The Structure of a Photo Prompt: Five Blocks That Decide Everything

A request to a model is neither a search query nor an incantation. It is a description of a scene, and it works better the closer its structure is to how a photographer would brief an assistant.

Five blocks

1. Subject: who or what is in frame. "A woman of thirty", "a ceramic cup", "a man in a coat".
2. Action and pose: what is happening. "Looking into the lens", "turned towards the window", "standing three-quarter".
3. Setting: where. "In a studio against grey", "on a morning street", "in a café interior".
4. Light: from where and of what kind. "Soft window light from the left", "backlit by the sun", "a studio setup with one source".
5. Technique and style: how it was shot. "Portrait focal length, shallow depth of field, light grain".

This order matches how a scene is built: essentials first, refinements after. The model distributes attention unevenly, and the start of a request weighs more than its end.

Why order matters

Elements compete for influence. If the most important thing is mentioned at the end of a long description, it carries less weight and may be ignored. The practical consequence: whatever the frame is pointless without belongs at the front.

This is easy to verify: swap the first and last elements of your request and compare results. The difference is usually visible.

Why a short request often beats a long one

Every requirement dilutes the others. A request with five conditions is followed more precisely than one with fifteen — not because the model "fails to understand", but because with fifteen requirements each carries less influence.

A working limit: no more than seven or eight meaningful elements. If you need more, split the task into two stages: get the scene right first, then refine details in a separate pass.

What must be specified

Silence about a parameter means the model will choose it — differently each time. The mandatory minimum is three:

• Light: without it you get neutral, sourceless illumination, the commonest tell of the AI look.
• Shot size: close, waist-up, full length. Otherwise the model picks the middle option.
• Background: even "a plain light background" beats nothing.

What not to add

• Evaluative words: "beautiful", "professional", "high quality". They carry no visual information and take up room.
• Camera parameters beyond what is needed: a long list of settings does not improve the result.
• Negations: "without a hat" introduces a hat into the description.
• Contradictions: "soft dramatic contrasty light" — the model will pick one.

A worked example

Bad: "a beautiful professional photo of a girl in the city, high quality, detailed".

Not a single visual instruction here: "beautiful" and "professional" are judgements, "high quality" is what the model already attempts, "in the city" is far too general.

Better: "a woman of thirty looking past the camera, an evening street, blurred warm lights behind, backlight from the setting sun, waist-up, portrait optics, shallow depth of field".

All five blocks are present: subject, action, setting, light, technique. Every element carries visual information.

Iterating

The first request is almost never the final one, and that is normal. A practical cycle: get a result, find the main divergence from what you intended, change only that in the request, rerun.

Changing everything at once is the most expensive strategy: you will not know which edit worked and will spend twice the runs.