Point2Part: Unified Promptable 3D Part Decomposition from Point Prompts

Forum topic · 小凯 · 2026-10-01

Summary

Point2Part is a promptable 3D part decomposition model that reformulates part decomposition as a joint partitioning of an entire shape, ensuring predicted parts are non-overlapping and collectively cover the full geometry. Built on a pretrained 3D generative model, it first extracts a shape latent code from an input image or mesh, then maps 3D point prompts to part tokens via a prompt encoder. A novel part decoder jointly scores the whole shape against all part tokens in a coarse-to-fine manner, assigning every position in the shape volume to exactly one part. Working in a shared shape latent space, the unified model supports image-to-part generation, mesh-to-part generation, and part segmentation. It outperforms prior work on all part-quality metrics across three tasks and improves part compatibility (no overlaps or gaps) by orders of magnitude. Code and models will be released. arXiv: 2609.38180.

Overview

Field: Computer Vision 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.

Key Ideas

Part decomposition should consider all desired parts jointly, rather than modeling each part independently.
The model is promptable: users specify desired parts via 3D point prompts, with one point prompt per desired part, enabling controllable decomposition from images or meshes.
Built on a pretrained 3D generative model: a shape latent code is first obtained from the input image or mesh.
A prompt encoder maps each 3D point prompt into a part token while attending to the shape latent code.
A novel part decoder jointly scores the overall shape and all part tokens in a coarse-to-fine manner, precisely assigning each position in the shape volume to a unique part.

Unified Model

Because decomposition operates in a shared shape latent space, a single model supports:

Image-to-part generation
Mesh-to-part generation
Part segmentation

Results

Outperforms prior work on all part-quality metrics across three tasks.
Improves part-to-part compatibility (avoiding overlaps and gaps) by orders of magnitude.
Code and models will be open-sourced.

---

*Auto-collected on 2026-10-01.*