Hacker News Reader

While I'm not expert on state of the art you should keep in mind there is huge difference with deepfakes created 100% from scratch and those created via face swap and style transfer.

Basically it's easier to create believable gesticulations when there is footage of actual person as raw material.

> deepfakes can't

There is a big difference between can't and don't. Every next generation will do more than what the previous generation did.

You think we can't generate a picture of a head facing sideways? Obviously incorrect.

The argument being that there is not very much video footage of people turning their heads, therefore not enough data to train deep fake videos / filters.

Videos are still very much in the baby phase. There are way, way easier tells when a video has been faked. We're talking about images. Turned head images are very much in scope.