Stable Diffusion is a family of open-weight text-to-image models from Stability AI. The first version was made public on August 22, 2022. You describe an image in words, and the model starts from random noise and removes it step by step until an image that matches the description is left.
Two things set it apart from closed image generators. The weights are published, so anyone can run it on their own GPU, fine-tune it or build a product on it. And a large community has trained thousands of variants on top of it, from photorealistic checkpoints to anime styles.
Stable Diffusion Web is an independent browser app that runs Stable Diffusion models, among others. It is not affiliated with Stability AI.
How Stable Diffusion works
Stable Diffusion is a latent diffusion model, the approach described by Rombach, Blattmann, Lorenz, Esser and Ommer in High-Resolution Image Synthesis with Latent Diffusion Models (CVPR 2022). Earlier diffusion models worked directly on pixels, which is slow and memory-hungry. A latent diffusion model does the expensive work in a compressed space instead.
The original model has three parts:
- A variational autoencoder (VAE). Its encoder compresses an image into a much smaller latent representation; its decoder turns a latent back into pixels.
- A text encoder. Version 1 uses a frozen, pretrained CLIP ViT-L/14 model to turn your prompt into numbers the image model can attend to.
- A U-Net denoiser. Starting from random noise in latent space, it predicts and removes noise one step at a time. At every step it reads the prompt embedding through cross-attention, which is how the words steer the picture.
When the last step is done, the VAE decoder turns the clean latent into the final image. During training the process runs the other way: noise is added to latents of real captioned images, and the U-Net learns to undo it. The first releases were trained on image-caption pairs from LAION-5B, a dataset built from web-crawled data.
What the main settings do
Every Stable Diffusion interface exposes a version of the same controls:
- Prompt. What you want to see: subject, medium, lighting, framing, style.
- Negative prompt. What you do not want, such as "blurry, extra fingers, watermark". The model is pushed away from it at every step.
- Steps. How many denoising passes run. More steps add detail up to a point, then mostly add time.
- Guidance scale (CFG). How strictly the image follows the prompt. Low values give the model more freedom; high values follow the words more literally and can look overcooked.
- Seed. The starting noise. The same seed with the same settings reproduces the same image, which is how you change one thing at a time.
- Size and aspect ratio. Each version has a native resolution it was trained at; far from it, quality drops.
- Strength (image-to-image). With a source image, the model adds noise to that image and denoises it toward your prompt. Strength sets how much noise is added: low values stay close to the source, high values let the prompt take over.
Stable Diffusion versions
| Version | Released | What changed |
|---|---|---|
| Stable Diffusion 1.x | August 2022 (1.5: October 2022) | The original latent diffusion release, trained at 512×512. Most community checkpoints, DreamShaper among them, build on 1.5. |
| Stable Diffusion 2.x | November 2022 | Added native 768×768 generation and a new text encoder. |
| SDXL 1.0 | July 2023 | A UNet about three times larger, a second text encoder and native 1024×1024 output; a 3.5B-parameter base model plus an optional refiner. |
| Stable Diffusion 3 Medium | June 2024 | A 2B-parameter model on a Diffusion Transformer architecture instead of a U-Net. |
| Stable Diffusion 3.5 | October 2024 | Large (8.1B parameters), Large Turbo (a distilled four-step version of Large) and Medium (2.5B parameters). |
Newer is not automatically better for every job. SD 1.5 and its fine-tunes are small and fast and have the largest library of community styles; SDXL is the common middle ground for quality at 1024×1024.
Licenses
The license depends on the version you run:
- Stable Diffusion 1.x uses the CreativeML OpenRAIL-M license.
- SDXL 1.0 is released under the CreativeML OpenRAIL++-M license.
- Stable Diffusion 3.5 uses the Stability AI Community License: free for research and non-commercial use, and free for commercial use by organizations with under $1M in annual revenue; above that, Stability AI requires an enterprise license.
The OpenRAIL licenses are open but carry use restrictions, so read the one that applies before shipping work. A hosted app adds its own terms on top: on Stable Diffusion Web, paid plans include a commercial license and free-tier output is personal-only (see pricing).
Which Stable Diffusion models run on Stable Diffusion Web
Two Stable Diffusion models run here today, with no install and no GPU on your side:
- SDXL, the SDXL 1.0 base model, text-to-image only. You set a prompt, an optional negative prompt, one of seven aspect ratios, a size from 512 to 2048, steps, guidance, a seed and the number of images. It is the default model in the playground on the home page.
- DreamShaper 8, a community fine-tune of Stable Diffusion 1.5, in its LCM (latent consistency model) build, also text-to-image only. The picker labels it Stable Diffusion.
For image to image, the app uses reference-image models such as Nano Banana and Flux Kontext Pro instead. They follow written instructions and have no strength slider, so you say in the prompt what must stay.
Stable Diffusion 2.x, 3 and 3.5 are not offered. The models page lists every image and video model the app runs, Stable Diffusion and otherwise.
How to try it
- Open the playground and keep SDXL selected.
- Write a prompt in the order subject, medium, setting, framing, lighting, style. The prompt guide explains the order and the Stable Diffusion prompts page has 20 templates to copy.
- Add a negative prompt only for things you can see and do not want, fix the seed, and change one setting at a time.
- When the composition matters more than novelty, open image to image and give a reference-image model the picture you want to keep.
Sources
- Rombach et al., High-Resolution Image Synthesis with Latent Diffusion Models, CVPR 2022.
- Podell et al., SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis, 2023.
- Stability AI, Announcing SDXL 1.0.
- Stability AI, Stable Diffusion 3 Medium.
- Stability AI, Introducing Stable Diffusion 3.5.
- Stable Diffusion, Wikipedia.