The World’s First AI Painting Sold Millions Dollars!

Before saying today’s topic, let’s guess how much the painting sold?

Before you announce the answer, let’s take a look at other paintings:

(Image courtesy of Wired, copyright TOM SIMONITE)

How about these abstract art seascapes?

Can you see what this is?

To say that these paintings have something in common, that is: they are all made by artificial intelligence! Yes, AI can also paint. And the price of this painting is not necessarily cheaper than famous paintings…

Not long ago, Christie’s, a world-renowned auction house, conducted an interesting auction, one of which was the first portrait of Edmond Belamy.

This man, who looks like a medieval man, has nothing special at first glance. Initially, Christie’s expectations for the painting were not high, and the auction price is expected to be between $7,000 and $10,000. But I did not expect that, in the end, I even shot $432,500 (fast 3 million yuan…)! You know, at the same time, another Picasso painting made by Christie’s is about the same price. Of course, this may be Picasso’s relatively unfamiliar paintings.

Why is artificial intelligence painting comparable to the price of famous paintings? Artificial intelligence can draw, who is this money? Today, Xiaoyan came to tell you about the things that AI is doing.

How is AI painted?

You may wish to take a closer look. What is special about the first painting?

The answer is in the lower right corner.

It stands to reason that whether it is a famous painting or not, there will usually be an author’s payment. But why is the line in the lower right corner of this painting a line of code?

In fact, the “author” of this painting is related to this line of code.

This painting by AI, the full name of the technology behind is “Generative Adversarial Networks (GAN)”, that is, “generating confrontation network.”

Simply put, this “generating confrontation network” is composed of two neural networks that play each other, one is the generator and the other is the discriminator. This was proposed by Google researcher Ian Goodfellow.

The generator is primarily responsible for inputting and generating data, and the discriminator is responsible for analyzing the data to distinguish whether the data is real (from the data set) or false (from the generator).

So again and again, once the “generator” successfully deceived the “discriminator”, let the discriminator think that the generated image is real, and a new painting made by artificial intelligence is born!

To make a simple analogy, it is like a girl making a photo of a male friend. It is not good to shoot. Come again, a little better, but not perfect, and then come back. Finally, only through the identification of the girl, the male friend just takes a picture that a sister wants…

In the process, the generator was input with more than 15,000 portrait datasets drawn between the 14th and 20th centuries. When the discriminator cannot distinguish whether the painting was done manually or by the computer, it is completed. The new image that everyone sees now is a copy of the painting printed on the canvas after passing through the discriminator.

It can be seen that artificial intelligence painting is not as simple as you think. Because the machine has to learn by itself and progress by itself.

Behind the painting is a French organization called Obvious, which consists of three 25-year-old French young people, Hugo Caselles-Dupré, Gauthier Vernier and Pierre Fautrel. Among them, Hugo is a Ph.D. candidate in the field of deep learning. His research direction is robotic reinforcement learning, and Pierre Fautrel is an artistic background. The three young people like to explore the combination of technology and new things. If so, why not try to combine artificial intelligence with art?

In fact, Obvious not only “painted” this painting, which was auctioned by Christie’s, but also “painted” the photos of the entire family. Obvious named the series “La Famille de Belamy” (Belamy is roughly translated into a “good partner” in French, a tribute to Ian Goodfellow, the creator of the GAN algorithm).

AI painting? Big companies are also mixing

Seeing this, everyone may say that the identification of famous paintings is still too difficult. Indeed, what makes you identify and identify celebrities?

(Image courtesy of Facebook AI Research)

In the pictures from left to right above, which photos do you think are real celebrity photos? What is the AI generated?

You may say it again, I don’t know one inside, but how do you feel that you have met before? For example, Xiaoyan thinks that the middle one feels like Beckham… The second child is like Beyonce…

That’s right.

Because the picture above is actually one of the results of NVIDIA’s use of the generated confrontation network. In fact, the celebrities in these photos are not real celebrities, but all the high-definition real photos generated by NVI based on the celebrity database.

Last year, NVIDIA published a related paper on how artificial intelligence creates real photos of “false celebrities”. Does it sound very winding? In fact, the truth is the same. Is generated by the continuous input celebrity database, and then improve the output from the discriminator identification, generator, until the last generation of this group of photos, to fool the machine’s “eyes” that they are real photographs, there above it A bunch of “celebrities” that do not exist but are familiar.

