search

johndpope / Emote-hack

Emote Portrait Alive - using ai to reverse engineer code from white paper. (abandoned)
https://github.com/johndpope/VASA-1-hack
172 stars 9 forks source link

Training Strategy - overfit training stages to 1 single video (mp4) #17

Closed johndpope closed 8 months ago

johndpope commented 8 months ago

With the training - I think we can just train on 1 video - and over fit it to begin with - no need just yet for 40gb videos. so given 1 video trained / overfit / layers saved - give 1 frame - and sounds - and have model generate video matching region + speeds. I have a crack at this.