How does Stable Diffusion work?
Stable Diffusion is a latent diffusion model. During training it learned to remove noise from hundreds of millions of images paired with captions. At generation time it starts from pure random noise and, step by step, denoises it in the direction of your text: after 20 to 30 steps the noise has become a picture that matches the prompt.
Three parts cooperate: a text encoder that turns your words into numbers, a U-Net (or, since version 3, a diffusion transformer) that does the denoising, and a decoder that converts the compact latent image into pixels. The denoising runs on that compact latent image, which is what makes the model fast enough for a consumer GPU.
Which versions exist?
- Stable Diffusion 1.5 (2022): 512×512 images, thousands of community fine-tunes and LoRAs.
- SDXL 1.0 (July 2023): 1024×1024 native, better hands, faces and composition, a refiner model for details.
- Stable Diffusion 3 Medium (June 2024): diffusion transformer, much better text rendering and prompt understanding.
- Stable Diffusion 3.5 (October 2024): the current release, in Large, Large Turbo and Medium sizes, with better prompt adherence and more varied styles.
See the detailed comparison in SDXL vs Stable Diffusion 3.
What can you do with it?
Text-to-image is only the start. The same models handle image-to-image (transform a sketch or photo), inpainting (repaint one area), outpainting (extend a picture), upscaling with dedicated super-resolution models, and, with ControlNet, generation guided by a pose, a depth map or edges. Add-ons called LoRAs, each a small separate file, teach the model a style or a character.
On this site you can try inpainting and upscaling without an account, generate images with SDXL or Stable Diffusion 3, and browse 1,731 artist styles and 100+ documented prompts to learn what a prompt does.
How is it different from DALL-E or Midjourney?
Two things: the weights are public and it runs locally. You can download the model, run it on your own computer for free, modify it, and keep your images private. Midjourney and OpenAI's image models (DALL-E, now GPT Image) are closed services: you pay by subscription or per image, and the model never leaves their servers. In exchange, Stable Diffusion asks more of you. You have to choose a model, write a precise prompt and a negative prompt, and understand the seed.
Do I need a powerful computer?
To run it yourself, a recent NVIDIA card with 6 GB of VRAM is enough for SD 1.5, and 8 GB for SDXL. AMD cards work through ROCm on Linux, and Apple Silicon through the MPS backend. If you have none of that, the tools on this site need no hardware: some run on our GPU, with the queue and progress shown live, others are hosted demos embedded in the page. Read how to install Stable Diffusion locally when you are ready.