train wav2lip288 in a low loss,but it not perform well in inference!

[killnice66]

i use hdtf for wav2lip288 training,nearly 170 0000 pictures ,16hours my syncnet eval loss is 0.3 and my L1 eval loss is currently 0.019051536196276822 and my sync eval loss is 0.18783933469512834 which seems good ,but my inference product the original frame with some blue dots on them ,what's the potential problem in it?

my inference video (which is the same lip as original video) https://github.com/Rudrabha/Wav2Lip/assets/74774674/8e794215-a9b6-4c08-ac16-170f8fefef70

[Ronincod3]

any updates?

[LiuJiaqi0505]

i use hdtf for wav2lip288 training,nearly 170 0000 pictures ,16hours my syncnet eval loss is 0.3 and my L1 eval loss is currently 0.019051536196276822 and my sync eval loss is 0.18783933469512834 which seems good ,but my inference product the original frame with some blue dots on them ,what's the potential problem in it?

my inference video (which is the same lip as original video) https://github.com/Rudrabha/Wav2Lip/assets/74774674/8e794215-a9b6-4c08-ac16-170f8fefef70

Hi br. The video looks like out of sync.