AI image generation

AI Image Prompts: How to Write Text-to-Image Prompts That Work

2026-07-10·5 min read

A good text-to-image prompt names the subject, the style, the lighting, and the framing — and says what to avoid. Vague prompts give generic images; specific, structured prompts steer the model. A reliable formula is: subject + style/medium + lighting + composition/aspect + what to exclude, then iterate a few times.

The prompt formula

Structure the prompt as: subject + style/medium + lighting + composition + negatives. For example: "ceramic coffee mug, product photo, soft daylight, top-down on marble, no text, no clutter". Each part gives the model a lever you can adjust independently.

Be specific, not verbose

Concrete nouns and settings beat a pile of adjectives. "Top-down, soft daylight, marble surface" steers better than "beautiful, amazing, high quality". Say the one or two things that matter most for this image rather than stacking vague praise words.

Control lighting and mood

Lighting changes everything. Name it: soft daylight, studio softbox, golden hour, moody low-key, bright and airy. For products, even diffuse light reads as clean and professional; dramatic light suits lifestyle and ads.

Framing and aspect ratio

Say the shot and the ratio: close-up, top-down, wide, eye-level; 1:1 for feed, 9:16 for stories, 16:9 for banners. Generating at the target ratio beats cropping later and avoids awkward compositions.

Iterate — do not expect one shot

Generate, then change one thing at a time: swap the lighting, tighten the framing, add a negative. Changing several things at once makes it hard to tell what helped. Two or three focused iterations usually beat one long over-stuffed prompt.

Prompts for products & e-commerce

For product images, ask for a clean or specified background, accurate color, even lighting, and realistic scale. If you already have a product photo, image-to-image often beats text-to-image because it keeps your exact product.

→ When to use image-to-image instead

Try it yourself

Write text-to-image prompts that work: name the subject, style, lighting, and framing, add what to avoid, and iterate. A simple prompt formula with examples.

Try text to image

FAQ

How long should a prompt be?

Long enough to name subject, style, lighting, and framing — usually one focused sentence. Detail beats length; avoid stacking vague adjectives.

What are negative prompts?

They tell the model what to leave out, e.g. "no text, no watermark, no clutter" — useful for cleaning up product and marketing images.

Why do I get a different image each time?

Generation is randomized by design, so each run varies. Lock in what you like by keeping the prompt and changing only one detail per iteration.