REFERENCE INSPECTION · 10 SEPTEMBER 2026
Product fidelity makes the difference.
The key lesson from these references is to treat the virtual piece as a product asset with its own construction, surface properties and fitting rules. A generic shape following a face is not sufficient to reproduce a real luxury product.
This report separates observed evidence from recommendations. Both interfaces were opened in a browser, consent was accepted with the user's authorization, and loaded assets were inventoried. Public source was inspected where relevant. Neither company's private source, training data or full proprietary renderer was accessible. Their assets were not reused in this demo.
Meta: what was verified
Meta's try-on page states that it renders 3D eyewear using estimated facial-feature locations. The interface exposed Wayfarer (Gen 1), separate frame/lens combinations, Standard and Large sizes, High bridge fit, Size guide, Fit guide, Take a photo and Refine fit. Opening Refine fit presented a face-calibration cutout. The shop link corresponded to the selected frame/lens variant.
The browser loaded WebAssembly modules, a binary resource and a GLB asset. Offline inspection of that GLB found one mesh with 382 position vertices, 760 triangles and one untextured material. Its role was not verified. It should not be presented as proof that the eyewear product itself has that polygon count or material setup; it may be a supporting mesh. Other opaque media resources were also observed, but their full pipeline was not established.
Practical lesson: store fit and finish as meaningful product variants. Provide visible refinement and fit guidance. Product selection, material appearance and dimensions should stay connected. The available evidence does not establish which proprietary tracking model or gemstone/material shader Meta uses.
Cartier: what was verified
The supplied Cartier try-on link passes a product ID, language and country. Cartier's product page identifies B6067517 as a LOVE bracelet in yellow gold, with a 6.1 mm width and a two-screw fastening system. It is a wrist experience, not a face or ear tracker.
The public integration loads a YMK SDK from plugins-media.makeupar.com and opens the supplied product using YMK.openWithProductId. Configuration includes country/language, zoom, consent and module options. The live screen displayed a wrist-position guide, capture control, Wrist size and Material controls. The camera opened, but a correctly positioned wrist was not available for validating the fitted bracelet. No claim of measured live fitting accuracy is made here.
| Observed resource | What it establishes |
|---|---|
Product-specific crb6067517_gold_150mm.obb | A dedicated bracelet asset is loaded. The container's exact geometry and physical calibration were not decoded. |
crb6067517_gold_albedo.png | A separate base-color texture is supplied for the product. |
crb6067517_gold_normal.png | A separate normal-map resource is supplied for surface detail. |
crb6067517_gold_orm.png | A packed material-map resource is supplied. ORM conventionally means occlusion, roughness and metalness; exact channel use was not verified in the proprietary renderer. |
crb6067517_gold_env_ibl.hdr | A dedicated HDR image-based-lighting resource is loaded. |
| Wrist occluder and bracelet/wrist collider resources | Separate assets support wrist-related visibility and geometric constraints. Their exact runtime algorithms were not inspected. |
| Hand/wrist tracking models and TensorFlow.js SIMD WASM | A specialized hand-tracking pipeline is delivered, rather than relying on a facial landmark model. |
Perfect Corp's own documentation describes physically based rendering, high-resolution textures and material reflections for its jewelry products. Those provider statements support the general approach; they do not prove that every feature in its product range is enabled in this particular Cartier deployment. See its earring rendering description and ring and bracelet technology description.
What changed in your demo
- Construction detail: modeled earring posts, backs, a hinge barrel, pearl cap, individual necklace links and milgrain beads.
- Stones: explicit table, crown, girdle, pavilion and culet geometry replaces the simple octahedron. Prongs and setting rims are separate metal geometry.
- Surface response: reflective metal materials, fine roughness variation and a higher-contrast studio reflection environment. This is a generated lighting environment, not a copy of Cartier's HDR map or an estimate of the live room.
- Consistent product presentation: the catalog, macro viewer and try-on now use the same 3D construction when WebGL is available. Illustrated fallback remains labeled.
- Detail inspection: each product has a 3D view with rotation, zoom, finish selection and a focused close-up of its main construction detail.
- Existing corrections retained: far-ear visibility checks, face masking, per-piece removal, Remove all and four model portraits.
What is still required for exact real-product reproduction
The current collection consists of original illustrative designs. To reproduce your inventory accurately, start with one real piece and supply a measured CAD/GLB model or sharp front, side, back and close-up photos with physical dimensions. A front photo alone cannot establish the hidden back, thickness, hinge or stone setting accurately.
The production asset should contain correct units and attachment origins, a high-quality mesh, detail-preserving textures, metal roughness and normal maps, and distinct gemstone materials. Validate its silhouette and surface detail against macro product photography. Then validate the same asset on a live wearer at multiple angles. Do not change the product's dimensions arbitrarily merely to make a screenshot look attractive.
The remaining tracking work is separate: actual piercing anchors, robust hair/hand occlusion, calibrated scale, full head orientation and dedicated neck fitting. Detailed 3D rendering alone cannot solve those problems. Similarly, a perfect tracker cannot invent details absent from the product asset.
Verification and limits
The product-detail dialog was visually inspected in the browser for a hoop and pendant; rotation and finish controls were exercised. The actual necklace Remove button was verified to preserve the earrings. Sample-photo landmark inference was observed working. WebMCP selection/read-back and intentional rejection of an unknown product were checked. Automated geometry, masking, removal and worker checks also passed.
No cross-device performance certification, millimeter-accurate fitting study, full optical gemstone simulation or production-equivalence claim is made. Cartier's wrist-fit and Meta's completed eyewear calibration were not quantitatively validated. Inspection was limited to the visible states and resources actually available in this session.