Photoshoot prompts: five of the eight you write yourself

A photoshoot brief settles eight things. Who is in the frame. The pose. The wardrobe. The light and where it falls. The backdrop. The camera and lens. The colour grade. The crop. Across 480 lines generated from 40 real photographs, the line an image to prompt tool hands you settles three of them — light, colour, crop — and carries zero words about the other five: not one of the 480 lines says anything about a pose, an outfit, a backdrop or a lens. This page is the accounting, plus the eight clauses that fill the gap and what they cost.

What these numbers are. Forty real files — 39 photographs at 640 × 360 and one code screenshot at 800 × 450 — were decoded on this machine, box-downsampled to the 192 × 192 grid the live tool measures on (36,864 cells regardless of file size), and composed into four output shapes at three lengths: 480 lines, twelve per file. No model was called and no image was generated. Every line was re-rendered a second time with each fragment tagged, and all 480 matched the real output byte for byte, which is what makes the word counts below attributable to slots. The eight lighting setups in the fourth section are pixel patterns generated on this machine, not photographs; they are described as what they are.

Here is a real line, 259 characters, and the eight things a shoot brief would have to settle next to it:

widescreen 16:9 composition, bright, high-key lighting, dominant palette of silver, moss green, cream, grey, muted, desaturated tones, neutral colour temperature, medium contrast, moderate detail and texture, evenly lit, balanced, natural atmosphere --ar 16:9

Everything in it is true of the picture and none of it is about a person. That is the whole problem with using a generated line as a photoshoot prompt, and it is a structural one rather than a tuning one: the measurement layer has no detector for people, and the data file that lists what is measured puts people faces objects brands scenes in a row tagged not measured and never claimed.

Eight things a shoot brief settles, three of them answered

Take the eight items above and ask, for each, which of the twelve slots in the composer could possibly speak to it, then count how many of the 480 lines actually do. The answer does not vary by output shape or length: all twelve shape-and-length combinations settle exactly 3 of 8. Three items are answered in every line; four are answered in none; one — who is in the frame — is answered in 48 lines and only because the tool has a word for a monochrome picture.

The last row is worth a sentence on its own. A long Midjourney line spends a median of 5 words on surface detail — moderate detail and texture and its two siblings — which answers nothing a photographer puts in a brief, while spending 0 on the lens that would actually produce it. And the three items that are answered are answered coarsely: item 4 is a brightness bucket plus at most two sentences about where light falls, item 7 is five slots that are really about colour statistics rather than a grade, and item 8 is the file’s own aspect ratio.

The shape matters for how much you get, not for what you get. On item 4, Midjourney’s long line spends 9 words and DALL·E’s 3, because DALL·E never writes a light-distribution phrase at any length. On item 8, Midjourney spends 3 and the others 2, because only Midjourney appends the aspect-ratio parameter. Neither difference changes the count of items settled: still 3 of 8.

The words a photographer reaches for, and where they live

Twenty-five words from ordinary shooting vocabulary were searched across the nine source files of this pipeline (18,714 characters) and across the 104,976 characters of generated output. Twenty of them returned zero on both sides.

Five did not. Three of those five are accidents of English rather than photography:

So the single photographic word the tool will write about your picture is black and white photograph, it appears in 48 of 480 lines, and it is triggered by a saturation share crossing 0.85 — not by anything being a photograph. Everything else you would say about a shoot — the pose, the wardrobe, the backdrop, the lens — you write.

Six clauses you will paste in, and what the line already claimed

The second problem is subtler than absence: on four of the eight items the line has already taken a position, and if your brief takes the opposite one, both are in the finished prompt. Six clauses a photographer would plausibly write were checked against all 480 lines.

Two of these are worth reading twice. Bright and airy is contradicted by 300 of 480 lines — and by 0 of them left unaddressed, because exposure is written at every length in every shape. The vertical crop is contradicted by all 480, but that is an artefact of this corpus: all forty files are 16:9, so framing never moved; a portrait file would flip it in one step. The parser, meanwhile, does not arbitrate. Paste your clause in and the finished prompt simply carries both claims.

Eight lighting setups, and how many the line can tell apart

What the line does carry is light — so it is worth asking how much light it can actually distinguish. Eight pixel patterns were built at 640 × 360: a high-key seamless sweep, a low-key seamless sweep, a rim-lit frame, light falling from the top, light bouncing from below, a vignette, a warm key and a cool key. All eight were measured on the same 192 grid.

