Stable Diffusion

A diffusion model was moved into a compressed latent space, which let it run on an ordinary graphics card, and the weights were released openly.

Why it matters

Image generation left the control of a few companies: anyone could run the model locally and change it.

Computing in latent rather than pixel space significantly lowered the computational cost, as the paper puts it. The consequences were social as much as technical: versions fine-tuned on particular artists' styles appeared at once, and the question of rights over training data moved into the courts. Open weights also meant any built-in restriction could be removed, an argument that continues. Stability AI's release post of 22 August 2022 gives the licence (Creative ML OpenRAIL-M, permitting commercial and non-commercial use) and a memory requirement of 6.9 GB of VRAM.

Event record

Event date

August 2022

Timeline date

Event date

Verification

Sources gathered automatically · September 17, 2026

Lines

ID

evt-0244

The method was published in 2021 and presented at CVPR in June 2022; the model weights were released in August 2022.

Sources

primary High-Resolution Image Synthesis with Latent Diffusion Models
CVPR 2022 · Published June 2022
primary Stable Diffusion Public Release
Stability AI · Published August 22, 2022

Related events

Earlier

Enables LAION-5B: 5.85 billion pairs, a third of them not in English 2022links to this record
The Stanford report names Stable Diffusion among the models trained on the images gathered here.
- Identifying and Eliminating CSAM in Generative ML Training Data and Models · Stanford Internet Observatory
Related The draft European AI Act 2021links to this record
The question of liability for training data and consequences entered law.
Related CLIP 2021
The text condition is supplied through a representation trained to match text and images.
Contrasts DALL·E 2021
The same task, but with the weights released openly.
- High-Resolution Image Synthesis with Latent Diffusion Models · CVPR 2022
Builds on Diffusion models 2015
The model belongs to the diffusion family.
- High-Resolution Image Synthesis with Latent Diffusion Models · CVPR 2022
Related Inpainting: a restorers’ word for a machine 2000links to this record
Latent Diffusion in 2021 claims the best result in inpainting itself but does not cite the 2000 paper; the link is only the shared task and name.

Later

Related LLaMA 2023links to this record
The second event of the period to put a capable model outside its developer's walls.
Related Llama 2 2023links to this record
A second model with released weights, and a second argument about what openness means.