AI Can Reproduce Popular Artworks When Trained on Just 200 Images
Generative AI can produce credible imitations of popular artworks when trained on as few as 200 images.
According to a report by Fast Company, new research has revealed how easy it is for AI technology to reproduce copyrighted images.
The number of images an AI model needs in its training data to reproduce a classic artwork varies depending on the system, typically ranging from just 200 to 600 images.
This requirement also depends on the subject the AI is aiming to depict. For example, capturing the brushstrokes of Vincent Van Gogh may require as few as 112 images, while replicating human faces can be done with as little as 234 images.
Fast Company reports that three versions of the Stable Diffusion model were examined by researchers to assess their ability to generate images resembling original works.
An “imitation threshold” was set using an algorithm that determined if a computer system could identify an image as an imitation. These algorithmic findings were then validated through human evaluation, revealing a strong alignment between computer and human judgments.
The findings were revealed in a study titled “How Many Van Goghs Does It Take to Van Gogh? Finding the Imitation Threshold” that was published in Cornell University’s preprint server arXi last month.
“Some people are surprised that it’s such a low number, and some people are also surprised that it’s a high number,” Sahil Verma, lead author of the study and a computer science PhD at the University of Washington, tells Fast Company.
The team’s research potentially demonstrates how AI programs might infringe copyright by generating outputs that resemble existing works.
“As we were working on it [the study], we realized this has a huge implication for the privacy and copyright stuff,” Verma adds.
The study comes as a flurry of copyright cases about AI were filed in the U.S. in the last year.
In August, a judge granted a group of artists the right to continue their copyright claims against four AI image generator companies. The makers of Stable Diffusion, Stability AI, are being sued along with Midjourney, Runway, and DeviantArt by 10 artists (including one photographer) for using their work to build AI image products without permission.
Image credits: Header photo licensed via Depositphotos.
