Jewelry AI

Why a mixed catalogue comes back at different scales

You shot every piece the same way, and the results still disagree about size: the rings sit close, the necklaces sit back. That difference is decided before the image is made, by the category you chose, and no amount of re-shooting moves it.

The difference is decided before the image is made

When a grid comes back uneven, the natural assumption is that the shooting was uneven: a different distance, a different surface, a different day. That assumption sends the seller back to the bench to repeat work that was already consistent.

The app works by category, and the category carries the instruction. Close framing is instructed for rings and bracelets alone; the bracelet page sets out what that close-framing instruction actually says. The other categories are not given that instruction, so a ring and a necklace prepared with the same care are not two attempts at one thing. They are two different requests, and they return two different frames.

That is the whole of the attribution. The variable that produced the drift is upstream of the camera, which means the camera cannot fix it — and it also means the drift is not a fault in the result.

Each category is aimed at a different part of the body

A ring is aimed at the hand, a bracelet at the wrist, a necklace at the neck, an earring at the ear. The mixing guide explains that the category decides the body region; what follows from that here is a matter of size. Those four regions are not the same size, and the region is the reference the frame is built around, so the piece occupies a different share of the frame in each case before close framing is considered at all.

The necklace instruction widens this further by naming two regions rather than one, the neck and the upper chest. The earring instruction does something different again: its extra clause is about whether a single piece or a pair is shown, and says nothing about framing. So the four instructions are not one instruction at four zoom levels. They differ in what they control.

This is also why a returned frame should first be compared with others in its own category. Inside a category the treatment is the same and a difference is worth investigating. Across categories a difference is the expected outcome of two different instructions. The separate question of a ring carrying no size cue of its own is set out on the ring page and is not repeated here.

Systematic, which is the useful part

Because the framing comes from a per-category instruction, it repeats. The same category returns the same treatment, which is why regenerating a necklace in the hope of a tighter result spends credits on an outcome that was already decided.

A repeating difference is easier to work with than a random one. It can be measured once per category and corrected once per category, and the correction stays valid for every future piece in that category. A difference corrected image by image, by contrast, produces a set that disagrees with itself in a new way each time.

There is one temptation to name here, because it looks like a shortcut and is not. Declaring a pendant as a ring to obtain the close-framing instruction also aims the result at a hand and replaces the clause written for that kind of piece. The framing is the smallest thing the category decides.

What comes back, and where the correction belongs

What returns is a single image at the app's fixed output resolution, 1K, framed according to the category that was chosen. There is no resolution selector, no background selector and no aspect-ratio selector rendered in the interface, so there is no setting that would have produced a different frame. Transparency and background belong to a different tool; the background remover has its own page.

That leaves one place for the correction: after the image comes back, while the set is being assembled. Define the adjustment per category, apply it to every image in that category, and check each crop against the features the listing text refers to — the clause instructing that openings, links and closures be preserved protects them during generation, not afterwards.

What a storefront grid then demands of the finished set, including its own image requirements, is covered on the catalogue page. Within a single listing, the images are read together, and the mixing guide sets out why the set rather than the individual photograph is the unit being judged.

Evening up a mixed grid

Results are grouped by category before anything is compared or cropped.
Rings and bracelets are reviewed together, since they share the close-framing instruction.
One adjustment rule per category, written down, rather than one decision per image.
Every crop is checked against the features the description refers to: openings, links, closures.
The finished grid is reviewed at the size a buyer sees it, not only at full size.
No piece has been declared as a category it does not belong to.

What this page does not cover

How the frame is produced. This page describes the instruction the app sends and the image that comes back. It makes no claim about the internal behaviour of the model that produces it.
Grid and aspect-ratio rules for a storefront, and the store's own image requirements, which the catalogue page owns and which are not restated here.
Anything not shipped. No planned capability is described, and nothing here should be read as an indication that a framing control is coming.
Output quality. Preservation of the piece is instructed with every generation; it is instructed, not promised, and the mixing guide sets out what that distinction means for a listing.
Reporting. The report sheet on a result is live, and its first listed reason is that the product changed. A frame that follows its category is not that, so scale drift between categories is not treated here as something to report. A result that came back with a changed stone count or a changed metal colour is a different matter, and the sheet lists those reasons by name.

Questions

Why do my ring photos come back closer than my necklace photos?

Because close framing is instructed for rings and bracelets alone. A necklace was never asked for a tight frame, and it is additionally aimed at a wider body region — the instruction names the neck and the upper chest, where the ring instruction names the hand. The difference comes from the category you chose, not from how either piece was photographed.

Can I change the framing in the app?

No control for it is presented. Output resolution is fixed at 1K, and no resolution, background or aspect-ratio selector is rendered anywhere in the interface. The frame is decided by the category, so the evening up has to happen after the image comes back.

Should I pick a different category to get a tighter photo?

That trade costs more than it returns. The category also decides which body region the result is aimed at and which clause is sent with it, so a pendant declared as a ring is instructed toward a hand. Declare the category the piece belongs to and adjust the frame afterwards.

Is a difference in scale between categories something to report?

It is the expected result of two different category instructions, so it is not a defect. The report sheet on a result is there for a different situation: its first listed reason is that the product changed, and it names problems such as a changed stone count or a changed metal colour. If the piece itself came back altered, that is what it is for.