Five pictures, one face. The prompt describes only the first of them — the flash-lit one against concrete. The other four came from prompts that were not kept.
Against the top image it is exact: dark brown hair in a soft messy bun with wispy strands, fitted light-grey long-sleeve mock-neck, high-waisted black trousers, standing beside a textured concrete wall, one hand holding the collar up near the lips and the other on the waist, direct on-camera flash blowing out the face against a dark hallway with a blurred luxury-apartment depth behind. Even the watch is there.
The demand this prompt makes is testable, and the set is what tests it. Face must remain 99% identical with absolutely no identity drift is easy to write and hard to check against a single image — one picture cannot fail it. Five can. Golden-hour field, warm bedroom, a lift with steel walls, a dark room, a flash-lit hallway: five different light sources, five different colour temperatures, five different distances. The eyes, nose, lip shape and jawline survive all of it. That is the thing the reference-image workflow is actually for, and this is the clearest evidence of it working that this collection has.
Where it slips is the printed furniture, again. The prompt negates watermark, logo, captions, UI elements, text overlay. The lift shot puts a small metal maker’s badge on the denim corset and fills the call panel with numerals that are not numbers. That is now the fourth entry here where a model has supplied plausible printed detail nobody asked for — a seal, a route number, a date stamp, and now a garment label. It seems to be a property of the genre rather than of the prompt: dress a picture as a fashion photograph and it will invent the little marks fashion photographs carry.