Mean luma across the eight runs from 45.81 to 231.83 and the edge figure from 0.00 to 3.16, yet the eight setups produce only 3 distinct exposure phrases, 2 contrast phrases, 2 saturation phrases, 3 temperature phrases, 3 distribution phrases, 3 vertical phrases and 1 detail phrase. The single detail phrase is the sharpest example: every one of the eight, from a flat field to a hard rim, comes out clean and minimal, smooth surfaces. Six of the eight are grey, so all six are also described as monochrome, black and white and pick up the black and white photograph subject — a grey seamless sweep reads as a black-and-white picture.

Taken as a whole twelve-line signature, all eight setups are different from each other: 8 distinct signatures, 0 identical pairs out of 28. But blank out the palette words — the one slot that acts as a fingerprint — and the differences collapse fast. A long Midjourney line still separates 7 of the 8; a short DALL·E line separates 4. Here are the two setups that a photographer would call opposites, as a long Midjourney line and a short DALL·E line:

Midjourney / long, top-lit   : … well-lit, balanced exposure, monochrome, black and white, neutral colour temperature, medium contrast, clean and minimal, smooth surfaces, evenly lit, brighter upper half, balanced, natural atmosphere --ar 16:9
Midjourney / long, bounce    : … well-lit, balanced exposure, monochrome, black and white, neutral colour temperature, medium contrast, clean and minimal, smooth surfaces, evenly lit, brighter lower half, balanced, natural atmosphere --ar 16:9
DALL·E / short, top-lit      : A widescreen 16:9 image dominated by black and white photograph, well-lit, balanced exposure, with a palette of silver, grey, slate, charcoal.
DALL·E / short, bounce       : A widescreen 16:9 image dominated by black and white photograph, well-lit, balanced exposure, with a palette of slate, grey, silver, charcoal.

The Midjourney pair differs by exactly one phrase, brighter upper half against brighter lower half. The DALL·E pair differs by nothing at all except the order of four colour names — silver, grey, slate, charcoal against slate, grey, silver, charcoal — because the vertical slot is one that only Midjourney and Stable Diffusion ever write. At short length, three more setups merge: low-key, rim-lit and vignette all produce the same line once the palette is blanked, since contrast and distribution are not written at that length.

What your eight clauses cost

So write them. Here is one eight-clause brief — 209 characters, 33 words, eight comma-separated clauses, every weight 1, no out-of-range term — placed in front of each of the 480 lines and run through the same parser the tool uses for weights:

two people seated close together, relaxed natural pose, matching neutral wardrobe, soft window light from camera left, plain mid-grey backdrop, 85mm lens at f/1.8, warm muted colour grade, vertical 4 by 5 crop

Your half is between 39.5% and 48.3% of the finished line at the long setting, and up to 66.8% at the short one. Nothing about the insert breaks the parser: eight clauses arrive as eight terms at weight 1, the out-of-range count does not move, and the round-trip through the explicit-weight converter still returns every Stable Diffusion, Flux and DALL·E line unchanged. Midjourney reads 0/40 before and after — not because the brief broke anything, but because its own --ar 16:9 tail is read as a weight of 9 in every line, including the untouched ones.

Put together, the way to build a photoshoot prompt from a picture is:

1. Keep the three the tool measured. Light, colour and crop are real measurements of your file, not guesses. They are also the only parts of the line that survive a change of picture.
2. Delete before you add on the four that clash. If your brief says bright and the line says low-key, or says vertical and the line says widescreen 16:9, one of them has to go — the parser keeps both.
3. Write the five that are missing, in front. Who, pose, wardrobe, backdrop, camera. None of them appear in 0 of 480 lines by accident; there is no slot for them.
4. Do not expect lighting setups to survive the trip. Eight setups collapsed to 4 distinguishable lines on a short DALL·E line, and the difference between two of them was the order of four colour names.
5. Check the grey trap. Six of eight synthetic setups came out monochrome, black and white. If your reference is a seamless sweep, expect the line to call it a black-and-white photograph and take the word out yourself.

What is left is a two-part prompt: a measured half you did not write and cannot argue with, and a written half of eight clauses that is 39.5% to 48.3% of the line. If you want the measured half explained slot by slot, what a photo prompt is is where the thresholds are listed, and what this tool measures is the underlying thirteen-row signal table. If you would rather start from your own words and see where they land in the line, how to write a picture prompt covers the insertion points, and good image prompt examples shows finished lines with the same accounting.