Diffusion model: creating an image by removing noise
A diffusion model is a type of generative AI that produces an image, a sound or a video by starting from random noise, like a TV with no signal, then cleaning it up step by step until a coherent result emerges.
Data as of October 22, 2026
Key figures
| Level | Intermediate |
|---|---|
| Origin | 2015 |
| Introduced by | Sohl-Dickstein et al. |
The principle
During training, millions of images are gradually damaged by adding noise until only random static is left. A neural network learns the reverse task: given a noisy image, guess the noise to remove. Once trained, it can start from pure noise and “denoise” it many times in a row, until an image that never existed emerges.
A concrete example
You ask an image generation tool for “a lighthouse in a storm, watercolor style”. The model draws a random noise image. At each step, it removes a little noise while being guided by your text: first blocks of color, then a vertical silhouette, then the sea, the light, the texture of the paper. Change the starting noise and you get a different lighthouse.
Origins and why it matters
The approach was described as early as 2015 by Sohl-Dickstein and co-authors, drawing on statistical physics. In 2020, work by Ho, Jain and Abbeel (DDPM) showed it could produce high-quality images. In late 2021, Rombach and co-authors proposed working in a compressed space (latent diffusion), which sharply reduces the computation required; this is the basis of Stable Diffusion.
These models have become a pillar of visual generative AI. They also explain why realistic images, including deepfakes, are now so easy to produce.
Common mistakes and myths
- “The model pastes together pieces of existing images.” It builds a new image from noise, drawing on what it has learned. Copyright questions around training images are still being debated, however.
- “Diffusion and transformers are opposites.” Diffusion is a generation method; the network that does the denoising can itself be a transformer.
- “It’s instant.” Each image requires several successive passes through the network, which makes generation more expensive than a single prediction.
Related terms: generative AI, image generation, deepfake, neural network.
Related terms
- Image generation · When an AI creates a brand-new picture from a written description (a prompt) or from another image.
- Generative AI · AI that creates new content such as text, images, audio, video or code from a simple written or spoken request.
- Deepfake · An AI-made or AI-altered image, video or voice recording that makes a real person appear to say or do something they never did.
- Neural network · Software made of many small connected computing units that learns by adjusting the strength of the links between them.
- Transformer · A neural network architecture built on attention, introduced in 2017, that underpins almost every large language model today.
Full glossary
Frequently asked questions
Where does the name “diffusion” come from?
From an analogy with physics: like a drop of ink spreading through water, the original image is gradually scrambled by noise. The model learns the reverse path, from noise back to a clean image.
Why does the same prompt give different images each time?
Because the starting point is a different random noise every time. Many tools let you fix that starting point (the “seed”) to reproduce an image.
Are diffusion models only used for images?
No. The same principle applies to audio, music and video. Images remain the best-known use among the general public, though.
Who invented diffusion models?
The idea was described in 2015 by Jascha Sohl-Dickstein and co-authors. It was made far more effective in 2020 by Jonathan Ho, Ajay Jain and Pieter Abbeel, then cheaper to run in 2021 with latent diffusion.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.