Therefore, this also confuses our human eyes.

In addition to NVIDIA, Facebook has also conducted research on generating confrontation networks. However, they turned GAN into CAN (Creative Adversarial Networks). What is the difference between the two?

As the name suggests, in the eyes of Facebook and Rutgers researchers in the United States, CAN is more creative than GAN. Because the GAN structure originally had limited ability to generate creative products, it was more like “imitation”. As a result, researchers have modified the goals of the network to prevent the final generation of content that is too similar to the original data. In this way, the machine maximizes the deviation from the established artistic style during the training process, thus creating creative images.

(Image courtesy of Facebook AI Research)

Abstract paintings like the above are machines that are generated according to the idea against the network. Hey, Xiaoxun can’t tell who painted this artwork. . .

Not only did artificial intelligence create paintings, but Facebook researchers also conducted tracking experiments to compare the reaction of people watching the images generated by the system with the paintings created by the artists. Is there a difference?

As a result, you guess.

Yes, these participants, like the little explorers, can’t tell the difference between the images produced by artificial intelligence and the paintings that contemporary artists show at top art fairs. It is even thought that some CAN-generated images are not much different in creativity from the artist’s work. (That is, if we don’t have real discriminating power, then you may not really know whether these paintings are machine-painted or artist-made.)

Is artificial intelligence painting infringing?

If you say those big companies do their own internal experiments. However, just as the first AI painting was auctioned, the voice of the discussion came along, that is: let AI paint, will there be copyright issues and infringement problems?

We all know that hand paintings are not prone to infringement problems. Because as long as you have your pen signature, the rest can only be called “impersonator.” Even the world-famous paintings such as Monet and Van Gogh, which we are familiar with, do not have infringement problems in terms of time. According to Article 7 of the Berne Convention, the protection period of the property rights of the general works of the member states of the conventions are the author’s lifetime and 50 years after his death.

However, when the author of this painting has changed from a person to a paragraph of “code”, once the code is still open source, who is the author? In this case, who is the income after the auction?

In fact, after the auction of Edmond Belamy, a 19-year-old boy, Robbie Barrat, jumped out. He thought the code in this painting quoted a lot of open source code on GitHub. Even on Twitter, he used the paintings trained by the neural network model more than a year ago to compare with the auctioned paintings.

(Photo from Robbie Barrat Twitter)

Robbie Barrat, an artist mentioned above, is currently doing AI and medical-related cross-disciplinary research at Stanford University. He used artificial neural networks to make works of art, including landscape paintings, nude paintings, and fashion T-pictures. The first nude paintings were his works.

Every once in a while, Barrat puts the old training dataset and training model on GitHub for open source for everyone to use. “My goal is that people will play with it like you, and then continue to do more,” Barrat said in an interview with the Yale School of Architecture magazine.

So, Obvious is thought to use Barrat’s data model directly, and finally selects the artwork they want from the generated photos, and then takes them to the auction.

According to the Obvious founder, in an interview, I really appreciate the contribution of Barrat. The team mainly uses Barrat’s code to complete the data capture and get all the artwork data. But the team itself adjusted the hyperparameters in the process to get the final result. Even without the use of Barrat’s code, the team focused on different datasets to ensure that the work on the Belamy family was completed.

The author of this painting seems to have fallen into a “Rashomon”. So the question is: If Barrat’s code is shared under an open source license, can he claim for ownership of the artwork? If it is really painted by AI in the future, once the painting is sold, who should the money be?

Everyone is welcome to leave a message to talk about it! Especially the code farmers friends~

If you want to try a friend of AI painting, Xiaoyan finally recommend a few URLs, you may wish to try:

DeepArt.io, users only need to upload photos and choose an artistic style, such as Picasso or Van Gogh. Artificial intelligence combines the content of uploaded images with artistic style, and even your selfies can be generated!

Get-art.work, this site focuses on creating Van Gogh-style paintings. It compares the submitted photos with Van Gogh’s nine paintings to find the best match between the two and generate the final painting. Can be customized to ensure no discoloration for 75 years, but the price is a little expensive.

Really not, we still have retouching software! I still remember the very hot retouching software Prisma that year, um, I will generate my own world-famous painting style for my photos~ Maybe I can sell the money too!