Drift goes one way: bigger, busier, shinier

When a generated film misses, the complaint it produces is almost always the same sentence: it looks nothing like my product. That sentence is useless for fixing anything, because it names no property. What is useful is a thing you notice only after looking at a number of these failures side by side — they rhyme. The errors are not scattered. They lean.

Three properties drift, and all three drift the same way. The product comes back bigger than it is. It comes back carrying marks it does not have. And it comes back shinier than the thing in the photograph. Nothing in any brief asked for any of that, which is exactly why it happened.

An unstated property is not a neutral omission

This is the part worth internalising, and it is reasoning rather than a measurement, so here it is in the open. A generator handed a description with a gap in it does not leave the gap empty. It fills the gap with the most likely value, and the most likely value is the average of everything it has been trained on. Product photography as a body of work is not a neutral sample: it is overwhelmingly advertising, and advertising products are shot to look more premium than they are. Bigger in frame, more decorated, wetter.

So the default is not neutral, and silence is not safety. Leaving the finish undescribed does not produce a finish-agnostic render; it produces a glossy one. Leaving the size unstated does not produce a size-agnostic render; it produces a product scaled to fill the composition the generator liked. Every property you do not state is a vote for the average, cast on your behalf.

Which of the three can be stated

They are not equally fixable, and the difference is about where the information lives.

• Size can be stated, if your packaging prints a net content. The quantity on the label is read back and converted into a band of plausible dimensions, which is stated on any frame holding a hand, a face, a model or a countertop. An earlier post, "a photo does not carry scale", covers the mechanism and its limits in full.
• Added marks can be forbidden outright, and are, on every product whether or not anything is known about it. Badges and seals, invented hexagons and starbursts, certification and award stamps, extra flanges and fins, a second logo or wordmark, a callout icon: seven prohibitions, and none of them costs any fidelity, because nothing true is ever lost by declining to add a decoration.
• Finish can be carried forward where it is recorded. A product read as matte — or satin, frosted, soft-touch, brushed, anodised, uncoated, kraft; sixteen words resolve to the same family, because a renderer does not distinguish satin from matte, it distinguishes whether there is a mirror highlight — additionally gets terms forbidding wet-look packaging and added speculars. A product with no recorded finish still gets the weaker instruction not to add a shine the photograph does not have.

The finish is read from the object, not from the label

There is a trap inside the third one, and it is the kind that would be invisible until it bit you. Where the size is concerned, the printed label is the authority — it is the only place a net content lives. Where the finish is concerned, the printed label is the least reliable thing in the frame, because a finish word on packaging is usually describing the contents rather than the pack. A high-gloss tube can quite correctly have MATTE LIPSTICK printed on it in large letters.

So the finish is taken from what the surface looks like in your photograph, and the printed word is only a fallback. That makes one thing you do when photographing the product matter more than it looks like it should: the surface has to photograph as the surface it is. A matte pack shot under a hard raking light photographs wet, and a wet-looking photograph of a matte pack teaches the shoot to protect the wrong finish — while gloss is the direction it was going to drift anyway. Flat, even light is not just easier to read. On a matte product it is the difference between the guard pointing at your finish and pointing away from it.

The one that is deliberately left unguarded

Of the defects this family of errors contains, the closure is the worst and it is the one not on the list above. A cap, a clasp, a flip-top, a spout: these come back redrawn more reliably than anything else, because a small piece of three-dimensional hardware is the part of a product a single photograph resolves least and the part a generator has the weakest prior for. Flat artwork printed on a panel survives; the hardware sitting on top of it does not.

It would be easy to assume that is an oversight. It is a decision, and the reasoning is worth knowing because it tells you what to do instead. A redrawn cap is a wrong part, not an added one. A prohibition against inventing a cap, sitting underneath an instruction to shoot a macro of the cap, is two contradictory instructions in one prompt — and handed a contradiction, a renderer does not stop. It picks the instruction it prefers, which leaves you worse off than the honest version, where nobody pretended the problem was solved.

Which means the closure stays yours to check. An earlier post, "check the cap before the colour", sets out the order to do that in, and the order is not arbitrary: it starts with the small hard parts for exactly this reason.

What the direction is actually worth

Four things follow from knowing drift leans rather than scatters, and none of them takes long.

• Review in the direction of the drift. Ask whether the product got bigger, then whether it gained anything, then whether it got shinier. Three questions, asked in the order they are likely to be wrong, beats looking at a frame and hoping something stands out.
• State what is cheap to state. A legible net content on the pack you photograph, and a surface photographed under light that lets it read as matte if it is matte. Neither costs you a composition.
• Treat the closure as unguarded, because it is. If a framing makes the cap the subject, that framing is the one to look hardest at.
• Know which complaint is which. A paid render where the object itself drifted can be reported, and a check against your reference returns that render's credits. A render you simply did not like is a different thing and is not that.

The honest limit

All of the above is prompt construction. A prompt is an instruction, not a mechanism — it makes the right render much more likely and it cannot make the wrong one impossible, and any claim stronger than that would be the same kind of overstatement this is meant to correct. What has changed is narrower and still worth having: the three properties that used to arrive as somebody else's average are now stated as yours, where your photograph gives us something true to say about them.

Which puts the useful question back where it belongs. Not whether the render will be perfect, but whether your reference photograph says enough about your product that the generator never has to guess. Every gap you leave, something fills.