A street-style prompt is mostly a list of nouns, and it is worth knowing which of them are working. Four frames went out that were identical except for one word or one sentence. The word was the fabric. The sentence was the light.
The frame, identical in all four. A full-length street-style photograph of a man standing in front of a plain concrete archway with wide shallow steps, shot from a low camera position so the vertical lines of the arch run straight up behind him… He wears a knee-length single-breasted overcoat, open, over a plain dark crew-neck and straight trousers. The overcoat is made of [heavy brushed wool / smooth black leather / technical nylon]. [Soft even daylight from directly behind the camera / Hard low sunlight raking across him from the left, almost parallel to the coat]… He fills most of the frame height.
The fabric words
To read the cloth I measured two things on the same panel of each coat: how much fine-scale variation the surface carries, which is what a weave looks like, and what fraction of it reaches the brightness of a light source, which is what a sheen looks like. The measure was checked first on synthetic surfaces — a smooth gradient, a gradient with a fine weave printed into it, and a gradient with a mirror highlight — and it separated all three correctly.
Same cut, same arch, same light. Only the fabric word changed.
Wool and nylon came back within two percent of each other on the surface measure, which for two materials that feel nothing alike is not a distinction anyone will see. Leather did separate, but not by texture — by sheen. It is the only one of the three with any pixels at light-source brightness, and it is the only one whose surface a viewer can name at a glance.
What the fabric word did buy was everything else: the nylon came back quilted with a padded collar, the leather came back black and glossy with a soft drape, the wool came back a heathered brown. Silhouette and color arrived. The surface did not.
The light sentence
Then the same wool, with one sentence changed from flat frontal daylight to hard sun raking across the coat.
The fabric word is identical in all three. The middle and right crops are the two sides of the same coat in the raking-light frame.
One sentence took the surface from 5.82 to 11.08, a 90% increase — against the 2% that changing wool to nylon managed. And the shadow side of that same coat reads at 10.03, still well above the flat-lit version, so the grazing light is not just carving one bright edge; it is giving the whole garment somewhere for its nap to catch.
That is the answer to the question this row starts from. The cloth is not illegible because the prompt failed to name it. It is illegible because flat frontal light has nothing for a weave to do.
The two things I was told to worry about, that did not happen
The usual complaints about outfit posts are a cluttered background and a subject too small in the frame. With the arch framing, neither turned up.
All four backgrounds are plain: against a reference of vertical lines every 60 pixels, which measures 12.83%, the busiest of these is 9.28%. The raking-light frame is the plainest of the lot, because the hard light dropped half the architecture into flat shadow, and shadow has no edges.
Headroom is the one that drifted. Every prompt said em>he fills most of the frame height/em>, and the empty space above his head came back between 15.6% and 20.5% — about a third more in one frame than another. Close enough to be the same shot family, not close enough to be a set. If you need a grid where the figure sits at the same height in every tile, you will be cropping afterwards.
What to do with this
Where this stops
One arch, one coat pattern, one run per condition. Three fabrics on one garment is not a survey of materials, and the wool-to-nylon result says those two words did not separate here, not that no fabric word ever will — leather separated in the same test. The light comparison is one pair, so the 90% is the size of one difference, not a rate.
The two lighting frames also differ in more than light: the model gave slightly different architecture each time, which is why the background numbers are reported for the whole frame rather than a hand-picked patch. An earlier version of that measurement used a small sample box and returned the opposite answer purely because of where the box sat.
Cloth is not written, it is lit
The instinct when a coat comes back looking like a flat shape is to describe the coat harder — more words about the fibre, the weight, the hand. The measurements say those words land somewhere else: they choose the garment, not its surface. What decides whether a viewer can tell wool from nylon is whether anything in the frame is raking across the cloth at a low angle, which is the same thing a photographer would do with a real coat and a real window. Write the fabric to get the right coat. Then write the light, because that is the sentence the surface is waiting for.
Configuration and cost. Four stills: Image, Seedream 4.5, 16:9, 2K, Standard — the panel declared 1 credit for each. Total 4 credits, 2026-09-22. Credit costs are tied to that configuration on that date.
Files. Four stills at 2560×1440. Prompt text is quoted as sent, abbreviated only where marked; the bracketed alternatives are the only parts that differed between runs.
How the surface was measured. Weave is the standard deviation of the image minus a blurred copy of itself, taken over the garment only and excluding any near-clipping pixels and their surroundings, so that a bright highlight cannot be counted as texture. Sheen is the fraction of garment pixels at or above light-source brightness. Both were checked on synthetic surfaces first: a smooth gradient read 0.03 weave and 0% sheen; the same gradient with a fine weave printed into it read 7.00 and 0%; the same gradient with a mirror highlight read 0.57 and 6.13%; and with both, 7.03 and 6.13%. Background edge density was checked against flat gray (0.00%) and against vertical lines every 60 and every 20 pixels (12.83% and 39.50%).
Every sampling box was drawn back onto every image and inspected. Three rounds of that were needed: the first set of boxes sat on buttons and lapel edges, the second put one of them on a sleeve rather than a front panel, and the third had to be shortened because two boxes reached down over a hand.
Two instruments were discarded. A subject-scale measure that looked for the largest dark region in a central strip returned the top of the coat rather than the top of the head, and the bottom edge of the frame rather than the feet. A replacement that looked for the first row significantly darker than its own surroundings was defeated by the dark arch openings behind the figure, returning the chest in one frame and the top of the frame in another. Head positions in this article were read off a labelled grid by eye and then drawn back onto all four images to confirm they land on hair.
One background measurement was withdrawn. A hand-placed sample patch made the raking-light frame look like the busiest of the four at 10.34%; the patch was straddling the edge of a dark arch. Measured over the whole background instead, that frame is the plainest at 6.52%. The figures in this article are the whole-background ones.