Most of my regular LoRA are trained with > 100 images (ideally). This is my test for a character that only has around 30 usable Garupa arts.

Usage

• This is a LyCORIS/LoCon. Can be used with with WebUI 1.5.0 and up like any other normal LoRA

• Activation tag: okusawamisaki (no space)

• Other recommended tags: black hair, blue eyes, medium hair, bangs, curtained hair

• LoRA ratio: should be around 0.7

• Recommended block weight: 1,0,0,0,1,1,1,1,1,1,1,1,1,1,0,0,0

• Recommended checkpoint: MeinaHentai v4

• You should inpaint the eyes for better results overall

Training

• 80 images from Garupa cards (full dataset available here). Data pre-augmented using waifuc's ThreeStageSplitter

• Base model: MeinaHentai v4 x0.7 + AnyLoRA AnimeMix x0.3

• Training was conducted in two steps:

  • Pre-train a Bandori art style LoRA with around 2100 card images (dataset available here) w/ around 3000 steps

  • Continue Misaki LoRA training from the previous LoRA w/ around 1200 steps

  • Both uses the same settings:

    • AdamW8bit, lr scheduler cosine_with_restart (3 restarts), 15% warmup

    • Unet lr 3e-4, TE lr 1e-5

    • LoCon dim 32/16, conv dim 12/6

    • min_snr_gamma=5

  ==> The LoRA still burns, especially the eyes