Aiodic

Back to the table

91 Creative tool Sd Stable Diffusion Creative Tools

Stable Diffusion

Open image-generation models.

What it is

Stable Diffusion is a family of AI models that turn a written description into a picture. You type "a lighthouse on a cliff at sunset, oil painting", and a few seconds later you have an image that never existed before. In that sense it does the same job as Midjourney or OpenAI Images. What sets it apart is that anyone can download the model itself and run it on their own computer, free, with no company in the middle deciding what you can make or keeping a copy of your work.

The name covers several generations of models released by the London company Stability AI since 2022, rather than one product. At the time of writing, the family includes the original Stable Diffusion 1.x models from 2022, Stable Diffusion XL (usually shortened to SDXL) from 2023, and Stable Diffusion 3 and 3.5 from 2024. Stability AI's website still describes 3.5 as its most powerful image model, but check for newer versions, as this changes.

Under the hood it is a diffusion model: it starts from random visual noise, like the static on an old television, and removes that noise step by step, guided by your words, until a picture emerges. Its particular trick came from a 2021 research paper by Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser and Björn Ommer, from a university research group called CompVis (then in Heidelberg, now at LMU Munich). Instead of working on every pixel of a full-size image, their "latent diffusion" method works on a compressed summary of the image, a bit like sketching on a small notepad before painting the big canvas. That made it far cheaper to run, which is why it fits on a home computer at all.

Open weights, not quite open source. A trained model is, at heart, a very large file of numbers called weights or parameters, which hold everything it learned. Stable Diffusion's weights are published for anyone to download. That is what "open weights" means. "Open source" traditionally means something stricter: software anyone can use for any purpose, with nothing held back. Stable Diffusion doesn't meet that bar, for two reasons. First, every version comes with a licence that restricts some uses. Second, the full training data and training process can't be reproduced from what was published. Open weights give you the finished model to run and adapt, but not the full recipe or unlimited freedom. The licences have also changed over time:

Licence terms change, so read the current version yourself before using the models for business.

Why it matters

It put image AI on ordinary computers. Before August 2022, the best image generators lived on company servers, and you used them on the company's terms. Stable Diffusion let anyone with a decent graphics card generate images at home, offline, for nothing beyond the cost of electricity. That changed who could experiment, and what they could build.

It grew the biggest community of add-ons in AI imagery. Because the weights are public, thousands of people have trained their own variations. Some specialise in anime, some in product photos, some in one particular artist's look. Free programs let you run it without writing code. For a creator, this means a level of control that closed tools don't offer: your own character in every picture, a consistent brand style, or a model trained on your own work.

It is at the centre of the biggest legal fights over AI art. Stable Diffusion was trained on billions of pictures collected from the public web, mostly without asking the people who made them. Photographers, illustrators and a major stock photo agency have taken Stability AI to court, and the outcomes will shape whether training AI on other people's work is lawful. Anyone selling AI-made images has a stake in those answers.

Openness cuts both ways. Nobody can take a downloaded model back. That protects users from a company changing its mind, but it also means safety filters can be removed. Models built on Stable Diffusion have been misused to make sexual images of real people without their consent. Understanding Stable Diffusion means understanding both sides.

Real-world examples

The August 2022 release, and what came next

On 22 August 2022, Stability AI announced the public release of Stable Diffusion. The model was developed by the CompVis group with the video company Runway, with computing power paid for by Stability AI and trained on images from LAION-5B, a public collection of links to billions of web images (its story is told on the datasets page). Stability's announcement said the model needed 6.9 GB of video memory, within reach of a good gaming PC, and described the licence as permissive for both commercial and non-commercial use.

What followed surprised almost everyone. Within weeks, volunteer programmers had built free interfaces around it. The best known, a web app published under the name AUTOMATIC1111, gave non-programmers a page of buttons and sliders instead of a command line. In January 2023 a developer known as comfyanonymous released ComfyUI, which lays out the image-making process as boxes joined by wires, so people can build and share complex "workflows". Hobbyists learned to train small add-on files called LoRAs, a fine-tuning method that teaches the model one new face, object or style without retraining the whole thing. A website called Civitai became the main place to swap them. Stability followed up with SDXL on 26 July 2023, which worked at a higher resolution and was better at hands and text, both notorious weak spots.

