ModelTC
[Bug] WAN 2.2 Quality Degrated Drastically
Description
Briefly describe the bug you encountered.
I'm comparing the Diffusers implementation by the Authors to your Implementation for WAN2.2. So the Non distilled version compared to the original diffusers implementation. Using the same parameters I notice that the output quality is way worse resolution wise and also content wise.
Steps to Reproduce
- I used the default script for I2V in your examples folder. https://github.com/ModelTC/LightX2V/blob/main/examples/wan/wan_i2v.py
- I used the default WAN2.2 Diffusers implementation.
Expected Result
Similar video content and quality with same seed and same parameters.
Actual Result
The video looks lower resolution and less realistic motion and/or content.
Environment Information
- Operating System: Ubuntu 22.04, using H100
What are the differences between your implementation and the one in Diffusers? What can cause quality degradation even though we do not use the distilled models? Funny thing is, often the distilled models generate a better result that your non distilled ones. How is the speedup compared to the original in diffusers obtained?
