Artificial Intelligence Writes Pictures in a Unique Style


From time to time in the world there are such bright artists (for example, Claude Monet or Pablo Picasso), which literally turn the world of art upside down. They invent their own styles in painting, resulting in the formation of new trends in art, for example, impressionism or abstract expressionism. Now the researchers asked the question: can the same breakthrough in art make a machine?

And such thoughts are completely justified. More recently, it was reported that neural networks learned to draw simple drawings. And now artificial intelligence (AI) is already able to create images in non-traditional art styles. Moreover, most of his works have already received public approval.

The development team, which included researchers from Rutgers University’s Facebook’s AI laboratory, changed the type of algorithm known as generative controversial networks (GANs). In it, two neural networks are “pitted” against one another in order to obtain a better result and to improve oneself. So, one network offers a solution, and another criticizes it. Thus, the algorithm chases all decisions until the desired result is achieved.

In the “artistic gifted” AI, the role of one of these networks is played by a special network generator, which creates an image, and another – a network recognizer. The specialists trained this network recognizer on more than 80,000 paintings so that he eventually learned to understand the difference between works of art and an ordinary chart or photo. He is also able to understand the difference between different styles in art – rococo or cubism, for example.

The interesting fact is that the generator creates a work that the recognizer recognizes as a work of art that does not fall under any of the existing styles.

After the AI wrote a series of pictures, scientists asked the members of the public to evaluate them along with those pictures that the person wrote. All respondents did not know what kind of pictures the car had created. All respondents were asked how difficult and new they found each job. In addition, the researchers were interested in whether they were inspired by the picture and whether the mood was rising.

To the astonishment of the authors of the study, in many cases the images of authorship of AI received higher ratings than human ones.

It is noteworthy that in the world there are already programs that can slightly correct the photo so that it imitates the style of famous artists – the same Monet. There are even applications (for example, DeepArt), which can easily cope with this task. But the new system is designed to create original works, as they say, from scratch.

The results of the survey are very interesting, says Kevin Walker from the Royal College of Art in London. “Images rated extremely highly contain an aesthetically pleasing combination of colors and patterns, and pictures that have received poor grades are more homogeneous,” he adds.

Moreover, Walker said that the creative machine is already creating works for the galleries. For example, two of his students are currently experimenting with AI who is able to adopt their drawing style to create their own works. One of the students used this technology to create frames for a 12-minute animated film.

The results of the study and the description of the system are published on the website of the preprints

We add that AI is endowed with other talents. So, Google turns text into images, AI also defeated the world champion in the game go. Moreover, he was even able to create sounds that no one had ever heard.