Fake AI-generated faces

Advances in artificial intelligence are coming fast and furious these days, and one of the most impressive is an AI bot created by Nvidia researchers that can generate images of random and extremely realistic human faces.

The AI bot is built around a generative adversarial network (GAN). GANs are a relatively new concept in the world of artificial intelligence, first proposed in 1992, and explored in practice for the first time in 2013.

How the AI bot works

GANs work by pitting two algorithms one against the other. One generates fake data while the other tries to guess if it's a fake or true.

Nvidia's AI bot is also made up of two algorithms. The first works by taking images of online celebrities from the CelebA database, and piecing together a new face from random sections of available images.

The second algorithm tries to guess if the image is a fake or a real person. If the first algorithm fools the latter, the image is accepted as valid output.

Nvidia says it trained its new AI bot by feeding it over 30,000 high-resolution images of human faces for 20 days.

Same algorithm can generate other types of realistic images

Furthermore, the AI bot is versatile. It all depends on what images researchers use to train it. In another experiment, Nvidia trained the algorithm with pictures of rooms and furniture, and the AI bot was later able to generate random images of furnished rooms.

Nvidia's new AI bot is not the only mind-boggling image manipulation algorithm that was created this year.

Earlier this year, a team of researchers from the University of California, Berkeley created pix2pix, an algorithm that takes random doodles and fills in the lines with predetermined content, such as faces, animal bodies, environment, and others. Below is an image of the algorithm generating a realistic cat from a doodle.

pix2pix demo

Similarly, UK researchers created an algorithm that generates 3D face models out of 2D faces.

Obama's 3D face

Nvidia researchers will present their work next year at the International Conference on Learning Representations (ICLR 2018). More details are available in their research paper, entitled "Progressive Growing of GANs for Improved Quality, Stability, and Variation."