- Unstable Diffusion is an AI image generator with minimal content restrictions.
- The platform allows users to generate NSFW content, including pornographic images.
Unstable Diffusion, the NSFW AI image generator that was booted off Kickstarter in December, is gaining quite a following: The program is generating more than 500,000 images every day, its CEO and cofounder Arman Chaudhry told Insider over email.
Unstable Diffusion is similar to other AI image generators that create images from text inputs, but it has minimal content restrictions and can create NSFW images, including pornographic ones.
It is based on Stable Diffusion, the popular text-to-image generator from Stability AI. The code underlying Stable Diffusion is public, so anyone can theoretically view and modify it to create new versions of the original model.
That's what Chaudhry did when starting Unstable Diffusion in August 2022, the same month Stable Diffusion was released to the public.
Unstable Diffusion began as a subreddit to share AI-generated porn. The group then moved to Discord, where members started building tools, eventually creating a bot that would become an early version of its image generator, TechCrunch reported.
"We founded the discord group as a refuge for artists who wanted to create AI art without limitations, including those related to adult content," Chaudhry, CEO of Unstable Diffusion, told Insider in February over email.
The community launched a Kickstarter campaign in December to raise money to build its image generator, but was removed from the platform 12 days later.
Still, Unstable Diffusion was able to raise the necessary funds — $30,000, Chaudhry told Insider in February — to roll out a standalone web app. Unstable Diffusion has a basic, free service, as well as a trio of paid products that charge between $14.99 and $59.99 per month. These premium services allow users to generate more images simultaneously, produce images faster, and use images commercially.
Of course, there is opportunity for foul play when it comes to AI porn, as users can create deepfake pornography or content depicting minors engaging in sexual acts.
In February, Atrioc, a popular Twitch streamer, was caught watching this pornography depicting deepfakes of female streamers. At the time, experts told Insider that being featured in non-consensual deepfake porn can be traumatic and considered abuse.
Chaudhry told Insider in February that Unstable's "aggressive filter and moderation system" prevents deepfakes and other undesirable content.
But it's not foolproof. TechCrunch was able to use the generator to create images that produced look-alikes of Donald Trump and Chris Hemsworth, for example.
When asked about the TechCrunch images, Chaudhry said that the look-alikes were created during a period where the service was experiencing "a bug with our deepfake content generation filter that temporarily disabled it for some users."
Bug or not, perfect content moderation, especially for a product that has few guard rails is difficult.
"It's not possible because of Gilb's law of unreliability — any system which depends on human reliability is unreliable," Dan Hendrycks, an AI safety expert and director of the Center for AI Safety, told Insider.
Content moderation for any system that is capable of producing inflammatory material is impossible, he added.
Even AI chatbots moderated extensively by tech companies like Google and OpenAI are at risk of having their guardrails bypassed.
The platform may also run into legal problems:
In 2019, Gavin Newsom, the Governor of California, signed two deepfake-related bills into state law, one of which allowed residents to sue anyone using deepfake technology that inserts them into pornographic material without consent.
In May, a bill was introduced to Congress that would make it illegal to share or spread non-consensual, AI-generated pornography.