Prompt with "sarasa"
Training details:
- Trained with TheLastBen's fast-DreamBooth notebook
- data set: 111 concept images after data augmentation + a number of cusomized reg images (cannot remember)
- learning rate 2e-6 for 5000 steps
- text encoder rate 15%
Problems:
- The generated images can get blurred easily; this may be due to the use of anime screenshot
- Some cropping does not include face and this is not properly tagged, and therefore it may generate such images
- The pendant is never properly generated
Example generations (with inpainting):