The same openness had a darker side. In December 2023, the Stanford Internet Observatory reported that LAION-5B contained links to suspected child sexual abuse material, and noted that Stable Diffusion 1.5 had been trained on a wide range of material from it. The report recommended that models based on Stable Diffusion 1.5 that had not had safety measures applied should be deprecated and their distribution stopped where feasible. LAION released a cleaned dataset in August 2024. Separately, in May 2025, Civitai banned models made to imitate real people, citing new laws and pressure from payment processors after 404 Media and others reported how such models were used to make sexual images of real people without consent.

A stumble, a licence U-turn, and a change at the top (2024)

2024 was a rough year for Stability AI. The company had been reported to be short of money, and several senior staff left, including key researchers behind the original model. On 23 March 2024, founder and chief executive Emad Mostaque resigned, saying he wanted to pursue decentralised AI. Several of the original researchers went on to found Black Forest Labs, whose image models are covered on the Flux page.

In June 2024 the company released the downloadable weights of Stable Diffusion 3 Medium. Users quickly posted images of people with twisted, broken-looking bodies, especially for prompts such as a woman lying on grass. Many in the community blamed heavy filtering of the training data, though Stability didn't confirm that. The new model's licence also worried businesses. On 5 July 2024, Stability published a new Community Licence, admitting that SD3 Medium "didn't meet our community's high expectations". The new licence made commercial use free below 1 million dollars in annual revenue and removed limits on how many images could be made.

New leadership followed. On 25 June 2024 the company appointed Prem Akkaraju, former chief executive of the visual effects company Weta Digital, as CEO, with Sean Parker as executive chairman. In September 2024 the film director James Cameron joined the board. Stable Diffusion 3.5 arrived on 22 October 2024, with Large and Large Turbo versions, and a Medium version a week later. Since then, the company has put more of its effort into business customers and the entertainment industry.

Getty Images v. Stability AI

On 17 January 2023, the photo agency Getty Images announced that it had started legal proceedings in London, saying Stability AI had copied millions of its photos without a licence to train Stable Diffusion. Users had noticed something telling: some generated pictures contained a garbled version of the Getty Images watermark.

The UK High Court gave judgment on 4 November 2025, and it went mostly Stability's way. Before the end of the trial, Getty had dropped its central claim about the training itself, because it couldn't show the training had happened in the UK. On the main claim that remained, Mrs Justice Joanna Smith held that the model isn't an "infringing copy" under UK copyright law, because it doesn't store or reproduce the photos it learned from. Getty won only a narrow trade mark point: in a small number of examples, some older versions of the model produced images carrying Getty's watermarks, and the judge found no such infringement after the 2.x versions. On 16 December 2025, the judge gave Getty permission to appeal the copyright question, calling it a pure question of law on which reasonable lawyers may differ. Getty also has a separate case in the United States, first filed in Delaware in February 2023 and refiled in California in August 2025. At the time of writing, neither had reached a final answer. The judgment is linked under resources on this page.

Andersen v. Stability AI: the artists' lawsuit

In January 2023, three artists, the cartoonist Sarah Andersen and the illustrators Kelly McKernan and Karla Ortiz, filed a class action in federal court in San Francisco against Stability AI, Midjourney and DeviantArt, whose image tool used Stable Diffusion. More artists joined later, and Runway was added as a defendant. They argue that training on their work without permission was copyright infringement, and that the model itself contains copies of their work in compressed form.

Judge William Orrick threw out parts of the first complaint in October 2023. In an order on 12 August 2024, he allowed the core copyright claims to go forward, including the argument that Stability had encouraged infringement by distributing the model. That is not a ruling that the artists are right. It only means their claims were strong enough on paper to go to the next stage, where both sides gather evidence. A jury trial had been set for 8 September 2026, but at the time of writing it had been moved to April 2027, with a key hearing on the defendants' fair use arguments expected earlier that year. Check for updates, as court dates often slip.

