SadTalker +DPE 生成效果较差

[canghaiyunfan]

将SadTalker 生成视频作为源视频，data/s.mp4 作为驱动视频进行pose 驱动，和提供的demo 相比结果比较差，中间需要做什么处理吗

第一个视频是SadTalker 生成的结果，第二个是DPE pose 驱动的结果

https://user-images.githubusercontent.com/21115878/236419311-63f501f8-d425-4e46-9617-f5630e6f7d4a.mp4

[Carlyx]

For the current solution, you can try to narrow down the distance between the head proportion of source video and driving video, such as according to the 'crop_video.py' to process video, to make the result better.

However, there are some problems:

1. The provided pre-trained model was only fine-tuned on expression, and we didn't fine-tune pose information. Actually, we have had a better model for pose, but now I just can't find it. Maybe I will train another version afterwards.
2. Although we have a better model for pose transfer, DPE focuses on the transfer of expression, which means that pose model still has some limitations, for example, the proportion of heads. In fact, we are going to solve this problem in the next version.

[canghaiyunfan]

Thanks!