Machine learning is all (unnerving) fun and (unsettling) games when it comes to swapping Freddie Mercury and Rami Malek’s likenesses, slapping Nick Offerman’s face onto the cast of Full House, or watching Bill Hader melt into Tom Cruise and Seth Rogen while doing impressions, but it gets even creepier when the digital brains of our increasingly sophisticated artificial intelligence start dreaming up new people entirely.
We probably need to get used to this. Just look at a new project like This Person Does Not Exist, which, like the name implies, consists of a series of extremely realistic portraits of people who do not, in fact, exist. This shit is only getting wilder with every passing year.
The website, when you first click over to it, seems to be comprised of nothing more than screen-filling photos of regular people. There might be a middle-aged woman who looks like she’s posing for a company portrait or a younger guy who seems to have had his picture snapped on a sunny day’s hike. The only hint that they’ve been generated by machine learning and don’t, y’know, actually exist on this planet, is the occasional patch of mottled skin texture that shows where the digital flesh has been sewn together.
A video about the process, which is the result of style-based Generative Adversarial Networks (GANs), shows how it works, outlining the way style-based GANs train on the data of real people’s faces, use levels of fine tuning to touch up the fake person, and output a hyper-uncanny result.
This Person Does Not Exist, specifically, is the creation not only of spooky technology, but also Phil Wang, whose Patreon page says he created it “to instantly educate the public on an important subject in artificial intelligence called [GANs].”
“This technique allows the machine to dream up new data indistinguishable to the training data,” he writes, mentioning that his Patreon is used to maintain the project site in order to “help inform more of the public on what is possible today [and to] keep people in the future from becoming victims to faked images.”
You can learn more by reading a technical outline of style-based GANs here, but it’s also totally fine for now to just rest your head in your hands, hold onto a solid object in front of you, and ponder the reality-obliterating shape of our future for a while first.
Send Great Job, Internet tips to firstname.lastname@example.org