Machine learning manipulates photos to change the subject's gaze

A poster talk at last year's European Conference on Computer Vision introduced "Deep Warp," a machine-learning based technique for "photorealistic image resynthesis for gaze manipulation" — that is, you hand the algorithm an image of a face, and tell it where you want the person in the face to look, and it moves the gaze realistically to have the person look in your desired location.


A browser-based implementation of this shows off the technique in impressive realtime, grabbing public domain portraits off Pixabay, then animating the faces' gaze so it always looks away from your pointer. It took me a minute to figure out what was going on, but when I compared the animation in the webtoy to the original still, I was blown away. It's not just the eyes that move — it's the whole face that follows the gaze. This will produce some amazing conspiracy theory artwork, mark my words.

Eyes Gaze Warping [Altered Qualia]

(via Waxy)