Research · October 2026
Jewellery photo to 3D: what one photograph can and cannot tell us
We tested Kubikly's jewellery reconstruction on 102 real photographs before and after rebuilding the pipeline. This page reports the method, the numbers and the failures, including the ones we have not fixed.
Summary
- On 94 jewellery photographs, the Precision pipeline identified the piece type correctly in 93% of cases (84% before) and found the major components in 89% (86% before).
- The mean error in counting the visible stones fell from 46% to 36%; 45% of counts were exact.
- Agreement between the photo and the model rendered from the same viewpoint (silhouette, 0–100) rose from 61.9 to 66.8.
- Non-jewellery photographs were refused 6 out of 6 times (0 of 6 before), with one false refusal.
- Freeform, sculptural pieces remain the weakest case: the outline matches, the volume does not.
- A single photograph still cannot establish real size; one known measurement is required for calibrated dimensions.
Method
102 photographs: 94 of jewellery and 8 traps (6 non-jewellery images and 2 pieces worn on a person). Sources: open-access museum collections (The Met, Cleveland Museum of Art, CC0), Wikimedia Commons, Flickr under Creative Commons, and earlier test uploads. Labels were frozen before testing (piece type, construction family, major components, stone groups with visible counts where countable, symmetry and repeated pattern). None of the photographs were used to develop the test fixtures.
The sample: rings 29 (solitaire, halo, cluster, three-stone, pavé, eternity/channel, signet, freeform), bracelets 25, necklaces and chains 11, pendants 11, brooches 11, earrings 7, plus 8 multi-photo sets of 2–4 views. Hard cases were included on purpose: macro close-ups, pieces draped over props, reflections, CGI renders, archaeological finds, three stacked rings, a piece in a box.
Every run was scored by the same deterministic scorer. The silhouette metric compares an independent mask of the photo with the model rendered from a camera fitted to the photo; the stone-count error is |predicted − labelled| / labelled, capped at 100%.
Results
| Metric (94 jewellery photos) | Before | Quick | Precision |
|---|---|---|---|
| Piece type correct | 84% | 85% | 93% |
| Major components found | 86% | 88% | 89% |
| Stone groups found | 85% | 80% | 89% |
| Visible stone-count error (mean) | 46% | 46% | 36% |
| Repeated pattern detected | 75% | 82% | 84% |
| Symmetry agrees | 72% | 70% | 71% |
| Scale reported honestly | 100% | 100% | 100% |
| Silhouette (0–100) | 61.9 | 64.1 | 66.8 |
| Non-jewellery refused | 0/6 | 6/6 | 6/6 |
| Median time per photo | 22 s | 32 s | 108 s |
| AI cost per photo | $0.034 | $0.034 | $0.124 |
“Before” is the previous pipeline with a crash fixed so every photo produced a model. “Scale reported honestly” means a size was never labelled as calibrated without real evidence.
Rings, by construction
| Family | Photos | Components found | Stone-count error | Silhouette |
|---|---|---|---|---|
| Solitaire | 7 | 95% → 100% | 0% → 0% | 62.2 → 64.0 |
| Halo | 9 | 87% → 96% | 33% → 6% | 57.7 → 64.4 |
| Cluster | 11 | 77% → 97% | 44% → 23% | 57.3 → 64.6 |
| Pavé | 4 | 77% → 100% | 25% → 25% | 55.7 → 58.8 |
| Three-stone | 5 | 80% → 90% | 35% → 20% | 63.1 → 64.0 |
| Eternity / channel | 5 | 70% → 90% | 68% → 74% (worse) | 68.3 → 68.6 |
| Signet | 4 | 88% → 100% | — | 73.4 → 70.4 (worse) |
Tennis bracelets
| Before | Precision | |
|---|---|---|
| Recognised as a bracelet | 70% | 100% |
| Stone run and clasp present | 85% | 100% |
| Visible stone-count error | 59% | 9% |
Before, three of five bracelets photographed draped over a prop were read as rings.
More than one photo
| 8 multi-photo sets | One photo | All photos |
|---|---|---|
| Silhouette, mean over all views | 59.9 | 66.2 |
| Silhouette on views the one-photo model never saw | 59.0 | 64.5 |
| Cost / time | $0.162 / 154 s | $0.189 / 180 s |
Five sets improved; three were flat or slightly worse (−0.5 to −0.8).
Other categories (silhouette, before → Precision)
Rings 63.8 → 67.1 · bracelets 67.8 → 69.0 · necklaces 39.9 → 50.0 · pendants 65.8 → 76.1 · brooches 63.9 → 70.3 · earrings 43.7 → 49.8. Chains and link pieces (20): component recall 70% → 80%, count error 63% → 46%, silhouette 52.4 → 59.7.
What still fails
- Freeform 3D form. On 14 sculptural pieces the outline improved (silhouette 61.4 → 67.7) but component recall fell from 100% to 79%, and the body is a flat extruded plate: a coiled snake ring becomes a box.
- Run-to-run variance. Repeating the same piece three times gave the same piece type every time (silhouette ±2.9), but the component list differed in 7 of 8 pieces, and inferred stone totals swung widely (one piece: 126, 359 and 163 stones). Visible counts are far more stable than inferred totals.
- Dense chains collapse into a blob; gem-set links render as solid discs.
- Half-hoop eternity rings did not improve (count error 68% → 74%).
- Several rings in one photo become one piece.
- Earrings are flat, and their stone counts were wrong before and after.
- One false refusal: a carved mask pendant was judged “not jewellery”.
- Physical size stays estimated from a photo, by design: nothing in a photo of a ring gives its real size. One measurement from the user rescales the whole model.
For the same reason, Kubikly does not show users a single “quality score”: its run-to-run spread (±6.7) is as large as the improvements it would advertise. Users see a confidence map (high, medium, low, unknown) for each aspect of the piece instead.
What this means if you use it
- For rings, bracelets, pendants and brooches, expect a recognisable, editable model with the main parts in place; check stone counts on dense pavé and eternity settings.
- Add a second or third angle when you can: it improved the views the model had not seen.
- Give one real measurement — ring size or band width — before relying on any size or gold weight.
- For sculptural pieces, treat the model as an outline to refine, not a finished form.
Related: jewellery photo to 3D · photo-to-3D approaches compared · gold weight calculator · Kubikly facts