How do they make those fake people on thispersondoesnotexist.com?

937 views

I tried to read what they explained (the GAN software) but I couldn’t quite understand it. Can someone explain it to me?

In: Other

2 Answers

Anonymous 0 Comments

The way I understand it, it works like this:

Imagine a robot painter and a robot art critic. The robot painter has all the tools and paint, but has never painted (or seen) a human portrait before. The robot art critic is incapable of painting or explaining what a face looks like, but has memorized hundreds of human portrait paintings he had seen earlier.

Now the painter begins by just splashing random paint on the canvas and passes it to the critic, who tells him it doesn’t even come close to any of the human faces he had seen before and to try again.
They repeat the procedure a hundred times, until one of the random paintings resembles the outline of a human face.
The robot critic tells him: “paint more like that one!”. So he does, again a hundred times until facial features appear, almost by accident, in one of them. The critic encourages him to paint even more like that one painting.
This process is repeated thousands of times and with every positive feedback, the robot painter is getting more precise in his brush strokes and his choice of colors, until he gets just every detail right and is able to produce a perfect portrait that leaves the other robot speechless. Neither of them really fully understands what an actual human face looks like or how to paint anything else, but they worked out a perfect way to imitate portraits as we know them.

In this case, the paint is random pixels. The painting robot is the “generator” and the art critic robot is the “discriminator”.

Anonymous 0 Comments

In the very basics, these networks work as follows:

1. Collect a ton (think hundreds of thousands samples if not more) of training data, in this case pictures of people.
2. Feed this data into the network to train it. Training allows the network to ‘learn’ what makes the picture a human.
3. After the training, you can ask the network to generate a random picture which according to the network still looks like a human.

Of course, training such a network is hard, but there’s libraries out there that can help you get started.