AI Image Prompt Structure: The Six-Part Formula Behind Reliable Results
Most bad image prompts are not badly written, they are incomplete. They describe the subject well and leave everything else to the model, which is where the generic look comes from. Six slots, filled in a fixed order, fixes that.
A prompt like a woman drinking coffee in a cafe will produce something, and it will look like the average of everything the model has ever seen about cafes. Add the medium, the light, the composition and the constraints and the same idea starts producing images you would actually pick.
The structure below is not a creative straitjacket. It is a checklist that makes sure the decisions you care about are stated rather than defaulted. You can skip a slot deliberately; the failure mode is skipping one without noticing.
Order matters more than people expect. Subject first and constraints last matches how these prompts get parsed in practice: the model commits to what the image is about early, and later clauses modify it rather than competing with it.
The core idea
Four principles behind every prompt on this page.
Subject first, always
One concrete noun phrase with one or two distinguishing details. A specific subject anchors everything that follows; a vague one leaves the model to choose for you.
State the medium
Photograph, illustration, 3D render, watercolour, oil painting. This is the single biggest fork in output style and it is usually left out.
Light is a slot, not an afterthought
Source, direction and softness. Leaving light unspecified is the most common reason an otherwise good prompt comes back flat.
Constraints go last
Aspect ratio, negative instructions, what to keep out of frame. Placed at the end they modify the image rather than fighting the subject.
Step-by-step
The order that keeps you from debugging five things at once.
Write the subject in one phrase
A matte black pour-over kettle on a walnut counter. Concrete noun, one material, one surface. Avoid adjectives about mood at this stage.
Add the medium
Commercial product photograph. Two or three words that tell the model what kind of image this is, not just what is in it.
Add the light
Large softbox from above and slightly in front, gentle gradient shadow. Direction plus softness is the minimum; a background clause is the upgrade.
Add composition and lens
Three-quarter angle, shot on 85mm, shallow depth of field, generous negative space on the left. This is where a snapshot becomes a composed frame.
Close with constraints
Square 1:1, no text, no logos, no people. Anything the model should keep out belongs at the end of the prompt.
Prompt recipes
Paste as-is, then swap the subject for yours.
The bare prompt
A woman drinking coffee in a cafe
Why it works: Every slot except the subject is empty, so all of them get filled by the model's defaults: centred, evenly lit, medium shot, stock-photo grade.
Subject + medium
A woman drinking coffee in a cafe, candid documentary photograph shot on 35mm film
Why it works: Naming the medium immediately moves the image away from the default render look toward a specific photographic tradition.
Subject + medium + light
A woman drinking coffee in a cafe, candid documentary photograph on 35mm film, soft morning window light from camera left, shadow side falling to warm grey
Why it works: Light is the slot that adds depth. One clause about direction and one about the shadow side gives the frame dimension.
Full six-part prompt
A woman in a linen coat drinking coffee at a corner table, candid documentary photograph on 35mm film, soft morning window light from camera left, seated by the window with the street defocused behind, shot on 35mm at f/2, vertical 4:5, no text, no logos
Why it works: All six slots filled. This is the version that reproduces consistently and looks directed rather than generated.
Product version
A brushed steel travel mug on a pale oak surface, commercial product photograph, large softbox from above and slightly in front with a soft gradient shadow, three-quarter angle, 85mm, shallow depth of field, generous negative space on the right, square 1:1, no text
Why it works: The same structure applied to a product. Negative space is named deliberately here because it leaves room for a headline later.
Illustration version
A lighthouse on a cliff in heavy fog, muted editorial illustration, flat gouache texture with visible paper grain, diffuse overcast light, wide composition with the lighthouse small in frame, limited palette of slate blue and bone, 16:9, no text
Why it works: Medium and palette carry the style. Note that the subject is deliberately small in frame — composition is doing the work, not detail.
Character sheet version
A courier in a weathered orange jacket, character design sheet, neutral studio light on a plain light grey background, front view, full body, even lighting, no props, no text, no watermark
Why it works: Plain background and even light make the subject reusable. This is the base image for consistent characters rather than a finished illustration.
