[Discussion] NVFP4 model quality issue

[lobo1586]

thanks for the great work on providing https://huggingface.co/lightx2v/Wan2.2-NVFP4-Sparse model.

upon comparing the output of this fp4 against https://huggingface.co/lightx2v/Wan2.2-Distill-Models, i noticed the fp4 demonstrated less accurate prompt adherence compare with the fp8_e4m3 version.

不确定团队是否aware这个现象，以及这种质量的下降是否是蒸馏模型的预期效果？谢谢

[chengtao-lv]

Thanks for your interest in our Wan2.2-NVFP4-Sparse model.

Since this model combines lower-precision quantization with sparse attention, a certain degree of quality degradation compared with the FP8_E4M3 version can be expected, including possible differences in prompt adherence.

Could you please share a concrete example where you observed this issue, such as the prompt, generation settings, and corresponding outputs? This would help us better analyze the behavior.

The LightX2V team will continue optimizing this model and improving its generation quality. Thanks again for your feedback!