Int8 quantized AnimaYume (hf link) (civitai link).

All credit belongs to the original model author. License is the same as the original model.

ComfyUI finally... added hardware int8 support in v0.27.

On my rtx 4xxx card int8 is 60% faster than bf16. The quality degradation of the int8 model is negligible unless you want to compare every pixel with a magnifying glass. (I did it for you)

Just ignore the "unet unexpected" warnings when loading the ckpt, all weights are loaded. It's a small bug in ComfyUI.