Why YouCam API.
Fine jewellery requires sub-millimeter precision, true facial likeness, and catalog product fidelity. Here is why specialized computer vision succeeds where generic text-to-image diffusion fails.
Technology reference · Perfect Corp YouCam API
Why specialized computer vision matters for virtual try-on.
Generic language models and text-to-image diffusion tools generate speculative pixels rather than true physical measurements. Darpan integrates Perfect Corp’s YouCam API to deliver deterministic landmark tracking, authentic identity preservation, and production retail accuracy.
| Evaluation dimension | YouCam API in Darpan | Generic LLMs & diffusion |
|---|---|---|
| Scale & physical proportions | Sub-millimeter biological calibration anchored to the 11.7 mm iris, sternal notch, and wrist width. True 1:1 physical millimeter rendering. | Arbitrary pixel scaling without spatial grounding. A 3 cm earring may render at 8 cm or shrink to a pinhead depending on random seed. |
| Facial identity & natural skin | 100% customer likeness preserved. Zero face-morphing. Natural skin pores, authentic bone structure, and ambient lighting remain untouched. | Regenerates or smooths facial features through diffusion inpainting, altering facial symmetry and creating an uncanny valley effect. |
| SKU & inventory fidelity | Renders the exact catalog item with authentic metal luster, gemstone facets, and physical drop physics from retailer inventory. | Hallucinates fictional decorative pieces that do not exist in inventory, leading to customer dissatisfaction and product returns. |
| Earlobe & hair occlusion | Specialized hairstyle simulation lifts hair into tied styles (braids, buns, updos) to uncover the earlobe and neck before earring try-on. | Lacks 3D occlusion understanding; overlays earrings on top of hair strands or paints them across the jawline. |
| Complexion & tone harmonization | Cosmetic lipstick simulation applies genuine cosmetic shades across the vermilion border to preview harmonious jewelry and beauty tones. | Over-saturates image colors and shifts background temperature unpredictably, unable to reproduce specific cosmetic shade codes. |
| eCommerce response time | Sub-second to 2-second cloud inference designed for high-conversion eCommerce checkout funnels and mobile shoppers. | Heavy diffusion generations take 20 to 45 seconds per image with substantial compute costs, causing checkout drop-offs. |
Explore live try-on with hairstyle simulation and lipstick harmonization.
Computer vision physics versus generative diffusion
Darpan separates geometric scaling and physical placement from conversational styling advice.
Why prompting an LLM fails for fine jewellery
When a diffusion model is asked to render jewelry onto a portrait, it resynthesizes pixels from noise rather than calculating spatial metrics. In retail, this causes critical failures:
- Unverifiable scale: Text prompts have no concept of physical millimeters, causing earrings or dials to appear at arbitrary sizes.
- Facial distortion: Inpainting alters natural facial geometry, skin texture, and lighting, creating an uncanny likeness.
- Non-existent inventory: Diffusion invents decorative details that do not exist on the retailer’s actual shelf.
Deterministic landmarks and physical anchors
Perfect Corp’s YouCam API decouples biometric recognition from synthetic imagery:
- Biometric grounding: Exact tracking of iris diameter (11.7 mm average), lobule attachment, and clavicle width.
- Zero face alteration: The customer's natural skin, pores, and facial structure remain completely authentic.
- eCommerce latency: Sub-second cloud execution suitable for active shopping sessions and mobile checkout.
Try-on is a visual guide. Explore our calibration benchmarks or try a piece on your own photo.