Learn AI /Generative Media /Image generation

Image generation

Starter
Generative Media
Communications teams

Also called: text-to-image, AI images, image generator.

Detail

In five seconds

What it is

Software can create new pictures from prompts or other inputs by learning visual patterns from many examples, then producing a fresh image rather than retrieving one file.

In plain words

You give the tool a description or other prompt, and it makes a new picture that matches the request as best it can from patterns it learned during training.

An everyday example

A school in Geneva, Amman and Sarajevo is preparing a cultural fair poster. The organiser gives three illustrators the same short brief: “children carrying lanterns at dusk, warm colours, joyful mood”. Each illustrator produces a different poster because each has learned different visual habits from past work. If the organiser adds more detail, such as “include mountains in the background and leave space for text”, the next versions usually move closer to what was asked. The posters are newly drawn, not copied as one existing poster from a drawer.

Where the story stops working

Human illustrators can explain their choices, use broad real-world judgement and may deliberately check facts. An image generator does not work like a person receiving instructions. It uses statistical pattern matching in numerical form. Human experience is not equivalent to model training, and many image generators are trained on large datasets and can produce results in seconds.

Remember thisImage generation creates new pictures from prompts or other inputs, and a convincing result can be useful without being true, accurate or appropriate.

Understand it

Why it matters

Non-specialists need to recognise two things at once: these tools can speed up creative work, and they can also create false impressions. That matters when you choose images for reports, campaigns, teaching, journalism, fundraising or internal presentations. A generated image may save time for a mock-up or concept sketch, but it may also include errors, stereotypes, invented scenes or visual clues that imply something untrue. Knowing the term helps you ask practical questions: What was the input? Was this generated or edited? Is it clearly labelled? Can viewers mistake it for documentary evidence? Was approved data handling followed? Those questions support better decisions about trust, communication, rights and reputational risk.

How text-to-image generation turns a prompt into a picture

After viewing this, the reader should be able to describe the main stages from prompt to generated image and explain how text-to-image relates to the broader idea of image generation.

01
1. Input
The user provides a prompt, and sometimes also a reference image, style choice, size or other settings.
Draw a rounded box labelled 'Input' containing a text line icon, a small image icon and short labels: prompt, optional reference, settings.
02
2. Start from noise
Many systems begin with random visual noise, which looks like static rather than a meaningful picture.
Draw a second box labelled 'Random starting pattern' with a grey speckled square. Connect from Input with an arrow.
03
3. Guided refinement
The model repeatedly adjusts the pattern so it better matches the prompt and learned visual patterns.
Draw a third box labelled 'Refine in steps' showing three tiny squares from blurry to clearer, with a side label 'guided by prompt'.
04
4. Output image
The result is a newly generated image. It may be useful, but it is not evidence that the depicted event happened.
Draw a final box labelled 'Generated image' with a clear thumbnail and a small caution badge reading 'realistic does not mean real'.

A left-to-right flow with four boxes connected by arrows. Read from input to output: a text prompt enters the generator, the model builds a rough image from noise, repeated refinement adds details, and the final image is shown with a note that it may look convincing without being real.

On hover or tap, each box reveals a one-sentence note with the explanation for that stage; keyboard focus moves left to right through the four stages.

Describe this without the diagram

This diagram explains image generation as a four-stage flow. First, the user gives input such as a text prompt, and sometimes a reference image or settings. Second, many systems begin with random visual noise. Third, the model refines that noisy pattern in repeated steps so it better matches the prompt and learned visual patterns. Fourth, it outputs a new image. The key message is that the image is generated from learned patterns and prompt guidance, so a convincing picture is not proof that the scene was real.

With reduced motion: A static four-box flow appears left to right with arrows between the stages: Input, Random starting pattern, Refine in steps, Generated image. The refinement stage shows three still thumbnails from noisy to clearer instead of animation.

Image generation means using software to produce a new image. The input might be a text prompt, another image, or both. When the input is text, people often say text-to-image. The output is often called an AI image. Image generator usually refers to the tool or system that makes it.

These terms are related but not identical. Image generation is the broad activity. Text-to-image is one common method inside that broad activity. AI images are the results. An image generator is the product or model used to create them. This differs from image classification, which labels an existing picture instead of making a new one. Generated images can be synthetic media, but synthetic media also includes heavily altered images, audio, video and text. A deepfake imitates a real person, scene or event. Watermarking and provenance provide origin information rather than generating the image.

In simple use, a person types a prompt such as “a red bicycle beside a market stall at sunrise, painted in watercolour”. The system turns that prompt into internal numerical patterns and then builds pixels that match the described features. Some systems also use a starting image, a style reference, or settings such as size and aspect ratio.

Many modern systems use a diffusion model. A model is a mathematical system trained on examples. During training, a diffusion model learns to reverse a process that adds noise. During generation, many such systems start with random noise and apply the learned denoising process step by step, guided by the input. Other approaches also exist, such as generative adversarial networks and variational autoencoders. Generation alone does not verify facts about the real world.

