AI Image Generation Prompts: How to Write Them (With Examples)

What an image prompt needs to describe

An image generator can't see what you picture; it only has your words. Type 'a cat by a window' and it fills in everything else on its own. Every detail you decide yourself is one less thing left to chance.

• Subject: what appears and what it looks like (approximate age, clothing, material, colour, pose).
• Setting: where it is and what surrounds it.
• Style: photograph, digital illustration, watercolour, 3D render, comic panel.
• Light and colour: soft afternoon sun, diffused studio light, cool tones, a limited palette.
• Framing: close-up, wide shot, top-down, low angle.
• Aspect ratio: landscape, portrait or square, depending on where it will be used.
• Exclusions: what must not appear (text, watermarks, extra fingers, a cluttered background).

Every tool (Midjourney, ChatGPT, Gemini, Adobe Firefly, Stable Diffusion and others) has its own controls for aspect ratio and exclusions. Check its current documentation before copying settings from another tool.

From a vague prompt to a complete one

The first prompt below works, but it leaves nearly everything to chance:

A nice coffee shop.

With the parts above filled in, the same idea becomes far easier to steer:

Interior of a small neighbourhood coffee shop in mid-morning, with a light wooden counter, hanging plants and a latte in the foreground on a small table.
Style: realistic photograph, natural light from a side window, warm and soft colours.
Framing: medium shot at eye level, slightly blurred background.
Aspect ratio: landscape 16:9.
No text, no logos, no people.

Realistic images: light, lens and believable detail

When you want something that looks like a photograph, borrow a photographer's vocabulary: type of light, rough focal length, depth of field and time of day. Ask for everyday detail (small creases, reflections, natural imperfections) rather than flawless perfection.

Realistic photograph of [subject, e.g. a middle-aged baker] kneading dough in a small bakery, flour on their hands and apron.
Light: large window on the left, soft directional light. 50 mm lens, f/2.0, gently blurred background.
Detail: natural skin and fabric texture, no heavy retouching.
Aspect ratio: portrait 4:5.
Avoid: a rendered look, distorted hands, text, watermarks.

An image that looks like a photo but isn't can mislead people. If you publish it and it could be taken for real, say it is AI-generated, and always do so where a platform or regulation requires it.

Artistic styles: describe the technique, don't copy a person

For illustration and art, define the technique (watercolour, linocut, oil, paper cut-out, pixel art), the mood and the palette. Asking for the style of a living artist raises questions about rights and authorship; describing the qualities you like is safer and gives you something more original.

Illustration of a lighthouse on a cliff at dawn, watercolour technique with soft edges and visible paper texture.
Limited palette: slate blue, soft orange and cream.
Composition: lighthouse on the right third, lots of sky, very simplified waves.
Mood: calm, slightly nostalgic.
Aspect ratio: landscape 3:2. No text or signatures.

Editing an image you already have

Many tools let you upload an image and ask for changes. The key is to separate what changes from what stays the same; otherwise the AI may alter more than you intended.

Use the attached image. Change only the background: replace the wall with a white brick wall in soft light.
Keep the person, their clothes, their pose and the foreground lighting exactly the same.
Don't add or remove objects. Keep the same framing and aspect ratio.

If the photo shows people, only use images of yourself or images you have permission to use, and never edit them to show someone doing or saying something they didn't.

Common mistakes and responsible use

• Changing several things between attempts, so you can't tell what improved the result.
• Asking for long text inside the image: many tools misspell it. Add the text afterwards in an editor.
• Stacking contradictory adjectives ('minimalist and highly detailed').
• Not checking the result: hands, teeth, backgrounds and objects can contain errors.
• Imitating real people, or using copyrighted characters and brands without permission.
• Ignoring each tool's terms: commercial use, ownership of the output and usage limits.

Terms, limits and features change often. Check each tool's official documentation and, if you plan to use an image commercially, read its terms first.

Skip the blank page with a template

PromptNow has three templates for this: one for realistic images, one for artistic styles and one for editing images. You fill in a short form (subject, setting, style, light, format) and get a complete prompt ready to paste into the tool you prefer.

A prompt can't guarantee a specific result, because each tool interprets text differently. Try it, adjust it and review the output before you use it.