Diffusion model explained simply: how AI turns noise into images and videos

The basic idea: noise becomes a picture

Imagine an old TV screen showing nothing but flickering snow. A diffusion model starts with exactly that kind of random noise and removes it in many small steps until a clear image remains. What comes out is steered by your description, for example: a lighthouse at sunset, painted like a watercolour.

The model learned this skill backwards, so to speak. During training, real images are made noisier step by step until only noise is left. The model practises reversing this process, turning a noisy image into a slightly clearer one. IBM describes this as a forward and a reverse process. In 2024, OpenAI explained the same principle for its video model Sora: the model receives noisy image patches and learns to predict the original, clean patches.

This leads to something you notice straight away when trying it: when a run starts with fresh random noise, the same description usually produces a slightly different image. If you fix the random starting value (the seed) and leave the other settings unchanged, a result can be repeated instead. According to the developer platform Hugging Face, it may still differ slightly on different hardware.

How your text steers the image

So that the model knows what to work out of the noise, your description is first translated into numbers by a language model. When Google presented its image model Imagen in 2022, it found that a larger language model for understanding the text improved quality more than a larger image model. Imagen first created a tiny image of 64 by 64 pixels and then enlarged it to 1024 by 1024 with further diffusion models.

An important trick came from Munich: Björn Ommer's research group at LMU Munich developed latent diffusion models, which became known as Stable Diffusion. The model does not work on the full-size image but on a heavily compressed version, from which the finished image is only created at the end. This saves a lot of computing power. In the model card for its video model Veo 3, Google calls latent diffusion the de facto standard for modern image, audio and video models.

Well-known models in October 2026

Google: according to Google, Imagen is a diffusion model. For video, Google uses Veo; according to its model card, Veo 3 uses latent diffusion and generates picture and sound together.

Black Forest Labs: the company, headquartered in Freiburg, Germany, presented FLUX.2 in November 2025 and describes its technique as latent flow matching. Since July 2026, FLUX 3 has been in early access, a single model for images, video and audio, also based on flow matching. Our assessment: flow matching is a further development of the same basic idea, where the result also emerges step by step from noise. The provider itself does not call it diffusion, though.

OpenAI: according to OpenAI, Sora was a diffusion model. OpenAI discontinued the Sora app and website on 26 April 2026 and the developer API on 24 September 2026. We could not reliably establish whether, and in what form, there is a successor video model. The image generation that OpenAI built directly into ChatGPT in March 2025 is, according to OpenAI, explicitly not a diffusion model but creates the image piece by piece, similar to how a language model writes text (autoregressive).

Midjourney: according to its documentation, the current default is version V8.2. We found no information on the underlying technique in the documentation.

What this means for you in everyday life

A diffusion model invents an image from scratch; it is not a photo of something that really existed. That is why details can be wrong. Google lists typical signs of AI images, including illegible or inconsistent text, errors in hands, teeth or the background, inconsistent shadows and lighting, and unnaturally repeating patterns. You cannot rely on such errors, though.

When trying it yourself: describe subject, style, light and framing as precisely as possible, and have several variants generated, because with a fresh random start every run usually turns out differently.

Many providers now add invisible watermarks to their images. How this works and how to check an image is explained in our article on AI watermarks. How to spot fakes in everyday life is covered in our article on AI fakes in images, voices and calls.