On Friday, the online art community DeviantArt announced DreamUp, an AI-powered text-to-image generator service powered by Stable Diffusion. Simultaneously, DeviantArt launched an initiative that ostensibly lets artists opt out of AI image training but also made everyone’s art opt-in by default, which angered many members.
DreamUp creates novel AI-generated art based on text prompts. Due to its Stable Diffusion roots, DreamUp learned how to generate images by analyzing hundreds of millions of images scraped off sites like DeviantArt and collected into LAION datasets without artists’ permission, a potential irony that some DeviantArt members find problematic.
this is how dreamup is being developed. it adds training on non-opt-out artwork on top of the stability diffusion model, which is ALREADY trained using unethically sourced data. they CANT UNLEARN those parts.
this whole discussion is not in good faith unless they address this pic.twitter.com/KnQZerCIol
— svlt ✈️ (@svltart) November 11, 2022
As we’ve reported frequently on Ars in the past, Stable Diffusion’s web-scraping nature ignited a huge debate earlier this year among artists that challenge the ethics of AI-generated artwork. Some art communities have taken hard stances against any AI-generated images, banning them completely.
Perhaps anticipating a backlash, DeviantArt is making overtures to pacify artists who might be upset about their work being used to train AI image generators. The site is providing a special “noai” flag that artists can check in their image settings to opt out of third-party image datasets. (Whether third-party image scrapers will honor this flag, however, remains to be seen.)
Also, DeviantArt will let artists opt out of letting their images train DreamUp in the future, but each artist must fill out a form that requires human review first. This policy has led to significant pushback among DeviantArt members, some of whom have threatened to delete all of their work and deactivate their accounts.
DeviantArt’s DreamUp information page also takes a defensive tone, stating that DeviantArt did not consent to third-party AI image models (such as Stable Diffusion) that scraped their site to make their models work. And further down the page, the site attempts to debunk common misconceptions about how AI image synthesis works.
As for DreamUp itself? We experimented with the service, which feels like a vanilla Stable Diffusion model. DeviantArt members get five free prompts to try it out, and members can acquire more prompt credits by subscribing to various CORE plans that range from US $3.95 to $14.95 a month.
Alternatively, you can also use Stable Diffusion locally for free if you’re handy enough to install a package from GitHub or if you download the Draw Things app on your iPhone.
Update: A few hours after press time, DeviantArt announced an update to its AI dataset policy stating that in response to community feedback, all DeviantArt-hosted images will be “automatically labeled as NOT authorized for use in AI datasets.”
At the moment, this appears to only include automatic opt-out of third-party AI datasets with the “noai” flag—a label that dataset scrapers will have to voluntarily honor to be effective (we know of no plans to honor this flag at this time). Artists who wish to opt-out their work from being included in DeviantArt’s own AI training for DreamUp will still need to fill out a form.