AI Image Generation

Define where the image will be used

A library wants an illustration for a fictional seed-swap poster. The image should show three paper seed packets on a table, with enough empty space above them for an event title. It will sit behind text in a portrait layout. This is a better starting brief than “make a beautiful poster” because you can check the subject, number of objects, placement and usable space.

Decide what the image must communicate and which details must be exact. In this example, packet shapes and a welcoming mood can be illustrative, while the date and location must come from the organiser's approved text. Keeping those words as editable text in the final layout makes correction and accessibility easier. A generated image can contain lettering, but its spelling, meaning and legibility still need inspection.

Write a short acceptance checklist before generating: three packets, no people, a simple table, a calm background and a clear upper area. Also record the intended crop and minimum useful dimensions. More pixels will not fix a composition that leaves no room for the title or places the important subject outside the final crop.

Understand what the generator is doing

An image generator produces pixel values using a learned model and the inputs supplied to it. A text prompt conditions the result: it influences what appears, where it appears and how it looks. The prompt is not a drawing program whose every instruction is enforced exactly. Different runs can satisfy some requirements while missing others, and different systems use different underlying model designs.

Text-to-image generation starts from a written description. Reference-guided generation adds images that communicate a subject, layout or visual treatment. Editing starts from an existing image and requests a change. Some tools support a mask to identify an area for replacement; mask meanings and edit controls depend on the tool, so check its instructions rather than assuming black and white always mean the same thing.

Generating an illustration is different from retrieving a photograph of a real event. A convincing image of a seed swap does not document an event that happened. For our poster, make the illustrative role clear and avoid visual details that imply a real venue, sponsor or person has participated when that has not been established.

Prepare references with clear roles

If you have an approved packet drawing, it can communicate the desired shape. A second reference might demonstrate the composition. Label their roles explicitly: “Image 1 supplies packet shape; image 2 supplies the arrangement only.” Without that distinction, the generator may copy irrelevant colours, combine backgrounds or change the identity of the main subject.

Use reference material you are authorised to upload and adapt. Check photographs for addresses, faces, school badges, document text and private objects in the background. Crop or redact unnecessary information before sending it. A public-looking picture can still contain personal details, and a prompt asking the service to ignore them does not remove them from the upload.

Choose a tool whose documented input, editing and output capabilities fit the brief. A chat application and its provider's API may expose different controls. The prompt below is a tool-neutral creative brief, not a runnable API request or a promise that a particular account supports every setting. No personal photograph is needed to practise it.

Describe visible choices rather than vague quality

Name the subject, composition, medium, lighting and exclusions. “Three plain paper packets on a pale wooden table, viewed slightly from above” gives more useful direction than “award-winning, highly professional, amazing quality”. State the aspect ratio using the tool's control when available and repeat the important composition in the prompt. Output settings and the content description have different jobs.

Run the brief only in an image tool you have chosen to use. If you are planning without access to one, sketch a thumbnail and apply the same acceptance checklist. The exercise is about translating a need into inspectable visual choices; the wording alone does not establish that an image was created.

Purpose: illustration behind editable text on a portrait event poster.
Scene: exactly three plain paper seed packets on a pale wooden table.
Composition: packets grouped in the lower half; upper third mostly empty.
Appearance: friendly hand-painted illustration, soft daylight, muted greens.
Keep the packet shapes distinct and the background simple.
No people, logos, lettering, dates or event details inside the image.
The illustration represents a fictional event, not documentary evidence.

Inspect meaning before polish

First check the image at the size and crop where people will see it. Count the packets, inspect the open space and ask whether the visual hierarchy supports the title. Then zoom in for malformed edges, merged objects, accidental lettering and inconsistent shadows. A pleasing overall style can hide a practical failure that becomes obvious once the poster text is added.

For diagrams, maps or product images, review relationships and identities more strictly than decoration. An arrow pointing to the wrong part is a factual error even if it is neatly drawn. A generated product label or connector shape may be plausible but incorrect. Use authoritative references or conventional drawing tools where precise geometry and labels are essential.

Keep the first useful version and record why it passes or fails. Compare candidates using the same checklist rather than choosing only by novelty. If one has the right composition but four packets, document that specific defect. This gives the next edit a clear target and prevents a sequence of increasingly complicated prompts from hiding the original goal.

Request one change and verify what stayed

Suppose the chosen illustration has the correct three packets but the upper area is crowded with leaves. Supply that actual image for editing and ask to remove the leaves above the packets while preserving the packets, table, colours and lower composition. Referring to “the earlier one” is ambiguous if several candidates exist; identify the exact input.

Compare the whole result with the previous version, not only the requested area. Generative edits can change supposedly preserved details. If a packet changes shape or a fourth object appears, the edit did not meet the contract. A mask can focus a change but is not a universal guarantee of pixel-identical surroundings. Exact preservation may require compositing an approved edit into the original with an ordinary editor.

Change one major variable at a time when diagnosing a problem. If you simultaneously change style, crop, lighting and object placement, you cannot easily tell which instruction helped. Save the input, prompt and selected result so a collaborator can reproduce the intended workflow, while recognising that identical settings do not guarantee identical output across every system or version.

Check the exported file in its destination

Place the selected illustration in the actual poster layout and add the approved title as text. Verify the crop, contrast and readability at the final display size. An image that looked good full-screen can become cluttered as a small card. Keep a usable text alternative for the illustration and ensure essential event information is available as text, not only baked into pixels.

Choose an output format appropriate to the destination. If a cutout needs transparency, inspect the exported file on different backgrounds; a painted checkerboard is not an alpha channel. Check edges and shadows as well as the centre. Preserve an editable working copy and the final version, and describe the asset as generated when its origin matters to the audience.

The result of this workflow is a checked visual asset for a stated purpose. It is not evidence that the illustrated event happened, that every depicted object exists or that every possible reuse is appropriate. Review the actual destination and reference-material permissions before publishing.

Sources & further reading

Primary references for the concepts in this guide. Examples and exercises are illustrative; verify provider-specific behaviour against the version you use.

• OpenAI: Image generation and editing
• OpenAI: Image prompting and result checks