How to Write Prompts for AI Image Generators

Image generators respond to description. A prompt such as "a nice logo for my café" leaves almost everything to chance, while a prompt that reads like a brief to a photographer or illustrator gets close on the first try. This guide works with the image tools in ChatGPT, Gemini, Midjourney and other generators.

The six-part formula

1. Subject — who or what, with the important details.
2. Setting — where and when.
3. Style or medium — photograph, watercolour, 3D render, pixel art, line drawing.
4. Lighting and colour — soft morning light, neon, muted earth tones.
5. Composition — close-up, wide shot, from above, subject on the left.
6. Extras — mood, aspect ratio, what to leave out.

You rarely need all six. Subject, style and composition do most of the work.

Examples

A photograph

A ceramic mug of black coffee on a worn wooden table beside an open notebook, soft morning light from a window on the left, shallow depth of field, 35mm photograph, warm muted colours, shot from a slight angle above. Landscape format.

An illustration

Flat vector illustration of a woman cycling through a city park in autumn, limited palette of orange, teal and cream, simple geometric shapes, no outlines, plenty of empty sky at the top for a headline. Square format.

A product shot

Studio product photograph of a matte green water bottle standing on a pale grey surface, plain off-white background, soft even lighting, sharp focus, centred, no text, no props.

A logo concept

Simple logo mark for a bakery called "Rise": a minimal rising sun made from three curved lines above a flat horizon, single colour, deep orange on white, flat vector, no gradients, no text.

Notice that each one says what the picture contains and how it should look. None says "beautiful", "amazing" or "high quality" — those words add little.

Useful vocabulary

Refining a result

• Change one thing at a time. If you rewrite the whole prompt, you cannot tell what made the difference.
• Say what to keep. "Same composition and colours, but make it evening."
• Edit in place. Most tools let you select an area and describe the change, which is more reliable than regenerating.
• Ask for variations once you are close.
• Specify the format — square, portrait, landscape — at the start. Cropping later loses detail.

Keeping a character consistent

The classic frustration is a character who looks different in each picture. Three approaches:

1. Reference images. Upload a picture and tell the tool to use it. ChatGPT now lets you save and reuse references; see ChatGPT Images 2.6.
2. A fixed description. Write a detailed description once and paste it into every prompt.
3. One conversation. Generate the set in a single chat so the tool has the earlier images in view.

Making animations

Tools that create GIFs or short clips need one more element: the motion. Describe a single simple movement, say what should stay still, and ask for a seamless loop. "A candle flame flickering gently, everything else still, looping" works better than a scene with five things happening.

Fixing common problems

Rules and rights

• Real people. Do not create realistic images of real people without their consent. Platforms restrict it, and the law is tightening: from 2 December 2026 the EU bans AI systems for non-consensual intimate imagery. See the EU AI Act changes.
• Copyrighted characters and logos. Tools refuse many of these requests, and using such output commercially is risky.
• Living artists. Asking for an image "in the style of" a named living artist is contested. Describing the qualities you want is safer and usually gives better results.
• Commercial use. Read the tool's terms. Rights to output differ between services and plans.
• Labelling. If you publish an AI image where people might take it for a photograph of a real event, say that it is AI-generated. Our guide to spotting AI images shows why that matters.

A template to copy

[Subject with key details], [setting], [style or medium], [lighting and colour], [composition]. [Format]. [What to leave out].

For text prompts, the method is different; see how to write AI prompts.

Frequently asked questions

How do I write a good prompt for an AI image?

Describe the subject, setting, style, lighting and composition as you would brief a photographer or illustrator, and state the format you need.

Why does AI get text in images wrong?

Image models generate shapes that look like letters instead of spelling words. Keep text short, put it in quotation marks, or add it afterwards.

How do I get the same character in several AI images?

Use a reference image, repeat an identical detailed description, or generate the images in one conversation.

Can I use AI images commercially?

It depends on the tool and your plan. Check the terms, and avoid images of real people, brands or copyrighted characters.