Welcome to Uncanny Valley: AI-Generated Selfies Invent People Out of Whole Cloth

When 3D computer graphics first started getting really big, some actors worried that it would put them out of a job.

Those fears were largely unfounded because there was always something a bit “off” about computer-generated faces.

That subtle difference you notice is called the “uncanny valley effect” and it is what makes doll-eyed computer characters fake to the human eye. Overcoming this is not only a goal of graphics designers but also an aspiration for other fields as well.

Now AI can generate selfies of people that don’t even exist, and they look amazingly real.

The geniuses over at NVIDIA unveiled their work a generative adversarial network (GAN) and the results speak for themselves. A technology that debuted in 2014, GAN’s in generating fake faces is well known. It’s just now that the tech has gotten to such a level that it is nearly impossible to distinguish the real faces from the fake.

The way PetaPixel describes it is that GAN combines elements from multiple pictures for a pastiche portrait that looks genuine. Because it is assembled from components of actual photographs and then augmented together the result is something that looks real.

It’s not just limited to selfies, however, as GAN can also create real-estate photos and other things by using much the same process by taking elements from multiple “source” photos for a final composite that looks completely authentic.

Probably the test case the Internet is most interested in involves GAN generating its own cat photos. The Internet can’t get enough of those so it is good to know that someone is literally developing a technology that could morph into a virtual cat meme factory.

Who knows what this means for the future but it is definitely equal parts cool and spooky.

NVIDIA even released a video explaining everything which you can watch over on YouTube by clicking here.