Accurate AI Ring Try-On Needs Millimetres, Not Just a Photo

A clean product photo can tell an AI image model what a ring looks like. It cannot, by itself, tell the model how large that ring should appear on a particular hand. This is why an attractive try-on image can preserve the general design while making a 7 mm face look like 11 mm, thickening the band or moving the setting to an implausible position.

New jewellery-specific research makes the problem unusually clear. JewelTry, a September 2026 research paper and benchmark, treats physical scale as an explicit input rather than asking the model to infer it from pixels. The practical lesson for sellers is simple: if a try-on workflow does not know the product’s measurements, use its output as styling inspiration, not as evidence of fit or size.

Why a reference photo does not contain real scale

A standalone packshot has no dependable unit of measurement. The ring could be 12 mm or 25 mm across because the camera may be closer, the image may be cropped, and the background may contain no object of known size. Two rings can fill the same square image while being very different dimensions in real life.

The hand photo creates a second uncertainty. Finger width changes with perspective, camera distance, pose and lens choice. A hand-landmark system can locate joints and finger regions, but locating anatomy is not the same as knowing the absolute size of the jewellery. Google’s MediaPipe Hand Landmarker, for example, returns 21 hand-knuckle coordinates. Those coordinates are useful for placement and tracking, but the product still needs scale information.

JewelTry addresses this by encoding physical product dimensions into a scale token. Its authors describe jewellery as especially difficult because it is small, rigid and sensitive to fine detail. Their benchmark covers rings, earrings, necklaces and bracelets, with more than 23,000 images and real-world scale annotations. This is a research system, not proof that every commercial try-on tool now handles measurements correctly.

Record the dimensions the image needs to preserve

Start with the dimensions a shopper can actually judge from a worn view. For a ring, record the width and height of the visible face, the band width, the setting height and the inner diameter or ring size. For a signet ring, the face dimensions may matter more visually than the inner diameter. For a solitaire, the stone diameter and setting height can be the details most likely to drift.

Keep these values beside the product assets rather than buried in a description. A useful file note might read: face 10 x 8 mm, band 3 mm, setting height 5 mm, UK size N. If you create several colour variants from the same design, share the geometry record but verify that plating, stone or construction changes have not altered the real dimensions.

Ask any try-on provider a direct question: can the workflow accept numerical product dimensions, or does it infer size only from the reference image? If it accepts measurements, check which ones and how they are used. If it does not, the result may still be useful for mood, composition and colour, but it should not be presented as a precise size preview.

Prepare a reference that protects ring geometry

Use one clean image containing one ring. Photograph it straight enough that the face shape, band opening, shoulders, stone count and prong layout are easy to read. Keep some margin around the product, but do not make the ring a tiny object in a large frame. JewelTry’s dataset preparation used tightly cropped product references with additional margin, reflecting the need to give a model clear product information without losing the outline.

Control the physical photograph before using AI. GIA recommends stabilising the camera, cleaning jewellery carefully and using diffused lighting for polished metal. Diffusion produces softer tonal gradients and reduces hard hotspots that can hide the band edge or setting. Capture a second angle when the side profile or setting height matters.

Give the hand image a scale anchor

The safest option is a real worn photograph of the actual ring. If that is not practical and you are making a controlled composite, photograph the hand with a small reference of known size in the same plane as the finger. A ruler segment or calibration card can be placed beside the hand, then removed from the final image after the ring has been scaled. The reference must sit at the same depth as the finger; a card closer to the camera creates a false measurement.

Use a neutral, open hand pose with the target finger visible, minimal motion blur and no existing jewellery near the placement area. Avoid wide-angle close-ups that exaggerate the nearest fingers. Keep the camera angle and hand pose consistent across a product family so that one ring is not made to look larger simply because the hand was photographed differently.

Check scale, placement and identity separately

A try-on image can pass one check and fail another. Review it in three passes.

First, check identity. Compare the ring with the packshot: face outline, band shape, openings, stone count, prongs, engraving, edge profile and metal colour. Reject a result that redesigns the SKU, even if it looks realistic.

Second, check placement. The band should wrap behind the finger rather than sit like a sticker on top. The near and far edges need believable occlusion. The setting should follow the finger angle, and shadows should agree with the hand lighting.

Third, check scale. Compare the visible ring face and band width with your recorded millimetres and the calibrated hand image. Do not judge only by whether it “looks about right.” Small jewellery can be materially misleading after a few pixels of enlargement.

Publish try-on images in the right role

Add the real measurements in the listing and keep them consistent with the visual. If the try-on is illustrative rather than calibrated, say so. Phrases such as “visual styling preview” or “shown for styling; see measurements for exact size” are clearer than implying a precise fit.

Ring size is also not the same as visual scale. UK size N describes the fit around the finger, while a 6 mm and 12 mm ring face can both be made in size N. Sellers need to communicate both.

Where AI is safe and where it is not

The important shift is to stop treating virtual try-on as one problem. It is at least four problems: product identity, anatomical placement, physical scale and photorealistic integration. A tool that solves three can still mislead on the fourth.

Until measurement-aware jewellery try-on becomes standard, keep the rule practical: provide dimensions when the system can use them, calibrate any manual composite, verify against the product record and publish a real dimension graphic beside the attractive image. A ring reference photo shows design. Millimetres establish scale.

Sources

Niu et al., “JewelTry: Mask-Free Scale Aware Jewelry Virtual Try-On,” arXiv, 15 September 2026: https://arxiv.org/abs/2609.16626

Google AI Edge, “Hand landmarks detection guide,” updated 1 October 2026: https://developers.google.com/edge/mediapipe/solutions/vision/hand_landmarker

Gemological Institute of America, “How to Photograph Gems & Jewelry”: https://www.gia.edu/gem-photography

Sources reviewed 8 October 2026. JewelTry is research, not a claim about the measurement controls or commercial availability of any current jewellery try-on product.