tested on WAI/Hassaku/Rouwei).Important Notesince all samples had white background I recommend adding simple background, white background to negative prompt, it should fix the problem with overfitted background (if you have them)quality tags like masterpiece, best quality etc are very strong so they can cause changes to style, you are free to use them but if you want the exact style as in this post i'd recommend remove all quality tags and stick with negative quality tags, it will clean up results but won't affect the style, you can increase lora's strength as well if you feel your checkpoint doesn't follow the style enoughUPDATE: Version for Anima Base v1.0Decided to try training an Anima LoRA, since it's my very first attempt, I didn't tweak any settings and left the preset ones. As expected, the model had gotten strong biases on trained data, especially white background, which was the most consistent part. I have picked the sweet sport between style and stability, yet you still should be more specific in your prompt in order to negate the biases. If you are using some finetune the biases might not affect you, yet I didn't test.