It would be amazing and very useful if it were possible to “tag” faces in an input video (place a square over a face), then for each of those, assign a good face picture of that person for each face ‘tag’…then the model can use the uploaded image to generate the facial features in the video, rather than getting whatever random face it pulls together from training data. You could even re-use images you already assigned by giving each one a name and simply clicking and assigning a person/picture you already uploaded.
Probably couldn’t build that off an existing model, but it’s something to think about.
Just saying that it can be a bummer when you’re upscaling home videos and Uncle Ted looks like some random person (or an alien).
