How Jewelbench Cut Jewelry CAD Modeling Time by 95%

Jewelry CAD modeling takes a trained designer days per piece. The startup was building a platform to put that capability in any jeweler's hands, not just specialists, and the one constraint that could not bend was accuracy: output that looked right on screen but could not be manufactured was not a product. It needed a pipeline that could take a rough starting point, a photo or a prompt, and turn it into a model precise enough to hand straight to production.

Ideas2IT built a two-stage pipeline. ClipDrop runs first on every image for enhancement and background removal, producing clean input before anything touches 3D; that cleaned image feeds Tencent's Hunyuan3D 2.0 on AWS to produce a base mesh, while text prompts skip ClipDrop and feed Hunyuan3D directly. Blender renders controlled 2D views and NeuS2 reconstructs a high-definition, manufacturing-grade model. Basic and HD generation are split into two sequential stages so basic models serve fast iteration and client preview while HD models go to the manufacturing handoff. The pipeline is wrapped in a web and mobile product with Google SSO, a Stripe token economy, and a superuser admin panel. A second phase in active development extends the pipeline to hand-drawn sketches via a preprocessing stage for line normalization and contour extraction.

The first decision shaped everything downstream: reference photos carry backgrounds and reflections that corrupt 3D geometry, so images could not feed generation directly. ClipDrop runs first on every image, handling enhancement and background removal to produce clean input before anything touches 3D. That cleaned image feeds Tencent's Hunyuan3D 2.0, hosted on AWS, which produces a base mesh; for text prompts, ClipDrop is skipped and the prompt feeds Hunyuan3D directly. Blender then renders controlled 2D views of the mesh, and NeuS2 uses them to reconstruct a high-definition, manufacturing-grade model. Splitting basic and HD generation into two sequential stages was deliberate: basic models serve fast iteration and client preview, while HD models go to the manufacturing handoff where accuracy actually matters. Ideas2IT wrapped the pipeline in a real product rather than an internal tool, with a web and mobile app, Google SSO for authentication, a Stripe token economy for payment, and a superuser admin panel for user, token, and pricing control. A second phase now in development extends the pipeline to hand-drawn sketches, adding line normalization and contour extraction before the same Hunyuan3D and NeuS2 flow.

Startups and manufacturers building generative design tools where the output has to be production-accurate, not just visually plausible, particularly in physical-goods verticals like jewelry, where a bad model costs real material and time to catch.