Point2Part: Unified 3D Part Decomposition from Point Prompts

Forum topic · 小凯 · 2026-10-01

Summary

Point2Part is a promptable model for 3D part decomposition that reformulates the task as a joint partitioning of the entire shape, ensuring predicted parts are non-overlapping and collectively cover the full geometry. Unlike prior methods that model each part independently—often producing overlaps or gaps that hinder downstream part-level applications—Point2Part considers all desired parts jointly. Built on a pretrained 3D generative model, the system first encodes the input image or mesh into a latent shape code. A prompt encoder maps user-specified 3D point prompts, one per desired part, into part tokens that attend to the shape latent. A novel part decoder then jointly scores the overall shape against all part tokens in a coarse-to-fine manner, assigning every location in the shape volume to exactly one part. Operating in a shared shape latent space, the framework unifies image-to-parts generation, mesh-to-parts generation, and part segmentation in a single model. The authors report that it outperforms existing methods on all part-quality metrics across the three tasks and improves inter-part compatibility by orders of magnitude. Code and models are planned to be open-sourced. arXiv: 2609.38180.

Paper Overview

Field: Computer Vision (CV)
Authors: Hao-Tang Tsui, Yu-Rou Tuan, Xiaoxuan Ma, Nicolas Ugrinovic, Takaaki Shiratori, Kris Kitani
Published: 2026-09-29
arXiv: 2609.38180

Abstract

Existing 3D part decomposition methods do not necessarily partition the original shape into non-overlapping parts that collectively cover the entire shape, allowing overlaps or gaps that hinder downstream part-level applications. Point2Part instead formulates part decomposition as a joint partitioning of the entire shape, where the predicted parts are non-overlapping and jointly recover the entire shape. The key insight is that part decomposition should consider all desired parts jointly, rather than modeling each part independently.

Method

The authors develop a promptable model for controllable 3D part decomposition from images or meshes:

Point prompts: Users specify desired parts through 3D point prompts—given one point prompt per desired part, the model outputs the corresponding parts as a complete segmentation of the whole shape.
Latent shape encoding: Built on a pretrained 3D generative model, the pipeline first obtains a latent shape code from the input image or mesh.
Prompt encoder: A prompt encoder maps each 3D point prompt into a part token that attends to the shape latent code.
Part decoder: A novel part decoder jointly scores the overall shape against all part tokens in a coarse-to-fine manner, assigning each location in the shape volume to exactly one part.

Unified Framework

Because part decomposition is performed in a shared shape latent space, a single model unifies:

image-to-parts generation
mesh-to-parts generation
part segmentation

Results

The method outperforms prior work on all part-quality metrics across the three tasks and improves inter-part compatibility by orders of magnitude. Code and models will be released as open source.

---

*Originally collected on 2026-10-01.*