Overfitting #24
Description
Hello, I noticed that the released checkpoint exhibit a huge degree of overfitting. The generated videos with the released checkpoint often come from the training dataset.
Here is an example of a generated video (in audio-to-video generation) and the corresponding training sample.
Generated video:
https://github.com/user-attachments/assets/1e0e8a4e-459b-45fb-a09a-15e304a0f618
Training sample:
https://github.com/user-attachments/assets/a5b17fa5-5c2d-405b-a415-57b3982ab43f