Stable Diffusion image generator
Released: August 2022
Stable Diffusion is a free, open source AI-based system that generates images from text descriptions or prompts.
Trained on pairs of images and captions taken from LAION-5B, the system was initially restricted to researchers as Stable Diffusion 1. It was publicly released as Stable Diffusion 2 in August 2022.
Stable Diffusion quickly became popular largely on account of its free availability, minimal usage restrictions, and the fact that it can run on ordinary consumer hardware.
But it is seen to suffer from a number of technical limitations, and to pose several actual and potential challenges and risks to organisations and individuals operating and using it:
Accuracy: The quality of images produced by the system degrade above 512×512-pixel resolution; it also has trouble generating images with human limbs and faces.
Bias/discrimination: Stable Diffusion was primarily trained on images with English descriptions, resulting in images that tend to reinforce western stereotypes and biases.
Copyright: Stable Diffusion draws on copyrighted works of artists, without the owner’s consent or any form of attribution, prompting concerns about plagiarism and the misuse, abuse and loss of IP. Furthermore, the user is given full rights to images they create on Stable Diffusion, and is free to use them commercially, resulting in questions about fair use. Responding to pressure, Stability AI says artists will have the chance to opt out of the training for the next version of the software. In January 2023, a group of lawyers filed a US class-action lawsuit against Stability AI, DeviantArt, and Midjourney on the basis that their use of Stable Diffusion illegally remixes the copyrighted works of millions of artists whose work was used as training data. In the same month, Getty Images announced it will sue Stability AI for alleged copyright violation, on the basis that it believes Stability AI 'unlawfully copied and processed millions of images protected by copyright' to train its software. And a debate erupted as illustrator Hollie Mengert discovered her artwork had been used to fine-tune Stable Diffusion to recreate her style. Furthermore, Google researchers have shown that Stable Diffusion is able to memorise and recreate specific copyrighted images.
Dual/multi-use: The relative accessibility of the model, including the release of its model weights, together with its relative paucity of filters, means Stable Diffusion can be used for more or less any purpose, including nefarious ones such as spreading spam, propaganda, deepfakes, misinformation and disinformation.
Employment: Artists' loss of copyright and the ability of users to create images in their unique style has prompted increasing concern that artists, photographers, illustrators, cinematographers and others stand to lose their commercial viability and careers. A French game developer's release of a Stable Diffusion-based tool to generate images in the style of late Korean artist Kim Jung Gi with a simple text prompt sparked a furious response from artists who saw the tool as theft, and accused the developer of cultural appropriation.
Safety: As TechCrunch notes, Stability AI bans lewd or sexual material, hateful or violent imagery, prompts containing copyrighted or trademarked material, and personal information such as phone numbers and Social Security numbers. But this has not stopped users generating abusive, violent, sexual and pornographic images, notably to 4chan.
Unlike DALL-E, Stability AI has made Stable Diffusion accessible courtesy of a public demo, a software beta called DreamStudio, a development notebook, and weights, model card and code.
The company also appears to have listened to feedback from users and artists as Stable Diffusion 2 has most NSFW content, celebrity images and famous artists’ names removed.
It has also made it easier to find out whose work is in the dataset, thereby making simpler for artists to opt-out.
Operator: Stability AI; Prisma Labs
Developer: Stability AI; CompVis LMU; Eleuther AI; RunwayML; LAION
Purpose: Generate images
Technology: NLP/text analysis; Computer vision; Text-to-image; Neural network; Deep learning
Issue: Safety; Bias/discrimination; Dual/multi-use; Mis/disinformation; Copyright; Employment
Transparency: Black box
Stability AI. Stable Diffusion launch announcement
Stability AI. Stable Diffusion public release
Stability AI. Stable Diffusion 2.0 release
Congresswoman Eshoo letter to National Security Advisor (NSA) and the Office of Science and Technology Policy (OSTP)
Carlini N., Hayes J., Nasr M., Jagielski M., Sehwag V., Tramèr F., Balle B., Ippolito D., Wallace E. (2023). Extracting Training Data from Diffusion Models
Somepalli G., Singla V., Goldblum M., Geiping J., Goldstein T. (2022). Diffusion Art or Digital Forgery? Investigating Data Replication in Diffusion Models
News, commentary, analysis
Published: December 2022
Last updated: February 2023