Trained on images from the movie The tale of the princess Kaguya by Isao Takahata
Klein 9B
Showcase images :
Distilled model
768x1152
Euler / normal / 4 steps / CFG 1
Lora strength of 1-1.2
Qwen
2512
It was trained on 2512, but I prefer to use the original Qwen-Image model for inference.
Showcase images :
Original Qwen-Image (BF16)
Lightning lora (4 steps)
5 steps / 1 CFG
DDIM & Beta
Lora strength 1
Flux
v2
Fixed the straigth line issue with a bigger and better dataset.
It seems better at following prompt with this version, but the style is a bit different from v1, colors are more vibrant.
Maybe adding "pastel colors watercolor" can help having less vibrant colors.
v1
I may retrain the lora one day because there is often a straight vertical line appearing (going to the sun/sky/moon) :