Interior version
A reading nook with a low armchair and a floor lamp, architectural interior photograph, warm late afternoon light raking across the floor from a window on the right, wide angle 24mm, symmetrical composition, vertical 9:16, no people, no text
Why it works: Aspect ratio and no people are doing real work. Vertical framing is stated because the image is destined for a phone screen.
Common mistakes
The four that account for most bad results.
Stacking style words instead of naming a medium
Beautiful, stunning, hyperrealistic, 8k, award winning are weight tags that mostly add contrast and sharpening. Photograph, gouache illustration or 3D render actually changes what the image is.
Leaving light out
An unspecified light source defaults to even, frontal, slightly top-down — the flattest possible setup. One clause about direction fixes it.
Contradicting yourself
Minimal clean background and busy detailed market in the same prompt. The model averages contradictions into mush. Pick one and let it hold.
Forgetting the aspect ratio
If the image has a destination — a phone screen, a banner, a square thumbnail — say so. Re-cropping afterwards throws away composition.
Why structure beats longer prompts
The model is filling empty slots whether you fill them or not.
Every image a model produces is fully specified. There is no such thing as an unspecified background or an unspecified light — if you do not name them, the model picks them, and it picks the most common option in its training data. That is exactly why un-structured prompts converge on the same look: they are all being completed by the same defaults.
Filling the slots yourself does two things. It replaces the average with your choice, and it reduces the space of valid images, which makes the output more reproducible from run to run. Reproducibility is the thing people actually want once they move past generating one-off images.
The order is a practical detail rather than a rule. Subject early because it anchors the image, style and light in the middle because they modify it, constraints last because they restrict it. Prompts that put constraints first tend to have them overridden by later subject details.
Practise it in OpenFilmo
Every recipe on this page runs in the same workspace.
One image costs 2 credits and a 4K upscale costs 4. The Free plan gives you 30 credits every month with no credit card, no watermark and commercial use included — roughly fifteen images before you pay anything. Paid plans start at $9 for 400 credits and add faster generation and parallel tasks.
Related guides
Longer reads on the same workflow, from the OpenFilmo blog.
AI Video Prompt Formulas That Hold Up in Production
Prompt structures that work for AI video: subject, motion, camera, time and style — with tested formulas, failure patterns and a m…
AI Product Video for Ecommerce: What Converts
How to make AI product video that converts: shot types ranked by impact, consistency requirements, and the workflow for generating…
Text in AI Video: Why It Breaks and What to Do Instead
Why AI video models garble text and logos, what actually works instead, and a production workflow for adding typography to generat…
AI tools
Single-purpose tools that run in the same workspace, on the same credit balance.
Vidu models
The Vidu Q1 and Q2 models this site runs on, and how they compare to Kling, Veo and Runway.
Related workflows
Closely related tools and guides on the same problem.
Frequently asked questions
What is the best structure for an AI image prompt?
Subject, medium, light, composition, lens or framing, then constraints — in that order. Filling those six slots replaces the model's defaults with your decisions, which is what makes output consistent.
How long should an AI image prompt be?
Long enough to fill the six slots, usually two to four sentences. Extra length beyond that adds little; a paragraph of style adjectives tends to add contrast rather than control.
Do quality tags like 8k or ultra detailed still work?
Mostly as sharpening and contrast. They rarely change composition or subject. Naming a medium and a lighting setup does far more for the result.
Why do my images all look the same?
Because the slots you left empty are being filled with the same defaults every time. Change the medium and the light direction, and the look changes immediately.
Should I include negative prompts?
Yes, but keep them short and specific — no text, no watermark, no extra hands. A long list of negatives can start to suppress things you wanted.
How much does one image cost?
One image costs 2 credits and a 4K upscale costs 4. The Free plan gives you 30 credits every month with no credit card, which is roughly fifteen images before you pay anything.
AI prompt collections
Copy-ready prompt sets for portraits, couples, selfies, anime, fashion and film.