Why are the model weights larger than the official ComfyUI repo?
#6
by
HappyLion2
- opened
e.g. Wan2_1-I2V-14B-720P_fp8_e4m3fn.safetensors (https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan2_1-I2V-14B-720P_fp8_e4m3fn.safetensors) is 17GB, but wan2.1_i2v_720p_14B_fp8_e4m3fn.safetensors (https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/diffusion_models/wan2.1_i2v_720p_14B_fp8_e4m3fn.safetensors) is smaller, at 16.4GB.
Repo referenced: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
I left more weights at the original fp32 precision, maybe unnecessary, but initially I was testing what works with fp8 and what doesn't.