This is ANIMA LORA version for Chinese Luka. The biggest difference is I'm using datasets of Grok-generated images which came from combining both actual cosplay photos and put it into Grok to generate couple of variants.

However, the LORA training ended up a little bit strong so you will need to put the prompt with weights eg. (prompts:2) to force the image generation output exactly like the prompts with weights. Works well with most of ANIMA-based models.

Image generation settings:

• Sampler: er_sde, gradient_estimation or Euler Ancestral

• Scheduler: Simple or any kind that works well

• Steps: 30

• CFG: 7.0

*Requires prompt with weights to force image output exactly with the prompt. Should be applied onto negative prompts as well. See image samples here for prompt examples.