Learn AI /Generative Media /Diffusion model
Diffusion model
- Intermediate
- Generative Media
- Technical readers
Also called: diffusion.
Back to the milestone you came from
In five seconds
What it is
A diffusion model creates media by learning how to remove noise step by step until a coherent image, audio clip or other output appears.
In plain words
A diffusion model makes new media by starting with random noise and gradually cleaning it up until it matches a prompt or another guide.
An everyday example
Imagine a design class in Addis Ababa, Bogotá and Panama City sharing the same exercise online. Each student starts with a screen full of television static and a note saying, "make this look like a red bicycle beside a market stall". On each round, they compare the noisy screen with examples they studied earlier and remove a little static while keeping what seems to fit the note. After many rounds, the static turns into a picture that roughly matches the description. If the note changes to "at night in the rain", the students steer the clean-up in a different direction and a different picture emerges.
Real diffusion models do not literally see a hidden finished picture inside the noise waiting to be uncovered. They generate a new output by applying learned numerical transformations step by step. Also, students in the story consciously interpret instructions, while the model only matches patterns in data and guidance signals.
Remember thisA diffusion model generates media through learned, iterative denoising, often guided by a prompt. Plausible output is not evidence that the depicted content is real or accurate.
Understand it
Understand it
Why it matters
Non-specialists may meet diffusion models in image, audio and video tools, procurement documents, media literacy discussions and authenticity checks. The term matters because it signals a specific way content can be generated: from noise through many refinement steps. That affects speed, controllability, cost and common failure modes. It also matters for risk assessment. Outputs can look convincing even when they contain errors, fabricated details or misleading context. If you recognise the term, you can ask better questions about training data, bias, consent, provenance, watermarking, human review and whether a tool is suitable for a real task.
How a diffusion model generates an image
After viewing this, a reader should be able to say that generation starts with noise and becomes an image through repeated denoising steps guided by a prompt.
01
Start with noise
The process begins with random pixels that contain no clear object or scene.
Draw a square filled with grey static dots. Label it 'random noise'.
02
Add guidance
A prompt or other input guides what kind of result should emerge.
Place a text box above the next panel reading 'Prompt: red bicycle beside a market stall'. Add an arrow from the prompt to the image sequence.
03
First denoising steps
The model predicts how to make the noise slightly less random while staying consistent with the guidance.
Show a second square with heavy noise but faint red and wheel-like shapes. Connect with an arrow labelled 'step 1...n'.
04
Repeated refinement
Many small updates gradually strengthen shapes, colours and layout.
Show a third square with a recognisable bicycle and stall, still blurry. Use another arrow labelled 'more denoising'.
05
Output
After enough steps, the result is a coherent generated image.
Show a final clear illustration of a red bicycle beside a market stall. Label it 'generated image'.
Show a left-to-right sequence of panels. Panel 1 is pure noise. Panel 2 adds the prompt. Panels 3 and 4 show intermediate noisy-but-emerging versions. Panel 5 shows the final image. Read from left to right as repeated refinement, not as one sudden jump.
Hovering or tapping each panel reveals a short caption about what changed at that stage; a slider can scrub through early, middle and late denoising states.
Describe this without the diagram
The diagram explains diffusion image generation as a sequence of five stages. First, there is only random visual noise. Second, a text prompt is introduced to guide the process. Third, the model performs early denoising, where faint shapes begin to appear within the noise. Fourth, repeated denoising steps make the scene clearer and more organised. Fifth, the process ends with a coherent generated image. The key idea is that the output is built gradually through many guided clean-up steps, not produced all at once.
With reduced motion: A static five-panel strip replaces animation. Each panel shows one stage: random noise, prompt added, early noisy structure, clearer intermediate image, and final output image.
A diffusion model is a kind of generative model. A generative model produces new content such as pictures, sound or video after learning patterns from many examples. In diffusion systems, the usual starting point is random noise, which looks like static.
During training, the system learns a reverse process. It sees clean data corrupted with different amounts of noise, and it learns how to reverse that corruption one step at a time. In simple terms, it practises turning mess into structure.
At use time, the model may receive a prompt, such as a text description, or another input such as a sketch, image or audio clip. It starts from noise and repeatedly refines it over many small steps. Each step updates the sample so that it better reflects learned patterns and any conditioning input.
What comes out depends on the type of model: an image, a short audio segment, video frames, or another kind of generated content. The output can look realistic or stylised, but it is still a statistical construction based on patterns in training data and guidance signals. It does not guarantee factual accuracy, authorship, consent or authenticity.
How “Revealing a picture from television static” maps onto the real thing
| In the story | In the AI system |
|---|---|
| screen full of static | random noise used as the starting point |
| note describing the scene | text prompt or other conditioning input |
| removing a little static each round | iterative denoising over many steps |
| examples studied earlier | patterns learned from training data |
A worked example
A communications officer in Geneva needs a non-photorealistic visual for a side-event concept note before it goes for clearance.
