In February, a set of AI-generated cosplay photos gained immense popularity on the internet for their impressive quality and accurate representation of characters. The creator, Kanyon Industries, received widespread attention for their use of AI technology in art.
Using a more natural shadow handling
— Kanyon Industries (@auditore_k) February 15, 2023
Do you smell the sea？#アズールレーン #AzurLane #碧蓝航线 #柴郡 #チェシャー #AI #stablediffusion pic.twitter.com/0gFqmCUyhc
The photos quickly garnered millions of views across various social media platforms, with some viewers mistaking the AI-generated images for real-life photos. Furthermore, the administrator of Bilibili initially flagged the work for being in the wrong submission category, as they were unaware that the images were not actual cosplay photos.
Screenshot of Bilibili’s message of “wrong submission category” notice.
Despite the excitement around AI-generated art, there are concerns about its impact on traditional cosplay enthusiasts and the potential for unauthorized use of AI-generated images. To address these issues, we spoke with Kanyon Industries about their thoughts on the matter.
Kanyon Industries did not start posting AI-generated art until early 2023, but they had been keeping an eye on AI drawings since the release of DALL-E 2. Kanyon Industries believes that AI technologies would greatly change workflows in many aspects, but the rapid development of the tech still surprised them. Instead of jumping into the trend immediately, they decided to observe for a while. "What now takes me a lot of time to achieve may become easily achievable in a few months," said Kanyon Industries. When the LoRA model was released, they knew it was the right time to start using AI-generated art.
LoRA, one of the AI model training methods, can easily and quickly fix details such as art style, characters, and actions.
For Kanyon Industries, experimenting on “AI Cosplay” was not just for fun, as they believes it was a technical challenge to reveal the potential of the AI drawing. Each target character has different features, some of which do not even exist in reality.
Describing the workflow of their art, they said, "I must train separate LoRA models for each character, and the effect of each model varies depending on the training conditions. Basically, adjusting parameters from scratch is required when switching to a different model."
When asked about the debate on whether "AI would replacing Humans," Kanyon Industries explained that it’s not their concern, stating, "AI won’t replace anybody, only humans can replace humans." In their mind, they would compare AI drawing as a more “advanced form of Photoshop”, a helpful tool that will significantly improve artists' workflows in the future, and not much more.
Citing a study by Harvard Business Review that surveyed over 1000 companies across 12 industries, Kanyon Industries reinforce their belief that improved collaboration between humans and artificial intelligence leads to higher performance improvements.
The chart in HBR’s research shows a higher number of collaborations results in better performance improvement.
Kanyon Industries acknowledge that, at the time, there are limits to what could be achieved with AI-generated art. Each of their works was composed of nearly a hundred images, and some still suffered from imperfections such as the "bad finger" problem.
According to Kanyon Industries, there are three critical areas where AI drawing technology needs to improve: fundamental performance, operability, and AI model training effectiveness.
They explain that fundamental performance is reliant on the development of the algorithm, which is the underlying principle of AI-generated images. A more efficient and effective algorithm can significantly enhance output quality and productivity.
Operability refers to the ability to fully control the AI and steer it in the right direction, similar to how a steering wheel controls a car. Without good operability, AI-generated art may only produce random images and cannot be integrated into industrial processes. Several control methods, including bone recognition, edge detection, and depth detection, have been developed based on ControlNet.
An example picture shows the potential operability of ControlNet
Training models primarily use the Seg plugin to bind semantics to color values and directly construct compositions in the image to specify different elements in different areas.
An example picture shared by Kanyon Industries about how the Seg plugin helped in AI drawing.
Kanyon Industries was confident in their stance on the controversies surrounding AI in legal, ethical, and copyright issues. They firmly believed that AI is just a tool, and that the responsibility for its use should fall on the humans behind it. They recognized that countries have comprehensive regulations in place for various crimes, such as the infringement of portrait rights in China when using AI to replace faces, which is covered in the "Management Regulations for Deep Synthesis of Internet Information Services" issued by the Chinese government in 2022.
Looking ahead, Kanyon Industries sees a bright future with AI. They envision the technology being widely used in the novel industry to provide visually stunning reading experiences with cheap and high-quality illustrations. In addition, they believe AI can optimize workflows in the animation and comic industry by assisting in making original sketches and editing. Moreover, AI can assist in the production of various artistic assets and automatic modeling in the gaming and film industry, enabling efficient and cost-effective output. This will speed up the development of games and movies, leading to more cultural and entertainment products that enrich people's lives.
Kanyon Industries hopes AI will enhance human creativity, resulting in more high-quality cultural products and enabling individuals to focus on their personal interests and creativity.