AI might be a good western painter, but will it perform well in Chinese ink painting?
The term Artificial Intelligence (AI) has infiltrated every level of social life. International grandmaster Arnold Denker studies the pieces on the board in front of him. He realizes there is no hope; he must resign the game. His opponent, Hitech, becomes the first computer program to defeat a grandmaster in a game of chess in 1988. Since almost that time, AI has not only existed in science fiction but has truly spread into the lives of every one of us.
In painting, in 2016, an AI model called “Image Style Transfer” was launched, which can transform any Image into the art style of the great masters such as Van Gogh and Monet. Furthermore, there are artificial intelligence models that can directly generate artistic paintings [1].
The application of AI in artistic creation not only brings surprises to people but also makes people involuntarily think and question: Do AI creations belong to artistic works? Can models of AI be called artists? Can AI models really “understand” art and “know” how they are creating it?
There are many definitions of “artificial intelligence.” As defined by the author of “Artificial Intelligence: A Modern Approach,” artificial intelligence needs to have four modes: thinking like a person, thinking rationally, acting like a person, and acting rationally. It makes a distinction between thinking and acting, so we can also think of “artificial intelligence” as a machine simulating human thinking and behavior.
The definition of “art”, according to Oxford Dictionaries and Merriam-Websters Dictionary: Art is a diverse range of human activities in creating visual, auditory or performing artifacts (artworks), expressing the author’s imaginative, conceptual ideas, or technical skill, intended to be appreciated primarily for their beauty or emotional power.
It can be seen that the prerequisite for a work of art is the need to contain “beauty”; Secondly, art is a kind of social activity, a kind of “social performance”. Artworks are the emotional communication link between artists and audiences as well as between audiences.
There is a professor of philosophy in a Chinese university who proposes the concept of “the Aesthetics of Artificial Intelligence”: “Aesthetics is the study of artificial intelligence in the process of artificial intelligence technology development of some problems related to aesthetics, its main contents include artificial intelligence of human perceptual (including emotional) and the style of the art of the simulation, artificial intelligence, artificial intelligence, and appreciation view the human emotion and artistic nature, its method is mainly the philosophy and aesthetics, and require a combination of many disciplines, such as brain science, neuroscience, biological evolution theory and recent advances in the field of artificial intelligence to study.” [3]
In his opinion, artificial intelligence needs to simulate two aspects of human beings in order to complete the artistic creation of real intelligence. One is the simulation of human artistic behavior; Second, it stimulates the emotional behavior of human beings in artistic activities.
The “Xieyi” style in Chinese ink painting
In 2018, an AI portrait hangs at Christie’s in New York, opposite a Paint by Andy Warhol, and next to a bronze work by Roy Lichtenstein. “Edmond de Belamy, from La Famille de Belamy” sold for $432,000, including commission, more than 40 times Christie’s initial estimate of $7,000 to $10,000. The buyer is an anonymous telephone bidder. It sold for more than twice the combined price of the two works next to it.
This made me start to think about a problem. Oil paintings in the Western world, especially portraits and landscapes, are mostly accurate descriptions of the details of models. The painter’s skills are reflected in the use of light, the accurate capture of the expression of the figures, and the accurate depiction of the figures’ body details. This “technique” allows the AI to learn a lot to imitate and “create” new works of art on top of that.
However, according to the means of expression, Chinese painting can be divided into two categories: the Xieyi style and the Gongbi style. The Xieyi style is marked by exaggerated forms and freehand brushwork. The Gongbi style is characterized by close attention to detail and fine brushwork. A famous painter once said: “ The Gongbi Style” can be learned in three years, but “ the Xieyi Style” may not be learned in 30 years.
The Xieyi style of classic Chinese ink painting is more than a technique that features a reduction in detailed brushstrokes. Essentially it describes a philosophy of Chinese culture to emphasize freedom, spirituality, individuality, and expressiveness.
Even in ancient times, Chinese artists were unwilling to be restrained by reality. A famous artist of the Jin Dynasty Gu Kaizhi (348–409) was the first to put forward the theory of “making the form show the spirit”. In his opinion, a painting should serve as a means to convey not only the appearance of an object but express how the artist looks at it.
This technique can be applied not only to portraits but also to animals and landscapes. Take wu Guanzhong, one of my favorite painters, as an example. The object of his painting is his hometown, but he only uses a few strokes to outline what his hometown looks like in spring. He does not paint it as it is in real life; only its essence has shown as a result of the artist’s long-term observation and profound understanding of the subjects.
So I’m wondering, can AI deep learn this Xieyi Style in Chinese ink painting?
My Chinese ink painting model
On the training model, Due to the limited materials, I only used less than 300 pieces in the end but used 3000 steps. The training of the model took three hours.
In this training, there were many shortcomings. For example, I needed to distinguish the Gongbi Style painting technique from the Xieyi Style by myself. This step of finding materials consumed a lot of time.
I can’t judge whether this transformation can be called successful, because the human face has undergone a transformation similar to Chinese ink painting, but it still lacks the core soul of Xieyi style.
There may be a few conveniences that I can improve in the future: for example, many of the paintings this time have text next to them, which may cause problems with the machine. Another example is that some of the drawings have yellow paper backgrounds instead of white, which may also affect the machine’s learning. Whatsmore is that my model contains landscapes, plants, animals, people, etc., which are too complex and could interfere with deep learning.
Still, whether the artistic creation of artificial intelligence has a sense of self
My model is not perfect, but there are already people in the world who can produce “perfect” Chinese ink paintings. Here’s the link： https://www.wallpaper.com/art/victor-wong-ai-gemini-making-moments-lane-crawford
But to me, even if the AI creations look just like humans, I still “feel” that such a system is not intelligent enough for humans. Just like bees unconsciously make beautiful trajectories in the air and roses unconsciously open their eye-catching flowers, these behaviors may be “beautiful” to humans, but I don’t think bees and roses are artists.
I hope that the creator of a true artistic creation possesses emotions, not just expresses them; I hope that the creator is self-conscious and knows exactly what he is doing, not just unconsciously simulating an artistic act (although there may already be emotions present in such an act).
From this point of view, we come back to the general philosophical question that all general AI systems face: do AI systems possess or are they capable of possessing, self-awareness?
— — — — — — —
References：
[1] Elgammal A, Liu B, Elhoseiny M, et al. CAN: Creative Adversarial Networks, Generating “Art” by Learning About Styles and Deviating from Style Norms[J]. arXiv: Artificial Intelligence, 2017.
[2]Artificial Intelligence: A Modern Approach Chapter 1, https://people.eecs.berkeley.edu/~russell/intro.html.
[3]陶锋. 人工智能美学如何可能[J]. 文艺争鸣, 2018, 5: 80–85.
[4] Gan training: RunwayML
