Image-generating AI can copy and paste training data, raising intellectual property issues
Image-generating AI models like DALL-E 2 and Stable Diffusion can — and do — reproduce aspects of images from their training data, show researchers in a new study, which raises concerns as these services come into widespread commercial use.
Co-authored by scientists from the University of Maryland and New York University, the research identifies cases where image generation models, including Stable Diffusion, "copy" from data public Internet (including copyrighted images) on which they were formed. .
The study has not yet been peer-reviewed, and the co-authors have submitted it to a conference whose rules prohibit media interviews until the research has been accepted for publication. But one of the researchers, who asked not to be identified by name, shared high-level thoughts with TechCrunch via email.
"Even though streaming models such as stable streaming produce some nice images, and often ones that look very original and customized for a particular text prompt, we show that these images can actually be copied from their training data, either in bulk or by copying only portions of training images," the researcher said. "Companies that generate data with diffusion models may need to reconsider intellectual property. It is virtually impossible to verify that a particular image generated by Stable Diffusion is new and has not been stolen from the training set."
Images of noiseState-of-the-art image generation systems, such as stable diffusion, are referred to as "diffusion" models. Broadcast models learn to create images from text prompts (e.g., "a sketch of a bird perched on a windowsill") as they work their way through huge training datasets. Models, trained to "re-create" images instead of drawing them from scratch, start with pure noise and refine an image over time to gradually bring it closer to the text prompt.
It's not a very intuitive technology. But it's exceptionally good at generating artwork in virtually any style, including photorealistic art. Indeed, broadcasting has enabled a multitude of interesting applications, from synthetic avatars in Lensa to artistic tools in Canva. DeviantArt recently released an app powered by Stable Diffusion for creating custom artwork, while Microsoft is leveraging DALL-E 2 to power an upcoming generative art feature on Microsoft Edge.
Above are images generated by Stable Diffusion from random legends in the model's training set. At the bottom are images that the researchers invited to match the originals. Image credits: Somepalli et al.
To be clear, it was no mystery that the broadcast models reproduced elements of training images, which are usually indiscriminately grabbed from the web. Character designers like Hollie Mengert and Greg Rutkowski, whose classic painting styles and fantastical landscapes have become one of the most commonly used prompts in Stable Diffusion, have decried what they consider poor AI impersonations. which are nevertheless linked to their names.
But it's...
Image-generating AI models like DALL-E 2 and Stable Diffusion can — and do — reproduce aspects of images from their training data, show researchers in a new study, which raises concerns as these services come into widespread commercial use.
Co-authored by scientists from the University of Maryland and New York University, the research identifies cases where image generation models, including Stable Diffusion, "copy" from data public Internet (including copyrighted images) on which they were formed. .
The study has not yet been peer-reviewed, and the co-authors have submitted it to a conference whose rules prohibit media interviews until the research has been accepted for publication. But one of the researchers, who asked not to be identified by name, shared high-level thoughts with TechCrunch via email.
"Even though streaming models such as stable streaming produce some nice images, and often ones that look very original and customized for a particular text prompt, we show that these images can actually be copied from their training data, either in bulk or by copying only portions of training images," the researcher said. "Companies that generate data with diffusion models may need to reconsider intellectual property. It is virtually impossible to verify that a particular image generated by Stable Diffusion is new and has not been stolen from the training set."
Images of noiseState-of-the-art image generation systems, such as stable diffusion, are referred to as "diffusion" models. Broadcast models learn to create images from text prompts (e.g., "a sketch of a bird perched on a windowsill") as they work their way through huge training datasets. Models, trained to "re-create" images instead of drawing them from scratch, start with pure noise and refine an image over time to gradually bring it closer to the text prompt.
It's not a very intuitive technology. But it's exceptionally good at generating artwork in virtually any style, including photorealistic art. Indeed, broadcasting has enabled a multitude of interesting applications, from synthetic avatars in Lensa to artistic tools in Canva. DeviantArt recently released an app powered by Stable Diffusion for creating custom artwork, while Microsoft is leveraging DALL-E 2 to power an upcoming generative art feature on Microsoft Edge.
Above are images generated by Stable Diffusion from random legends in the model's training set. At the bottom are images that the researchers invited to match the originals. Image credits: Somepalli et al.
To be clear, it was no mystery that the broadcast models reproduced elements of training images, which are usually indiscriminately grabbed from the web. Character designers like Hollie Mengert and Greg Rutkowski, whose classic painting styles and fantastical landscapes have become one of the most commonly used prompts in Stable Diffusion, have decried what they consider poor AI impersonations. which are nevertheless linked to their names.
But it's...
What's Your Reaction?
