Last week, the internet stood in awe of the artificial intelligence-powered website This Person Does Not Exist that generates lifelike portraits of completely made-up people. The site was created by 33-year-old software engineer Phillip Wang, who tells Inverse it has been visited more than 4 million times since he first posted it online on February 11. It was only a matter of time, then, that its popularity prompted Wang to create a sister site, titled This Cat Does Not Exist.
Much like his original page, the copycats are all based on Nvidia’s StyleGAN algorithm published in December of last year. The program is a special kind of A.I. algorithm known as a generative adversarial network (GAN) that can be trained using hundreds of thousands of images to create its own renditions of whatever images you’ve fed it. To make his first, ahem, copycat site, Wang simply gave StyleGAN images of felines instead of humans.
“I [made] that because some kids on Reddit wanted to pay me for it [but] I told them I’d do it for free,” he explains. “The generator isn’t really good [but] with a bigger net and better training data, we could have infinite cats and dogs too.”
In other words, it’s a rush job that’s less refined than Wang’s than the GAN which first captured so much attention. Wang trained the Nvidia algorithm using completely random cat pictures, which included memes, outdoor shots, indoor photos, and various environmental changes that affect the quality of the final image. With a cleaner dataset — a more consistent one which features cat portraits and cat portraits only — the A.I. could generate even more scarily high-quality kitty pics.
But perhaps more interestingly, the algorithm’s replicability has since inspired other computer scientists to create versions of their own, including one StyleGAN which creates fake Airbnb listings and one that creates anime characters.This Airbnb Does Not Exist and This Waifu Does Not Exist. The Airbnb version, in particular, is the most chilling (whether you’ve stayed in one or not).
“None of the pictures, nor the text, came directly from the real world,” writes Google software engineer Christopher Schmidt. “The listing titles, the descriptions, the picture of the host, even the pictures of the rooms: They are all fevered dreams of computers.”
Every time the page is refreshed, StyleGAN creates five images of a imaginary room along with a pretty coherent description from its (also fake) AirBnb host. While the overall depiction is a bit clunkier, and less refined, than the original script first released by Wang, they do illustrate the wide range of artificiality StyleGAN’s are already capable of. The gap between fake and real, whether you’re talking about a headshot or an advertisement, grows ever-thinner.