Deviant Art Bans AI Models from Using Website Images, Launches its Own Image Generator
DreamUp is the latest Stable Diffusion variant
Deviant Art, the online art community, owned by Wix.com, has been busy recently on the AI-generated image front. It has launched its own Stable Diffusion-based text-to-image generator trained on the LAION image dataset that accepts natural language inputs to create art. And it has taken steps that it believes will protect artists’ works from becoming training data for new AI-based image generation models. Let’s tackle both topics.
Default AI Dataset Opt-out
AI image generation models are primarily trained on large image datasets scraped from the internet. The datasets have sometimes included a selection of the more than 500 million images hosted in Deviant Art. The company believes its users should have more control over whether or not their images are used in AI-model training datasets.
A recent Deviant Art blog post stressed that images from its site found their way into these datasets without anyone’s permission. It also outlined a new solution to help reduce the incidence of posts on the site from winding up in any future datasets.
To help prevent your work from being used by third party AI datasets without your consent, we implemented a flag to tell AI datasets not to use your art.
All deviations on the platform are not authorized for inclusion in third-party datasets used to train artificial-intelligence models — unless you choose to opt in.
DeviantArt has introduced the noai and noimageai directives to unambiguously communicate absence of such authorization. When an artist declares that their content may not be used as input to AI models, an HTML tag with the noimageai directive is placed on the page and an HTTP header with the noai directive is sent when the image is directly downloaded from DeviantArt’s servers.
In order to remain in compliance with DeviantArt's updated Terms of Service, third parties that continue to use DeviantArt-sourced content to train machine-learning models of any kind must ensure their training data set excludes all content for which either of these directives are present.
The website claims to be “the only platform giving creators the ability to tell third-party AI datasets and models whether or not their content can be used for training.”
Are Deviant Art Users at Significant Risk?
How big is the problem for images hosted on Deviant Art? A recent analysis by internet pioneers and bloggers Andy Baio and Simon Willison sampled a set of 12 million images in one of the popular image training datasets used for Stable Diffusion. Baio and Willison were able to identify about 67,000, or 0.06%, came from Deviant Art. If their estimates are correct, there may be around 3.3 million Deviant Art posted images in the latest Stable Diffusion training dataset.
Pinterest showed up in this study as the source for 8.5% of the images. The risk is not zero that a Deviant Art submission will be used to train AI models. However, it is also extremely low.
Then again, a low risk is not no risk. AI-image dataset collections that want to act as good citizens now have a way to comply with Deviant Art’s unilateral move and potentially have a model to suggest to others interested in providing similar protection. This does not necessarily mean that images in existing datasets will be removed.
What is the big deal?
As long as generative AI models are not replicating existing images but instead creating new images based on what they learn from analyzing a variety of images, what is the big deal? This is unclear. However, popular artist Greg Ratkowski spoke to Technology Review of his experience with AI art generated using his name as a prompt entity:
[Ratkowski] tried searching for his name to see if a piece he had worked on had been published. The online search brought back work that had his name attached to it but wasn’t his.
“It’s been just a month. What about in a year? I probably won’t be able to find my work out there because [the internet] will be flooded with AI art,” Rutkowski says. “That’s concerning.”
This suggests there is a search engine optimization issue already. Eliminating or making artist discovery harder is a harm, though that is a problem across many domains based on the way search algorithms are designed. There is also the issue of making a distinctive art style less rare. With that said, you cannot copyright a style, so this issue will surely persist.
Another angle could be that art creators are not being properly compensated for the use of their copyrighted work as a raw material input into training the AI models. Then again, LinkedIn just lost a court decision in U.S. courts that “scraping publicly available data from sources on the internet is not a violation of the Computer Fraud and Abuse Act,” according to Technology Review.
We have a situation where the severity of harm is hard to calculate and a legal framework that does not favor artists looking to restrict use of their creations hosted online. Deviant Art is being responsive to its community, nonetheless, and offering them a default that may make some of them feel better. It is unlikely that it will have a material impact on the general use of images to train image generation models. This Reddit thread hits on the futility angle.
https://www.reddit.com/r/singularity/comments/ystjji/deviantart_ai_update_now_artists_will_be_opted/
However, some creators are clearly interested in this type of protective feature, and someone needed to be first to take the first step. Deviant Art has put forward a model that could mitigate or resolve this conflict. It may be a futile gesture, but it may also start a trend. Time will tell.
If artists truly want to keep their data out of these training datasets, there is a simple step they can take. Do not add alt text descriptions to your image. This is suboptimal for SEO, but the AI image datasets will ignore images without this alt text. In addition, there is emerging interest in using AI to automatically add descriptive alt text to images to save time and improve SEO. Still, there appears to be a way to make your images undesirable to AI dataset web crawlers.
Deviant Art DreamUp
Deviant Art is not at all against AI-generated art. It is supportive of the innovation and has deployed its own flavor of Stable Diffusion. The company says it will remove images from training the model that originated on Deviant Art and are listed as opt-out. However, it is unclear how they would untrain a model that already used Deviant Art and many other images that were used without explicit permission from the creators.
Aside from this detail and integration with Deviant Art’s online marketplace, DreamUp is not that much different from many other websites with text-to-image generators. Some suggest it is less robust, and that may be true overall. But, hey, here are some pretty nice polar bears created with the solution.
Deviant Art automatically produces the highest resolution version as an output. Unlike other text-to-image generators, where you often need to upscale the ones you like best, I needed to actually downscale these images to post here. This is a smart choice, given the nature of Deviant Art’s community.
The pricing seems a bit high compared to other services, but it actually comes with the other Deviant Art services, such as the ability to list the creation for sale and track activity around your portfolio. You only have five free creations before being required to subscribe, but the cost is essentially the same as subscribing to Deviant Art for its different tiers of service. That means you can get the AI generation for free for a limited number of images each month if you value the other website services.
This again brings up the recurring theme of whether generative AI will ultimately be more of a standalone product or a feature of a related service. Deviant Art is following a similar path to Canva and Picsart in making it the latter. That convenience factor may be appealing to existing Deviant Art users, but it is not clear that it will attract new users to the platform. Then again, maybe it will coax some free Deviant Art users to cross the line into paid subscriber territory. Interesting times.
