Stable Diffusion is released with downloadable weights

Stable Diffusion 1.4 was released publicly on 22 August 2022 with its weights freely downloadable, by Stability AI with CompVis at LMU Munich, Runway, EleutherAI and LAION. Because the diffusion runs in a compressed latent space, it generates 512 by 512 images in around 6.9 gigabytes of video memory, within reach of a consumer graphics card.

Background

Two capable text-to-image models had arrived earlier that year. DALL-E 2 in April and Imagen in May, and both stayed with their owners: you sent a prompt to a company and a company sent back a picture.

The method behind Stable Diffusion was not new either. Rombach, Blattmann, Lorenz, Esser and Ommer had published latent diffusion in December 2021 and presented it at CVPR the following summer. Their contribution was to run the diffusion in a compressed latent space rather than on pixels directly, which cut the computation enough to matter.

What happened

Stable Diffusion 1.4 was released publicly on 22 August 2022, by Stability AI working with CompVis at LMU Munich, Runway, EleutherAI and LAION. The weights were made freely downloadable.

The model generates 512 by 512 images using around 6.9 gigabytes of video memory. That is a consumer graphics card rather than a data centre, and it is the consequence of the latent space approach rather than an optimisation applied afterwards.

Twelve days earlier the weights had gone to roughly a thousand researchers. The public release is the event here, and the earlier date is worth knowing.

What followed

Within days it was running on ordinary laptops. It was reportedly downloaded more than two hundred thousand times shortly after release, a figure from secondary reporting rather than from Stability.

The training data has been contested since. LAION-5B is a set of image and text pairs collected from the open internet, and the artists whose work it contains were not asked. Litigation on that question was still running years later, and the entry records the release rather than settling it.

Why it mattered

Text-to-image models already existed. DALL-E 2 had arrived in April 2022 and Imagen in May, both reachable only through their owners. What was new was that anybody could download this one and run it on their own machine, and it was reportedly downloaded more than two hundred thousand times within days.

Sources

• High-Resolution Image Synthesis with Latent Diffusion Models. arxiv.org. Primary source
• Stable Diffusion launch announcement. stability.ai. Official
• Stable Diffusion. en.wikipedia.org. Secondary

Cite this page

AI Achievements. (2022). Stable Diffusion is released with downloadable weights. Retrieved 2026-08-29, from https://achievements.ai/milestone/stable-diffusion-released-with-open-weights

@misc{achievements_stable_diffusion_released_with_open_weights,
  title  = {Stable Diffusion is released with downloadable weights},
  author = {{AI Achievements}},
  year   = {2022},
  url    = {https://achievements.ai/milestone/stable-diffusion-released-with-open-weights}
}