search

ali-vilab / UniAnimate

Code for Paper "UniAnimate: Taming Unified Video Diffusion Models for Consistent Human Image Animation".
https://unianimate.github.io/
982 stars 52 forks source link

Some questions to paper and the implementation #16

Open zhysora opened 3 months ago

zhysora commented 3 months ago

Thanks to your great work.

  1. Temporal Mamba: In paper, it suggests to use Temporal Mamba as the temporal module. But in this code repo, it seems there is no related implementation about it.
  2. Long video generation: In paper, it shows that the first-frame conditioning solution is better than the slide window strategy. But in this code repo, the default config is still slide window strategy.
wangxiang1230 commented 3 months ago

Hi, thanks for your attention.

  1. Firstly, temporal mamba is an alternative. We released the temporal transformer version since we find that junior users need to install many python packages for mamba and may counter some problems.

  2. The first frame conditioning strategy can also be achieved by using this codebase actually. You can input the last frame of the segment to the model as the first frame.