Skip to content

Stockertim Dictionary · Artificial intelligence

Diffusion model

Also called diffusion models, image diffusion model

In short

A type of generative model that creates an image by starting from random noise and removing that noise step by step, guided by a written prompt.

  • Intermediate
  • Technology
  • Checked

In detail

A diffusion model is a kind of generative AI that learns to turn random visual noise into a coherent picture. During training it is shown a very large number of images that have had increasing amounts of noise added to them, and it learns to predict and remove that noise. Once trained, it can run the process in reverse: it begins with pure static and refines it, step by step, into an image.

The written prompt steers each step. Text is converted into a numerical representation, and the model uses it to decide which shapes, colours and textures are more likely at every stage. Settings such as the number of steps, the image size and a starting number called the seed affect the result. More steps usually add refinement but take longer, and the same prompt with a different seed gives a different image.

Because the model builds images from learned patterns rather than from an understanding of objects, it can struggle with things that need exact structure, such as readable lettering, the right number of fingers or consistent logos. It can also repeat biases found in its training images. Knowing how the process works helps a designer write clearer prompts and expect where errors are likely.

At work, diffusion models power most image generators used for concept art, mock-ups and early ideas. Designers treat their output as a draft to be reviewed, edited and checked for rights and brand fit before it reaches a client.

Worked example

A designer at Northwind Bikes generates four options for a poster background.

  1. The tool runs 30 refinement steps for each image: 30 x 4 = 120 steps in all.
  2. Each image starts from a different seed, so the four results differ even though the prompt is the same.
  3. The designer notices that lettering on a shop sign in one image is garbled and plans to add real text later.

Four distinct backgrounds are produced; the designer keeps one as a starting point and adds the headline as real type.

In practice

Designers meet diffusion models whenever they use an image generator for concepts, backgrounds or variations. Teams note which outputs were generated, keep the prompt and settings with the file, and replace any generated text, logos or faces that need to be exact or properly licensed.

Watch out

  • A diffusion model does not search for or copy a stored picture; it builds a new one from learned patterns, though output can still resemble training images.
  • Garbled text in a generated image is a known weakness, not a sign the prompt was wrong.

Learn it properly

Official sources

Facts last checked on 3 November 2026.

Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.

English