When it comes to artistic creation, is AI a friend or foe?

Arts & Design

Published: 26 Feb. 2023, 13:30 Updated: 26 Feb. 2023, 16:19

When it comes to artistic creation, is AI a friend or foe?

A text-to-image AI Dall-E2 generated image with a prompt: Edward Hopper-style image of a man and a woman sitting at a cafe table, both absorbed in their smartphones without speaking. [MOON SO-YOUNG]

“A man and a woman sitting at a cafe table, both absorbed in their smartphones without speaking, in Edward Hopper style.”

Upon entering the command in English to text-to-image generation website Dall-E2, four images were created in just over a minute. While evoking the wide expanses of color, spaciousness and a lonely atmosphere of the British painter’s iconic work “Nighthawks,” the resulting image was more than a mere imitation but a unique artwork.

It was a moment that confirmed the ongoing concern that AI is posing a threat to the field of artistic creation.

Will an AI be smart enough to know the Queen Seondeok of Korea’s Silla Dynasty (57 B.C. to A.D. 935) and generate her image correctly?

Midjourney generated image of Queen Seondoek of Silla in Alphonse Mucha style [MOON SO-YOUNG]

“Queen Seondeok Silla, an ancient kingdom of Korea, wearing a red hanbok [Korean traditional dress] and a golden crown of Silla, in an Art Nouveau painter Alphonse Mucha style.”

This time, this reporter utilized Midjourney, another text-to-image generation app similar to OpenAI’s Dall-E2. A faint image started to appear as soon as the enter key was pressed. Within 40 seconds, four complete drawings were generated, depicting intricate details such as facial expressions, hair and even the big folds in drapery. The quality of the images was so high that they could be used immediately as illustrations for a fantasy web novel.

The emergence of ChatGPT, which can write speeches for politicians or solve medical exam questions with ease, has sparked a global phenomenon, with over 100 million users utilizing its services. Different AI programs that challenge human activities have been emerging one after another in recent years.

“AI has reached a stage where it seems to have its own sensibility by predicting elements that humans like and are moved by, based on vast data learning to create text and images,” said Yiyun Kang, a media artist and an assistant professor at KAIST. ChatGPT may still be in its early stages, but it is already writing poetry and novels that touch on human emotions. There’s even AI that composes music.

The paintings created by AI are ultimately a collection of imitations based on data learning. However, even among creative artists, there are those who use AI or collaborate with it in their creative activities.

“The world where AI does everything has arrived.”

This is the title of a YouTube video uploaded by a famous Korean webtoon artist Joo Ho-min last October. As AI that can draw pictures from text input surpasses the realm of creative activities once considered unique to humans, Joo initiated a live discussion with his viewers to talk about how much of a threat this poses to creators such as webtoon artists. Video content related to such issues has been rapidly increasing on social media over the past few months.

The big three AI image generators — Dall-E2 and Midjourney from the U.S. and Stable Diffusion from the UK — made a stir as they were introduced to the public one after another last summer. In particular, webtoon and game industry insiders in Korea were shocked by the emergence of Noble AI, which produces high-quality Japanese anime-style images based on Stable Diffusion technology.

Is the quality that high to shock the professionals?

According to the official website of Dall-E2, the AI developed by OpenAI, a U.S.-based research institute that also created the conversational AI service ChatGPT, is named after the surrealist artist Salvador Dali and Pixar’s Wall-E. Dall-E uses the underlying AI language model GPT-3, which was trained on a massive amount of text data to learn how to use language in various ways and applies this knowledge to generate images. It was first released in early 2021 and an upgraded version has been available to the public since the second half of 2022.

As for the quality of the images produced by the AI, it’s definitely high enough to threaten professionals in the webtoon and game industry. Dall-E can be easily accessed if you already have a Google ID. In the first month, users are provided with 50 credits for free, which allows them to generate 50 images. After that, 15 credits will be provided every month for free. If users wish to use more, they can purchase additional credits. To generate an image, users can simply enter a prompt describing the desired image in the input field and press the “generate” button. As the AI on the website does not understand Korean, prompts should be written in English to generate images.

Will the AI be able to generate editorial cartoons for newspapers? Let’s give it a try.

An illustration of members of parliament generated by Dall-E2 [MOON SO-YOUNG]

Upon entering “Illustrations of bickering members of parliament” and clicking on the generate button, it took about a minute for four illustrations in the style of editorial cartoons to appear. While the details such as the hand gestures were a bit exaggerated, the overall atmosphere was well captured and could be used in editorial pages or for political commentary videos. If you don’t like any of the images, you can click on the generate button once again to get a new set.

Images of Geobukseon (turtle ship) that turned into a mechanical and futuristic spaceship generated by Midjourney [MOON SO-YOUNG]

This reporter decided to challenge the AI and entered prompts such as “A cute rabbit wearing hanbok” and “A geobukseon [turtle ship] that turned into a mechanical and futuristic spaceship.”

