Closed wanghzhRun closed 3 years ago
Your L1 and sync loss are quite low and decreasing rapidly. The GAN is hard to train due to the pose changes and artifacts that occur during training. We are working on an improvement for this at the moment.
your means is we should train gan without pretrain wav2lip model when we want to use gan to train wav2lip?
No, I didn't mean that. I just told the reason why perceptual loss is increasing.
why Percep loss keep increasing when using gan?![image](https://user-images.githubusercontent.com/24540965/100046633-87297c80-2e4b-11eb-81a6-11ad690f3afd.png)