About the project

A project page for interview-visible engineering depth.

CargoForge3D presents a port-oriented image-to-3D system that connects algorithmic acceleration, low-VRAM inference, multi-GPU serving, domain data preparation, and digital twin asset production. The page is intentionally written like a technical project report rather than a marketing landing page.

Project context

The work is motivated by breakbulk cargo terminals, where wooden crates, steel structures, pipe bundles, covered cargo, and irregular machinery components need to be reconstructed as reusable 3D assets. The target workflow is non-contact image capture, mesh generation, GLB/PBR export, and digital twin ingestion.

Engineering scope

The implementation includes Windows adaptation for Hunyuan3D-2.1, local model cache handling, HiCache++ shape acceleration, low-VRAM texture experiments, FastAPI serving, Redis queues, PostgreSQL job state, dual-GPU workers, watchdog recovery, and pressure-test metrics.

Research scope

The domain adaptation component converts teacher-generated GLB assets into watertight meshes, SDF samples, and multi-view render conditions. Rank-16 LoRA is used to improve open-source Hunyuan3D-2.1 on cargo-specific geometry.

Author

Developed by Zezhong Li, Tianjin University. The project is part of a broader effort on industrial AI, generative 3D reconstruction, GPU inference services, and port digital twin asset generation.

GitHub: cargo-forge3d Hunyuan3D Windows fork

Technical keywords

Stack covered by this project

Hunyuan3D-2.1 Image-to-3D DiT diffusion Flow matching HiCache++ PyTorch CUDA VRAM profiling PEFT / LoRA DINOv2 SDF sampling Watertight mesh GLB / PBR Blender / bpy FastAPI Redis PostgreSQL GPU workers Prometheus metrics Pressure testing

中文说明

这个页面服务于面试场景

面试官通常不会只关心“模型跑没跑通”，而会追问业务约束、显存瓶颈、并发调度、指标口径、失败恢复和数据闭环。 CargoForge3D 页面把这些点拆开呈现，便于围绕系统设计、深度学习推理、数据制备和微调实验展开讨论。