Stable Diffusion
A text-to-image model whose weights anyone could download and run on a gaming PC, while its best-known rival sat behind a waitlist; within a month a member of Congress asked the White House to act against it.
What it was
On 22 August 2022 Stability AI released Stable Diffusion, a model that turns a text prompt into an image, and published its trained weights for anyone to download under the CreativeML OpenRAIL-M licence. The licence allowed commercial use but listed prohibited uses. The model was built on the latent diffusion method developed by Robin Rombach and colleagues at the CompVis group in Munich, with Runway, and the company said it needed 6.9 gigabytes of graphics memory, so it ran on an ordinary gaming PC.
The release shipped with what Stability called an "AI-based Safety Classifier included by default" to filter explicit output, but because users had the weights, they could switch it off. OpenAI's DALL-E 2, the best-known rival, was then available only through OpenAI's own interface, with a waitlist that OpenAI did not remove until 28 September 2022.
What it changed
Uptake was fast and so was misuse. TechCrunch reported on 24 August 2022 that the model had leaked early on 4chan, where several threads were dedicated to generated nude images of celebrities, and that services including Midjourney and NovelAI had already adopted it. Chief executive Emad Mostaque called the leak "unfortunate."
On 22 September 2022 Representative Anna Eshoo, a California Democrat, wrote to the National Security Advisor and the White House Office of Science and Technology Policy asking them to act on Stable Diffusion's release. She wrote that "Unlike OpenAI's DALL-E 2, Stable Diffusion's model is available for anyone to use without any hard restrictions," and asked officials to address "this and similar unsafe releases using any authorities and methods within your power, including export controls." When Senators Blumenthal and Hawley wrote to Meta about its leaked LLaMA model in June 2023, their letter cited Stable Diffusion's use for sexual deepfakes of real people as its first example of how open models had been abused.
The arguments it moved
Who should have access to powerful models?
Stable Diffusion brought the open-weights question to Congress in 2022, months before ChatGPT. Eshoo's September 2022 letter proposed export controls as a remedy for a model that had already been published to the whole world. Blumenthal and Hawley's June 2023 letter to Mark Zuckerberg about the LLaMA leak used Stable Diffusion's misuse as its lead example, arguing that "centralized AI models can be more effectively updated and controlled to prevent and respond to abuse compared to open source AI models." Stability's side of the argument was the release itself: a model small enough to run at home, whose safety filter was optional because users controlled the code.
Sources
- Stable Diffusion Public Release, Stability AI, 22 August 2022
- Stable Diffusion repository and model card, CompVis (GitHub)
- Deepfakes for all: Uncensored AI art model prompts ethics questions, TechCrunch, 24 August 2022
- OpenAI removes the waitlist for DALL-E 2, TechCrunch, 28 September 2022
- Eshoo Urges NSA & OSTP to Address Unsafe AI Practices, press release of 22 September 2022 (LegiStorm archive)
- Letter from Senators Blumenthal and Hawley to Mark Zuckerberg, 6 June 2023