This AI-Assisted Aging Software Looks Spookily Realistic

It's nearly impossible to accurately imagine what a person will look like when they're older. Just look at the makeup on Michael J. Fox in "Back to the Future II," or Linda Hamilton in "Terminator II." Even when you see images created by age-progression software, they never quite look right — more like a bunch of wrinkles projected onto a picture of a young person. But now, a new AI-assisted approach is putting out pictures of aged celebrities that a lot of people think are weirdly accurate.

Aging Faces

This new kind of age-simulation software was the brainchild of a global team from Beihang University in China and Michigan State University in the United States. The key was realizing that for an age-progression program to work, it's got to fulfill two conditions: First, it needs to realistically portray the age it's trying recreate, and second, it needs to stay true to the person's original face. Managing the balance between those two is the key to creating a synthesized face that looks like an older version of a person you know, not a ventriloquist's dummy version.

Don't believe it? Just look for yourself. They tested their new technique on two different sets of faces: images from the MORPH mugshot database, and red-carpet snapshots of celebrities from the Cross-Age Celebrity Dataset. As it turns out, Michael Cera has got the face of a proto-Michael Douglas, and Kirsten Dunst may one day look a bit like Sharon Stone. At least, according to some of the images. They ran different pictures of the same celebs through to get a broader range of results, and it's clear that they can't all be accurate. But what sets this method apart isn't just its potential for accuracy — it's also that it just looks right in a way that other types of simulators don't.

Two Computers Are Better Than One

What sets this team's software apart from others? Simple: it's got GANs. Short for "Generative Adversarial Network," a GAN is actually a system of two neural networks that sort-of "argue" with each other in order to get to the best answer. One network tries to modify a picture of a person's face, and the other one makes a judgment call about how well the first one did. That second network is focusing on the believability of the image, and the feedback it provides helps the first network perfect its techniques. Obviously, they haven't quite perfected the technology that lets them compare the results to a real-life 75-year-old Leonardo Dicaprio. However, when human participants were asked to judge if the GAN-assisted faces or faces made from older methods looked more like their favorite celeb aged up, they chose the new method up to 70 percent of the time.

Curious about how your own brain's software sees faces — and what happens when it goes wrong? Check out Oliver Sacks' "The Man Who Mistook His Wife for a Hat," where you'll also meet a man who has been unable to form new memories since World War II, and a pair of mathematical savants on the autism spectrum who seem to have an innate sense of prime numbers.

Your Brain's Facial Recognition Technology

Written by Reuben Westmaas July 5, 2018

Curiosity uses cookies to improve site performance, for analytics and for advertising. By continuing to use our site, you accept our use of cookies, our Privacy Policy and Terms of Use.