search

JianhongBai / UniEdit

UniEdit: A Unified Tuning-Free Framework for Video Motion and Appearance Editing
https://jianhongbai.github.io/UniEdit/
89 stars 4 forks source link

about DDIM inversion #5

Open Meiqi-Gong opened 7 months ago

Meiqi-Gong commented 7 months ago

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?

JianhongBai commented 7 months ago

hi @Meiqi-Gong , some videos indeed cannot be perfectly reconstructed, you can try to modify the beta value in the scheduler for better reconstruction. Besides, we are currently not using null-text inversion for optimization, we will take UniEdit+Null-text inversion as an extension, just like prompt-to-prompt.