These people do not exist. Why do websites produce fake images of people (and cats)?


But it does not exist in real life. It is a face generated on a website – which is aptly named: this person can not exist – by artificial intelligence. If you reload the page, it will be replaced by another face as convincing as unreal.
Launched earlier this month by software engineer Phillip Wang as a personal project, the site builds on a newly developed AI system developed by researchers at computer chip maker Nvidia. Called StyleGAN, artificial intelligence is able to offer some of the most realistic faces of non-existent people that machines have produced up to now.
Thispersondoesnotexist is one of the many websites to appear in recent weeks that use StyleGAN to create images of people, cats, cartoon characters and holiday homes that are increasingly close to reality and sometimes indistinguishable. from the average viewer. These sites show how easy it has become for people to create fake images with a likely real appearance – for better or for worse.
Wang, like many researchers and AI enthusiasts, is fascinated by the potential of this type of AI. To such an extent that he has created a second site called thiscatdoesnotexist that generates fake felines. But he is also worried about how he might be misused.
These three faces seem real, but were created by AI.
This makes sense, because the AI ​​tactics behind StyleGAN have also been used to create "deepfakes", which are persuasive (but dummy) audio and video files meant to show a real person what they are. She did not say.
These concerns are relayed by personalities of the sector. Earlier this month, OpenAI, a non-profit research company on artificial intelligence research, decided not to publish the AI ​​system it had created, fearing to be so skilled at composing a text that it could be misused.

However, even if the images that appear on Wang's website could be used, for example, to help a scammer create realistic online characters, he hopes this will help raise awareness of AI's emerging capabilities. .

"I think those who do not know about technology are the most vulnerable," he said. "It's a bit like phishing – if you do not know it, you could crack it."

The attractiveness (and tells) of fake people

Many people do not really know how to feel about such easy access to fake faces. But they are interested in seeing them.

Wang, a former software engineer at Uber, had been studying AI for six months when he created his website in February – shortly after Nvidia made StyleGAN publicly available. He posted an article on the site on an Amnesty International Facebook group on February 11th. Since then, about 8 million people have visited it.

"I think a lot of people look at it and say," Wow, The Matrix! Is it a simulation? Are people really in the computer? "Wang said.

The generator creates a new face every two seconds, says Wang, which you'll see when you refresh the page.

"You can think of that while artificial intelligence creates a new face every two seconds on the server and the world's poster," he said.

The faces that visitors see vary infinitely, with a multitude of eye colors, face shapes and flesh tones. Some wear lipstick or eye shadows; a handful of sports glasses. From time to time a guy with facial hair appears; any of them even looked sweaty.

This cat was also generated by AI.

They have all kinds of facial expressions. Some smile, others pout or look serious. The youngest faces appear to be young children, but none appear to be older than the average age.

As realistic as these faces may appear, there are still many details that suggest that they are not real people. For example, teeth often look a little strange and give the impression that they are in great need of scrap, and accessories such as earrings may appear on one ear. Frequently, a person will seem to have an other world skin problem or severe scarring in the face. Clothes may look blurry, have swirls of color, or just a little weird.

How are faces made

In order to generate such images, StyleGAN uses an automatic learning method called GAN, or Generative Accession Network. GANs are made up of two neural networks – which are algorithms inspired by brain neurons – that face each other to create real images of everything from human faces to impressionist paintings. One of the neural networks generates images (of a face, for example, that of a woman), while the other tries to determine if this image is a fake or a true face.
Although the field of artificial intelligence extends over decades, GANs have only existed since 2014, the year in which tactics were invented by Ian Goodfellow, a research scientist at Google. They quickly became important to many researchers, which is a major breakthrough in the field.
StyleGAN is particularly good at identifying different features within images – such as hair, eyes and face shape – allowing people who use it to have more control over the faces created. . It can also give better images.
GAN manufacturing can be fun – if you know what you're looking at – and potentially a big business. A start-up called Tangent, for example, said it uses GANs to change the faces of real-world models so that online retailers can quickly (and realistically) customize catalog images for different buyers. country rather than using different models or Photoshop. A video game company could use GANs to create new characters or iterate over existing ones.

This is not an Airbnb

Christopher Schmidt, a software engineer at Google, was one of the millions of people who saw Wang's site shortly after it was launched. He noticed that Nvidia researchers had also trained StyleGAN to create realistic bedroom images and had the idea of ​​creating his own website, thisrentaldoesnotexist, in order to combine ersatz 's. images invoked by AI with text generated by AI. The text generator that he used was trained on a host of Airbnb listings.

Nvidia declined to comment on this story. A spokesman said it was because the company's StyleGAN research is undergoing peer review.

Resembling odd and confusing versions of vacation rental listings, the results of Schmidt generated by the AI ​​are much less credible than the faces of the Wang site. (One included the picture of a Dali-esque dining room table, another included the line "Woods Minutes, and there was a garden or summer glow or relaxation for all electrical products . ")

However, Schmidt also hopes that sites like his will encourage users to question what they see online.

"Maybe we should all think a few more seconds before thinking that something is real," he said.

Source link