search

TMElyralab / MuseTalk

MuseTalk: Real-Time High Quality Lip Synchorization with Latent Space Inpainting
Other
3k stars 372 forks source link

How many minutes of Chinese data were used when training the model? #146

Closed XH-B closed 4 months ago

XH-B commented 4 months ago

The model you trained works very well, but we use our data （about 2000 minutes) to train our model，the lip is bad and image is blur. Do you have any suggestions? Thanks

huyduong7101 commented 4 months ago

Hi @XH-B , may I see your loss visualization on tensorboard? I trained MuseTalk with other language, but loss did not decrease much and the outputs are out-of-sync

XH-B commented 4 months ago

the result is not saved, but i can tell you the changing of it, from 0.70 to 0.22

huyduong7101 @.***> 于2024年7月31日周三 16:26写道：

Hi @XH-B https://github.com/XH-B , may I see your loss visualization on tensorboard? I trained MuseTalk with other language, but loss did not decrease much and the outputs are out-of-sync

— Reply to this email directly, view it on GitHub https://github.com/TMElyralab/MuseTalk/issues/146#issuecomment-2259950104, or unsubscribe https://github.com/notifications/unsubscribe-auth/AL4PQCEJDC35H5KKJWFH3S3ZPCNVHAVCNFSM6AAAAABK7XPBYSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENJZHE2TAMJQGQ . You are receiving this because you were mentioned.Message ID: @.***>