This is a test model, mostly to see if the split hair was actually possible to train with a small dataset.

Trained on 9 images, mostly recolours of other characters. Mostly masked to focus on the hair.

The available SD1 models do not seem to have much comprehension of "split-color hair" as a concept, and so you get a more reliable split hair result with the overtrained models. Try playing with the LoRA alpha value.

Trainer: OneTrainer
Base model: AnythingV3
Prompt: "paisonchan"
Optimiser: ASAMW_8BIT
LR Scheduler: CONSTANT
LR: 0.0005
Batch size: 1
Accumulation steps: 1
CLIP Skip: 1

Recommended starting prompts:
"masterpiece, best quality, paisonchan, 1girl, solo, heterochromia, yellow eye, blue eye, multicolored hair, (split-color hair:1.4), blue hair, (blonde hair:0.80), blunt bangs, ponytail, glasses, labcoat, necktie, office lady, <lora:paison-chan_1800-step_20-epoch:0.9>"
"masterpiece, best quality, paisonchan, 1girl, solo, heterochromia, yellow eye, blue eye, multicolored hair, (split-color hair:1.4), blue hair, (blonde hair:0.80), blunt bangs, ponytail, glasses, labcoat, necktie, office lady, <lora:paison-chan_3600-step_40-epoch:0.7>"