This year is a year of outstanding achievements in AI. Write articles, make videos, play music, and be a painter. You give the information and it really learns.
This year is also a year when many art practitioners are trembling. Humans' double fists or four hands are no match for the rapid progress of AI.
Now, the pressure is on the cartoonist side, and the two-dimensional people quit.
AI imitates dead artist, causing public outrage
On October 3, Kim Jung-ki, a famous Korean cartoonist and illustrator, passed away unexpectedly. He was called a "human flesh printer" because he did not draw on the scene and drew huge and complex scenes directly. He once painted "Marvel Heroes Brawl" for Marvel. ".
▲ Works by Kim Jung Ki.
Today, when computer painting is popular, Kim Jung-ki still adheres to the traditional hand-painting method:
I have always insisted on hand-painting, and I am more proficient in freehand creation. I believe that using my own hands and pen can better express the taste of the work. If I have time, I would like to learn computer painting and try more painting methods, but if I rely on computers, my works will have a standardized feeling, so I am somewhat evasive.
Then on October 7th, Twitter user @BG_5you, who is keen on procedural generation and AI art, fed his work to the AI model Stable Diffusion, ironically filling it up.
@BG_5you just googled Kim Jung Ki's work without any permission, as we're used to.
▲ AI generative art by @BG_5you.
What is the final result? At first glance, it seems quite realistic, but a closer look is a bit sloppy.
Unsurprisingly, @BG_5you has a lot of opponents, with cartoonist @Dave Scheidt writing:
It's been less than a week since Kim Jung Ki left us, and the AI brothers are already "copying" his style and hoping for the honor. This is a vulture, and a loser without backbone and talent.
Of course @BG_5you also saw the criticism from netizens. He responded that he likes AI technology, but even if AI-generated art is perfect, it cannot replace the artist's thinking and painting skills. It's just a new way of exploring the artist's style, a toy shared with the masses that doesn't need to rival a lifetime's worth of talent.
▲ Kim Jung Ki.
It seems that he has lowered his posture to the dust, and has not forgotten to pay tribute to the artist, but the words are scornful.
Kim Jung-ki himself is a talented and hard-working painter. He painted 12 hours a day in high school, and he averaged 10 hours a day after he became famous. Some netizens pointed out that treating other people's art as "just a toy" is the problem: "His work has been distorted, and children in five years will see this shit, not his real work."
Another cartoonist, @Kori Michele, couldn't help but refute:
An artist is not just a "style". They are not products, but people who breathe and experience, learn, feel and grow, and create art. @BG_5you's AI model represents the worst way to look at an artist.
The AI generator is not a tribute, but a theft of Kim Jung Ki's work, such views flooded the comment section that @BG_5you even claimed to have received death threats from Kim Jung Ki's loyal fans.
▲ AI generative art by @BG_5you.
He didn't win, but the AI model Stable Diffusion was out again. Stability AI, the company behind Stable Diffusion, raised $101 million in mid-October at a valuation of about $1 billion.
Competitors like Stable Diffusion and Dall-E are a little different: it's open source, all code is publicly available on GitHub, and anyone can use it, provided it follows the original project's CreativeML Open RAIL-M license .
You can use it any way you like, from performing text-to-image tasks, to developing stand-alone applications or services.
This means that, without the need for beta invites or subscriptions , engineers can train models on any image dataset to produce the desired art style, except for content involving crime, defamation, harassment, and more.
Stability AI consultant Aemish Shah once said: "We are very happy to support the democratization of AI." The attributes of open source not only make AI models no longer a game for a few people, but also the basis for the explosion of AI art.
But for artists, its popularity may not be good news, and being infringed is an unavoidable problem.
▲ The output of Stable Diffusion under the same topic and different prompt words. Image from: wiki
Since the artistic style and composition are not copyrighted, images generated using Stable Diffusion are often not considered copyright infringements of "visually similar works" . However, if a real person or brand identity is involved, then personality rights or intellectual property rights are still protected .
According to the Stability agreement, the copyright of the pictures generated by Stable Diffusion belongs to the creators' economic collective , not to any individual or company, however, you can use the pictures you generate for commercial or non-commercial use. The anger of Kim Jung-ki's fans may be more empathetic.
"AI Learning Prohibition"
It is Japan, where the comics and animation industries are particularly prosperous, that makes the controversy over AI comics really come to the forefront.
At the end of August this year, Japanese AI startup Radius5 launched a beta version of the AI-generated art tool "Mimic".
The content generated by Mimic is dominated by two-dimensional avatar illustrations. Its original intention is that artists upload their own works, generate works with their own styles, or use them as communication tools for social media and fan communities . Provide Mimic with more than 30 illustrations, and you can train your own drawing AI.
In order to give Mimic a good start, Radius5 also recruited five artists to experience, but the next thing is not as expected.
These artists are seen as accomplices of AI – their support for AI may lead to the theft of other artists' works, after all, "this kind of thing can't be abused."
When several artists came under fire, Radius5 issued a statement strongly opposing the practice. Not only that, but a few hours after the announcement, Radius5 froze the beta version and deleted all images, less than a day and a half before it went live.
▲ Mimic's "original intention".
