CompVis: Latent Diffusion Research Behind Stable Diffusion

Björn Ommer’s computer vision research group, previously at Heidelberg and now at LMU Munich, developed latent diffusion and collaborated on the original Stable Diffusion release.

kindResearch lab

Updated

Latent diffusion as a research contribution

CompVis is the research-group identity associated with Björn Ommer's computer vision work, earlier at Heidelberg and now at LMU Munich. Its latent-diffusion project lists academic and Runway collaborators; the group is distinct from the commercial organizations involved in releasing Stable Diffusion.

Latent diffusion runs the diffusion process in an autoencoder's compressed representation instead of directly in full-resolution pixel space. That design reduces computation while retaining a route back to images. It does not establish that the model was trained on a single consumer GPU.

The 2022 Stable Diffusion release was a collaboration involving CompVis, Stability AI and Runway. The repository credits Stability AI's compute and LAION support and links public checkpoints under CreativeML OpenRAIL-M, a license with use restrictions. Downloadable weights should not be described as unrestricted simply because the code and model files are public.

Sources

  1. [1]
    Prof. Dr. Björn Ommer

    Computer Vision & Learning Group, LMU Munich · retrieved Oct 23, 2026

  2. [2]
    High-Resolution Image Synthesis with Latent Diffusion Models

    Computer Vision & Learning Group, LMU Munich · retrieved Oct 23, 2026

  3. [3]
    CompVis/stable-diffusion

    CompVis · retrieved Oct 23, 2026

CompVis history

2022

●
Stable Diffusion puts image generation on a home GPU

With Stability AI, Runway

Stable Diffusion released downloadable weights for text-to-image generation, letting developers run and adapt the model themselves. It generated images through a compressed representation that reduced compute demands, with responsibilities and distribution requirements under its release license.

Stability AI
Model