Result of hq_wav2lip model training with LRS2 dataset

[linhnv97]

Thank you so much for your model. I tried to train the hq_wav2lip model after training the expert discriminator (get loss ~ 0.23) but the result was not good when inference. This is a training loss:

and this is the video result after running the inference:
https://github.com/Rudrabha/Wav2Lip/assets/140048495/3b6255c8-20c0-4c94-8eb2-3981f5f2780c
Please let me know, what are the reasons for this issue.

[xingdi1990]

Hi, Did you try the officially published checkpoints for inferencing the same face video? If the same issue happens, I would like to believe the issue comes from the video itself instead of the model. Based on my experience, your sample video contains lots of pose variations and lighting changes. Those could make the degradations on the synthesized result.