This model is trained with short clips of human speech. There is enough statistical information to "guess" how to fill the gap created by opened lips. I'm still

This model is trained with short clips of human speech. There is enough statistical information to "guess" how to fill the gap created by opened lips. I'm still amazed how it conserves temporal coherence (what it looks like from frame to frame).