sourced story
22 August 2022Primary source · 2 sourcesWell documented

Stable Diffusion open-sources text-to-image generation

A latent diffusion model small enough to run on a single consumer GPU goes public

On the timeline · around 22 August 2022 · Transformers and the Generative AI WaveTransformers and the Generative AI WaveRecognition and ReasoningStable Diffusion open-sources text-to-image generation202120222023

Quick facts

Researchers
Robin Rombach, Andreas Blattmann, et al.
Released by
Stability AI, with CompVis, Runway, LAION
Method
Diffusion in a compressed latent space

What happened

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer, working across LMU Munich, Heidelberg University, and Runway, published 'High-Resolution Image Synthesis with Latent Diffusion Models,' describing a diffusion model that generates images by learning to reverse a process of gradually adding noise to an image, but crucially performing that denoising in a compressed latent space produced by a pretrained autoencoder rather than on full-resolution pixels directly. That compression cut the computational cost enough to make training and running the model practical on far more modest hardware than earlier diffusion approaches required. Stability AI, in collaboration with the original research groups, EleutherAI, and LAION, publicly released the resulting model, called Stable Diffusion, on August 22, 2022, under a license permitting commercial use, after an earlier restricted release to about 1,000 researchers. It could generate 512-by-512 pixel images using under 7 gigabytes of GPU memory, putting it within reach of a single consumer graphics card.

Why it matters

Stable Diffusion was the first capable text-to-image model whose full weights were released publicly rather than kept behind a company's API, letting anyone run, modify, or fine-tune it, which rapidly spread diffusion-based image generation into open-source tools, apps, and research well beyond what closed systems like DALL-E allowed.

How we know

The original latent diffusion paper documents the compressed-latent-space method directly; Stability AI's own public release announcement confirms the licensing terms and release date.

Sources

See something wrong? . Corrections with a source get fixed fastest.

Part of a timelineArtificial Intelligence30 events · From a wartime theory of neurons to machines that write, paint, and fold proteinsView all →