Point2Part Makes 3D Parts Promptable
Point2Part presents a unified model that uses 3D point prompts to decompose images or meshes into complete, non-overlapping parts. It supports image-to-part generation, mesh-to-part generation, and mesh segmentation within one framework. [arXiv paper](https://arxiv.org/abs/2609.38180)
Point2Part targets a real bottleneck in 3D pipelines: parts that look separate but fail as a coherent assembly. Its global partitioning approach is more compelling than another standalone segmenter, though practical adoption still depends on the promised model and inference releases.
- –Joint assignment prevents gaps and interpenetration by construction, improving downstream editing, simulation, and asset workflows.
- –One prompt interface spans generated and existing geometry, giving developers consistent control over part granularity.
- –The project reports state-of-the-art part quality and substantially better inter-part compatibility than prior methods.
- –Checkpoints, inference tools, evaluation scripts, and training code remain planned releases, making this promising research rather than a ready production API. [GitHub repository](https://github.com/henrytsui000/Point2Part)
DISCOVERED
3h ago
2026-10-04
PUBLISHED
3h ago
2026-10-04
RELEVANCE
AUTHOR
AI Search