AI Image Prompt Guide: Structure, Style, and Control

This page is a durable, model-neutral foundation for writing AI image prompts. It covers creative objective, subject, environment, composition, perspective, lighting, color, materials, style, constraints, on-image text, reference images, editing versus fresh generation, iteration, and a Portable Image Brief (PIB) checklist created for this guide.

It does not rank image products, promise perfect adherence, typography, hands, realism, or consistency, treat camera terms as real optics, claim that every model supports negative prompts, or claim that longer prompts are always better. Capabilities differ by product and version; verify details in current vendor docs.

For OpenAI ChatGPT Images and GPT Image API workflows, see the GPT Image prompt guide. For Midjourney parameters and Nano Banana model limits, see Midjourney and Nano Banana prompts. For Gemini surfaces, see How to prompt Google Gemini.

Before style adjectives, state what the image is for: ad, product page, thumbnail, concept art, UI mock, editorial still. Intended use steers polish, empty space, and how strict the constraints should be.

Name the audience and deliverable in one sentence. A cafe homepage hero needs negative space for a headline. A catalog shot needs empty backdrop and crisp materials. A moodboard concept can stay looser. OpenAI's cookbook and Google's Imagen prompt guide both treat purpose and structure as part of a controllable brief.

Priority stack (not a word-count rule): purpose and subject first, then environment and action, then composition and lighting, then color, material, and style, then optional camera-language shorthand, then tool controls such as aspect or quality. Clarity and non-contradiction beat length. Add detail only for elements you must control.

Every image prompt needs a clear subject. Layer from broad category to defining traits to action or pose. Say how many subjects there are. If people appear, note gaze and pose when those matter.

Weak:

realistic photo of a woman

Improved:

Photorealistic three-quarter portrait of a woman in her 30s with short silver hair, looking just past the camera, soft window light from camera-left, shallow background blur, neutral gray wall, calm editorial mood. No watermark, no extra text.

Why this is more controllable: identity, pose, light direction, environment, style word, and constraints replace a vague genre label. Camera brand names are omitted; "shallow background blur" states the look without promising real optics.

Weak (product):

cool sneakers product shot

Improved (product):

Studio product photo of white leather high-top sneakers on a matte light-gray seamless backdrop, 3/4 front angle, soft diffused overhead light with gentle contact shadow, crisp material detail on stitching and rubber sole, catalog e-commerce style. Empty background, no props, no logos other than the shoe's own.

Why this is more controllable: angle, light, backdrop, material cues, output purpose (catalog), and explicit emptiness replace hoping the model "knows" product-shot norms.

Watch clutter: two or three elements usually work better than five competing focal points. If the scene is complex, generate pieces separately or simplify the brief.

Setting, time of day, weather, and load-bearing props give the subject a place to exist. Omit props that do not serve the goal; extra nouns invite extra objects.

Weak:

cyberpunk city

Improved:

Wide establishing view of a rainy near-future city street at night, neon signs reflecting on wet asphalt, one empty crosswalk in the foreground, dense midground traffic lights, towering glass buildings fading into fog, teal-and-magenta color grade, cinematic still.

Why this is more controllable: depth layers, weather, color grade, and a single foreground anchor replace a genre keyword alone.

What you leave unspecified is filled by model defaults and training priors. That is not "random malice," and it is not literal word-for-word obedience. Unspecified details are simply uncontrolled.

Composition language steers arrangement: close-up, wide shot, eye-level, low angle, top-down, over-the-shoulder. Say what should read first (visual hierarchy). Use negative space when you need room for a headline or UI overlay.

Perspective and depth are descriptive craft: foreground, midground, background; shallow versus deep look. Do not promise real depth-of-field physics. Framing words often travel across tools; aspect ratio usually does not live only in prose.

Set aspect ratio or canvas size in the product controls when available (UI picker, API size / aspectRatio, or Midjourney --ar). Generating square and cropping later often wastes the composition.

Lighting is broadly influential, but defaults vary by product. Prefer source, quality, and direction over empty mood words alone: soft window light from camera-left, harsh overhead fluorescent, overcast diffusion, neon rim light.

Color: name a palette ("muted teal and coral," "desaturated earth tones") rather than "beautiful colors." Materials and textures make surfaces concrete: matte ceramic, brushed steel, cold-press paper, film grain. Prefer concrete visual facts over empty quality adjectives like "8K ultra detailed."

Camera and lens language is optional photoreal shorthand. Framing and look words help. Exact brand and aperture strings may be interpreted loosely for high-level look and composition, not as exact physical simulation. Prefer describing the visible result over brand drops.

Style is medium plus movement plus visual qualities: botanical watercolor with ink outlines, flat UI illustration, 35mm candid photo, low-poly 3D. One coherent system beats a soup of conflicting styles.

Weak:

watercolor flowers

Improved:

Botanical watercolor illustration of a single poppy with ink outlines on textured cold-press paper, soft pigment blooms, generous white negative space on the right for a caption, field-guide plate aesthetic, muted greens and coral red.

Why this is more controllable: medium, technique, composition purpose, and palette stack on purpose instead of a vague medium word.

Conflicting instructions reduce control: photoreal plus cartoon without a deliberate hybrid, a style reference plus opposing style adjectives, or "change everything" while also "keep everything identical." Choose one style system and one primary subject.

Many important knobs are not adjectives inside the prompt. When your tool exposes a control, use it instead of burying the same idea only in prose.

