The game industry regards AIGC as an artifact, but more than half of the artists commented that it was "not helpful". Why?

Source: gamelook

GameLook report/In the era of big AIGC, from overseas Blizzard, Ubisoft, Unity to domestic industry leaders such as NetEase and Tencent, the global game industry has taken the lead and started a comprehensive exploration of the implementation of AI large-scale model technology.

Among them, Vincent diagram technology is the closest to popularization. Open source graphic tools such as Stable Diffusion have now become the right-hand man of game art workers. Companies such as Blizzard have also launched self-developed generative models to help developers create art more conveniently. The industry has spread its arms and generally accepted AIGC technology, but in some sister industries, the promotion of AIGC is still bumpy.

Artist research report: AI still can't do high-end routes

The fine art industry is the most representative example. Recently, Playform AI, an overseas generative AI art company, sent a questionnaire survey to 500 artists and digital designers. The results showed that more than half of the artists felt that they could not help them after trying DALL-E, Midjourney and other Vincent AI. to create.

Playform AI is positioned as an art generation platform, so most of the creators interviewed this time come from more traditional industry backgrounds such as fine arts, digital design, and photography. Also in the survey, only 18% of creators said they would use AI generation tools in their daily creation. In addition, 60% of artists believe that the image quality generated by AIGC tools does not meet their quality standards.

On the one hand, the game art world is in full swing, and on the other hand, the pure art world is relatively indifferent. From GameLook's point of view, the differences are quite interesting. On the surface, both game painters and digital artists are engaged in drawing pictures, but behind the scenes, the needs of the two industries for the final delivery products are quite different. For commercial game artists, the output of art resources that meet the needs of the project is the top priority. For artists, the final product they need is more "high-end". Not only the quality of meticulously crafted paintings, but also the artistic expression and style with personal touch are more important.

The current AI can generate exquisite pictures from massive original training materials. But in principle, AI is still at the level of collage of the elements in the picture, and it is difficult to form an "understanding" of the elements in the picture. Therefore, the paintings generated by AI often have problems with messy details, light and shadow relationships that cannot withstand scrutiny, and even distorted limbs. For industries that require higher detail quality, AI is undoubtedly unable to meet the needs of artists.

This lack of understanding naturally makes it very difficult to adjust skills such as picture composition. More often than not, all AI does is combine input text and parameters. Fundamentally speaking, AI painting is difficult to form ideological expression, so the artists interviewed this time also said that only 30% think that the products generated by AI can reflect the artist's own style.

In addition, artists have various criticisms of AI. For example, only 30% of artists think that there is no problem with the copyright of AI drawings, while more than half of the artists think that there are still copyright concerns. Artists are also dissatisfied that the current AI model is still difficult to precisely control the final output, and the resolution and detail control generated by AI cannot meet their needs.

Compared with the highly streamlined game industry, the fine craftsmanship market has not been greatly impacted by AIGC. However, artists also have a certain degree of identity crisis.

Another Dimension Beyond Business

AIGC has brought us a great improvement in productivity, but it is followed by endless anxiety. Remarks such as "Painters will be replaced!" were once popular, which is why practitioners in the art industry are more wary of AIGC.

While anxious about unemployment, instead of sitting still, we might as well take this opportunity to break the inherent cognition and reevaluate the competitive value of human beings. We can see that the current AIGC still has too many shortcomings and shortcomings. For example, it cannot think like a human. For artistic creation, this is a fatal injury, so excellent creativity and humanistic ideas will only become more precious in the era of AI.

Ahmed Elgammal, the founder of Playform AI, recently wrote an article "Why the era of AI art has passed", sharing his interesting thinking with us.

Ahmed Elgammal is an interdisciplinary talent with a unique resume - as the head of Rutgers University's Artificial Intelligence Art Lab, he has received both a solid art education and a profound understanding of artificial intelligence technology. In Ahmed Elgammal's view, instead of saying that the AI era is coming and replacing humans, it is better to say that the current AI model is too familiar and "imitates human thinking", which stifles the creativity that belongs to AI. Ahmed Elgammal believes that the era of "AI art" is effectively over.

GameLook compiled this great article by Ahmed Elgammal:

Everyone is talking about creative artificial intelligence and "AI art" now, about the arrival of a new era of creative AI that will take over the jobs of artists. We've seen a huge backlash from artists and the art world. Yet the opposite is true: the era of "AI art" may actually be over.

What happened? First, let me clarify what I mean by "AI art".

AI doesn't create art, it creates images. What makes these generated images art are the human artists behind the AI—those who feed data into the machine, manipulate knobs, and curate the output. Therefore, I use the term "AI art" to talk about human art that uses AI as part of the creative process, with varying degrees of autonomy. We are entering an era where such tools are heavily used. However, the days when these tools sparked artistic genius may be over.

What makes art spark? When Picasso created "The Maiden of Avignon" in 1907, the painting caused controversy and was opposed by friends close to him. Even Picasso's Cubist colleague George Braque didn't like the painting. It wasn't until 1939, when the painting was exhibited at the Museum of Modern Art in New York, that it was accepted and recognized by the public as a forerunner of Cubism. Writing on the Guardian's centenary, Jonathan Jones wrote: "Works of art eventually settle and become respectable. But, 100 years on, this Picasso is still so fresh, So disturbing, to call it a 'masterpiece' is an understatement."

