Model sizes between paper and code implementation

[Maxwell-Zhao]

Thank you for open-sourcing the code. This project is a fantastic contribution to the video generation community! I have a question regarding the model sizes used for the teacher (real_score), critic (fake_score), and generator.

In the paper, it is mentioned that the teacher model is the larger Wan2.1-T2V-14B, while the generator and critic are the smaller Wan2.1-T2V-1.3B. This setup, where knowledge is distilled from a much larger teacher to a smaller student, is a key highlight of the method. However, when I look at the implementation in dmd.py (specifically in the BaseModel), it seems that the default model names for all three components point to the smaller 1.3B model.

Any clarification on this would be greatly appreciated. It would be very helpful for those of us trying to fully understand and reproduce your excellent results.

[AndysonYs]

Hi. Sorry for the confusion. We did use Wan14B as the real_score model. The default training setting are all set to 1.3B is for memory reason. Using 1.3B as the teacher model leads to worse results.