Input video instead of image for inference.

[loic-combis]

Hi there,

I'm trying to test style avatar to correct the lip sync generation from Wav2Lip.

(I currently have an issue with the preprocessing in Faceverse LizhenWangT/FaceVerse#36)

But regardless, we aim to do the following:

• (1) Train Style avatar on the original video/face.
• (2) Run Wav2Lip on the original video with a new audio, to get a new video with but poor quality lip sync.
• Question for this step: Can we use the fine-tuned style avatar model (from step 1), to correct the lip movement on the video generated step 2?

Note that we try to avoid inputing a single image and recreate the whole speech on the avatar, and rather trying to simply correct the video.

Thanks for your help and your work!!