The image model was never asked to leave room for the copy

01The omission

The text-free branch of the production prompt does not ask for negative space. It asks for the opposite.

THIS AD CARRIES NO TEXT AT ALL. Render a pure product image… The photograph alone has to do the work.

That closing sentence pushes the model toward a self-sufficient, filled frame, and a separate PRODUCT SCALE AND FRAMING rule pins the product to roughly the same share of frame as the reference. So every plate in the earlier bake-off arrived as a complete photograph, and the compositor then had to hunt for clean regions that were never created.

The architecture already knows how to reserve space — there is a reserved-corner rule so a real logo can be composited without the model drawing one. It simply was never extended to copy. For this test that closing sentence is replaced (not appended to, which would leave the prompt arguing with itself) by a clause asking for one generous, even band, clearly light or clearly dark.

02Does the clause change the pixels?

Not a matter of opinion. A band scan slides a 20%-height window down the safe box and scores every candidate on evenness and value. “Usable” means spread ≤ 28, mean outside the 110–138 dead zone, and best available ink clearing 4.5:1.

The headline rate is the weaker number. The median is the real signal: it moves from 0 to 4 usable bands per plate. When the clause lands, the model does not leave one marginal gap — it opens most of the frame. When it misses, it misses completely.

03Seed, incumbent, plate, and both arms

Every row leads with the original seed, outlined in amber, so any claim about fidelity can be checked rather than taken on trust. Column two is what ships today: gpt-image-2 setting its own type. The two arms are the ones that measured best so far — gpt-5.4 directing from the plate, and gpt-5.4 shown gpt-image-2’s own finished ad and told to reproduce its type treatment.

Black leather shoe dark product

PELAGIC Torrent Jacket

Gymshark Flutter Tee

Gymshark Campus Crest Tee

04The contrast finding that closed a false wall

A separate zero-cost test: four director specs applied across all 16 portrait plates, 192 type elements, local background luminance 48–218.

This corrects something written into the handoff as a hard limit. White clears 4.5:1 up to background ≈118.7; pure black clears it from ≈116.1. Those ranges overlap, so every background is coverable and no dead zone exists. A designer’s near-black only works from ≈126.3, which manufactures a gap at 118.7–126.3 — exactly where all nine failures sat. “No-scrim is not always satisfiable” was wrong: the constraint was satisfiable, the fallback palette could not reach it.

05Band scan, every plate

06What is still wrong

• Aspect is 2:3, not 9:16. Verified live against the schema: gpt-image-1.5/edit offers only 1024x1024 / 1024x1536 / 1536x1024. The prompt’s own FORMAT block declares 1152×2048, so the plate geometry and the declared geometry disagree. Unavoidable on this model.
• n = 8. Two plates per seed. Enough to show the clause moves the pixels; not enough to put a compliance rate on it.
• The copy is synthetic. Ratings, quotes and names on this page are invented test strings for the typographer to set. The real path hard-gates fabricated proof through quoteProvenance; none of this is review data.
• Gymshark’s typeface was not obtainable from their site, so those rows are set in PELAGIC’s Archivo. Fine for comparing type direction; not brand-accurate.
• The incumbent column is square. No gpt-image-2 run exists at portrait, so type-to-frame scale is comparable but the crop is not.

07The decision

Exactness is no longer the question — composited copy is exact by construction, and the renderer now guarantees 4.5:1 on any background. The question is entirely taste: read down each row and decide whether the composited arms are ones you would ship next to what gpt-image-2 produces on its own.

If they are, the work that remains is making the reservation clause reliable — it is obeyed about half the time, and the misses are measurable before a single peso of typesetting is spent, which means a plate without a usable band can simply be regenerated.