Study Finds Generative AI Can Imitate Famous Artworks Using Just 200 Images

Recent research has uncovered that generative AI has the capability to create believable imitations of celebrated artworks, demonstrating the technology’s efficiency and raising pressing questions regarding copyright laws. A report from Fast Company detailed how AI systems can reproduce copyrighted images, with findings revealing that the number of images required for training an AI model to replicate a classic artwork can be surprisingly low. Specifically, it can range from as few as 200 to up to 600 images, depending on the complexity of the subject being depicted.

This groundbreaking study tracked the performance of three versions of the Stable Diffusion model, which is widely used in generative AI image generation. Researchers established an “imitation threshold” using an algorithm designed to discern whether a generated image could convincingly be identified as an imitation. This threshold was then corroborated through human evaluation, showing a notable alignment between computer assessments and human judgments.

The research, titled “How Many Van Goghs Does It Take to Van Gogh? Finding the Imitation Threshold,” was published on the preprint server arXiv last month by a team led by Sahil Verma, a computer science PhD student at the University of Washington. Verma commented on the surprising nature of the findings, explaining, “Some people are surprised that it’s such a low number, and some people are also surprised that it’s a high number.”

Particularly fascinating is the detail that replicating the distinct brushstrokes of iconic artists like Vincent Van Gogh can be achieved with as few as 112 images, while producing human likenesses can be done with only 234 images. This efficiency raises serious considerations regarding the implications for copyright infringement as AI technologies become more powerful and ubiquitous.

As the landscape of digital art and its legal underpinnings shifts, the study reflects a growing concern within the art community. Verma noted, “As we were working on it [the study], we realized this has a huge implication for privacy and copyright issues.” This revelation comes at a time when there has been a surge in copyright litigation concerning AI-generated content in the U.S., emphasizing the urgent need to navigate the intersection of technology and intellectual property rights.

In August, a notable ruling allowed a group of artists to proceed with their copyright claims against several AI image generators, including Stability AI, the creator behind Stable Diffusion. The lawsuit, filed by ten artists, claims these companies used their works to develop AI art tools without obtaining the necessary permissions, highlighting a critical area of contention as the technology evolves.

As generative AI continues to push the boundaries of creativity and art replication, the stakes for artists and their rights are set to rise. With the advent of such sophisticated techniques revealing that just a handful of images may suffice for imitation, the conversation surrounding copyrights and AI’s role in the creative landscape is more urgent than ever.