Point2Part

Unified 3D Partitioning from Point Prompts

Hao-Tang Tsui · Yu-Rou Tuan · Xiaoxuan Ma · Nicolás Ugrinovic · Takaaki Shiratori · Kris Kitani

Carnegie Mellon University

Point to the parts you want.

Point2Part formulates 3D part decomposition as a partition of the whole: users specify the desired parts with 3D point prompts, and the model jointly decomposes the entire shape among them.

Image → Parts   ·   Mesh → Parts   ·   Mesh → Segmentation

---

Overview

• Point-prompted control
  Directly specify the desired decomposition with 3D point prompts.

• Partition of the whole
  Parts are predicted jointly rather than independently, producing a complete and mutually exclusive decomposition by construction.

• One model, three tasks
  Point2Part handles image-to-part generation, mesh-to-part generation, and mesh part segmentation within the same framework.

• Strong part quality and compatibility
  Point2Part achieves state-of-the-art part quality while reducing inter-part penetration by more than an order of magnitude.

---

Controllable Decomposition

Point2Part lets users control the desired granularity simply by specifying point prompts.

---

Unified Tasks

---

Code & Models

Planned release:

• Pretrained checkpoints
• Mesh → Parts inference
• Image → Parts inference
• Mesh segmentation inference
• Evaluation scripts
• Training code

⭐ Star the repository to follow future releases.

---

Acknowledgements

This work builds on Hunyuan3D 2.1, TripoSG and HY3D-Bench. We thank the authors for releasing their work.

---

Citation

If you find Point2Part useful, please consider citing:

@article{tsui2026point2part,
  title   = {Point2Part: Unified 3D Partitioning from Point Prompts},
  author  = {Tsui, Hao-Tang and Tuan, Yu-Rou and Ma, Xiaoxuan and
             Ugrinovic, Nicol{\'a}s and Shiratori, Takaaki and Kitani, Kris},
  journal = {arXiv preprint arXiv:2609.38180},
  year    = {2026}
}