Nvidia's AI Bot creates the face of a real person

Artificial intelligence has grown rapidly. And one of the most impressive products is when an Nvidia AI bot is able to create a very real random face image.

This AI bot is built on a system called GAN (Generative Adversarial Network). This is a relatively new concept in the AI ​​world, first introduced in 1992 and put into practice since 2013.

How does bot work?

GAN works by using two opposing algorithms. An algorithm generates fake data, the other will have to guess whether the other data is fake or real.

Nvidia's AI bot also generates two algorithms. The first algorithm takes photos of celebrities' networks from the CelecA database, then creates new faces from random areas on the image.

The second algorithm guesses whether the image is true or false. If the first algorithm can trick the second algorithm, the image will be accepted as the output.

Nvidia says it teaches the new AI bot with more than 30,000 high-resolution face shots within 20 days.

Similar algorithms can create other realistic images

This AI bot is also very flexible, depending on the image the researcher uses to teach it. In another experiment, Nvidia taught algorithms using room images and furniture to create AI images of rooms with furniture.

This is also not the first AI bot that uses algorithms to create images. Earlier this year, researchers from the University of California also created pix2pix, an algorithm that takes random lines and predicts image content, creating human faces, animals .

See more: Turn doodle pictures into colors with Pix2Pix

This is an algorithm image when creating a cat from a drawing with only lines.

Nvidia's AI Bot creates the face of a real person Picture 1
Create images from simple strokes

Similarly, UK researchers also use algorithms to create 3D human faces from 2D images.

Nvidia's AI Bot creates the face of a real person Picture 2
Create 3D images from given 2D shapes

See also: Invite you to create 3D faces only from a quick photo with this AI

Nvidia will present its product next year at the ICLR conference 2018. Interested readers can read the research details here.https://arxiv.org/pdf/1710.10196v1.pdf