Names, versions and owners change fast. Checked October 2026.
In plain words
Stable Diffusion is a text-to-image model: describe a picture in words and it creates one. It is not a website or an app on its own. It is the model itself, a file anyone can download and run, which is called open weights.
That openness is what made it famous. When it was released in August 2022, the first public version needed about 7 GB of graphics memory, so it could run on a gaming PC rather than only in a company's data centre. Many free apps, plug-ins and art tools were built on top of it.
It was developed by Stability AI, a company in London, with researchers from a university group in Munich and the video company Runway. New versions have followed the 2022 original; the box above shows the latest.
Everyday example
Developing a photo in a darkroom
In an old darkroom, a blank sheet sits in the tray and a picture slowly appears, faint at first, then sharper with each minute. Stable Diffusion works in a similar way. It starts with a square of random noise and, over dozens of small steps, clears the noise into a picture that matches your words. And the darkroom is your own computer.
Where the example breaks: A darkroom reveals a photo that already exists on film. Stable Diffusion invents a new image at every step, guided by patterns it learned, so nothing was "there" to begin with.
Latestas of October 2026
- Newest image model
- Stable Diffusion 3.5 (October 2024), in Large, Large Turbo and Medium sizes; Stability AI's 2026 model releases have been for audio, such as Stable Audio 3.0 (May 2026)
- Licence
- Community licence free below $1 million in yearly revenue; enterprise licence above that
- Company
- Stability AI closed a Series B funding round in August 2026, which it described as under new leadership, with backers including Electronic Arts and major music companies
These details change often. The rest of this page explains what it is, which stays true.
How it actually works
The name tells you the method: it is a diffusion model, with one trick that makes it light enough for home computers.
Work in a compressed space
Instead of handling every pixel, an encoder squeezes images into a much smaller grid of numbers, a latent space. All the heavy work happens there, which cuts the computing cost. This is why it is called a latent diffusion model.
Start from noise, remove it step by step
Generation begins with random noise in that small space. A neural network predicts which part is noise and takes some away, again and again, for tens of steps.
Words steer every step
Your prompt is turned into an embedding. Through a mechanism called cross-attention, it guides each denoising step toward what you described.
- Made by
- Stability AI, with CompVis (LMU Munich) and Runway
- Released
- 22 August 2022
- Versions
- Several generations since 2022, each released as downloadable weights
- Built on
- The 2021 paper "High-Resolution Image Synthesis with Latent Diffusion Models"
Because the weights are public, Stable Diffusion became a base for many community-made variants trained on particular art styles.
Questions people ask
Is Stable Diffusion free?
The model weights can be downloaded at no charge and run on your own computer. Recent versions come under Stability AI's community licence, which is free for individuals and smaller businesses, while larger companies need an enterprise licence. Check the current terms before commercial use.
Is Stable Diffusion the same as Midjourney?
No. Both make images from text, but Midjourney is a service whose models stay on its own servers. Stable Diffusion's weights are public, so you can run and change the model yourself.
Can Stable Diffusion run on my computer?
Usually, if it has a reasonably recent graphics card. Smaller versions need only a few gigabytes of graphics memory. Free desktop tools let you install it and type prompts without any coding.
Who makes Stable Diffusion?
Stability AI, a company based in London. The original model grew out of research by the CompVis group at LMU Munich, with help from the video company Runway.