The Anatomy of a Great Image Prompt (Six Parts)

TL;DR: A complete image prompt has six parts: subject, setting, composition, light, medium and style, and constraints. That grouping is mine, built one piece at a time below and checked against what Google, OpenAI and Ideogram publish. Unlike a video prompt, there's no action or sound to describe, so composition and constraints carry more of the weight instead.

What Is Image Prompt Structure, and How Is It Different from Video?

Image prompt structure is the order and grouping of information you hand a text-to-image model. Like video prompt structure, no vendor enforces it at the parser level: every model still takes one free-text field, or a field plus a handful of container parameters, and none of them require six labeled slots. OpenAI's own cookbook prompting guide says as much directly: "Minimal prompts, descriptive paragraphs, JSON-like structures, instruction-style prompts, and tag-based prompts can all work well as long as the intent and constraints are clear" (developers.openai.com/cookbook/examples/multimodal/image-gen-models-prompting-guide, accessed September 3, 2026). Structure is a discipline you impose, not a rule the model checks.

What's genuinely different from video: there's no Action part, because nothing moves, and no Sound part, because nothing plays. In their place, two categories carry more weight than they do in a seven-part video prompt. Composition does the job Shot does in video, but as a single static arrangement rather than a camera move through time; the full vocabulary for that part lives in a dedicated composition and framing glossary rather than here. And Constraints (what to exclude) becomes its own numbered part rather than a side note, because whether "no X" is a real field or just prose you write is one of the more inconsistent things across image tools, which the constraints section below documents directly.

Does Any Vendor Publish an Official Image Prompt Structure?

One comes close. Google's Gemini API image-generation guide publishes a reusable template for photorealistic scenes:

"A photorealistic [type of shot] of a [subject description] in a [setting description]. [Description of the light]. Shot from a [camera angle] with a [lens type]." (ai.google.dev/gemini-api/docs/image-generation, accessed September 3, 2026)

That's four of my six categories in one sentence (subject, setting, light, composition as shot type and camera angle), presented as one example template among several, not a required schema. OpenAI's guide takes the opposite shape: a bulleted list of best practices (prompt format, specificity, composition, people and pose, constraints) rather than a single formula, and it explicitly tells you the prompt format itself doesn't matter as long as intent is clear.

So: two vendors, two different postures. Google publishes something closer to fill-in-the-blanks. OpenAI publishes categories to think about and states outright that the container format is your choice. Neither commits to six parts, or to this order. What follows is my synthesis, tested against what both actually document.

The Six Parts, Built One at a Time

Subject, Setting, Composition, Light, Medium & Style, Constraints. This site doesn't render images: Prompt Architects writes the prompt, not the picture. So what follows shows what each addition changes in the prompt text and what it's documented or reasoned to control, not a claim about a specific generated result.

Part 1 — Subject: What or Who Is the Shot About?

The subject is the thing the whole image exists to show, described with fixed, concrete details rather than a vague label.

Omit it and you get a competent but generic environment shot, or a subject the model invents on its own, different on every regeneration. OpenAI's cookbook prompting guide is specific about what "concrete" means here: "Be concrete about materials, shapes, textures, and the visual medium" (developers.openai.com/cookbook/examples/multimodal/image-gen-models-prompting-guide, accessed September 3, 2026).

A weathered stone lighthouse, its lamp lit, paint peeling near the base.

Part 2 — Setting: Where and When Is This?

Setting is place, time of day, weather and season: the world the subject sits inside.

Omit it and the background reads as an unplaceable blur or a plain studio backdrop, which is sometimes exactly what you want and sometimes a missed opportunity to establish mood cheaply.

A weathered stone lighthouse, its lamp lit, paint peeling near the base,
perched on a rocky cliff above a churning sea at dusk in late autumn.

Part 3 — Composition: How Is the Frame Arranged?

Composition is shot distance, camera angle, and any framing device or layout principle, everything that decides where the implied camera stands and how the frame is divided. The full vocabulary for this part (wide shot, Dutch angle, negative space, rule of thirds, and 22 more) lives in composition and framing terms for image prompts; this section covers only where it slots into the whole prompt.

