There’s an AI that can make up faces that don’t exist, male or female, young or old, and can almost be messed up.

The most powerful face generator on the planet, NVIDIA’s StyleGAN2, has been around for three months since it was born. It’s been three months since it was born, a long time, and a virus that’s driving a lot of people crazy at home, and no one’s going to play bad with this AI?

No way. At least some of the deadheads who are addicted to sex-turning won’t let it go. A handful of fighting teenagers from Russia took StyleGAN2’s gender reassignment skills and unleashed them on their own. With this, Harry Potter Mother was born.

As a result, the teenagers slammed the skin color/ethnicity and age-changing Dafa with the same principle:

The little AI that performs each task is blissfully on par with the best performing algorithms of the moment.

How to tap into sexual transfer skills

How did three teenagers from Yandex, a Russian Internet company, and the Moscow Institute of Physics and Technology, go from a casual The algorithm that generates the face separates this particular function of sex transfer? They are using the Distillation technique.

The idea behind Distillation is to use StyleGAN2 as a teacher, which generates images containing knowledge that can be used to train the A student model. The sex-turned-AI is this student, acting as a small model that can quickly perform its own specific tasks.

The problem, however, is that StyleGAN2 generates just a single graph, not a set of male and female control charts. But what about the biased training data that the student needs in pairs?

Let’s start with the way face generators work. The generator actually relies on something called “latent code” to produce images. The latent code is a randomly generated vector (with length and direction) that is input to the generator to obtain a corresponding image… So what needs to be explored is what kind of latent code corresponds to what gender.

This problem can be easily solved with an image classification AI that identifies men and women. For every portrait entered, the AI outputs a gender.

The AI not only gives you the classification results, but also tells you how certain it is of the judgment. The Confidence score is between 0 and 1. The higher the score, the more certain the AI is that a graph is male or female. To ensure the quality of the training data, graphs with too low a score will be discarded Gas Gas Ref. 2

Of course, it’s not enough to judge the gender of each picture. Because the goal is sex transfer, not to turn a boy into a girl who looks completely different.

The next step is to put all the boys together and all the girls together and observe them. Each person is a vector that corresponds to a position of their own in space. Figure out the center of the boy and the center of the girl to get the difference between the two positions, which is still a vector. And the direction of this vector is the direction of the sex turn.

Now that we have the direction, it’s easy to create a set of graphs corresponding to each sex: still randomly generate the latent code and add this vector to the latent code , subtract this vector, plus 1/2 of the vector, and subtract 1/2 of the vector. this gives you 5 graphs per photo, and then The “most male” and “most female” images (with the highest AI confidence scores) are selected according to the judgment of the image classifier, and paired sex transfer datasets are available. That’s how it’s done.

Right = most male & most female Gas Gas Ref. 2

50,000 samples in the dataset, each sample includes 2 source maps (before gender transition), and 1 target map (after gender transition) . It was taken to train student models with remarkable results.

In addition to gender alteration, skin tone alteration and age alteration were distilled using the same principles, and the results of several of the little AIs were comparable to the most powerful algorithms available today.

Note that the AI’s training data are all synthetic images produced by StyleGAN2, but the tests are modifications of the real Photo. the AI performed well, and indeed the teenagers should be proud of it; it also means that they chose to trust StyleGAN2 in the first place! The image-generating capabilities of the company were not chosen incorrectly.

On the shoulders of giants

Some of you may not be familiar with StyleGAN2, but here’s a serious introduction to this shiny AI.

As you can see from the name, it’s a GAN, which is a Generative Adversarial Network (GAN). (Adversarial Network). By confrontation, we mean that a painter (generator) and a connoisseur (discriminator) live in the body of an AI and love each other to death. The painter looks at a human painting or photograph and tries to generate a realistic image so that the connoisseur thinks it is the work of a human; the connoisseur The effort not to be fooled motivates the painter to produce more realistic images. As both become stronger, GAN’s paintings become more and more realistic.

StyleGAN also has this dual personality. It’s just that the painter in its body is different, using the style of one image (e.g. tone, texture) to transform another.

The method of changing the style of an image is called “style migration” and it is the brainchild of Leon Gatys, who can turn a cat film into an oil painting.

The developers of StyleGAN have divided the style of portraits into three scales: on the coarse scale, they fix the face shape, orientation, and hairstyle; on the medium scale, they fix the hair color and facial features; and on the fine scale, just the color scheme of the picture changes and the person doesn’t become another person anymore.

Upper left = coarse scale, middle left = medium scale, lower left = fine scale Stella Ref. 3

The initial generation of StyleGAN, in fact, has become the most powerful face generator, even supporting 1024×1024 HD! Big picture generation. But the developers still picked out the flaws in the algorithm and effectively filled them in the second generation of the algorithm.

For example, the strange “water drop” flaw is gone.

Upper = first generation, lower = second generation Stellar Reference 1 The problem of the facial features being out of sync (e.g. the orientation of the face changed, but not the teeth orientation all together) was also improved by:

Down = second generation, teeth and eyes can rotate together with the orientation of the face Stella reference 1

In addition, scientists say, the fit between the three scales is more natural than before.

And so a new pinnacle emerged. It’s only natural for tech geeks to seize the opportunity to flirt with it, which is why we see the scene at the beginning.

The Sexual Turning of the Lifetime

In fact, even without StyleGAN, the act of humans taking AI for sex turns has been going on for a long time.

In 2016, a mobile app called FaceApp arrived and shocked almost the entire world. “Gender Swap” is probably its proudest feature.