Fantastic work!

[agwmon]

Thank you for contributing to the community!

I’m also a user trying to solve this issue. I'm not very familiar with ComfyUI—would it be alright if I asked for a brief explanation of how it works or any reference materials you could recommend?

Thanks!😊

[TaekyungKi]

same here.

[raindrop313]

Thank you for contributing to the community!

I’m also a user trying to solve this issue. I'm not very familiar with ComfyUI—would it be alright if I asked for a brief explanation of how it works or any reference materials you could recommend?

Thanks!😊

Okay, the main focus is on VAE and mask. I'll draw a simple model diagram over the weekend, you can keep an eye on it. @TaekyungKi

[raindrop313]

Thank you for contributing to the community!

I’m also a user trying to solve this issue. I'm not very familiar with ComfyUI—would it be alright if I asked for a brief explanation of how it works or any reference materials you could recommend?

Thanks!😊

The official technical report has already been published, so it seems unnecessary for me to draw a simple model diagram. The structure of the VAE and other details are thoroughly presented in the report. Regarding the first and last frames, as shown in Figure 6 of the paper. In short, the processing method for the last frame is the same as for the first frame. Additionally, the mask, image latent, and noise latent of Wan2.1 are concatenated together and fed into DiT. The denoised noise latent will continue to be concatenated with the mask and image latent for iteration.
Paper link: https://files.alicdn.com/tpsservice/5c9de1c74de03972b7aa657e5a54756b.pdf
If you have any questions, feel free to discuss further. Good luck! @agwmon @TaekyungKi