[RFC]: Trying to reproduce DMD distillation for WAN-Animate

[tellurion-kanata]

Motivation.

Trying to reproduce DMD2 distillation on Wan-animate for 4-step denoising (following fast-video pipeline, but currently w/o GAN loss).
The first frame looks fine, but all subsequent frames in the generated video are completely black.

Anyone knows what's the problem? Or any open-source repos for distilling Wan-animate? Thanks!

Training sample and loss curve here:

Proposed Change.

4-step distillation pipeline for Wan-Animate.

Feedback Period.

No response

CC List.

No response

Any Other Things.

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues.

[jzhang38]

What training script are you using? We do not officially support Wan-Animate DMD?

[tellurion-kanata]

What training script are you using? We do not officially support Wan-Animate DMD?

I made one following your Wan-i2v's implementation (fastvideo/training/wan_i2v_distillation_pipeline.py), as Wan-Animate is more likely to be based on Wan2.1 instead of 2.2.
I just wondered if any similar collapses observed in reproducing DMD training on Wan series.