How to Write Prompts That Produce Photorealistic Images, Not Renders

There is a recognisable look to AI imagery used badly: a glossy, weightless quality, colours a shade too saturated, lighting that comes from nowhere in particular, everything in focus at once. People clock it in under a second, and once they do, the page reads as filler.

The interesting part is that this look is not a limitation of the models. It is what you get when a prompt describes what is in the picture and says nothing about how the picture was made. As of July 2026 the fix is entirely on the prompt side, and it is more structural than stylistic.

Why the default output looks synthetic

Image models are trained on enormous quantities of internet imagery, and the internet's visual centre of gravity is commercial: stock photography, product renders, digital illustration, wallpaper. Ask for "a locksmith working on a car door" with no further constraint and the model produces the statistical average of all such images — which is a clean, evenly lit, slightly plastic composite.

A real photograph is the opposite of an average. It was taken with a specific lens at a specific aperture in specific light at a specific moment, and every one of those decisions leaves evidence: shallow depth of field, directional shadows, a colour cast from the light source, motion in something, a frame cropped by physical constraint.

So the job of a photorealism prompt is to specify the constraints a real camera would have imposed. Not to ask for realism as an adjective.

The five-part prompt structure

Everything below is one repeatable order. Cover all five and the output changes noticeably; skip the camera and lighting parts and you are back to renders.

1. Subject and action. Who or what, doing something specific. Actions beat poses — "measuring a key blank against a worn original" is more photographic than "holding a key".

2. Setting and context. Where, with two or three concrete environmental details. Specific beats generic: "a cluttered workbench with a bench vice and coiled cable" over "a workshop".

3. Light. The highest-leverage clause in the whole prompt. Name the source, direction and quality: "hard afternoon sun through a roll-up door, long shadows across the concrete", "diffuse overcast light from a north window", "a single work lamp raking across the surface".

4. Camera. Focal length, aperture, distance, angle. "35mm at f/2, waist height, three-quarter view." This clause does most of the work in pushing output away from illustration.

5. Mood and grade. Two or three words on tone and colour treatment: "muted, documentary, slight warm cast". Restraint matters here — this is where over-specification produces the saturated look you are avoiding.

Put together:

A technician cutting a replacement car key on a bench-mounted key machine, brass shavings on the steel surface, a cluttered workbench behind with a vice and coiled cable, hard late-afternoon sun through an open roll-up door casting long shadows across the concrete, 35mm at f/2 from waist height in three-quarter view, shallow depth of field, muted documentary tone with a slight warm cast, photorealistic editorial photography, no text, no logos, no watermarks.

Compare it with "a professional locksmith cutting a key, ultra realistic, 4K, HDR". Same subject, entirely different output.

What to include and what to drop

The instinct to add quality words is strong and mostly counterproductive. A rough guide:

Instead of	Use	Why
ultra-realistic, hyperrealistic	35mm, f/2, shallow depth of field	Camera specifics produce photographic cues; adjectives do not
4K, 8K, HDR	natural light from a window, overcast	Resolution words map to processed wallpaper aesthetics
beautiful lighting	hard side light, long shadows	Direction and quality are what make light read as real
professional, high quality	editorial photography, documentary	Names a genre with actual visual conventions
in the style of [living artist]	the visual qualities you want	Policy risk, litigation risk, and it is imprecise anyway
detailed background	two or three named objects	Models render specifics; "detailed" produces clutter
perfect, flawless	slightly worn, used, imperfect	Imperfection is the strongest realism signal available

That last row deserves emphasis. Real working environments have scuffs, uneven stacks, cables that are not coiled neatly, surfaces with wear. Asking for imperfection is counter-intuitive and works better than almost anything else on the list.

Composing around known weak points

Some subjects remain unreliable. Rather than retrying twenty times, compose so the weakness is out of frame — which is what a working photographer does anyway.

Text. Do not ask for it. Generate a clean plate with negative space and set real type over it in your design tool. You get correct typography, editable copy, and no uncanny lettering.

Hands in close-up. Move them to the middle distance and give them a task. Hands operating a tool at working distance are convincing; hands filling the frame are a coin flip.

Faces. Three-quarter profile, seen from behind, cropped at the shoulders, or softly out of focus behind an in-focus foreground object. All are standard editorial framings and all remove the problem.

