How to Write Better AI Image Prompts: A Practical Guide with Examples

The same idea can produce a forgettable picture or a striking one, and the difference is usually the prompt. This guide gives you a simple structure for prompts, shows how to describe style, light and composition, and explains how to fix the most common problems.

A structure that works

Most good prompts answer five questions, in roughly this order:

1. Subject: what is in the picture? ("a ceramic coffee cup")
2. Details: what makes it specific? ("matte sage-green glaze, steam rising")
3. Setting: where is it? ("on a wooden café table by a rainy window")
4. Style: what should it look like? ("product photography, shallow depth of field")
5. Light and mood: how is it lit? ("soft morning light, calm and cozy")

Put together:

A ceramic coffee cup with a matte sage-green glaze and steam rising, on a wooden café table by a rainy window, product photography with shallow depth of field, soft morning light, calm and cozy mood.

You do not need every part every time, but each one you add removes a guess the model would otherwise make.

Be concrete

Vague words leave the model to choose for you. Compare:

• "a beautiful landscape" → "a mountain lake at sunrise, mist over the water, pine trees in the foreground"
• "a cool logo" → "a flat vector logo of a fox head, orange and white, on a plain background"
• "a person in a city" → "a woman in a yellow raincoat crossing a busy Tokyo street at night, neon reflections on the wet road"

Nouns and visible details (colors, materials, objects, places) help more than adjectives like "amazing" or "stunning".

Describe style with references the model knows

Style words change the whole image. Useful families include:

• Medium: oil painting, watercolor, pencil sketch, 3D render, vector illustration
• Photography: portrait, macro, aerial, product shot, film photography
• Genre: anime, manga, cyberpunk, fantasy, minimalist
• Look: cinematic, editorial, isometric, flat design

The AI art styles library collects examples you can borrow from.

Control the camera and the light

For photo-style images, a few photographic terms go a long way:

• Framing: close-up, medium shot, full body, wide angle, top-down
• Lens and focus: shallow depth of field, macro, sharp background
• Light: golden hour, studio lighting, backlit, overcast, neon

Also choose the aspect ratio in the editor before you generate: 16:9 for banners and thumbnails, 9:16 for stories and phone wallpapers, 1:1 for feeds (GPT Image 2 also offers 4:5).

Text inside images

If the image has to contain words, such as a poster title or a label, put the exact text in quotes and say where it goes:

A minimalist concert poster, the title "SUMMER NIGHTS" in bold white letters at the top, a silhouette of a guitarist against an orange sunset.

Nanobanana Pro is the strongest model on Banana Art Lab for readable text, posters and infographics.

Iterate instead of starting over

1. Start broad with a cheap model such as Z-Image Turbo or Nanobanana (10 credits).
2. Change one thing at a time, such as the light, then the style, then the framing, so you can see what each change does.
3. Keep what works. When an image is close, upload it as a reference image and ask only for the next change. The reference image guide explains how.
4. Finish on a stronger model such as GPT Image 2 or Nanobanana Pro at the resolution you need.

Fixing common problems

• Too busy: remove details and name one clear subject.
• Wrong style: move the style words to the start of the prompt.
• Odd hands or faces: use a closer framing or a stronger model, and keep poses simple.
• Unwanted elements: say what the background should be ("plain white background") rather than what it should not be.

Shortcut: start from a template

Every template comes with a prompt that already works. Open one, read how the prompt is written, and change the subject to your own. It is the fastest way to learn what a model responds to.