Dall-E2-generated image of hanbok-wearing rabbits [MOON SO-YOUNG]

The hanbok-wearing rabbit prompt generated relatively satisfying images, including photographs and 3-D rendering images. However, the images generated for the spaceship-turned-geobukseon prompt did not reflect the real appearance of the geobukseon, or the Korean warships otherwise known as turtle ship, that are well-known to all Koreans.

Lee Tak-yeon, an assistant professor of Industrial Design at KAIST, who also leads an AI-Experience Lab, explained that this may be due to the AI not having any data on Korea’s geobukseon or due to an inaccurate description of the prompt. One user also commented that the text-to-image AI programs seem to lack “data training on Asian culture.”

The most high-quality images were generated for the following two prompts: “A modern city covered in volcanic ash clouds after a volcanic explosion nearby in oriental painting style” and “A man and a woman sitting at a cafe table, both absorbed in their smartphones without speaking in Edward Hopper style.”

But not all AIs showed the same competence for the same prompts.

When the result for the latter prompt on Dall-E2 was jaw-dropping, the same prompt on Midjourney generated a one-dimensional composite image of a man and a woman holding smartphones.

When the same prompt to generate an image of a “geobukseon turned into a spaceship” was asked on Midjourney, it only generated four images that looked like a turtle with a submarine-like structure — far from anything that resembled a spaceship or geobukseon. Despite adding detailed descriptions of a turtle ship, such as a ship with “sharp iron spikes protruding from hexagonal plates covering the top of the ship,” or “a dragon head on the bow of the vessel,” the AI kept generating increasingly bizarre images, clearly showcasing it did not understand what geobukseon was. One notable feature of Midjourney was, however, that even the strangest images it generated had sophistication.

One of the defining characteristics of Midjourney is its ability to create illustrations that resemble those found in science fiction and fantasy games. In August, an image created by Midjourney won first place in the digital art competition at the Colorado State Fair, sparking a debate whether AI-generated art can be considered a form of creative work.

Video artist Jason Allen's Midjourney generated piece called “Théâtre D’opéra Spatial” won first place at the Colorado State Fair. [JASON ALLEN]

The quality of the image generated by the AI exceeded the expectations of the people and had the potential to significantly disrupt the illustration industry.

“Illustrators who only perform relatively simple tasks are at high risk of losing their jobs due to such AI but at the same time, they have great potential to utilize AI to do better,” said Prof. Lee. “They can use the images generated by AI as assets and add their final touch, which can help them complete five to 10 projects a day compared to doing one a day.”

Webtoon artist Joo also said in the video he posted that he doesn’t think creators of webtoons will be threatened by such AI.

“However, it is highly likely that assistant illustrators who help webtoon artists to draw backgrounds and color are likely to be replaced by such AI,” he said as he introduced an “out painting” function in Dall-E2.

This function allows the user to upload a painting, such as the 17th-century masterpiece “Girl with a Pearl Earring,” with an input text describing a scene where the girl is standing in a kitchen outside the original painting. The AI will then generate a new image that shows the entire figure of the girl standing in the kitchen with similar shading and texture to the original.

Professor Lee explained that the activities of content creators such as authors, educators, YouTubers and web novelists are expected to become much more abundant. In the past, even simple illustrations for content had to be commissioned from illustrators or purchased one by one from image stock companies, but now they can be created with AI. Of course, this will affect professional illustrators, but they should actively utilize AI to increase the quality and quantity of their works, or focus on creating works tailored to clients’ detailed demands, which AI is not yet fully capable of.

Professor Lee has been insisting through his research that AI can be a partner to human creators rather than a competitor.

“In our lab, we have invited six beginner illustrators and four hobby illustrators and four professionals and asked them to work with Dall-E2. We’ve been studying whether they perceive the image-generating AI as a colleague, assistant or just a tool,” said Lee. “The professional illustrators tend to regard Dall-E2 as a search engine that provides reference materials and they really like it when the AI creates images that they had not thought of as they will use them as material to create their final works,” Lee added, explaining that this may be the way for AI and professional illustrators to ultimately coexist, allowing humans to preserve their creativity.

Collaboration between humans and AI, however, can raise copyright issues.

Kathleen Kim, a New York-qualified attorney who specializes in art law, said, “Only humans can be the subject of creation under the copyright law, therefore, AI is unlikely to be a copyright holder and may not be protected as a copyrighted work. However, if a human artist is the subject of creative activity and simply uses AI’s capabilities as a tool, the copyright for the resulting work may belong to the artist. Therefore, it can be determined whether the work is protected as a copyrighted work depending on the extent of human intervention and AI intervention.”

However, Kim added that AI has a degree of agency that is incomparable to past computer programs, so the legal profession is struggling with how to accommodate it.

“It is safe to say that we’re heading into a completely new era,” Kim said.

BY MOON SO-YOUNG [moon.soyoung@joongang.co.kr]