How to write AI image prompts that actually work
A practical structure for text-to-image prompts: subject, setting, light, lens and finish. With examples you can paste and adapt.
Most weak renders come from prompts that name a subject and stop there. The model then has to guess everything else - where the subject stands, how it is lit, how close the camera is - and a guess is usually the most average answer it knows. A good prompt makes those decisions for it.
This guide gives you a five-part structure that works for almost any still, then shows how to tighten a prompt when the first render is close but not right.
The five parts of a strong prompt
- Subject - what is in the picture, with one or two concrete details. "A ceramic teapot" is a start; "a matte white ceramic teapot with a chipped spout" is a picture.
- Setting - where it is. A surface, a room, a landscape, or a plain background if you want the subject isolated.
- Light - the single biggest lever on mood. Name a source and a direction: "low window light from the left", "a hard spotlight from above".
- Camera - distance and lens. "Macro close-up", "wide shot", "85mm portrait lens, shallow depth of field".
- Finish - the look of the final image: "editorial photography", "film grain", "clean studio product shot".
You do not need all five every time, but each one you leave out is a decision you hand to the model.
A matte white ceramic teapot with a chipped spout on a weathered oak table, low window light from the left, soft shadows, 50mm lens, shallow depth of field, quiet editorial still life, film grainOrder matters more than length
Image models pay the most attention to what comes first. Put the subject at the front, then the setting, and leave style words for the end. A prompt that opens with "cinematic, 8k, masterpiece" spends its strongest position on words that describe nothing in particular.
Length helps up to a point. One or two sentences of specific detail beat a paragraph of adjectives. If you find yourself listing five synonyms for "beautiful", delete four.
Describe what you want, not what you do not want
"No clutter" still puts the idea of clutter in the prompt. Say what should be there instead: "a bare concrete floor", "an empty white background". Positive description is more reliable than negation.
Iterate one change at a time
When a render is close, change a single part and run it again. Swap the light, keep everything else. Then the lens. If you change three things at once you cannot tell which one helped.
In the studio, Reuse prompt on any result puts its prompt back in the composer so you can edit one phrase, and Vary runs the same idea again for a different take. The price is shown before each render, so iterating never surprises you.
Pick the frame before you write
A tall 4:5 frame and a wide 16:9 frame want different compositions. Decide where the image will be used first - a feed, a banner, a phone wallpaper - then describe a scene that fits that shape. Our guide to aspect ratios covers which frame suits what.
Three prompts to adapt
A glass perfume bottle on polished black stone, a single softbox reflection, violet rim light, dark background, macro product photographyAn abandoned greenhouse at dawn, broken panes, mist between overgrown ferns, pale gold light through the roof, wide shot, 24mm lens, cinematicA bowl of ramen seen from directly above, steam rising, dark slate table, warm tungsten light, food editorial photographyQuestions
How long should an AI image prompt be?
One or two sentences of concrete detail is usually enough. Past that, extra adjectives tend to blur the result rather than sharpen it.
Do style words like 8k or masterpiece help?
Rarely. They describe nothing specific. Naming the light, the lens and the kind of photography does far more.
