Image Prompt Style Guide

A good image prompt is a specification, not a wish. Vague prompts get vague, generic results; specific prompts about the wrong things waste the model's attention on details that don't matter. Front-load what actually changes the image, and be explicit rather than hoping the model infers your intent.

Structure of a prompt

• Lead with the subject: what is actually in the frame, described concretely (not "a person" but "a woman in her 30s wearing a rain jacket, mid-stride"). Vagueness here produces the most generic possible interpretation.
• Follow with style: medium and aesthetic (e.g. "35mm film photograph," "flat vector illustration," "oil painting, impasto texture"). Style words are doing more work than most people expect — they set the model's entire visual vocabulary for the rest of the prompt.
• Add composition: framing, camera angle, and what's foreground vs. background ("close-up," "wide shot," "shot from below," "rule-of-thirds, subject left of frame").
• Add lighting: direction, quality, and mood ("soft window light from the left," "harsh midday sun," "golden hour backlight"). Lighting changes the emotional read of an image more than almost any other single variable.
• Add lens/technical detail last, and only when it matters to the result: focal length, depth of field, film stock, aspect ratio. These are refinements, not the foundation — putting them first buries the actual subject.

Order and weighting

• Put the most important elements earliest in the prompt; most models weight earlier tokens more heavily, especially in longer prompts. If lighting is the whole point of the image, don't bury it at the end.
• Keep each concept to a short, concrete phrase rather than a long descriptive sentence — models parse discrete visual concepts better than flowing prose, and a run-on sentence dilutes which words are actually doing the work.

Consistent style tokens across a set

• When generating a series of images meant to look related (a brand's visual identity, an illustrated set, a storyboard), lock a fixed block of style/lighting/lens tokens and reuse it verbatim across every prompt in the set, changing only the subject and composition lines.
• Keep a running "style sheet" of the exact tokens that produced the desired look (e.g. "isometric, muted pastel palette, soft ambient occlusion, 3:4 aspect ratio") so the set stays visually coherent even across separate generation sessions.

Negative prompts

• Use negative prompts to suppress recurring unwanted artifacts (extra limbs, warped hands, text/watermarks, oversaturated color) rather than fighting them by rephrasing the positive prompt repeatedly.
• Keep the negative prompt short and specific to what's actually going wrong in your results — a bloated generic negative-prompt list (copied wholesale from someone else's setup) can suppress details you actually wanted.

Aspect ratio

• Choose aspect ratio based on the final placement, not habit: square or 4:5 for social feed posts, 16:9 for hero banners and video thumbnails, 9:16 for stories/reels, 2:3 or 3:4 for print and portrait subjects. Decide this before generating, since re-cropping after the fact wastes composition the model already spent on the wrong frame.

Iteration protocol

• Change one variable at a time between generations (lighting, then composition, then style) so you can tell which change caused which result — changing three things at once makes every result unexplainable.
• Keep the seed fixed while iterating on wording, and vary the seed only once the prompt itself is locked — this isolates "did the prompt change do what I expected" from "did I just get a different random draw."
• Save the exact prompt text alongside every kept result. A great image with a lost prompt is a dead end the next time you need something similar.

Ethics note: artist names

• Avoid prompting with a living artist's name to imitate their specific, recognizable style — it raises real consent and attribution concerns, and many platforms restrict it. Describe the visual qualities you actually want (palette, brushwork, composition style, era, movement) instead of the name of the person who made them famous.