Conversation

Quite a striking result. Given a camera mounted in one’s glasses, pointed at one’s face, this system can synthesize the view of a camera in *front* of you, pointed at you. You could video chat while walking around without holding anything in front of you. youtube.com/watch?v=atzPvW

115

472

SpaceNerduino

@snetmonster

Nov 19, 2020

How the invisible parts of the face like hair are generated

Replying to

and

Quote Tweet

· Nov 19, 2020

Replying to @moonriseTK

It requires per-subject training, which presumably conditions the hair.

7:26 PM · Nov 19, 2020TweetDeck

Replying to

and

Ok, fair enough. If it is a per subject training, this is plausible. But the network should do a lot of fill in the blancs, epecially to generate the background. So when you move in a scene, it has to come up with background based on the chunk it sees which woudnt be accurate.

Replying to

and

Unless, the training set is done exactly in the same place where the inférence is done.