jialetao/mrfa

[NeurIPS 2023] Learning Motion Refinement for Unsupervised Face Animation

mrfa's Issues

Possibility of animating the face of a full body picture?

Hi! Many thanks for your work! I'm very interested to apply your work to full-body images, would it still produce good results or would the full-body image need to be cropped around the head first? Also, is it possible to animate the face of a video (a sequence of frames), assuming we only use one reference face image plus a full-body video of this person moving and talking? Thanks a lot!

demo.py 帧参考问题

在处理视频时，我有驱动视频 post_10s.mp4 通过拆分成两个视频 post_5s.mp4 和 post_5s.mp4。
在对这两个视频进行处理时，使用 --relative --adapt_scale 参数在 python demo.py 运行，最终结果合并后的动作不衔接。
去掉 --relative 参数后，五官会扭曲。
如何实现一开始就匹配参考视频的动作，以确保这两个视频的运动可以无缝衔接？

absolute keypoint coordinates

Hi, I have tried using absolute keypoint coordinates by disabling the "--relative" option, but the output appears to be very blurry.
Could you please explain the reason behind this? Are there any specific areas in the code that require modification to address this issue? Thank you.

Can't run demo.py

num_repeats Parameter in Config Files

Hi, I saw that you use different values for num_repeats parameters in config files.
While it is set to 150 for VoxCeleb1 training, it is set to 5 for CelebV-HQ training.

What is the exact reason for this specific setting?

Initially, I thought that you are setting num_repeat values to get rid of the shuffling burden per epoch.
However, since you set such a low value for CelebV-HQ, I could not further make sense out of it.

Hope you can answer soon, thx in advance!

Code Release

Hey Jiale, thank you for the great work!
When do you plan to release the code?

Thx!