A lab, not a company
CompVis is an academic research group, currently the Computer Vision and Learning Group at LMU Munich, led by Björn Ommer. It carries the name of an earlier stop: the Computer Vision Group Ommer ran at Heidelberg University before moving to Munich. It has no product to sell and no customers, which sets it apart from most other organizations on the timeline. [1][3]
The group's GitHub organization hosts the code behind its published papers, including taming-transformers, a 2021 image-generation method (VQGAN combined with a transformer) that predates and feeds into the diffusion work it later became known for. [4]
The paper behind the product
In December 2021, Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser and Björn Ommer, all working out of CompVis, posted "High-Resolution Image Synthesis with Latent Diffusion Models," later presented at CVPR 2022. The method's central move was running the diffusion process, the step-by-step denoising that produces an image, inside a compressed latent space rather than on full-resolution pixels. That cut the compute a usable model needed by enough to make training and running one on a single consumer GPU realistic instead of theoretical. [5]
Atlas interpretation: That paper is the actual invention. What most people call Stable Diffusion is CompVis's architecture, trained on a larger dataset and released to the public as a downloadable model. [5]
From a paper to a public release
Stability AI announced the public release of Stable Diffusion on August 22, 2022, describing itself as coordinating the release in cooperation with Hugging Face's legal, ethics and technology teams and CoreWeave's infrastructure. That announcement documents Stability AI's role in the release; it does not itemize CompVis's research contribution or Runway's part in it, so the specific division of labor between the three is better read as reported by outside coverage than as stated word for word in Stability's own post. [6][2]
Atlas interpretation: The pattern is one the timeline shows elsewhere: a university lab does the research, a company with more compute and a distribution plan turns it into something millions of people can download and run. CompVis's continuing relevance to AI is less about a single release and more about being one of the labs whose architecture choices, not just whose papers, ended up running in production. [5][6]
Sources
- Prof. Dr. Björn Ommer
Computer Vision & Learning Group, LMU Munich
- CompVis/stable-diffusion
GitHub
- CompVis
Computer Vision & Learning Group, LMU Munich · Sep 9, 2026
- CompVis (GitHub organization)
GitHub · Sep 9, 2026
- High-Resolution Image Synthesis with Latent Diffusion Models
arXiv · Dec 20, 2021
- Stable Diffusion Public Release
Stability AI · Aug 22, 2022