search

primepake / wav2lip_288x288

MIT License
524 stars 135 forks source link

Generated bottom half face always blur. #144

Open wangjio opened 1 month ago

wangjio commented 1 month ago

I use nearly 4w high quality talking head data to train wav2lip288x288. I found that during training, the generated bottom half face always blur. I try to use GAN loss and perceptual loss, but it doesn't help. What should I do. Is the wav2lip model structure has limitation on image quality? Or any loss can improve this?

tailangjun commented 1 month ago

预处理完的图片清晰度怎么样，你先检查一下

wangjio commented 1 month ago

预处理完的图片清晰度怎么样，你先检查一下

GT data after s3fd detection and resize is high quality.