This is a INT8 (Row-wise, Optimized via Prodigy) quantization of AnimaYume Base Final 1.0 with pre-merged Turbo Lora. On my setup, I noticed a significant increase in generation speed on the base model — so why not speed up the inference even more?

number of steps, Euler/Simple

I've noticed that
Euler - is better with prompt adherence
ER_SDE - is better quality, styles and shading


Thanks circlestone_labs for amazing Anima model, duongve13112002 for great AnimaYume finetune, kai for excellent revision of turbo lora.