SadTalker
SadTalker copied to clipboard
Question on custom wave2lip model.
Hi:
We prepare to train a custom wave2lip model on CMLR dataset, and use the custom dataset in sadtalker, since we find that the chinese voice to lip movement is not quite accurate. I'm new in this area, and i notice that CMLR and LRS2 (which the wave2lip model originally trained on) have different image resolutions. CMLR is 64 * 128, where LRS2 is 112 * 112, if I'm right. I'm wondering could this difference cause any trouble in direct model substitution?
Thanks.