Starting from: Prompt: "Flat square illustration of a red bicycle beside a fruit stall at dusk, warm lights, no people, no logos"
- The system converts the prompt into conditioning signals that guide image generation.
- Sampling begins with random noise in the model's working representation.
- At each step, the trained model predicts noise or a related quantity and uses that prediction to make the sample slightly less noisy. Guidance pushes the evolving image towards the prompt.
- Across repeated steps, broad composition appears first, followed by colours, edges and smaller details. If the model works in a compressed representation, the final sample is decoded into an image.
Result: The officer receives a new illustration broadly matching the prompt. It may still invent awkward spokes, fruit shapes or lettering, so the image requires review before it is inserted into the concept note.
In a United Nations workplace
During a records retention review at a duty station working in three languages, a monitoring and evaluation officer finds AI-generated header images in several monitoring reports. Before the files are transferred, the officer asks the information management focal point whether a diffusion model produced them and what prompts, source files and provenance records were retained.
Understanding diffusion helps the officer distinguish generated illustrations from documentary images. The model began with noise and iteratively produced a plausible output under guidance, rather than retrieving a photograph of an actual event. That distinction matters when preserving the reports and recording how their visuals were made.
Before you try thisGenerated images must not be treated as monitoring evidence merely because they look realistic. Personal data, protection-sensitive images or unpublished report content must not be entered into a service that has not been approved for such material, because prompts and uploads may be retained or used by the provider.
Where you may hear it
- A vendor demo before an ICT governance board
- A procurement unit reviewing requirements for a media tool
- Clearance of a background note on synthetic media
- An inter-agency working group discussing image provenance
Avoid confusion
Avoid confusion
Often confused with
Generative adversarial network
A training method with two neural networks, one making synthetic examples and one trying to spot fakes, with the aim of making generated output more realistic.
Variational autoencoder
A generative model that compresses examples into a regularised probabilistic latent space, then reconstructs inputs or samples from that space to create new outputs.
Autoregressive model
A model that assigns each token a probability based on earlier tokens in the same sequence, and generates by predicting and adding one new token at a time.
Common misconceptions
Common misconception: A diffusion model finds a hidden real picture inside random noise.
Correction: It does not uncover a pre-existing photograph. It starts from noise and applies learned transformations that gradually create a new output consistent with the prompt or other guidance. The process is constructive, not a search for something already there.
Common misconception: If the result looks realistic, it must reflect something true or somewhere real.
Correction: Realistic appearance only means the output matches visual or audio patterns the model learned. It may combine details that never existed together, invent objects, or imply events that did not happen. Authenticity still needs separate checks such as provenance, source review and human verification.
Common misconception: Diffusion is just another name for a generative adversarial network.
Correction: They are both generative approaches, but the mechanism is different. GANs usually train a generator and a discriminator in opposition. Diffusion models usually learn a denoising process and generate through many iterative refinement steps. People confuse them because both can produce synthetic media.
Quick check
Which statement best describes how a diffusion model usually generates an image?
The usual diffusion process begins with noise and refines it over many steps. The first statement describes retrieval-based editing. The third describes autoregressive generation. The fourth describes adversarial training rather than diffusion.
Go deeper
Go deeper
The simple version
A diffusion model makes new media by starting with random noise and gradually cleaning it up until it matches a prompt or another guide.
The technical version
A diffusion model learns a parameterised reverse-noising process. During training, clean samples are progressively corrupted by a forward diffusion process that adds noise over multiple time steps. The model is trained to predict the noise, the clean sample, or a related quantity at each step, often conditioned on text, images, audio or other inputs. At generation time, sampling begins from noise and iteratively applies the learned reverse steps, sometimes with guidance methods that push the sample toward the conditioning signal. In practice, many modern image systems use latent diffusion, where denoising happens in a compressed latent space rather than directly on full-resolution pixels, which improves efficiency. Compared with autoregressive models, which generate one token or patch after another, diffusion models typically refine the whole sample over repeated steps.
Reviewed external resources
Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
This UNESCO recommendation gives policy and ethics context for AI use and suits readers concerned with governance, rights and responsible deployment rather than model mechanics.
Intermediatelink checked 2026-07-31
Primary sourceContent Credentials (opens in a new tab)Content Authenticity Initiative
This Content Authenticity Initiative resource explains provenance and content credentials concepts and suits readers assessing how to signal or verify the origin of generated media.
Starterlink checked 2026-07-31
EncyclopaediaDiffusion model (opens in a new tab)Wikipedia
This Wikipedia page gives a broad overview of the term and suits readers who want a general starting point with some technical vocabulary.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Diffusion model (opens in a new tab)YouTube search
This YouTube search results page helps readers find video explanations in different styles and depths, and it suits people who learn better through visual walkthroughs; results will vary over time.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Diffusion model (opens in a new tab)Reddit search
This Reddit search results page helps readers see community questions, examples and debates about the term, and it suits readers who want informal discussion; as a search page, quality and accuracy will vary.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
A diffusion model generates media through learned, iterative denoising, often guided by a prompt. Plausible output is not evidence that the depicted content is real or accurate.