Neural networks have learned to create humanfaces. The corresponding project was launched by one of the Uber developers, who used a combination of two neural networks to create a site-generator of realistic individuals. Moreover, the service is able to independently create both an image of a man and an animal.
The project is based on the generative algorithmneural networks StyleGAN, developed by Nvidia. One of the neural networks is directly responsible for creating the image, after which the second analyzes the image for realism. When generating a face, a neural network uses the features of several people at once, and then adapts them with the right age, facial expressions, hair length and other parameters.
These are all not real people.
To generate a random person, enoughgo to the site - every time the page is refreshed, the neural network will show another image. We checked, and in some cases it’s really hard to distinguish faces from real ones. The generator first determines the attributes of a human face (from smile to hair and even freckles), and then connects them together - but so that the image looks like a real one.
However, sometimes the fact that the image was created by the algorithm is immediately apparent: this person, for example, is no longer so realistic.
Also, a neural network can make mistakes by giving out several faces in one frame or by combining the features of an elderly person with the image of a young man.
In this video, you can see how a neural network creates new faces from its data.
Although the algorithm itself works on a powerful computer with11 GB of RAM and several video cards, you can go to thispersondoesnotexist.com and get your own generated image, both from a computer and from a smartphone. Have tried? Write what you got in our Telegram chat.