Prompting Image Models: Structure Beats Adjectives

Image prompts fail most often because they describe a mood instead of a scene. Models cannot infer composition from enthusiasm.

A structure that works

• Subject. What is in frame, described concretely.
• Composition. Shot type and framing: wide, close-up, eye level, overhead.
• Lighting. The single highest-leverage element. Soft window light, hard midday sun, neon at night.
• Style. Medium and reference: editorial photograph, watercolour, technical illustration.

What to leave out

Negations are handled inconsistently — describing "an empty street" works better than "a street with no cars". Long lists of quality adjectives mostly consume tokens. Repeating a concept three ways does not make it more likely.

Iterate one variable at a time

Change the lighting while holding everything else fixed. Randomised prompt edits make it impossible to learn what actually moved the result.

Use references where available

Style and character reference features do more for series consistency than any amount of prompt engineering. When you have an image you like, reference it rather than describing it in words.