ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›AI MEANINGS›WHAT IS A DIFFUSION MODEL?
AI MEANINGS · MEANINGBEGINNER

What is a diffusion model? AI that turns noise into pictures

A diffusion model is the AI behind many image makers: it starts from random static and cleans it up, step by step, into the picture you described.

READ6 min
WORDS1,108
SECTIONS7
TYPEMEANING
CHECKED1 OCT 26

A diffusion model is a kind of AI that makes pictures (and video) by starting from random static and cleaning it up, step by step, until an image matching your description appears. It is the technology behind many AI image makers.

◈ IN PLAIN TERMS

Think of a steamed-up bathroom mirror. Each wipe clears a little more fog, and a picture slowly appears. A diffusion model starts with pure “fog” (random noise) and wipes it away in many small steps. Your words decide what picture it is wiping towards.

In a sentence: “The image tool is a diffusion model, so you can watch the picture sharpen out of grey noise while it works.”

How it works, simply

It learns by watching pictures get ruined, then learns to undo the damage. In training, the model takes real images and adds a little random noise at a time, until nothing is left but static. It learns how to reverse each step. IBM puts it plainly: diffusion models add noise, “then refine the mess they created into new, similar data.”1

Once trained, it can start from fresh static and “clean” it into a brand-new image that never existed. Your text prompt steers each step. A widely cited 2020 research paper showed the method could make high-quality images.2

Where the word comes from

The idea was set out in 2015. A paper by Jascha Sohl-Dickstein and colleagues, drawing on ideas from physics, described it: slowly destroy the structure in data with a “forward diffusion process”, then train a model to run it backwards.3 A 2020 paper by Jonathan Ho, Ajay Jain and Pieter Abbeel showed it could produce high-quality images.2

A simple example

You type a description; static turns into your picture. Say you ask for “a red bicycle leaning against a blue door, morning light”. The model begins with something like the snow on an old untuned TV. Step by step, shapes form, then colours, then fine detail, until the bicycle and door are there. Run it again and you get a different bicycle, because it started from different static.

FROM STATIC TO PICTURE · HOW A DIFFUSION MODEL WORKS
Like wiping fog off a mirror, one small wipe at a time.
Learned in trainingIt watched real pictures turn to static, and learned to undo it.
Start from staticRandom noise, like an untuned TV.
Clean up, step by stepShapes, then colours, then detail. Your prompt steers each step.
A new imageOne that never existed. New static, new picture.
Reasoning, 1 Oct 2026 — summarises this page’s sections on how it works (IBM; Ho, Jain & Abbeel) and the bicycle example.

IBM names well-known tools that work this way: Midjourney, Stable Diffusion and OpenAI’s DALL-E 2.1 IBM also notes the same family of models is used for video.4

Why it matters to you

Knowing how it works explains its odd habits.

  • Every run is different. New static means a new picture. Many tools offer a “seed” setting so you can repeat a result you liked; the prompting terms glossary explains seeds and other settings.
  • Small details can go wrong. Hands, text on signs and fine patterns are where the clean-up often slips. Why AI output looks fake covers the common tells.
  • Describe the picture, not a story. The model is steering towards an image, so concrete words about subject, light and setting help most. How to use reference images shows another way to steer it.

Often confused with

TermHow it differs
Chatbot models (LLMs)Differs: chatbots write text one piece at a time, left to right. A diffusion model refines a whole image at once, over many passes.
Stable DiffusionDiffers: that is the name of one family of image tools. “Diffusion model” is the general kind of technology.
Photo filtersDiffers: a filter changes a photo you already have. A diffusion model makes a new image from scratch.

Questions people ask

Is Midjourney a diffusion model?

IBM lists Midjourney, Stable Diffusion and DALL-E 2 among the image tools that diffusion models power.1 That list dates from 2024; for how a specific tool works today, check its maker’s own pages.

Why does a diffusion model start from noise?

Because that is how it was trained: it learned by watching images get buried in noise, step by step, and then practising the reverse. Making a new picture is running that clean-up on fresh random noise.1

Learn more on this site

Related on this site: Negative Prompts for Stable Diffusion and SDXL · What Is CFG Scale? · What Is ControlNet? · What Is Img2img in AI? · What Is Latent Space?

QUICK CHECK · 3 QUESTIONS

Pick an answer to see if you got it. Nothing is recorded.

1. How does a diffusion model make a picture?

  • It starts from random static and cleans it up step by step
  • It searches the internet for a matching photo
  • It writes the picture one piece at a time, left to right
Show the answer

Answer: It starts from random static and cleans it up step by step. Like wiping fog off a mirror, it starts from random noise and clears it in many small steps, steered by your words.

2. Why do you get a different picture each time you run the same prompt?

  • The model forgets your words between runs
  • It starts from different random static each time
  • Your internet connection changes the picture
Show the answer

Answer: It starts from different random static each time. New static means a new picture, which is why each run differs unless a tool lets you repeat the seed.

3. What kind of words does the page say help most in an image prompt?

  • A long story about what happens next
  • Words about how you feel today
  • Concrete words about subject, light and setting
Show the answer

Answer: Concrete words about subject, light and setting. The model is steering towards an image, so concrete words about subject, light and setting help most.

Related terms: AI upscaling · CFG scale (guidance scale) · ControlNet · Image-to-video · Img2img (image-to-image) · Inpainting and outpainting

SOURCES
  1. 1IBM, What is a generative model?, 11 Nov 2024, read at source 1 Oct 2026“Diffusion models gradually obfuscate or diffuse input data by adding noise, then refine the mess they created into new, similar data.” ↑a“Diffusion models power mainstream image generation and image synthesis tools such as Midjourney, Stable Diffusion and OpenAI’s DALL-E 2.” ↑b“Diffusion models power mainstream image generation and image synthesis tools such as Midjourney, Stable Diffusion and OpenAI’s DALL-E 2.” ↑c“Diffusion models gradually obfuscate or diffuse input data by adding noise, then refine the mess they created into new, similar data.” ↑d
  2. 2Ho, Jain & Abbeel, Denoising Diffusion Probabilistic Models, arXiv, 19 Jun 2020, read at source 1 Oct 2026“We present high quality image synthesis results using diffusion probabilistic models”. ↑a“We present high quality image synthesis results using diffusion probabilistic models”. ↑b
  3. 3Sohl-Dickstein, Weiss, Maheswaranathan and Ganguli, Deep Unsupervised Learning using Nonequilibrium Thermodynamics, arXiv:1503.03585, submitted 12 Mar 2015, read at source 12 Oct 2026“systematically and slowly destroy structure in a data distribution through an iterative forward diffusion process.” “We then learn a reverse diffusion process that restores structure in data” ↑
  4. 4IBM, AI model types: past, present and predictions for the future, 13 Oct 2025, read at source 1 Oct 2026“Then came generative adversarial networks and diffusion models for image and video generation.” ↑

The comparisons and the example on this page (the steamed-up mirror, the TV static, the red bicycle) are this site’s own ways of explaining, not claims from these sources.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE