Stable Diffusion (SD)

basicwidespread

The Stable Diffusion family of AI models that generate images from text descriptions, whose model files anyone can download and even run on their own computer.

In plain terms

Stable Diffusion is an image generation model: you describe in words what you want to see, and depending on the machine, it creates an image in a few seconds or minutes. It can also transform an existing image, complete it, or replace a part of it.

Imagine it like a sculptor chipping away unnecessary pieces from a shapeless block of stone until the statue emerges. The model starts from a completely noisy, 'grainy' image and step by step removes the noise while adapting to your description.

What distinguishes it from closed services, such as Midjourney or ChatGPT image generation, is that the model itself is publicly downloadable. You do not need to subscribe to a website. If you have a sufficiently powerful graphics card, you can run it at home after downloading, even without internet, and your images will not leave your computer.

Because of this, a huge community has formed around it: users further train it on their own styles, characters, and themes, and share these variations with each other. This freedom is simultaneously its greatest advantage and the source of most debate.

How it works

Training: the model learned on countless images with captions which words correspond to which shapes, colors, and styles.
Text interpretation: your entered instruction (the prompt) is converted by a text-understanding component into a sequence of numbers that guides image generation like a compass.
Image from noise: the model starts from random noise and, over dozens of steps, removes a bit at a time so that the result increasingly resembles your description.
It works in compressed space: it does not calculate the full, pixel-based image but a reduced, condensed version of it. Therefore, a more powerful home computer is sufficient.
Decoding: at the end, a decoder section generates the full-resolution, viewable image from the compressed version.

An everyday example

A small business owner wants an illustration for their online shop's autumn promotion but has no budget for a graphic designer. They download a free program based on Stable Diffusion, type "pumpkins and chestnuts on a wooden table, warm light", and within minutes can choose from several variations.

Who is affected

marketerYou can quickly and cheaply create illustrations, architectural visualizations or advertising image variations with it, but legal and quality checks remain your responsibility. → Marketing

developerSince the model is downloadable, you can build it into your own application and train it further on your own images. → Code

entrepreneur, executiveRunning on your own machine keeps your confidential plans and product photos from reaching external providers, but you must read the license terms before commercial use.

everyoneSome images seen on the internet are already created with such models, so it is worth treating overly perfect or shocking photos with caution.

A common misconception

MythStable Diffusion cuts and pastes pieces of existing images from a massive image database.

In realityThe model does not store training images but learns patterns from them, generating every image anew from noise. It is rare for a frequently seen training image to be reproduced almost exactly, but this does not eliminate copyright concerns.

Limits and risks

It often makes mistakes in details: distorted hands, nonsensical text, and strange proportions occur, especially in older versions.
Copyright status remains unsettled. Several lawsuits have been filed over images used for training, including by Getty Images and a group of illustrators. A November 2025 UK ruling largely favored Stability AI, but the issue may be judged differently by country, so verify before commercial use.
In the version running on your own machine, the built-in filters can be switched off, so it can be used to create misleading fake images or fake images that harm real people.
'Free' does not mean unlimited: the license for each variant differs, and commercial use may become fee-based above a certain revenue threshold.

Similar, but not the same

Diffusion model – The diffusion model is the method itself, while Stable Diffusion is a specific, named implementation of it.
Midjourney – Midjourney is a closed, subscription-based service whose model cannot be downloaded. Stable Diffusion, however, can be run and modified on your own machine.
DALL·E – DALL·E is OpenAI's closed image generator family, which you can only access through the company's services, and which in ChatGPT has largely been replaced by newer, built-in image generation. The Stable Diffusion model files, on the other hand, are publicly downloadable.

Related concepts

Broader

Generative AI, Diffusion model, Image generation

Narrower

SD 1.5, Stable Diffusion XL, Stable Diffusion 3

On the edge

Midjourney, DALL·E, Flux, ControlNet, LoRA, Prompt, Deepfake

A brief history

The model is based on the latent diffusion process, published jointly by the CompVis research group (then at Heidelberg, now part of Ludwig-Maximilians-Universität Munich) and Runway researchers in December 2021. Stable Diffusion itself was made public in August 2022. CompVis and Runway researchers developed it, while Stability AI provided computing capacity and funded the work. The model was trained on a portion of the LAION nonprofit organization's public image database (LAION-5B). Its open accessibility led to rapid spread. Later major versions, SD 2.0 (2022), SDXL (2023), and Stable Diffusion 3 and 3.5 (2024), were developed by Stability AI.

Where things stand

Stable Diffusion and its community-based variants are widely used tools, integrated into countless programs. In terms of image quality, several newer closed and open models have surpassed it, such as Flux, and community interest has partially shifted toward these. However, due to the rich ecosystem of add-ons for older versions (SD 1.5, SDXL), they remain a foundation for locally run image generation.

Status: 2026 November

Also known as

Stable Diffusion · SD · SDXL · Stable Diffusion XL · SD 1.5 · Stable Diffusion 3 · SD3 · SD 3.5 · Stable Diffusion

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.