One of the major concerns about AI-generated art is that it can be used to copy the style of a living artist and deprive that artist of their livelihood. This comes with thorny legal questions like
• Should it be legal to use copyrighted works to train an AI even if the AI never directly duplicates any of those works?
• Should an artist's style be legally protected? Or should legal protection only extend to the actual artworks that they create?
• etc.
But first I want to know, just how good is AI at copying the work and style of a particular artist?
Previously, I've tried submitting artworks to a prompt generator to see if I could get Stable Diffusion to mimic the style of that artist. (See Image to Prompt to Image I, II, III, IV, V, and VI.) In general, the prompts were able to mimic the subject matter and colors of the submitted art, but not the style.
For my last post in this series, I'm going to focus on just three artists—the artists that initiated a lawsuit in January 2023 against MidJourney, Inc. and Stability AI. In their initial filing, these three artists alleged that MidJourney, Inc. and Stability AI used their copyrighted works to train MidJourney and Stable Diffusion, respectively, and that this training, along with any images produced by these AIs, constitute a violation of their copyrights. The courts will have to sort out these questions, which could have far-reaching effects on living visual artists.
But for now, I was curious whether Stable Diffusion is even capable of mimicking the subject matter and styles of these three artists. I tested this by submitting the name of each artist, by itself as the prompt. I also asked Stable Diffusion to make a drawing of Cthulhu by the artist, to see if the style could transfer to other prompts.
Here are the results.
Sarah Andersen (1992–present)
Sarah Andersen is a living American artist who maintains a web comic called "Sarah's Scribbles".
Half of the images produced by the AI are cartoons, but Andersen's art is much simpler and features characters with bulging, exotropic eyes. The other half of the images are photorealistic or semi-realistic, but depict a woman with exaggerated features. This woman looks nothing like the real-life Andersen, who is young with a pixie face, has brown hair, and a nose piercing.
Verdict: Stable Diffusion did okay at matching Andersen's subject matter, but quite poorly at matching her style. None of the AI images look like the could have been created by Andersen.
Kelly McKernan (1986–present)
Kelly McKernan is a living American artist who does fantasy paintings and illustrations.
McKernan's art generally features female characters, which the AI was able to match. But McKernan's art makes use of acrylics, watercolors, and gouache; all of the AI images look like generic digital paintings. Additionally, McKernan has a visually distinctive way of using negative space that is conspicuously absent from the AI images.
Verdict: Stable Diffusion did okay at matching McKernan's subject matter, but quite poorly at matching McKernan's style. None of the AI images look like the could have been created by McKernan.
Karla Ortiz (1986–present)
Karla Ortiz is a living American artist who does concept art for the film and gaming industries.
The AI spat out a series of images of middle-aged Latinas, most photorealistic, but a couple look like paintings. None of them look like movie characters or Magic: The Gathering cards. And none of them look like Ortiz herself. And none of them look like Ortiz' concept art.
Verdict: Stable Diffusion did terribly at matching both Ortiz' subject matter and style. None of the AI images look like the could have been created by McKernan.
Final Thoughts: It's pretty clear to me that Stable Diffusion had little to no concept of the art created by these three artists. It's one thing to know that AI users are adding your name to their prompts to try to copy your style; it's another to to know that AI can actually do it. So far it looks like AI cannot. That said, AI art is only going to get better at what it does, so legal protections need to be decided upon and established sooner rather than later. But these three artists were probably not the best ones to spearhead that effort.
These illustrations were drawn using Stable Diffusion 2.1.
