What is a diffusion model? An AI that makes a picture by starting from random noise and cleaning it up step by step

An AI that makes a picture by starting from random noise and cleaning it up step by step.

A diffusion model doesn't draw a picture in one go. It starts from pure static, like a TV with no signal. Then, over and over, it asks: which bits of noise should I remove to get closer to the description? After many small steps, a clear image appears.

Training works the other way round. The model sees real photos with more and more noise mixed in, and practises guessing exactly which noise was added so it can take it away. Many well-known image generators since about 2022 use this idea, and it is used to make video too.

Think of it like this

Think of staring at clouds. Once you decide one looks like a dog, you start to see the ears and the tail. A diffusion model does something similar with random dots: guided by your words, it nudges them step by step until the dog is really there.

For example

Type "a cat on a windowsill on a snowy night" and the model works its way from static to that scene over a series of steps.

Common mix-ups

Mix-upAn image AI draws the picture in one go.
ActuallyThe method widely used today doesn't polish the full-size picture directly: it removes the noise over many rounds on a small compressed "summary" version, then expands it at the end. This "latent diffusion" approach from 2021 cut the computing a lot and became the basis of Stable Diffusion.
Mix-upA diffusion model cuts and pastes photos into a new picture.
ActuallyInside the model there are no photo files, only the numbers it uses to guess the noise. That's why it can draw combinations no photo has ever captured, like "a cat swimming in a spacesuit".

Try it in class

For teachers at the start of a lesson — or for students working on their own.

Question: Fewer steps would be faster — so why take many?
Activity: Give each group a sheet covered in random dots and a description (say, "a cat by the window on a snowy night"). In five stages, rub out dots and add lines to match it, photographing each stage. Line the photos up and compare when the shape started to show. Open the Latent Diffusion (E0134) and Sora (E0329) records to follow the path from still pictures to moving ones.