In February 2023, Getty Images filed a lawsuit against Stability AI, accusing the company of infringing on their intellectual property rights by using over 12 million copyrighted images as training data for their AI art generator, Stable Diffusion. This legal battle marks a crucial turning point in the ongoing struggle between AI startups and rights-holders.
In light of these allegations, we decided to conduct an investigation to uncover more instances of AI-generated watermarked images produced by Stable Diffusion.
Editor’s note
At Photutorial, we staunchly advocate for proper image licensing and the respect of intellectual property rights. We firmly condemn any misuse of intellectual property, including the use of watermarked images, without obtaining the necessary licenses. The purpose of this article is solely to inform and document our research on how Stable Diffusion has allegedly utilized unlicensed images from various platforms. Our intention is to raise awareness of the issue and contribute to the ongoing conversation around AI-generated content and intellectual property rights.
If you have feedback or questions, get in touch with us.
6 Types of watermarks we found
1. Getty Images
2. Depositphotos
The reproduction of the Depositphotos watermark was notably the simplest and most common occurrence. We were able to replicate it even on standard prompts that lacked specific keywords like "stock photo" or "Depositphotos".
3. Alamy
4. Dreamstime
Dreamstime watermarks are available in two distinct styles. When previewing an image, a diagonal grid featuring their logo appears, which is similar to the "unknown watermarks" shown below. Conversely, thumbnail previews display only the logo centered within the image, as demonstrated in the examples provided below:
5. iStock
The iStock watermark is less prominent in comparison to other watermarks and is more challenging to replicate.
6. Undetermined
We also found several watermarks we couldn't identify with 100% certainty, but we suspect they're either a mix or belong to Dreamstime's full-screen previews.
Our methodology
To investigate the extent of Stable Diffusion's alleged copyright infringements, we devised a systematic approach to reproduce watermarked images using the AI image generator. Our aim was to test whether the AI model could generate images containing watermarks from various sources, thereby demonstrating the unauthorized use of copyrighted material in its training data. We selected a diverse range of subjects, styles, and image types as input to ensure a comprehensive analysis.
We used the Stable Diffusion 2.1 model available through HuggingFace. The first watermark was generated by accident when we were testing Stable Diffusion for our article about the best AI image generators. Back then, we were already aware of the Getty Images' watermarks created through Stable Diffusion, but the one we saw resembled Depositphotos or Dreamstime. From there, we tested several prompts, including those that included keywords such as "editorial", "[various stock image sites]", "[various celebrities]", and more.
Throughout our investigation, we carefully documented each instance where Stable Diffusion generated an image containing a watermark. By tracking the occurrences of different watermarks, we aimed to uncover patterns and gain insights into the extent of Stability AI's unauthorized use of copyrighted material in their AI model. After analyzing the AI-generated images, we compiled our findings, including the number of different watermarks discovered and the frequency of their appearance in the output images.
Background story
Stable Diffusion is an AI image generator developed by Stability AI, a startup company specializing in artificial intelligence technologies. By leveraging machine learning algorithms, Stable Diffusion can generate realistic images based on input data, which usually consists of a large number of photographs, illustrations, and artworks. These AI-generated images have found applications in various industries, such as advertising, entertainment, and design.
In February 2023, Getty Images, a leading stock photography company, filed a lawsuit against Stability AI, alleging that the company had copied more than 12 million images from its database without permission or compensation. According to the lawsuit, Stability AI used these copyrighted images as training data for Stable Diffusion, thereby infringing on both copyright and trademark protections. This legal action has intensified the ongoing battle between AI art generator creators and rights-holders, who argue that such AI tools require permission to use copyrighted materials for their training data.
The lawsuit filed by Getty Images is part of a broader legal struggle involving AI startups and rights-holders. As AI art tools depend on a vast amount of visual data to function effectively, they often scrape this data from the internet without obtaining the consent of the creators or owners. This practice has raised significant concerns regarding the unauthorized use of copyrighted materials and has led to a series of legal disputes in recent years. Stability AI is also facing a class-action lawsuit in the United States, filed by a group of artists who allege copyright infringement by both Stability AI and another AI art startup, Midjourney. The outcome of these legal battles will have far-reaching implications for the future of AI-generated content and intellectual property rights in the digital age.
Comments