Picture: Picasso's "Girl of Avignon"

The role of disturbing challenges in artistic development is well explained by the theory of aesthetic psychology pioneered by Colin Martindale in his 1990 book The Clockwork Muse. He believes that the main force behind the evolution of art is the artist's struggle against habit through innovation. However, if the artist innovates too much, their art will be too shocking and the audience will not like it. Good artists find that sweet spot between being innovative and not too shocking. Great artists are those who go further.

Can artificial intelligence go beyond "good" to "great"? When generative adversarial networks (GANs) emerged, some artists took notice of this new AI technique. You can train these models on lots of images and they can generate new images for you. When we trained a GAN on classic portraits in Western art in 2017, it produced some disturbingly deformed portraits that reminded me of Francis Bacon's 1963 portrait of Henreitta Moraes. However, there is one fundamental difference between the two: Bacon's intent was to distort portraits, whereas AI is simply disobedient in its generation.

Picture: Francis Bacon, "Three Studies for Portrait of Henrietta Moraes"

With the advent of GANs, we have entered the era of the "aesthetics of failure" for machines. Some critics associate it with "glitch art". In fact, the surprises brought about by GAN have made artists interested in it. Many in the field call this the "uncanny valley effect."

It’s this uncanny valley and serendipity that makes AI art interesting between 2017 and 2020. In 2019, I did a study with art historian Marian Mazzone in which we interviewed several artists who pioneered the use of artificial intelligence in their creative process. We found that "artists understand artificial intelligence as the main driver of their creative process". In particular, artists have found AI useful in two ways: creative inspiration and creative volume. Creative Inspiration is where artists find AI giving them new ideas, new directions, and new ways to make art.

Figure: A portrait of a person generated using GAN, made in 2007

Different from the current atmosphere of condemnation, artificial intelligence art has been welcomed by the art world from 2017 to 2020.

In October 2018, Christie's auctioned off a GAN-generated artificial intelligence portrait similar to the above-mentioned deformed portrait. In March 2019, Sotheby's auctioned off works by artist Mario Klingemann. HG Contemporary in Manhattan had an exhibition in February 2019 featuring my own work. In the summer of 2019, the Barbican Center in London exhibited the work of different artificial intelligence artists. AI art was welcomed at art fairs such as Scope Miami in 2018 and Scope New York in 2019. The National Museum of China in Beijing held a month-long AI art exhibition in November 2019, which attracted 1 million viewers.

During this period, AI art was actively reported in the media. The art market welcomes AI artists, and no one is calling for AI art to be banned. But what happened then?

Picture: Mario Klingemann portrait work "Memories of Passersby"

A fundamental difference between early AI models and today's large, cue-based models is that the early models were trained on smaller image sets. This allows artists to train their own AI models based on their own visual references. Today's large models are pre-trained on billions of images taken from the internet without the artist's consent. This brings up a host of copyright issues. This sprawling system erases the identity of the artist. The difference between my work and yours simply depends on which keywords we use in our prompts to guide the system. No wonder the Copyright Office refuses to grant copyright to such system-generated artwork. Capturing the identity of the artist was the main reason photography was able to obtain court copyright in the late 19th century.

Over the past few years, artificial intelligence has gotten better and better at generating high-quality and realistic images. It is also improving in its ability to mimic training data. A new way of interacting has been introduced, mainly using text prompts to control spawning. Today, text has become the main way artificial intelligence generates images. These advances in AI generation technology have made it possible for AI to do a good job of generating any image we want, be it a photograph or an illustration, following the instructions we utter in carefully crafted text prompts. Surprises are limited to the variations of ideas we might get. With multiple iterations, we can get the stunning high-fidelity, high-resolution images we want.

Text input helps the AI out of the "uncanny valley," but it kills surprise. This is because the models are trained on both text and images, and learn to associate visual concepts with linguistic semantics. This makes models better at creating characters and imitating styles that can be described in words.

Picture: Refik Anadol's "Unsupervised" collection in the Museum of Modern Art, New York

But on the other hand, including language as part of training makes the model very limited in creating inspired visual deformations. The visual output created by AI is now limited by our language, losing the freedom to manipulate pixels visually without being influenced by human semantics.

In a sense, artificial intelligence is becoming more and more like us, and it can no longer complement and challenge our way of seeing the world.

Of course, artificial intelligence still fails surprisingly in the generative process. We still see figures with four fingers and three legs. Such silly failures aren't necessarily funny, however. Creative inspiration isn't the only thing missing from this new generation of AI. The idea of using text to generate images can limit the inspiration of artists, because artists are visual thinkers. Using words to describe what they want adds an extra layer of unnatural linguistic abstraction.

AI is becoming a tool for generating massive images, not a co-creative partner that artists get excited about. AI is becoming very good at following the rules, but the artistic spark is missing. Artists have to dig deeper, go beyond literal prompts, and use artificial intelligence differently to find their artistic spark.