Stable Diffusion (2022) › 2022: Stable Diffusion

Stable Diffusion, released by Stability AI with researchers from LMU Munich and RunwayML in August 2022, made text-to-image generation accessible on consumer GPUs through an open-source latent diffusion model. Contemporary systems such as DALL-E 2 and Imagen required cluster-scale GPUs and access through proprietary services. Stable Diffusion instead allowed individual creators and developers to run or modify the model on their own hardware.

By 2022, DALL-E 2 and Midjourney could produce high-quality images from prompts, but both remained closed systems with limited access. Research on latent diffusion suggested a more efficient approach: operate in a compressed representation instead of directly on pixels. The CompVis group at LMU Munich had developed such models for several years, reducing computation while preserving image quality. These models, however, had not yet been released as open systems that individuals could run locally.

Local execution changed who could experiment with image generation. Artists and game developers could use the model without a proprietary API, while developers could fine-tune it for a style or domain and integrate it into other software. The release connected an efficient architecture with weights that users could inspect and modify.

The release showed that an open model could approach proprietary systems in image quality while giving users local access and control. Rapid adoption and community development around Stable Diffusion influenced later model-release decisions. Broader access also let more people test and modify the system or build applications around it.