Koharu_Experiment

Test version trained with 20 CGs and character sprites, using both high and low noise levels for 20 epochs each, with repeat=10. I feel the results lean somewhat toward a 3D style, which is not the 2D effect I wanted. It might need to be used together with a style LoRA. Since the base model’s capability leans more toward realism and 3D, it’s likely that the results for character LoRAs will work better in 3D or realistic styles.

The model has a very good understanding of prompts, but if no videos are included in the training of the character LoRA, the results are not as good. All outputs were generated using Light 1.1. I’ve seen people mention that Light2 v1.1 causes slow character motion, but I’m not entirely sure about that. It could also be due to my training process — the training set might have been too limited in variety.

type = 'adamw_optimi'

lr = 2e-5

betas = [0.9, 0.99]

weight_decay = 0.01

stabilize = false