Hunyuan3D Part (huggingface.co)

Tencent’s Hunyuan3D Part introduces a two-stage pipeline for converting holistic meshes (including image-derived or scanned shapes) into semantically segmented, high-fidelity parts. The system pairs P3‑SAM, a native 3D part detection and segmentation module that produces semantic features, part masks and bounding boxes, with X‑Part, a part-completion generator that produces coherent, high-detail part geometry. Both modules are released with code, demos, and arXiv papers; P3‑SAM is presented as a general mesh-capable part segmenter, while the current X‑Part release is a “light” version optimized for inputs from scanned or AI-generated meshes (e.g., Hunyuan3D V2.5/V3.0). This work matters for AI/ML practitioners building shape understanding, reconstruction, and content-creation pipelines because it automates structure-aware decomposition and part completion—key capabilities for editing, simulation, robotics, and AR/VR asset production. Technically, the pipeline separates part localization/semantic reasoning (P3‑SAM) from geometry synthesis (X‑Part), enabling modular improvements and mixed-input workflows: P3‑SAM can accept any mesh, X‑Part yields higher fidelity when fed cleaner scanned or generated meshes. The projects include reproducible code, Hugging Face demos, and arXiv preprints, making them immediately useful for research and production prototyping in 3D understanding and generative modeling.