Why AI Jewellery Try-On Gets Size Wrong (and How to Fix Ring and Bracelet Scale)

AI can put a bracelet on a wrist or a ring on a finger in seconds, but one of the easiest mistakes to miss is scale. The product may look realistic while being noticeably too large, sitting too high on the finger, or wrapping around a wrist in a way the real item never would.

This is a specifically difficult problem for jewellery. Research on ornament virtual try-on notes that rings, bracelets, earrings and necklaces contain tiny repeated structures and often need to survive large changes in pose and scale between the product reference and the person wearing it. A recent virtual ring try-on paper takes the problem even more literally: it detects hand landmarks and rescales the ring using finger thickness before placement.

For ecommerce sellers, the lesson is simple. Do not ask the AI to guess size from appearance alone. Give it visual and numerical evidence whenever scale matters.

Why scale errors happen even when the image looks good

A normal product photograph tells the model what the jewellery looks like, but it may not tell the model how large it is in the real world. A 15 mm pendant and a 25 mm pendant can occupy almost the same number of pixels if they were photographed at different distances. The AI sees shape and appearance, but physical scale is ambiguous.

The same problem appears on the person side. A close-up hand photograph may make the wrist look wider than it is. Perspective, focal length, hand pose and camera distance all change the apparent size of fingers and wrists. If the model has to infer both the jewellery size and the body scale at the same time, small errors are unsurprising.

Use one reference to define the product and another to define scale

Modern image models can accept several references, so use them for different jobs rather than uploading a pile of similar photos. Google's current Gemini image documentation supports multiple high-fidelity object references, and it also recommends pose references for complex character poses.

For a bracelet, a useful set might be one clean hero photo, one side or clasp view, and one image that provides scale. That scale reference could show the bracelet beside a ruler, on a known wrist, or include a clear measurement such as 18 cm inner length and 5 cm extender. For a ring, use the exact ring diameter or size rather than relying on the model to infer it from the photograph.

Tell the model what each image is for. For example: "Image 1 is the exact bracelet design. Image 2 shows the clasp. Image 3 is only for physical scale. Preserve the design from image 1 and match the bracelet size shown in image 3." This reduces the chance that the model blends different references together without understanding their roles.

Measurements are more useful than adjectives

Instructions such as "small bracelet" or "delicate ring" describe style, not physical size. If you know the measurement, include it. A prompt can state that a pendant is 20 mm wide, a bracelet is 6 mm across, or a ring has a 17.3 mm internal diameter.

Measurements do not guarantee mathematically exact rendering, but they remove ambiguity from the brief. They are especially useful when the same product needs to appear across several people or poses, because the jewellery should not grow or shrink simply because the camera framing changed.

Match the hand or wrist pose before asking for the jewellery

A good try-on image is not only a jewellery-placement problem. It is also a body-geometry problem. The ring needs to follow the finger angle. A bracelet needs to sit around the wrist rather than across it. Parts of the jewellery may need to disappear naturally behind the hand or wrist because of occlusion.

This is why pose references help. If you need a specific wrist angle, start with a real hand or model pose that already has the composition you want. Adobe Firefly's current composition-reference tools explicitly use outline and depth from a reference image, while its Object Composites workflow lets users position and resize an uploaded object before generating a surrounding scene. The broader principle is useful even when you are working in another model: solve position and scale deliberately instead of leaving both to chance.

A practical bracelet try-on workflow

1. Start with the exact product reference. Use the sharpest image that clearly shows the bracelet design, clasp and finish.

2. Add a scale reference. Include the real bracelet measurements or a photograph that shows its size relative to a known object or wrist.

3. Use a deliberate wrist pose. Choose the target hand or model image before generation. Avoid asking for a random wrist pose and exact product placement in the same step if consistency is important.

4. State the fit. Say whether the bracelet should sit close to the wrist, hang slightly loose, or show part of an extender. This is more useful than simply saying "wearing the bracelet naturally."

5. Check scale before checking beauty. Compare the finished image with the real dimensions. Then inspect clasp shape, chain path, engraving, thickness and metal finish. A realistic hand does not make an oversized bracelet accurate.

Where Lustra Studio fits

This is one reason jewellery-specific workflows benefit from saved model measurements and hand references. Lustra Studio can use product references together with model and wrist information, so the task is more constrained than a completely open prompt. The aim is not to make the model more creative. It is to reduce how much it has to guess about the person, product and fit at the same time.

The takeaway

AI jewellery try-on often fails quietly. The image can be attractive, anatomically plausible and still show the wrong product scale. The most reliable fix is to treat size as its own input: provide measurements, use a dedicated scale reference, choose the body pose deliberately and tell the model exactly how the jewellery should fit.

For rings and bracelets, physical scale is part of product accuracy. If a customer would notice the difference when the item arrives, it deserves the same verification as the clasp, engraving or metal colour.

Sources: Google AI for Developers, Gemini image generation documentation, updated August 2026; Adobe Firefly Help, Match image composition to a reference image, updated 16 June 2026; Adobe Firefly Help, Object Composites, accessed 20 August 2026; Miao et al., "Shining Yourself: High-Fidelity Ornaments Virtual Try-on with Diffusion Model," 2025; Burkhawala et al., "Virtual Ring Try-On," 27 June 2026.