This AI generates realistic human faces from pixelated images

It's not unmasking the faces in the photos, but creating something much, much creepier.

Rudin Lab, Duke University

AI systems are getting frighteningly good at fabricating human faces.

A Duke University team reported this week that it’s developed a tool that can produce photo-realistic human faces with nothing to go on but a heavily pixelated portrait — essentially a blurry underpainting lacking eyes, lips, a nose, or any other recognizable details. While it can’t spit out a replica of the person in the original image, it’s able to create a pretty plausible HD doppelganger.

For some time now, researchers have been exploring the ways in which AI smarts can be used to sharpen up blurry or otherwise obscured objects in images — think Blade Runner’s infamous “enhance” scene. A good enough system might be the difference between an unidentifiable bird in a nature photo and that perfect shot you’ve been chasing. Or, from a more worrying surveillance perspective, it could mean more accurately identifying distinct faces in shoddy CCTV footage. But generating facial features from scratch is something else entirely.

A new face every time — To be clear, Duke’s "PULSE" AI isn’t designed to uncover the exact face beneath the pixels, and can’t be used for the purpose of identification. The face it generates is instead something entirely new that could work just as well within that space based on what it’s learned about human features. Each time it tries, the outcome will be different — and for the most part, you’d never know it wasn’t a real person. The team says the resulting portrait is 60 times sharper than the input image.

Rudin Lab, Duke University
Rudin Lab, Duke University

“Never have super-resolution images been created at this resolution before with this much detail,” computer scientist Cynthia Rudin said in a statement.

But why... and how? — Where the traditional approach is for the AI to begin with a low-resolution image and guess its way up from there, the PULSE method in a sense works backward from the top, starting with high-resolution images and matching the shrunk-down versions to the input. This is refined using a generative adversarial network (GAN), a machine learning tactic that relies on competing neural networks to produce the most realistic outcome.

With this approach, the image can go from a 16-by-16-pixel input to 1024-by-1024. According to study co-author Alex Damian, the algorithm “still manages to do something with” a photo in which it can’t make out the eyes and mouth, “which is something that traditional approaches can't do.”

Its potential is vast when you consider it beyond the task of creating human faces; this type of image enhancement could be a game-changer for areas like satellite imagery or microscopy. Just keep it far, far away from law enforcement.