Discrepancy between paper's model (Wan 2.2 5B) and code's model (14B)

[Maxwell-Zhao]

First of all, thank you for your amazing work on this project! I notice a potential discrepancy. The paper mentions the use of a Wan 2.2-5B model for the experiments. However, when looking at the provided training script, it seems to be configured to use a 14B parameter model.

[stdstu12]

Feel free to try this code. I apologize, but since I currently don't have access to any servers, improvements to code readability and reliability will have to wait. I will update it later.

train-5b.py

[bg25452]

Hi, thanks for the great work on Yume 1.5!

I have a question regarding the training procedure described in the paper:

“We then convert the multi-step diffusion model into a few-step generator G_\theta by minimizing the expected KL divergence between the diffused real data distribution and the generated data distribution across noise levels t.”

From the description, it seems:
• the student model learns a few-step generator
• while the generator G_\theta is trained with a consistency-style objective

However, it is unclear to me how these two objectives are combined during training.
Specifically:
1. How are these two learning objectives integrated into one final model ( in the paper, it was described as a G_\theta, generator)?
2. Is the generator G_\theta actually the same model as the student model, or are they separate components?

I also noticed that the released code does not seem to explicitly include a teacher transformer—so I’m wondering whether the distillation is implicit or pre-baked into initialization.

Would really appreciate any clarification on this part!