It happens every week in creator groups. Someone posts a reference photo they love — a portrait with perfect window light, a product shot with a beautiful hard shadow — and asks, “what prompt gives me this?” Then they show what they typed, and it describes a different photograph entirely. Not a worse photo. Just a different one. Because their words reported the subject of the image and missed everything the image was actually doing.
Translating a photo into prompt words is a learnable skill, and it’s one I practise constantly, in reverse. In a storyboard-first pipeline, I’m always turning approved frames into words and words into frames. The looking protocol below is the one I use when a reference — a still I shot, a frame from my own tests, a look a client sent — needs to become a prompt that can reproduce its decisions in a new generation.
It isn’t about copying a photograph. It’s about reading one properly.
Look Longer Before You Type
The single biggest improvement is unglamorous: time. Most prompt translations fail because the looking lasted four seconds and the typing lasted forty. The eye grabbed the subject — a woman, a café, a jacket — and the prompt became a description of nouns, while the photograph’s actual character lived in light direction, lens compression, depth, and colour relationships nobody paused to name.
Set a small discipline. Before typing a word, look at the reference in three passes. First pass: what is in the frame — the inventory. Second pass: how is it lit, honestly — where’s the light coming from, what’s it doing to faces and surfaces. Third pass: how was it captured — how close does the camera feel, how much of the scene is sharp, what’s the colour doing. Thirty seconds of this produces a prompt that resembles the photo. Skipping it produces a prompt that resembles your memory of the photo’s topic.
Which is a very different thing.
Zoom in. Details that hide at full-frame view dominate prompts once named: catchlights in eyes tell you where the key light was; the blur gradient behind a subject tells you the lens more honestly than any metadata; the edge quality on a shadow tells you whether the source was hard or soft. Models respond to these specifics because they’re consistent features of photographs like yours. Your translation improves in direct proportion to how many of them you caught.
Also look at what the photo refuses to show. Cropped space. Blown highlights accepted rather than rescued. Background allowed to fall dark. A reference’s restraint is part of its look — and prompts that tidy everything up (“well-lit, detailed background, everything in focus”) translate the restraint away. Describe the omissions when they’re beautiful. “Background falls into shadow” is a rendering instruction. And it’s often the one that made you love the photo.
The Five Things Every Photo Is Telling You
Every reference photograph, whatever its subject, makes five reportable decisions. Extract all five and your translation stops drifting.
1. The subject and its state. Not just who or what, but in what condition: this age, this wardrobe state, this expression at rest, this product angle with this much wear. State is where “a man” becomes filmable.
2. The spatial arrangement. Where the subject sits in the frame, how close, what shares the frame and at what distance. Foreground, subject, background are distances, and photographs state them precisely. “Subject large in frame, background far and soft” translates. “Nice composition” doesn’t.
3. The light. Source, direction, hardness, and colour temperature relative to the scene. This is the decision most responsible for mood — and the one most often replaced in translation with the word “moody,” which is where translations go to die.
4. The capture. The lens character and medium impression: compression or openness, depth of field behaviour, grain or cleanliness. You’re not identifying the exact equipment. You’re describing what the capture did to space and texture.
5. The colour relationships. Not a list of colours present — the relationships. Warm subject against cool background. One saturated accent in a muted field. Shadows that go green, highlights that go gold. Colour in good photographs is a small set of decisions, and models reproduce decisions better than inventories.
Miss one of the five and the model invents it — and its invention becomes the first thing you notice in the output. In my own translation misses, light is the most common omission and capture the second. The inventory, decision one, is almost never the problem. Everybody can list what’s in a photo. Fewer people can say how the photo saw it. And that is the whole transferable value.
Subject, Light, Lens: In That Order
When you write the translation, order it the way a render builds: subject first, then the light that models it, then the capture that records it. This mirrors the subject-first discipline used across this site’s prompting guides for a practical reason — each layer only makes sense applied to the one before it.
Subject sentence first, with state included. “A woman in her thirties, dark hair pulled back, seated in profile at a café table, hands around a cup.” That places the person and her one held action. Then the light, named like a source and a direction, not an adjective: “Window light from camera left, soft, falling off to shadow on the right side of her face.” You’ve just described the photograph’s modelling — why one cheek is bright and the other isn’t — in words a model can re-render on a new face in a new café.
Then the capture sentence: “Shot at a close focal length, shallow depth of field, background dissolved to colour.” Three clauses, each visible in the reference, each doing separable work. Note what the order buys you. If the generation is wrong, you know which layer to interrogate. Subject wrong is a wording problem. Light wrong is a source problem. Capture wrong is a lens problem.
A translation written as one blended paragraph gives you no such diagnostics. It’s why the companion guide on checking generated footage properly starts by separating what a frame is doing into layers just like these.
Setting belongs with the subject layer, briefly, as concrete nouns. The café is “marble table, dark wood, one window” in the reference — not “a cosy European café atmosphere.” Atmosphere is the sum of the light and capture sentences doing their jobs. Let it be a result, not a request.
Colour Words Beyond ‘Nice’
Colour is where translations get compliment-heavy and information-light. “Beautiful colours” and “vibrant tones” ask the model for its default palette, which is nobody’s reference. Working colour language is specific in three ways: it names the dominant relationship, it places saturation deliberately, and it describes what the shadows and highlights do.