• Aspect / size: UI picker or API fields; Midjourney --ar is an example, not universal syntax.
• Quality / steps / CFG: latency versus fidelity is a setting, not an adjective pile.
• Seed: reproducibility tool when the product offers it; not portable prose.
• Style presets / content type: use when they match intent; still describe the medium in words when needed.
• Reference weights / edit modes / policy flags: product features; check current docs.

Do not paste Midjourney flags into Gemini prose as if they execute, and do not assume every API shares the same parameter names.

Describe what you want first. Exclusion support is not universal. Mechanism depends on the model and interface; verify in current product documentation.

• Constraint-in-prompt (many GPT Image workflows): "no watermark," "no extra text," preserve lists.
• Parameter exclusions (Midjourney): --no item1, item2.
• Negative field (Imagen / Stable Diffusion lineage): separate negative prompt or negative weights; some docs prefer plain unwanted items without "no/don't" inside that field.
• UI remove / fill (Firefly-class): mask and describe the fill.
• No negatives (Flux-class): unsupported; rewrite exclusions as positive scene descriptions.

Practical workflow: generate once, note concrete defects, apply the tool's exclusion mechanism, avoid giant generic ban-lists on the first try. Prefer one exclusion method so prose and UI do not fight each other.

Put exact copy in quotes. Keep strings short. Specify placement and font character (bold sans-serif, centered upper third). Ban extra labels. Do not expect perfect long documents or exact named fonts. Iterate and verify lettering visually.

Weak:

poster with text about summer

Improved:

Minimalist poster, solid sun-washed cream background, centered bold sans-serif title "SUMMERLAND", smaller subtitle underneath "Open late", large clean lettering, plenty of margin, no other text, no watermark.

Why this is more controllable: quoted strings, hierarchy, placement, and "no other text" make failures diagnosable. Long copy and busy scenes still fail often; simplify and regenerate when needed.

Use references when brand look, character identity, or product geometry must stay locked; when a style is easier to show than to name; when compositing subject from one image and style from another; or when prose keeps drifting.

Weak: long prose trying to reinvent a brand look from memory.

Improved:

Image 1 is the product photo to keep. Image 2 is the style reference only. Keep the product geometry, label layout, and camera angle from Image 1. Apply Image 2's soft daylight, film grain, and muted pastel grade. Do not copy objects from Image 2.

Why this is more controllable: indexed roles, copy versus non-copy rules, and clear non-goals. Attachment UI and syntax differ by product; see the GPT Image, Midjourney and Nano Banana, and Gemini guides for product details.

Fresh generation needs a full brief. Editing needs a change-only list plus preserve invariants. Global restyle keeps content geometry and changes grade or medium. Masked or brush edits describe the fill for the selected region when the product supports it.

Weak:

make it better and more premium

Improved:

Edit only the background: replace the busy kitchen with a clean white quartz counter and soft daylight. Preserve the cereal box design, typography, camera angle, crop, and product proportions exactly. Do not change colors on the box. No new props, no logos, no watermark.

Why this is more controllable: surgical change plus preserve list reduces drift from vague "premium." Selection tools can still bleed; identity can still drift. Verify after each step. Do not promise near-zero drift.

Change one variable per iteration. Compare the render to your written brief, not to a vibe. Prefer short labeled segments when a prompt is hard to scan.

Preflight checklist:

• Purpose is stated
• Subject, count, and action are unambiguous
• Environment includes only load-bearing props
• One primary focal point
• Framing / perspective described
• Lighting has source + quality + direction (if mood matters)
• Style / medium is coherent
• Must-not items use the correct mechanism for this tool
• Aspect / size / quality / seed / presets set in controls when available
• On-image text is quoted, short, and placement-specified
• References attached and role-labeled if needed
• You know whether this is a fresh generate or an edit

Troubleshooting:

• Wrong subject or extras: tighten count and action; remove competing nouns; state positive emptiness.
• Flat / generic look: add light direction, materials, and medium; drop empty "8K" spam.
• Style soup: one style system; simplify.
• Exclusion ignored: switch to positive scene or the correct exclusion tool.
• Text wrong: quote short strings; simplify background; iterate.
• Edit drifts: change-only + invariants; one change per turn.
• Composition crushed after crop: set aspect before generating.
• Nothing like the brand: add a style or identity reference.

Portable Image Brief (PIB) is a 12-point checklist created for this guide. It is not an industry standard, ISO, or vendor framework. Use it before you generate:

1. Purpose: What is the image for (ad, thumbnail, concept, product page)?
2. Subject: Who or what, how many, defining traits?
3. Action: What is happening; gaze or pose if people?
4. Environment: Where; which props are load-bearing?
5. Composition: Framing, camera height or angle language, hierarchy (what reads first)?
6. Perspective / depth: Foreground / mid / background; shallow vs deep look?
7. Lighting: Source, quality, direction, time of day?
8. Color + material: Palette; surfaces and textures that matter?
9. Style / medium: One coherent system?
10. Constraints: Must-have / must-not; which exclusion mechanism does this tool support?
11. Controls: Aspect / size / quality / seed / preset set in the UI or API?
12. Proof plan: Reference needed? Edit vs generate? What will you check after the first render?

Practice with the free image prompt tool. For OpenAI-specific surfaces, sizes, and edits, open the GPT Image and ChatGPT Images prompt guide. For Midjourney parameters and Nano Banana model limits, open Midjourney and Nano Banana prompts. For Gemini product surfaces beyond image craft, open How to prompt Google Gemini.

Keep the creative brief portable. Let product guides own the syntax.