At Sunday brunch last weekend, my friend told me about trying to use Midjourney to generate art for her newly-furnished guest room. To match the aesthetic of the new furniture, she wanted to emulate the style of Asher Brown Durand, an 18th century artist of dramatic landscapes.
Here’s blueberry and ricotta pancakes, the style of Asher Durand
This led us to the question:Is it legal to use a prompt "in the style of X," where X is a specific artist or artwork?
For Durand, the answer is yes, this prompt does not infringe copyright, because his works are in the public domain. In the United States, copyright protections end 70 years after the author’s death (for the most part). Durand painted in the early 19th century, so his paintings are now in the public domain. As a result, it is legal to reproduce and emulate these styles, and most (all?) image generators will allow you to generate in the style of that artist. My friend still had some troubles – a fairy in the foreground of the image had three wings – but said that the outputs she wanted are pretty close to something she’ll hang on the wall.
It’s easier to prompt an AI model if we can refer to reference works because it’s often easier to name an artist than describe their common works in natural language. For my friend, she liked the bucolic, idyllic scenery and the dreamlike lighting of Durand’s photos that she’d seen on display in a museum and wanted to incorporate this into her home art collection.
There are many use cases for why you would want to generate multiple images in a similar visual style. Generating in the style of a single artist is also helpful for maintaining visual consistency. Let’s say you have a series of articles, podcast episodes, or videos for which you want to generate imagery that goes together and looks cohesive. You might want a generator to produce works in one visual style, even when portraying very different subjects. Book illustrators, video game artists, animators, and character drawers may also find this stylistic cohesion compelling and valuable.
At my first company, Kapwing, we ran into this problem when trying to generate images in the style of our in-house illustrator. She wanted to move on to different, more complex projects rather than spending time illustrating blog posts for our team. But we wanted to maintain the aesthetic of our website. Which led us to the question: is it possible to generate images in the style of a single artist? We wanted to generate cover art in the style of Emily’s branded illustrations.
Our platform, Tess, is an AI image generator that generates images in a specific visual style. A model is trained on 10-20 reference images, and every produced image appears in their style.
Tess is properly licensed, meaning that every Tess model is trained with permission from the artist, and they get paid every time it is used. Tess is the first properly-licensed generator that ensures you can apply a consistent visual style to generated images with copyrights ensured by the artist.
Here’s a few images generated by Tess, in the style of Kemal Sanli:
If you don’t have a license to Tess, you can generate in a specific visual style on some AI platforms. DALL-E 3, OpenAI’s image generation platform, uses prompt transformation to block users from producing works in the style of a living artist. Dall-e will generate an image, but using a prompt that tries to describe the artist’s style in words.
Below are some example of what happens when you prompt AI Image Generators in the style of a living artist. In these examples, I've referred to Shepard Fairey, a famous artist who is alive, and his works are protected by copyright.
When given a prompt in the style of, Dall-e transforms my initial "in the style of" prompt to a different prompt (“inspired by the striking and bold aesthetic of early 20th-century political propaganda art, blending festive themes with a powerful and assertive style"). I'm guessing that OpenAI queries Fairey’s most famous works and generates a description for it in words, then projects that caption to transform the initial prompt. This is a roundabout way of blocking stylist generation.
As ChatGPT explained: "Creating images in the style of specific contemporary artists, like Shepard Fairey, falls under the guidelines that protect intellectual property and respect the rights of artists."
As of April 2024, Dall-e prevented users from generating images of copyrighted like "Kim Possible."
Meta recently launched Llama3, a new generative model that creates images and short animations. If you prompt for the name of a living artist, Meta will either 1) not generate in the artist's style or 2) respond with an error "Oops! I can't generate that image." with no explanation.
Midjourney will produce these images with a more varied and representative style, drawing from many of the artist's works:
However, the artist was not consulted when these styles were constructed, and there are ongoing lawsuits related to the legality of Midjourney's product.
Google’s Gemini has been struggling with creating images, but it seems that in their first version, Gemini also will produce images in the style of a living artist:
Adobe's Firefly will block stylistic prompts, although it will allow you to upload a reference image when generating. Adobe has argued that impersonating an artist’s style should be illegal and has proposed legislation for a mechanism for artists to sue violators.
Copyright law protects original works of authorship, including paintings, photographs, and sculptures, among others. However, the concept of an "artistic style" falls into a gray area. While individual artworks are protected, an artist's style—being a broader, more abstract concept—typically isn't. The reason this is true is that courts want to protect creative freedom and expression, allowing artists to incorporate influence from various works they’ve studied and learned from, without fear of retaliation from past artists.
When an image generator creates content "in the style of" an artist, it does not replicate a copyrighted work directly. Instead, it synthesizes new images that evoke the artist's unique techniques or visual aesthetics, in the same way that a human artist would.
In a recent landmark court case Andersen et. al., v. Stability (2023) in California, a judge dismissed most copyright claims against AI Generators, saying that prompting an AI image generator with specific references to names does not "produce an “AI-generated image similar enough that people familiar with Plaintiffs’ artistic style could believe that Plaintiffs created the image.”
While using a prompt "in the style of" an artist is not outright illegal, it navigates a complex web of copyright laws, ethical considerations, and evolving norms that can lead to infringing activity. A stylistic prompt risks undervaluing the artist’s work if you rip off art that they could have produced themselves without paying them, asking them, or crediting them.
Generating art work in an artist’s visual style is risky legal territory and comes with liability. From a copyright perspective, prompting an AI to generate in a certain style has a high risk of violating copyright projections because the machine is likely to reproduce copyrighted elements, especially if the machine was trained on a small number of works from that author.
Some have argued that we need new laws to regulate copyright of AI Generators. After all, unlike humans, AI can mimic an artist's style in seconds, not hours or days and years of stylistic study. Given that copyright laws were designed in a pre-digital era, they are often ill-equipped to address the nuances of AI creations.
The only way to ensure that you’re not committing copyright infringement is to use a copyright-safe image generation platform like Tess which is trained with the artist’s permission. Technologists must also stay informed about new and evolving court cases and legislation involving AI generation and copyright.