Question about Solver used in Wan2.1

[Jinxi-Li]

Hi,

Thanks for your great work! May I ask what kind of sample solver you are using in Wan2.1? Their default is UniPC (2), are you following the default or using a Euler solver? Or DPM++?

The second question is, in Flow Matching (like wan2.1), are you using $X_0=X_t - v(X_t, t) * t$ to get the corresponding $X_0$, then masking the results, then resample the edited $X_0$ back to t?

Thanks again!

[Chenxi-Song]

Hi, Thanks for following our work! Here are the answers to your questions:

1. Solver: We use the native UniPC (2nd order) solver from the Diffusers framework, with some minor tweaks specifically for the resampling process.
2. Logic: You're absolutely right. In the IRR stage, we first get vpred​ from the DiT, calculate the one-step denoised X0​, and then fuse it with the video_ref (warped video) in the latent space using the mask before feeding it back to the DiT.

Regarding the code release—it was delayed due to some unforeseen issues, but it's back on the schedule now. We’ve also added some new updates and optimizations in the meantime.

Stay tuned, and thanks for your patience!