Crowds. Groups multiply the failure surface. One or two people, or none at all with evidence of their presence — a half-finished job, tools set down mid-task.

Brand-identifiable objects. Generic is safer and legally cleaner. See our post on commercial use and copyright for why generated logos and likenesses are a category to avoid entirely.

Consistency across a set

A batch of images that each look good individually can still look wrong together, because they read as eight different photographers. The fix is mechanical.

Write the style block once — camera, lens, light quality, palette, grade — and treat it as immutable. Vary only the subject and setting clauses. Resist the urge to "improve" the shared block partway through a set, because that is precisely what breaks the visual through-line.

STYLE (unchanged across the set):
35mm at f/2, waist height, shallow depth of field, diffuse overcast
daylight, muted palette with a slight warm cast, documentary editorial
photography, photorealistic, no text, no logos, no watermarks

SUBJECT (varies):
1. A technician programming a key fob with a diagnostic tablet in a car footwell
2. A row of uncut key blanks in a shallow tray on a workbench
3. A hand steadying a door panel while a trim tool releases a clip

This is also the moment to decide filenames and alt text, because the prompt already contains the description. Which brings up the part most workflows get backwards.

Iterating without starting over

The expensive habit is rewriting the whole prompt when one thing is wrong. Because the five clauses map to distinct visual properties, you can usually diagnose which clause is at fault and change only that.

What's wrong with the output	Which clause to change	Concrete edit
Looks like a render or illustration	Camera	Add focal length and aperture; add "editorial photography"
Flat, weightless lighting	Light	Name a source and direction; add shadow behaviour
Too clean, showroom-like	Mood	Add "used", "slightly worn", "working environment"
Wrong subject emphasis	Subject	Move the key noun earlier; cut competing detail
Over-saturated, wallpaper-ish	Mood	Remove quality adjectives; add "muted", "natural colour"
Background is generic mush	Setting	Name two or three specific objects instead of "detailed"
Everything in focus at once	Camera	Widen the aperture; add "shallow depth of field"
Composition too centred and static	Camera	Specify angle and height: "low angle", "three-quarter view"

Change one clause per attempt. Changing three at once means you learn nothing about which mattered, and prompt-writing skill is entirely a matter of building that cause-and-effect intuition for the tool you use.

One caveat: how much of this transfers depends on the model. Prompt behaviour differs meaningfully between generators, and our comparison of AI image generators for SEO covers where each one sits.

Two further habits save time. Keep the prompts that worked — a short library of five or six proven style blocks is worth more than any tips article, because they are calibrated to the generator you actually use and the look your brand actually wants. And stop at good enough: past three or four attempts on the same brief, you are usually fighting the model's interpretation rather than refining a prompt, and rewriting the subject clause from scratch beats a fifth variation.

Negative instructions and what they can do

Most generators respond to exclusions inside the prompt even without a dedicated negative field, and three are worth appending to essentially everything: no text, no logos, no watermarks.

Those three cover the failure modes that make an image unusable rather than merely imperfect. Generated lettering is almost always subtly wrong, invented logos carry trademark risk, and a hallucinated watermark makes an image look stolen when it is not.

Beyond those three, exclusions are less reliable than positive description. Asking for "no clutter" tends to work less well than asking for "an empty workbench with a single tool". Say what should be there rather than what should not — the positive form gives the model something to render.

Prompt specificity is a metadata advantage

Nobody crawling your site reads your prompt. But a vague prompt produces an image you then cannot describe precisely — and honest, specific alt text is one of the few image-SEO items Google's documentation names directly.

If your prompt said "a locksmith working", your alt text will say roughly that, and it will be weak. If the prompt named the action, the tool and the setting, the alt text writes itself and the filename does too. The alt text guide covers what belongs in the attribute; the filename post covers turning a description into a URL-safe name.

The connection is why SEOpix derives the filename, alt text and EXIF block from the generation itself rather than asking you to describe the image afterwards from memory. The description exists at the moment of creation, when it is accurate — and for teams producing images at volume, the bulk workflow is where that saving compounds.

Try the five-part structure on your next three images and keep the style block identical across them. The difference between that and a bag of quality adjectives is not subtle. Ten images a month are free to test it, and the pricing page has the volume tiers.