Omit it and the model picks a framing itself, and it won't be consistent across regenerations: one run reads as an establishing wide shot, the next as a tight close-up. This is also the part vendors are most direct about being imperfect: OpenAI's own limitations note that the model "may have difficulty placing elements precisely in structured or layout-sensitive compositions" (platform.openai.com/docs/guides/image-generation, accessed September 3, 2026). A composition instruction is a real lever. It is not a guarantee of exact placement.

A weathered stone lighthouse, its lamp lit, paint peeling near the base,
perched on a rocky cliff above a churning sea at dusk in late autumn. Wide
shot, low-angle, subject positioned along the right third-line, leading
lines from the jagged rocks converging toward its base.

Part 4 — Light: What's the Source, Direction and Colour?

Light is where it comes from, how hard or soft it is, and what colour it casts. The full vocabulary (golden hour, Rembrandt lighting, volumetric light, 40 terms in total) is in lighting vocabulary for AI image prompts.

Omit it and you get the model's default lighting guess for the setting you described, which is often serviceable and rarely matches a specific mood you had in mind.

A weathered stone lighthouse, its lamp lit, paint peeling near the base,
perched on a rocky cliff above a churning sea at dusk in late autumn. Wide
shot, low-angle, subject positioned along the right third-line, leading
lines from the jagged rocks converging toward its base. Warm lamp glow from
the lighthouse itself against a cool blue dusk sky, thin rim light along the
cliff edge from the last of the sunset.

Part 5 — Medium & Style: What Is This Pretending to Be?

Medium and style is the format claim: photograph, painting, 3D render, and any film stock, lens character, or artistic movement layered on top. The camera and lens half of this vocabulary (focal length, bokeh, anamorphic, film grain) is in camera and lens terms that change your AI images.

Omit it and you get whichever house style the model defaults to for the words you used, which shifts by model and by month. OpenAI's cookbook prompting guide treats this as load-bearing enough to call out directly: "For photorealism, include the word “photorealistic” directly in the prompt to strongly engage the model’s photorealistic mode" (developers.openai.com/cookbook/examples/multimodal/image-gen-models-prompting-guide, accessed September 3, 2026).

A weathered stone lighthouse, its lamp lit, paint peeling near the base,
perched on a rocky cliff above a churning sea at dusk in late autumn. Wide
shot, low-angle, subject positioned along the right third-line, leading
lines from the jagged rocks converging toward its base. Warm lamp glow from
the lighthouse itself against a cool blue dusk sky, thin rim light along the
cliff edge from the last of the sunset. Photorealistic, shot on 35mm film,
fine grain, muted desaturated palette.

Part 6 — Constraints: What Should the Model Leave Out?

Constraints are exclusions and invariants: what must not appear, and what must stay unchanged on an edit. This is the part that splits hardest by vendor, covered fully in its own section below.

Omit it and the model is free to add whatever commonly co-occurs with your subject in its training data: a lighthouse prompt with no constraints has a real chance of adding a boat or gulls, not because you asked for one, but because photos of lighthouses often have them nearby. Constraints are the only one of the six parts that can't make an otherwise-correct shot worse; they only remove things, which is why the debugging section below adds this part last.

A weathered stone lighthouse, its lamp lit, paint peeling near the base,
perched on a rocky cliff above a churning sea at dusk in late autumn. Wide
shot, low-angle, subject positioned along the right third-line, leading
lines from the jagged rocks converging toward its base. Warm lamp glow from
the lighthouse itself against a cool blue dusk sky, thin rim light along the
cliff edge from the last of the sunset. Photorealistic, shot on 35mm film,
fine grain, muted desaturated palette. No boats, no modern signage, no
people, preserve the peeling paint texture on the base.

Do Negative Prompts Work the Same Way Across Image Tools?

No — it's a real, three-way split, verified from each vendor's own documentation rather than assumed.

Ideogram exposes a literal field. Its own docs describe the feature plainly: "The Negative Prompt feature guides the AI away from generating specific types of images or content. It instructs the AI on what to avoid or exclude in the generated images" (docs.ideogram.ai/using-ideogram/generation-settings/negative-prompt, accessed September 3, 2026). The same page is direct about its limits, too: "The content of the regular prompt will always be favored over the negative prompt" when the two contradict.

