How to Stop AI From Changing Jewellery Chains and Clasps in Product Photos

A jewellery image can look excellent at first glance and still show the wrong chain. One link pattern becomes another, the clasp changes shape, a pendant connects to the chain differently, or a fine section quietly disappears when the product is moved into a new scene.

This is not a minor cosmetic problem. For a necklace or bracelet, the chain, connectors and clasp are part of the SKU. If a generated listing image changes them, the image can misrepresent what the customer will receive.

The safest workflow is to treat chain geometry as something the AI needs evidence for, not something it should be allowed to infer.

Why chains are unusually difficult for generative models

Jewellery contains exactly the kind of structure that is easy to lose during generation: tiny repeated shapes, narrow gaps, polished edges and details that occupy only a small number of pixels. Research presented at CVPR 2025 on ornament virtual try-on identifies intricate small patterns and repeated geometric substructures as a major reason rings, bracelets, earrings and necklaces are harder to preserve than larger fashion items.

A fine chain makes the problem obvious. The model has to preserve not only the overall silhouette but also the local structure of many small connected elements while changing pose, scale, background or placement on a person. A result can therefore preserve the idea of a delicate necklace while changing the actual construction.

Use separate references for separate jobs

One hero photograph rarely shows every important chain detail. The front may clearly show the pendant but hide the clasp. A coiled bracelet may show the link pattern while making the connector difficult to read.

Current Gemini image models support multiple high-fidelity object references, which makes a more deliberate setup possible. Rather than uploading several similar photos, give each reference a purpose.

Reference 1: overall product. Use the clearest full view as the primary source of truth for proportions, finish and overall design.

Reference 2: link pattern. Add a sharp close-up that shows whether the chain is curb, rope, box, snake, cable, figaro or another construction. The point is to show the real repeating structure rather than describe it from memory.

Reference 3: clasp and connections. Photograph the lobster clasp, spring ring, toggle, extender, jump rings, bail or end caps clearly. These small parts are often hidden in the main listing photo but are exactly what a model may otherwise invent.

Tell the model what must remain continuous

A useful prompt should describe the scene and separately define the product constraints. For a necklace, do not only say "preserve the necklace." Name the parts that are easy to lose.

For example: "Use image 1 as the exact necklace reference. Image 2 shows the exact chain-link construction and image 3 shows the clasp and pendant connection. Preserve the chain as one continuous piece with the same link pattern, thickness, clasp, extender and bail. Do not add, remove or redesign connectors. Change only the scene and placement."

That wording does not guarantee perfect fidelity, but it removes ambiguity. The model knows which reference defines the overall item and which images contain structural evidence.

Do not ask for too many structural changes at once

A chain is most likely to drift when the generation also requires a large change in pose or viewpoint. Turning a flat necklace into a model-worn image means the chain must curve around a neck, change apparent scale, pass behind the body and still reconnect correctly to the pendant.

The CVPR ornament try-on work found that pose and scale have a significant influence on fitting quality and used a dedicated wearing mask to improve alignment. A normal seller does not need to reproduce that research system, but the practical lesson transfers: placement should be constrained as much as possible.

If you only need a lifestyle flat lay, keep the product viewpoint close to the source. If you need a worn image, choose the target neck or wrist pose deliberately and give the model measurements or placement guidance where available. Asking for a random person, a new camera angle, different lighting and a fully re-draped chain in one step gives the system more opportunities to rebuild the jewellery.

Inspect the chain like a path, not a silhouette

When reviewing the output, zoom in and visually trace the chain from one end to the other. This catches errors that are easy to miss when you only compare the overall shape.

Check whether the link pattern stays consistent, whether any section disappears behind the pendant incorrectly, whether the clasp is the right type, whether an extender has gained or lost parts, and whether the chain enters the bail or connector in the same way as the real item. On a bracelet, check the same path through every connector and charm.

This is also where high resolution can be misleading. A beautifully sharp generated clasp is not useful if it is the wrong clasp. Compare against the real reference rather than judging realism on its own.

A practical workflow for chain-heavy jewellery

1. Photograph the entire product clearly. Avoid heavy overlap where possible and keep the fine chain in focus.

2. Capture two detail references. One should show the repeating link structure and one should show the clasp, extender and important connections.

3. Assign a role to every reference. State which image controls the overall product, link construction and hardware so the references are not treated as interchangeable examples.

4. Keep the first generation focused. Change the background or placement you actually need without adding unnecessary structural changes.

5. Trace the finished chain. Inspect the full connection path, then compare the clasp, bail, extender, link pattern, thickness and finish against the real product.

6. Reject structural drift. If the output has changed a connector or link construction, regenerate from the real references rather than treating the altered AI image as the new source of truth.

Where Lustra Studio fits

Lustra Studio supports multiple jewellery references because one photograph often cannot describe an entire piece. For chain-heavy products, the most useful setup is not the largest possible reference set. It is a small set where the primary image establishes the SKU and extra views reveal the exact hardware or geometry that would otherwise be hidden.

That same principle applies in any image model that accepts several references. Give the system evidence for the difficult parts, constrain the change, then verify the output against the physical product.

The takeaway

Fine chains and clasps are difficult because they combine tiny repeated geometry with parts that are often hidden in the main photograph. The answer is not simply a longer prompt. Show the AI the information it cannot see.

Use one authoritative full-product image, add sharp references for the link pattern and hardware, state what each image controls, avoid unnecessary pose changes and trace the finished chain before publishing. In jewellery ecommerce, a plausible chain is not enough. It needs to be the chain attached to the product you actually sell.

Sources: Miao et al., "Shining Yourself: High-Fidelity Ornaments Virtual Try-on with Diffusion Model," CVPR 2025; Google AI for Developers, Gemini image generation documentation, updated August 2026.