Closed mudassirkhan19 closed 4 months ago
do you apply syncnet?
btw wav2lip is not good enough to train the general model, you have to figure it out on yourself.
Hey @primepake,
do you apply syncnet?
yes, I'm applying syncnet, but this happens even before syncnet is started.
btw wav2lip is not good enough to train the general model, you have to figure it out on yourself.
That's too bad, I was excited after seeing the Chinese demo video, I guess DINet is much better, looking forward to seeing your take on it with the melspectogram attack, thanks for your contribution though, your comments on various issues helped me a lot in training syncnet which was a critical piece of the puzzle.
wav2lip is just good on specific person, for sure a person need at least 60 minutes to get good result
wav2lip is just good on specific person, for sure a person need at least 60 minutes to get good result
Hey, can you help me? Percep, Fake, and Real are always 0.0 during the Wav2Lip training. Please
Hi @mldev-stack ,
Thank you for your awesome work, I've been trying to train the wav2lip model (with SAM) that is present in the repo, I modified the models to work with images of size 288x288 (I don't have enough images with the resolution 384x384). However the model tends to not learn anything, it seems like it just copies the reference frames. Have you seen this before or can you point out where I can be going wrong?
Modified sam model
Dataset
Dataloader output shapes
Batch size: 16
Training pytorch lightning module