Better fp8 scaled models (when measured against fp16) based on quantization code from https://github.com/Tencent-Hunyuan/HunyuanVideo/blob/main/hyvideo/modules/fp8_optimization.py
Can be used with: https://github.com/kijai/ComfyUI-WanVideoWrapper (latest version) and ComfyUI native WanVideo nodes.
14B-T2V comparison test without LoRAs, 25 steps, 832x480x81
2.2 A14B-T2V test
The e5m2 marked as v2 is the one uploaded here and these are all scaled even if I forgot to label properly.