'This person does not exist' which can easily generate images of people who do not exist in this world with a single touch



Uber software engineer Phillip Wang creates a website called " This person does not exist " that will create a face of a new "no one exists in the world" one by one simply by updating the site by pressing the F5 key etc. I made it public.

This person does not exist
https://thispersondoesnotexist.com/

This Person Does Not Ex Is Is the Best One-Off Website of 2019 | Inverse
https://www.inverse.com/article/53280-this-person-does-not-exist-gans-website

When you visit the official page of "This person does not exist", the face of a woman who is not amusing somewhere in the world is displayed. When you refresh the page by pressing F5 key ......


This time the pretty child's face was displayed.


Asian women laughing gently ......


A white male with white smile with a smile.


Each time you update the page, such as a double youth, a new person's image is displayed.


In fact it seems like a face photo that seems to be on a corporate officer introduction page or the like, but in reality all these images are images of fictitious figures generated by machine learning algorithms that do not exist in this world.


At first glance it can not be distinguished from real face photographs, but sometimes strange light floats in the corner of the image ......


Slightly stuffy sweet images were also generated, such as strange distortion occurred behind the ears.


Phillip Wang who created this person does not exist creates these facial pictures using a hostile generation network (GAN) which is a type of artificial intelligence algorithm. The GAN is an artificial intelligence algorithm used in " unsupervised learning " in which what should be output is not decided, using two neural network systems competing with each other to improve learning accuracy.

Originally, in machine learning, it was common to teach computers questions and answers at the same time. For example, let's have a computer that reads a dog's image analyze "characteristics of dogs" and train to identify "images of dogs" based on the features found.

On the other hand, GAN uses two network systems, "Generator" and "Discriminator". For example, if the purpose is to generate images close to real, Generator generates images similar to the real one after another, Discriminator finds whether the image is genuine or fake. The Generator raises the generation precision so that the discriminator does not see the image generated by himself, and on the contrary Discriminator raises the accuracy of the appraisal so that it can overlook the image generated by the Generator and learns each other It is said that it is becoming. Continuing this, the image that the Generator eventually generates will grow to a level where it can not be distinguished from the real thing.

As a new mechanism of such a GAN, the research team of NVIDIA , an American semiconductor maker, has created a code " StyleGAN " that makes it possible to control the process of image synthesis by adding improvements to the Generator architecture. Based on this StyleGAN, Mr. Wang said that he created this person does not exist. Regarding the purpose of creating and publishing the site, Mr. Wang says, "I want to raise my own awareness and increase public awareness of this technology."

in Web Service, Posted by log1h_ik