Usage

• Activation tag: yonememei

• Recommended extra tags: red hair, blue eyes, single hair bun

• Recommended ratio: 0.75

Training

• Base model: AnyLoRA

• Training data here

  120 anime screenshots, augmented using waifuc ThreeStageSplitter to get around 150 total training images

• Optimizer: AdamW8bit, cosine_with_restart, 3 restarts

• 2 repeat x 10 epochs

• Unet lr 5e-4, TE lr 1e-5

• LoRA dim 32/16

Different from previous LoRAs which used fully automatic tagging, for this I use a hybrid approach: auto tag and then half-manually curate the tags.

• Filter tags with frequency lower than 10

• Ensure there's no tags of other characters (ex. the auto tagger might tag "blonde hair" if parts of Sumire is visible in the image)

• Use more specific tag: single hair bun instead of hair bun, pinafore dress instead of dress (for the uniform), ...

• Manually go through the image and add other tags where relevant, such as from side, close-up, ...

This one performs better than my (unpublished) Natsumi LoRA, which used over 200 training images but the tags were not prepared as carefully as this one.