jianhongbai / uniedit Goto Github PK

I'm curious about the reconstruction path. I can't reconstruct the original video very well only using empty text for noise and denoise. As I notice that you use the experience from the null context, which need to be optimized in the original paper, did you optimize the null context in the reconstruction process as well?

Is this article based on stable video diffusion？

请问你们有开源代码的计划吗，什么时候可以使用这个工具呢

Hello 👋

Thank you for your amazing work!

I have a few questions concerning your paper, typically the Mask-Guided Coordination (Section 4.3)

1. Is the mask-guided coordination scheme also implemented during "appearance editing"?
2. Is masked attention applied in the spatial self-attention block or the temporal self-attention block, or both?
3. When and where is masked attention applied in terms of denoising timestep $t$ and attention layer $l$?
   Is it only during content preservation $t&gt;t_0, l&gt;l_0$ (resp. structure control t<t_2, l>l_2)? In other words, is the $V$ (resp. $Q, K$​) in formula (6) from the reconstruction branch?
4. For the mask $M$, do you use the same mask for all video frames (if so, could you elaborate how this mask is generated?) or do you concatenate all the frame masks?

P.S. What's the exact source prompt you use to generate the results in Figure 1? I attempted 'A raccoon is playing guitar' but it didn't quite nail that cartoonish and detailed background vibe as in your demo

Your guidance on these queries would be immensely valuable, many thanks!

你好，请问什么时候可以提供web页面体验一下啊