Skip to the content
AI Explainedpreview Learning Center

AI Explained

How an AI makes an image

It does not paint. It starts with noise and takes it away.

From noise to picture

Removing noise, many times

An image model does not paint. It starts with random noise and takes a little of it away, again and again, steered by your words.

steersYour wordsA descriptionText encoderTurns words intonumbersPure noiseA random startDenoiserRemoves a little noise, manytimesDecoderSmall picture to fullsizeThe imageFull size

Press Play to walk through it, or Listen to have it read to you.

Read all the steps
  1. Your description is turned into numbers the image model can use.
  2. The picture starts as pure random noise.
  3. The denoiser removes a little noise, steered by your words.
  4. It repeats that many times. Each pass, the picture gets a little clearer.
  5. All of this happens on a small, compressed version of the picture, which is much cheaper than working on every pixel.
  6. A decoder turns the small result into the full-size image.

Try it

Watch a picture come out of noise

Drag the slider from noise to picture, or press Play. Each step keeps what looks like the picture and throws away some of the noise.

An illustration of the idea, drawn by blending noise with a finished picture. A real model does not have the picture: it has learned what pictures look like and predicts the noise to remove.

The idea

Four things to know

Text and images are made differently

A language model writes one token after another. An image model usually starts from a whole picture of noise and improves all of it at once, in many passes.

Denoising

Diffusion models break the making of an image into a sequence of denoising steps [1]. The 2020 paper that made them practical describes them as a class of latent variable models inspired by thermodynamics [2].

A small picture first

Working directly on every pixel is very costly, so latent diffusion runs the process in the compressed space of a pretrained autoencoder and enlarges the result afterwards [1].

How your words steer it

Cross-attention layers let the model take a condition such as text into account at every step [1].

Sources

Where this comes from

Read 2026-11-15, unless a source gives its own date. A number in brackets points to the page a statement was read on. Sentences with no number, the small demonstrations and the tools are this center's own explanations.

  1. Rombach and others: High-Resolution Image Synthesis with Latent Diffusion Models (2021)
  2. Ho, Jain and Abbeel: Denoising Diffusion Probabilistic Models (2020)