Kijai's picture
Create README.md
c886c23 verified

WanAnimate fp8_scaled models

v2 bugfix update for native workflows

The initial uploads has the face encoder layers in bf16, as back then I was unsure if they can be quantized.

When running these models in the WanVideoWrapper those layers are used in the base_precision, but when running these in native ComfyUI, they get downscaled to normal fp8, which seems to introduce pretty bad noise in the outputs in a grid pattern.

The v2 update addresses this by using fp8 scaled on the face encoder layers as well. For the wrapper you can keep using the initial models.