From Manual CAD to Generative 3D Modeling: A Case Study in AI-Driven Jewellery Design Engineering

JewelBench, a jewelry tech startup, needed production-ready 3D CAD models generated from images and prompts in minutes. We built the two-stage AI pipeline connecting ClipDrop, Hunyuan3D, and NeuS2 to make it possible.

Client

JewelBench

Industry

Manufacturing

Service

Artificial Intelligence

App Development

Team

10+ engineers, AI-assisted delivery

Duration

1 year Active · Ongoing

01 Challenge

Jewelry CAD modeling takes a trained designer days per piece. JewelBench was building a platform to put that capability in any jeweler's hands, with one non-negotiable constraint: output accurate enough for manufacturing.

02 Solution

The pipeline runs in two stages. A reference image or text prompt enters ClipDrop for enhancement and background removal, then feeds Tencent's Hunyuan3D 2.0, hosted on AWS, to produce a base 3D mesh. Blender then renders 2D views of that mesh, which NeuS2 uses to reconstruct a high-definition production-quality model.

03 Outcome

Model generation that previously required days of specialist design work now completes in minutes. The platform handles both image uploads and text prompts, and produces output at two quality tiers: basic mesh for iteration, high-definition for manufacturing handoff.

Phase 01

Core generation pipeline: from reference image to production-ready 3D mesh in a single automated flow

The first decision shaped everything downstream: images could not go directly into 3D generation. Reference photos carry backgrounds and reflections that corrupt geometry output.

‍

  1. ClipDrop handled enhancement and background removal as the mandatory first stage, producing clean input for Hunyuan3D 2.0, running on AWS.
  2. For text prompts, ClipDrop was bypassed; prompts fed directly into Hunyuan3D. Blender then rendered 2D views of the base mesh at controlled angles, producing the source images NeuS2 required for high-definition reconstruction.
  3. Separating basic and HD generation into two sequential stages was a deliberate capacity decision: basic models serve iteration and client preview, HD models go to manufacturing handoff.
This phase produced
  • ClipDrop API integration (image enhancement and background removal)
  • Tencent Hunyuan3D 2.0 on AWS (basic 3D mesh generation)
  • Blender render pipeline (2D source images for NeuS2)
  • NeuS2 integration (high-definition model reconstruction)
  • Web/mobile application with Google SSO authentication
  • Token economy with Stripe payment integration
  • Superuser admin panel (user management, token controls, pricing configuration)

Phase 02

Sketch-to-3D conversion pipeline: extending the generation surface from photos to hand-drawn input

The Phase 1 pipeline required a clean reference photo or a text prompt. Jewelry designers in early ideation work from hand-drawn sketches, and those sketches carry none of the photographic clarity the existing pipeline assumed.

  1. Phase 2 extends the generation surface to accept sketch input directly. The engineering shift is in interpretation: extracting geometric intent from a hand-drawn line rather than a clean image edge.
  2. A sketch preprocessing stage, handling line normalization and contour extraction, feeds into the existing Hunyuan3D and NeuS2 flow, preserving the two-tier output model. This phase is in active development.
This phase produced
  • Sketch preprocessing module (line normalization and contour extraction)
  • Extended input layer accepting hand-drawn sketch uploads
  • Sketch-to-mesh generation pipeline connecting to existing Hunyuan3D flow
  • Two-tier output preserved: basic mesh and HD model from sketch input

The Outcome

What the pipeline changed for the people building with it

Category Metric Description
Generation time reduction 95% A model that required days of specialist design work generates in minutes, without designer intervention
Input types supported 3 Image upload, text prompt, and sketch input (Phase 2). Any starting point a jeweler works from is a valid generation trigger
Output quality tiers 2 Basic mesh for client iteration and preview; high-definition NeuS2 reconstruction for manufacturing handoff
Engineers on platform 10+ Delivered with AI-assisted development practices embedded throughout the build
The generation speed came directly from the pipeline architecture: two discrete stages with a clean handoff between them, each optimized for its specific task. ClipDrop's preprocessing removed the variable that most corrupted 3D generation quality. The Blender intermediate layer gave NeuS2 exactly the input geometry it needed. Speed and output quality were consequences of those structural decisions, not separate goals pursued alongside them.