MIT Technology Review
Paper
The faces above don't seem particularly remarkable. They could easily be taken from, say, Facebook or LinkedIn. In reality, they were dreamt up by a new kind of AI algorithm. Nvidia researchers posted details of the method to produce completely imaginary fake faces with stunning, almost eerie, realism (here's the paper). The researchers, Tero Karras, Samuli Laine, and Timo Aila, came up with a new way of constructing a generative adversarial network, or GAN.
GANs employ two dueling neural networks to train a computer to learn the nature of a dataset well enough to generate convincing fakes. When applied to images, this provides a way to generate often highly realistic fakery.
In the most recent work, the researchers took inspiration from a technique known as style transfer to built their GAN in a fundamentally different way. This allowed their algorithm to identify different elements of a face, which the researchers could then control. A video produced by the researchers shows how the approach can also be used to play with, and remix, different elements, like age, race, and gender--or even freckles.
"It surely seems like another big quality leap for GANs," says Mario Klingemann, an artist and coder who GANs in his work. "It also appears to be amazingly controllable, unlike GANs so far where you have to experimentally figure out how to steer the results into a certain direction (like making a face smile or age it)."
Klingemann says he is keen to get his hands on the code, and to experiment with it for artistic purposes. "I am very interested to find out how to make that model do 'wrong' things," he says. GANs are likely to change the way video games and special effects are generated. The approach could conjure up realistic textures or characters on-demand. Nvidia recently showed a project that uses GANs to synthesize the appearance of objects in a scenes in realtime within a driving game.
Or as one tweet put it, "The end of photography as evidence"
