Is the complete distillation script (DMD) available?

[happyw1nd]

Hello, great work on this project!

I have a question regarding the distillation process. Is the complete distillation script currently provided in the repository?

I noticed that the YUME 1.5 paper mentions the use of Distribution Matching Distillation (DMD). My understanding is that this requires a fake model to fit the generator's distribution. However, in the current scripts, I only found the discriminator; I could not find the implementation of the fake model or the DMD loss calculation.

Please correct me if I missed something or if my understanding is incorrect. Thanks!

[stdstu12]

We previously attempted to integrate the Self-Forcing codebase into Yume, but encountered significant conflicts. Unfortunately, I currently do not have access to any GPUs. As a result, I plan to upload a train_dmd.py script later, which should cover basically all the functionalities.

[taoranyi]

Hello @stdstu12 .Thank you for the explanation regarding the conflicts. Since a full integration might take time, could you perhaps open a temporary branch for the Yume stage 2 DMD training or upload a standalone train_dmd.py script as a preview? That would be great for reference.
Best wishes