It turns out that the most worrisome thing is indeed happening, a large number of users are simply uploading works by other artists that they do not have permission to use. Although this behavior violates Mimic's Terms of Service, Radius5 is powerless to prevent it.
What makes the artist feel that the platform "has a heart to punish" is that Radius5 stated before the launch of Mimic that the ownership of new works belongs to the creator , and the ownership dispute is thrown to the user.
The surface of the water seems to have stopped temporarily, and the undercurrent is still surging. The phrase "AI learning prohibited" spread on Japanese Twitter, and artists who marked "unauthorized use" in their social media profiles also began to add "AI learning prohibited" ”, and some artists even deleted all their works on social media to avoid being regarded as a ladder on the road to AI progress.
▲ Picture from: [email protected]_umi
Contrary to the outrage are the loopholes in Japanese law regarding copyright and data scraping.
For AI generative art, legal issues only arise when the output image is identical or very close to the image that the model was trained on. The other way around, at least for now, is that AI-generated art is legally feasible as long as it is not identical to the input image.
This loose restriction is also a risk for AI-generated art tools such as Mimic, which may lead to infringement if only the images of a certain author are caught for training.
Some people are not optimistic. Kazuyasu Shiraishi, a partner of a Tokyo law firm, believes that it is unlikely to successfully defend against AI, because in 2018, Japan revised its national copyright law to allow machine learning models to be used without permission. To scrape copyrighted data from the internet, artists can only request that the work not be uploaded to some specific AI website.
To some extent, the legal status of AI and artists is not equal.
In short, the ambiguity of the law and the rapid development of AI art, like a cloud over the mind, make artists worry about their livelihood and the future.
However, when opposition to AI learning has become a trend, it can't help but feel a little stunned. For example, in order to prove that he has never used AI, Japanese artist @Yrui596 had to post screenshots of the software used .
On October 3, NovelAIDiffusion, the AI-generated art tool of AI company NovelAI, also lay down a gun and was rumored to "completely copy the anime illustrations found online."
NovelAI had to publish a blog explaining how NovelAIDiffusion works :
“Instead of copying and mixing existing images, our AI generates original images from scratch with the help of deep learning algorithms, and the AI has basically learned how to create images — just like a human being.”
NovelAIDiffusion is based on the aforementioned Stable Diffusion model, which was trained on about 2 billion images of the LAION dataset (~150TB), while the former uses a fine-tuned dataset consisting of about 5.3 million images and detailed text labels data composition.
The implication is that NovelAI Diffusion builds on Stable Diffusion's existing knowledge of anime.
Of course, NovelAI also emphasizes the original intention of AI-generated art tools: as a powerful creative tool, allowing artists to get inspiration, or understand what will happen when their own works are adjusted in style.
NovelAI explained that the AI model is not so stupid as to directly copy, but the AI's eagerness to learn is a certainty.
There are things that AI can't do
It may make people feel at ease. AI also has several mountains that cannot be climbed in the second dimension.
In arguing against AI models, cartoonist @Kori Michele noted:
I can't convince AI capitalists that this is not about taking our jobs, or making art more "effective" and "cheap" (those are things that happen). What is truly valuable is your experience of creating things that make the world a better and richer place – the unique fruits of your hands, your heart, your history, your community, and your future.
The so-called "I write my heart", the paintings themselves are a person's story, and the person who created them is a person who experienced something.
AI is not good at painting some paintings, but also because of less knowledge.
In the circle of Japanese painters, someone has summed up three major AI drawing problems: Sephiroth swimming, Higuchi Yenka eating ramen, and crying beautiful girl eating cake.
It sounds easy to imagine the scene, and AI painted them into: Sephiroth standing on the water, Higuchi Yenka grabbing ramen, and the beautiful girl turned into a cake.
▲ Picture from: togetter
The AI's misunderstanding is mainly due to insufficient learning materials or deviations in the materials themselves.
For example, grabbing ramen with hands is probably because the AI draws something that looks like noodles, but does not know that the noodles should be eaten with chopsticks, not only Higuchi, but also other characters that AI draws to eat noodles.
However, it is one thing that AI is not good at drawing. It does not mean that AI cannot draw at all. There are also a few successful examples. Although ramen and chopsticks are very abstract for AI, with more trials or fine-tuning, "survivor bias" appears.
At least for now, AI is still a good and diligent student. In order to achieve human expectations, it is inseparable from careful materials and patient guidance.
For example , AI-generated art platforms such as 6pen require users to "describe clearly and clearly, include the objects appearing in the picture and their information, and avoid words that are not related to the picture", and "use various keywords reasonably and skillfully, including style modification, screen type, etc.".
▲ AI Art Award.
The triumph and failure of AI in the second dimension also shows that the relationship between people and AI is changing subtly.
For things that are easy for AI, the human brain is difficult to solve. At this stage, AI can be regarded as a quick painter, but it is not so easy for humans to enter the threshold of art.
Conversely, what the human brain can do easily is not the case for AI. A few years ago, it was how to distinguish a dog or a cat, and now it is how to distinguish between chopsticks and ramen. For AI creation to be truly established, it is temporarily inseparable from the logic of human beings.
#Welcome to pay attention to the official WeChat account of Aifaner: Aifaner (WeChat: ifanr), more exciting content will be brought to you as soon as possible.