The generated hand is convincing and the ring on it is a curtain ring. Knuckles fold where knuckles fold, light sits correctly on the nail bed, and the band on the third finger is wide enough to hold a napkin. Garment try-on has been allowed to guess at size for years, because a jumper rendered one size too large still reads as a jumper. Jewellery has no such slack, which is why the categories carrying the fattest margin have been last to get a try-on that works.

JewelTry, posted to arXiv on 15 September by a team from the Australian National University and Amazon, treats that slack as the whole problem. Jewellery try-on has until now leaned on a mask to pin the item in place, while general try-on frameworks, the authors write, “do not explicitly model real-world product scale.” JewelTry drops the mask and hands the model the number instead: the product’s physical dimension in inches, written into the text prompt and projected through a small adapter into a single scale token, so the model learns how a stated size relates to the anatomy around it.

The scoring is where the honesty sits. The paper introduces ScaleErr, the absolute log-ratio between the rendered item’s dominant dimension and the real one, where zero means the size is right. JewelTry posts 0.169 against 0.208 for Qwen-JVTON, 0.256 for OmniTry, 0.397 for InsertAnything and 1.158 for Any2AnyTryOn. Read as a natural log, the winning score is still around eighteen percent off the largest dimension of the piece. That is the state of the art, and on a solitaire it is a different ring.

JVTO-Bench, built alongside the model, runs to 23,662 training triplets and 375 test samples across four categories. Existing try-on datasets, the paper notes, “rarely provide jewelry-specific annotations or reliable real-world product-scale information.” So the team used Claude Sonnet 4 to read dimensions off scale-informed product photographs and had humans check the output. We argued in May that try-on’s bottleneck had moved off the model and onto catalogue metadata; here the researchers had to manufacture the metadata before they could train on it.

Jewellery try-on shipped years ago, in one sense. Perfect Corp has been overlaying AR rings and earrings on live camera feeds for years, and its ring and watch pipeline asks a brand to import “industry-standard 3D model CAD files” for each SKU. Its earring product advertises try-on “from a 2D image, no expensive 3D modelling required”. The cheap route exists for the thing that hangs off a lobe and not for the thing sized to a knuckle. A manufacturer holds the CAD file; a marketplace holds a photograph and a caption, which is presumably why Amazon’s name is on this paper.

The benchmark carries the same tilt: 12,149 training pairs for earrings, 2,908 for rings.

The scale problem is older than the rendering. In 1932 Chanel showed Bijoux de Diamants at 29 rue du Faubourg Saint-Honoré, put the pieces on wax busts with real hair rather than in jewellers’ trays, and charged twenty francs at the door for charity. A comet of diamonds laid flat on velvet is a quantity of stones; laid across a forehead it is a proportion. Chanel’s answer was to build the body and let the buyer measure against it.

Signet is still paying for a version of that. The group reported $1.53 billion in sales for the quarter to 1 August, same-store sales up 2.2 percent and gross margin at 39.4 percent, and raised its full-year guidance on the strength of it. Chief executive J.K. Symancyk told analysts on 9 September that the redesigned Kay and Jared sites carry “better imagery and product presentation that includes more realistic on-model photography to help customers buy with confidence,” with Zales to follow later in the month. That is the wax bust, ninety-four years on and shot in a studio. It is also the most expensive answer on the menu, priced per shoot and per SKU.

What changes on screen is the grammar of accessory imagery. Jewellery e-commerce has lived on the macro shot and the flat-lay because both hide scale by convention, and the worn image has been the costly exception. A model that takes a dimension as an argument makes the worn image cheap enough to be the default view, which moves the category’s look from close-up glitter to proportion on a body. Eighteen percent off is not there yet. The number is in the table now, which is where it starts moving.