Hunyuan Video

• Converted to safetensors.

• Llava Llama was serialized and BF16 and FP8 versions uploaded

• VAE was converted to Safetensors

• For CLIP-L they recommend using the full vision model.

• NOTE When using the FP8 Model use FP8 Scaled

• BF16 should likely be used when downcasting to INT4 (Kijai Nodes)

• GBlue has posted a working FP8 workflow using native COMFY nodes on a 12GB card.

Comfy Native Prunes the Vision Tower - I don't know what effect that has on quality but reference Kijai Nodes

Even on a 8GB card I can render in FP8 at a decent speed (5-30 Seconds Per IT) however like many of you I am getting rainbow static. I am not sure if this is do to bad attention such as SAGE (An old Build) or a OOM issue that is not prompting the OOM error message