Type | |
Stats | 103 0 |
Reviews | (12) |
Published | Sep 5, 2024 |
Base Model | |
Training | Steps: 750 Epochs: 10 |
Usage Tips | Clip Skip: 2 |
Trigger Words | ppasuka |
Hash | AutoV2 A6E0FBE4EE |
Edit: v1.1 should have addressed a lot of existing issues of the previous version! You still need to prompt as below, but it should be much easier to draw good quality, faithful EVA anime style with this version.
Original notes for v1.0:
This is my first attempt for training a flux1 dev based anime LoRA. While flux1 reacts very well with realistic concepts, it is very difficult to train for anime. Why? flux seems to have strange bias in its anime outputs - it feels a lot like hand drawn. It also feels like more efforts to get flux to learn about an anime character comparing to SD 1.5.
This LoRA works, but you would need patience and luck to get the exact visual style you want, especially if you want to avoid the built-in anime style.
A few notes on using this LoRA:
You need to tell flux you want an anime picture. Otherwise it will draw a realistic girl that resembles Asuka. Use keywords like "anime style" to trigger that.
Sometimes the picture feels blurry. I trained the model on 768x768, so not sure if it was an issue.
Avoid prompting "Asuka Langley" directly. Doing this will bring out the stock Asuka from the model which, has a slightly different art style. This is the very reason I didn't put Asuka in my LoRA tags. Instead, use the trigger word "ppasuka". Example: say you want to draw something about Asuka playing tennis. You would prompt "ppasuka, in tennis outfits, swinging a tennis bat".
I am hoping to see an anime finetuned flux1 model appearing to address all these training issues. For now, feel free to experiment with this LoRA, and leave feedback. Remember, flux1 dev does not necessarily draw better anime than SD 1.5 finetunes. But you can use it to draw what was not possible before, since it understands your prompt much better now.