Dominant relationship first, in the form warm-against-cool or muted-with-one-accent. “Warm amber skin tones against a cool blue-grey street” is a decision a model can hold across a generation. Then saturation placement: “muted overall, one saturated red sign.” That single accent is often the real reason a reference works — and naming its loneliness protects it. Left alone, models distribute saturation democratically and the accent drowns.
Shadow and highlight behaviour is the connoisseur layer, and it’s worth your vocabulary. Shadows that stay neutral read modern and clean. Shadows that go teal or green read as certain film habits. Highlights that roll off gently read as film or good digital; highlights that clip hard read as phone or camcorder honesty. You don’t need technical colour science to write this. You need to look at your reference’s darkest corner and brightest spot and report what they’re doing in plain words.
Plain words that have carried colour reliably in my prompts include: muted, desaturated pastel, high-contrast with crushed blacks, warm highlights and cool shadows, faded like an old print, one bold accent colour. What these share is that each excludes something. “Muted” excludes punch. “Crushed blacks” excludes shadow detail. Exclusion is information. “Rich, beautiful, cinematic colours” excludes nothing — so the model’s average walks in. The same average discussed in the lighting prompting guide from the light side. Translation is the art of excluding everything the reference is not.
One honesty note on colour: if part of your reference’s palette comes from the subject itself — a red coat, a yellow product — separate wardrobe colour from grade colour in your prompt. State the coat as subject description. State the grade as capture behaviour. Otherwise the model may grade the whole frame toward the coat or desaturate the coat toward the grade, and either way a decision you loved gets negotiated.
What To Deliberately Leave Unsaid
A translation is judged as much by its omissions. Leaving things unsaid isn’t vagueness. It’s protecting the model’s freedom exactly where the reference’s value doesn’t depend on control.
Leave the exact identity unsaid when what you love is a construction, not a person. Translate the construction — age range, build, hair, wardrobe state — and let a new face arrive. Insisting on a specific recognisable identity from a photo you don’t own creates likeness problems no prompt phrasing solves. And it’s unnecessary: the look transfers without the person.
Leave micro-detail unsaid. Fabric weave, distant signage lettering, individual leaves — detail at that scale is either inherited beautifully in image-to-image workflows or invented noisily in text-to-video. Naming twenty small objects bloats the prompt and dilutes the five decisions that matter. Name the three objects that carry the scene’s meaning and trust the model’s invention for the rest. Invention is a feature when the big decisions are locked. It’s only a bug when it’s making the big decisions too.
Leave the backstory unsaid. “A woman who has just received bad news waits in a café” imports a narrative the photograph doesn’t contain and taxes the face toward melodrama. The photograph contains a woman, a window, a held cup, a gaze. Translate what’s visible. The story you project onto the reference is yours — and it’ll re-project itself onto a faithful translation without being written in.
And leave quality bait unsaid: no resolution numbers, no masterpiece language. Your reference is a photograph, not a contest entry. The words that reproduce it are observational — and observation is a quieter register than praise. When a draft translation starts sounding like an award citation, rewrite it as a location scout’s notes. Scouts describe. That’s why their descriptions can be rebuilt.
A Worked Translation, Photo To Prompt
Let me run the full protocol on an ordinary, non-famous kind of reference: a portrait many creators will recognise — a person at a kitchen table in morning light, laptop open, shot by a friend who happened to get it right.
The lazy translation writes itself: “A woman working on a laptop at home in the morning, cosy, natural, cinematic.” Subject-ish, mood-heavy, no light, no capture, no colour decisions. The generation will be a well-lit stock cousin of the photo.
Now the looking protocol. Inventory: woman, thirties, hair up, plain sweatshirt, both hands near a laptop, mug beside it, table worn wood, background a dim kitchen. Light: one window, camera left, low and warm, modelling her face from the side, background falling dark, a soft highlight on the mug’s edge. Capture: close-ish lens, background gently dissolved, texture clean but not glossy, nothing tack-sharp except the near plane. Colour: warm subject against a cool, dim room; palette muted, the only saturation a faded blue sweatshirt and the warm wood.
The translation, in subject-light-capture order: “A woman in her thirties, hair pulled up, faded blue sweatshirt, seated at a worn wooden kitchen table with a laptop open in front of her, one hand resting beside a mug. Low warm window light from the left, the right side of her face in soft shadow, background kitchen dim and out of focus. Close lens, shallow depth of field, muted natural colour, warm highlights against cool shadows. Photorealistic.”
Count what that prompt did. It made all five decisions explicitly. It stated the subject and her held state. It placed her spatially — near plane sharp, kitchen dissolved. It named a light source, a direction, a falloff, and a consequence. It described the capture’s distance and depth behaviour. It reported the colour relationship and protected the muted field. And it left unsaid everything the look didn’t need: her identity, the laptop’s contents, the kitchen’s small objects, her story.
Generate that, and the output won’t be the photograph — a new face, a new kitchen, a new morning arrive every time. But the decisions will match. And decisions are what made you stop scrolling on the reference in the first place.
Compare layer by layer, adjust one layer at a time, and save the translation that survives into your prompt library next to the reference itself. Over a season, those pairs — photo on one side, words on the other — become the most personal prompting asset you own. The broader prompting system on this site is built to hold exactly such pairs, because the skill being trained here isn’t typing.
It’s looking. The prompt box is just where looking becomes visible. Slow down on the photograph, report its five decisions in order, say less than you want to about everything else — and the words will start deserving the images they came from.