An everyday case: a small shop's product backgrounds

Picture Priya, who sells handmade candles online and needs seasonal backgrounds for her product photos. She tries the free Stable Diffusion demo in her browser, gets some usable autumn tabletops, and then installs ComfyUI on her gaming PC so she can make as many as she likes without paying per image. Her business earns well under the licence threshold, so the current terms allow commercial use. She reads them anyway, keeps her prompts away from named artists and brands, and composites her real candles into the scenes rather than letting the model invent the product.

Try it yourself

Button names and menus change, so treat these steps as a guide to the process.

  1. Start in your browser. Open the Stable Diffusion 3.5 Large demo on Hugging Face, listed under resources on this page. At the time of writing, it runs without an account, with a small daily allowance of free generations. Signing up for a free Hugging Face account raises the allowance.
  2. Write a clear, specific prompt. Describe the subject, setting, style and lighting. Try: A ceramic mug of coffee on a wooden windowsill, rain on the glass, soft morning light, photograph, shallow depth of field
  3. Generate it twice. You'll get two different images, because each one starts from different random noise. That random starting point is called the seed.
  4. Fix the seed and change one thing. Open the advanced settings, switch off "randomize seed", and note the seed number. Now change a single word, such as "morning" to "evening", and generate again. Because the starting noise is the same, you can see exactly what that word did.
  5. Try a negative prompt. In the negative prompt box, list what you don't want, for example blurry, text, watermark. Compare the results.
  6. Adjust the guidance scale. This setting controls how strictly the model follows your words. Try a low value and a high one with the same seed. Too low drifts from your prompt; too high can look harsh or overcooked.
  7. Optional: run it at home. If you have a Windows PC with an NVIDIA graphics card that has at least 8 GB of video memory, or a recent Apple Silicon Mac, you can install ComfyUI from comfy.org, then download a model from inside it. SDXL is a comfortable starting point on 8 GB. Stability says SD 3.5 Medium needs about 9.9 GB, not counting its text-reading parts. The downloads run to several gigabytes. Only get models and add-ons from sources you trust. The parameters page explains why model size decides what your computer can run.
  8. Check the licence before selling anything. Read the licence of the exact model you use, and of any add-on, since community models can carry their own terms.

Common mistakes and misunderstandings

"Open means I can do anything with it." Every Stable Diffusion licence restricts some uses, and the current Community Licence requires a paid licence above 1 million dollars in annual revenue. Add-ons on community sites have their own terms too.

"Stable Diffusion is one app." It's a family of models. The website, app or program you use it through is made by someone else: Stability AI's own platform, a free interface like ComfyUI, or a design tool that runs Stable Diffusion behind the scenes. Quality, safety filters and terms depend on which model and which app.

"It pastes together bits of existing pictures." It doesn't keep a library of photos to cut from, and the UK High Court accepted that the model doesn't store the images it learned from. It can still produce near-copies of things it saw many times, such as famous paintings, logos or watermarks. Treat that as a real risk, especially when you prompt with an artist's or brand's name.

"The court cases are settled." They aren't. The UK copyright question is under appeal, and the US cases were still under way at the time of writing. Be wary of anyone who says confidently that training is either plainly legal or plainly theft.

"A downloaded model is safe because it's popular." Model files and add-ons from unknown uploaders can be poorly made, trained on material you'd never want to use, or packaged in older file formats that can carry harmful code. Prefer well-known sources and the safer "safetensors" file format.

"I need a powerful computer to try it." You don't. The online demo takes a minute. A home setup only makes sense once you want unlimited images, privacy or custom add-ons.

"If it's possible, it's allowed." Making sexual or humiliating images of real people without their consent is harmful and, in a growing number of places, illegal, whether or not the software has a filter. The licence forbids harassment and harm to minors outright.

Related elements