We’ve seen some very weird and disturbing images created by AI before. Google AI developer Mike Tyka works on his own AI projects in his spare time. This project has something that the Google DeepDream AI lacks – the ability to create original faces. DeepDream simply takes parts from other images and cobbles them together.
The portraits that Tyka’s own AI created aren’t excessively realistic, but Tyka says that is fine. What’s important is resolution. The AI uses something called a “generative adversarial network” or GAN to refine the ability of the AI over time. The GAN draws a picture from scratch (the generative portion) and then another part of the AI (the adversarial portion) tries to tell if the image is real or AI generated.
The idea is for one part of the AI to trick the other into saying its photos are real. As both portions of the AI get better, more accurate images can be created. One thing you can see in the images created by the AI is that the hair is far from realistic. This is apparently because the AI hasn’t seen hair in enough detail to identify individual strands. It also seems to dislike symmetry – the faces often have less than perfect eye location.
You can check out more face images and read all about the technology over on Tyka’s blog.