Google recommends prose instead of a field. Its Gemini API image-generation guide has no negative-prompt parameter and instead advises using "semantic negative prompts": Instead of saying "no cars," describe the intended scene positively: "an empty, deserted street with no signs of traffic." (ai.google.dev/gemini-api/docs/image-generation, accessed September 3, 2026).

OpenAI has neither a field nor that specific framing, and instead folds exclusions into ordinary prompt language: "State exclusions and invariants explicitly (e.g., “no watermark,” “no extra text,” “no logos/trademarks,” “preserve identity/geometry/layout/brand elements”)" (developers.openai.com/cookbook/examples/multimodal/image-gen-models-prompting-guide, accessed September 3, 2026).

Three vendors, three different mechanisms for the same intent. Write your constraints as prose that works everywhere, and treat a literal negative-prompt box, where one exists, as a second, optional place to reinforce the same list.

What About Aspect Ratio, Resolution and Output Format?

None of them are one of the six parts. They're the container the six parts sit inside, and, like duration and resolution on a video prompt, they live in separate fields rather than prose you write into the description.

If your image keeps coming back the wrong shape regardless of what you write in the prompt, that's very likely a container-field problem, not a wording problem: why did my image come out the wrong size goes deeper on that specific failure.

Where Does Each Part Actually Land, Per Surface?

What Order Should the Six Parts Go In?

Subject and setting first, since composition, light and style all describe them. Composition and light next: they're the two parts most commonly left out and the two that most reliably change how a static image reads, unlike video where an omitted action or sound is more forgiving. Medium and style can move earlier if you want it to color everything written after it, the same logic the video-prompt anatomy applies to its own Look part. Constraints go last, because excluding an element can't break a shot that's already correctly composed and lit.

Nobody publishes evidence that a different order fails outright. The value of a fixed order isn't correctness, it's that a consistently ordered prompt is diffable: change one part, regenerate, and you know what caused the difference.

How Do You Debug an Image Prompt That Isn't Working?

Subtract, don't add. The instinct when an image comes back wrong is to pile on more description, and more prose usually confuses the model about which detail matters most.

1. Strip to Subject plus Setting. Two parts, one sentence. If the subject itself is wrong, nothing added later fixes it.
2. Add Composition. Regenerate. This is where an unwanted framing or a flattened, centered default usually shows up.
3. Add Light and Medium & Style together. They interact more than any other pair (a lighting choice can quietly fight a style choice), so splitting them wastes a generation.
4. Add Constraints last. They only remove things, so they belong at the end of the ladder, the same reasoning the video-prompt debugging ladder applies to sound.

A weathered stone lighthouse, its lamp lit, paint peeling near the base,
perched on a rocky cliff above a churning sea at dusk in late autumn.

That's step one, worth keeping around as a control prompt: if the subject and setting alone already look wrong, no amount of composition, light or style vocabulary layered on top will fix it.

Where to Keep the Six-Part Template

Six parts is easy to hold in your head for one image and hard to hold across a shoot, a campaign, or a client's whole brand library. The practical failure isn't forgetting the structure: it's retyping it, and losing the version that actually worked.

A prompt template only compounds when the skeleton is saved once and the parts that change are marked as variables:

[SUBJECT, fixed identifying details]. [SETTING: place, time, weather].
[COMPOSITION: shot distance, angle, framing device]. [LIGHT: source,
direction, colour]. [MEDIUM & STYLE: photo, painting, render, film stock].
[CONSTRAINTS: what to exclude, what to preserve].

Image prompt generation is available from the Pro plan up, and the free plan includes 5 prompt enhancements per day, forever, per our FAQ.

To be clear about what this page is: we build the prompt layer. We do not generate images, we do not host any of these models, and we are not affiliated with Google, OpenAI or Ideogram. Running anything above needs an account and credits with that vendor. Every specification on this page was read from the vendor's own documentation on September 3, 2026. This category updates fast enough that the vendor's page is always the final authority over this one.