That is why these tools can be useful and risky at the same time. They can help with sketching ideas, storyboards and illustrations, but they can also produce misleading, biased or impossible images. A polished result does not prove that the scene really happened.

How “A poster brief for three illustrators” maps onto the real thing

In the story	In the AI system
the organiser's brief	the prompt or other input
the illustrators' past practice	patterns learned during training
different posters from the same brief	multiple possible outputs from one prompt
adding more detail to the brief	prompt refinement and settings that guide the result

A worked example

A public information officer in a regional bureau needs a draft visual for a side-event concept note before it goes to the front office for clearance.

Starting from: Prompt: "Landscape illustration of an empty conference table with interpretation headsets, nameplates without text and a projection screen showing abstract data patterns, blue and gold palette, no logos, no identifiable people."

The officer enters the prompt and selects the required landscape format.
The system converts the prompt into numerical guidance. A diffusion model begins with random noise and repeatedly predicts how to remove noise in ways that better match the requested objects, composition and colours.
The model produces several newly synthesised options. It has not simply retrieved one existing conference-room image.
The officer rejects versions with invented lettering and implausible headset cables, then adjusts the prompt and generates another set for human review.

Result: The officer places the strongest draft in the concept note as a clearly marked placeholder. Before any public use, the visual still requires clearance, accessibility checks, rights review and confirmation that it cannot be mistaken for a photograph of an actual meeting.

In a United Nations workplace

At a liaison office, an ombudsman case officer is completing an evaluation matrix for three bids that include image-generation functions. Before close of business, the officer tests each proposed tool with the same fictional prompt for an illustrated workplace dialogue scene.

The test shows whether each system creates a fresh image that follows the prompt, rather than retrieving a stock file. Comparing the outputs also reveals common defects, such as distorted hands, invented text, inconsistent objects and stereotyped visual choices.

Before you try thisThe officer must not enter real case narratives, names, correspondence, witness information or other identifying details into a bidder's tool. Ombudsman material is confidential, and demonstration systems have not been approved to process it. The bids and evaluation comments are also procurement-sensitive and must not be uploaded. Testing should use wholly fictional prompts and follow the approved evaluation procedure.

Where you may hear it

During a vendor demonstration to an ICT governance board
At clearance of visual assets for a side event
In a newsroom verification briefing after a synthetic image circulates
During a procurement unit's review of creative-software bids

Avoid confusion

Often confused with

Common misconceptions

Common misconception: If an AI image looks photographic, it probably shows something that really happened.
Correction: Realistic appearance is not proof. Image generators are designed to produce plausible visual patterns, so they can create scenes that never existed. Always check source, context and provenance before treating an image as evidence.
Common misconception: Text-to-image and image generation mean exactly the same thing.
Correction: Text-to-image is one type of image generation where words are the main input. Image generation is broader and can also include systems that start from another image, combine image and text, or create visuals with different controls.
Common misconception: The system simply searches the internet and pastes together copied pictures.
Correction: Typical systems generate new pixel arrangements from learned patterns rather than retrieving one stored image file. However, training data and memorisation concerns are active topics, so originality, rights and similarity still need careful review.
Common misconception: Generated images remove the need for human designers or editors.
Correction: They can speed up drafts and variations, but people still need to set goals, review outputs, check for bias and errors, handle rights issues, and decide whether an image is suitable for the audience and context.

Quick check

Which statement best distinguishes text-to-image from image generation more broadly?

Go deeper

The simple version

You give the tool a description or other prompt, and it makes a new picture that matches the request as best it can from patterns it learned during training.

The technical version

Image generation is the synthesis of new visual outputs from learned data distributions, often conditioned on text, images or control signals. In text-to-image systems, the prompt is encoded into numerical representations that guide the generation process. In many current systems, a diffusion model starts from random noise in pixel or latent space and iteratively applies a learned denoising process, with prompt conditioning guiding each step. Other architectures, including generative adversarial networks and variational autoencoders, can also generate images. The output is not a factual record by default; it is a probabilistic construction shaped by training data, model design, prompting and decoding choices.

Reviewed external resources

Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
This UNESCO recommendation provides an official ethics framework for AI use and suits readers who need policy and governance context rather than technical detail.
Intermediatelink checked 2026-07-31
Primary sourceContent Credentials (opens in a new tab)Content Authenticity Initiative
This Content Authenticity Initiative resource explains content credentials and provenance concepts and suits readers concerned with identifying and labelling generated or edited media.
Starterlink checked 2026-07-31
EncyclopaediaText-to-image model (opens in a new tab)Wikipedia
This Wikipedia link gives a broad overview of text-to-image models and suits readers who want a quick background summary with related terms.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Image generation (opens in a new tab)YouTube search
This YouTube search results page helps readers find different video explanations of image generation and suits people who learn best by watching, but the quality will vary by creator.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Image generation (opens in a new tab)Reddit search
This Reddit search results page shows community discussion and examples around image generation and suits readers who want to see common questions and debates, but posts are not authoritative.
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

Image generation creates new pictures from prompts or other inputs, and a convincing result can be useful without being true, accurate or appropriate.