How to Write Prompts for Image Generators
An image prompt is not a caption. A caption describes what is already there; a prompt has to describe what should be there, including the parts a viewer would notice without being told: the framing, the light, the material, and the mood. The model fills every gap you leave, and it may fill it differently on the next attempt.
The AI Prompt Optimizer can turn a rough idea into a structured image brief. The order below is designed to make the result easier to control.
Quick answer
State the subject and the action first, then the setting, the framing, the light, and the style. Add what must not appear. Change one variable at a time between attempts, and write down the prompt that produced each result. Image models interpret language differently, so a prompt that works in one tool may need different wording in another.
Describe the image, not the idea
“A feeling of loneliness” is an idea. “A single figure at a bus stop at dusk, seen from across the road, with a wet pavement and a lit window behind” is an image. The second gives the model something to draw. If you need an abstract concept, translate it into objects, weather, distance, and light.
Start with the subject and what it is doing. A subject without an action often produces a static portrait or a product shot. Then add the setting and the relationship between subject and camera: close, wide, above, below, in front, behind.
Composition and camera language
Terms such as wide shot, close-up, eye level, low angle, and shallow depth of field are useful because they describe the frame rather than the subject. They give the model a way to organise the scene. Use them only if you know what they mean; a vague “cinematic” is less useful than “wide shot, low angle, subject small in the frame”.
If the image needs space for text, say where: “subject on the right, empty sky on the left”. If it needs to match an existing layout, give the aspect ratio and the direction of the crop. These instructions are easier to follow than a request to “make it work for Instagram”.
Light, colour, and material
Light is often the difference between a flat image and a convincing one. Describe the source and the quality: soft window light, hard midday sun, overcast, a single lamp, backlit. Then describe the colour: warm, cool, muted, high contrast. Material matters too: matte, glossy, brushed metal, linen, wet stone.
Style references can help, but they are blunt instruments. A named artist or photographer may produce a pastiche rather than the quality you wanted, and living artists raise ethical and legal questions. Prefer descriptive terms: editorial photograph, watercolour, flat vector, grainy film still, technical diagram.
What must not appear
Negative instructions are unreliable in some models, but they are still worth including when the problem is specific: no text, no logos, no extra fingers, no watermark. If the model keeps adding something, try describing the desired state instead: “a plain background with no objects” rather than “no clutter”.
Iterate and record what changed
Change one variable at a time. If you change the subject, the light, and the style together, you cannot tell which change produced the result. Keep a short log: prompt, model, seed if available, and the one thing you changed. That log is more useful than a folder of near-identical images.
Use the AI Prompt Optimizer to turn the final description into a repeatable brief. For text prompts, read Prompt Structure: Role, Context, Format, and Rules. For the general method, see How to Write a Prompt That Gets a Useful Answer.
Common questions
What should an image prompt include first?
Start with the subject and what it is doing, then add the setting, the framing, the light, and the style. A prompt that begins with a mood or an abstract idea gives the model little to draw. Concrete nouns and a clear camera position are more controllable than adjectives alone.
Why does the same prompt give different images?
Image models often use randomness, and some services change the underlying model without notice. A prompt is a set of probabilities, not a command. Record the model, the seed if available, and the exact wording so you can tell whether a change came from the prompt or from the model.
Can I ask an image model for text in the image?
Some models can render short text, but accuracy is unreliable, especially for long phrases, small type, and non-Latin scripts. Treat generated text as a design element to be replaced in an editor. Check spelling and do not use it for labels, prices, or legal notices without correction.