Point2Part: Unified 3D Partitioning from Point Prompts

TL;DR

The paper introduces Point2Part, a unified framework for 3D part decomposition that generates non-overlapping, complete partitions from 3D point prompts. It operates in a shared shape latent space derived from images or meshes, using a prompt encoder and a coarse-to-fine part decoder to assign every point to exactly one part. The method achieves state-of-the-art performance across image-to-part, mesh-to-part, and part segmentation tasks.

Why It Matters

Accurate and complete 3D part decomposition is essential for downstream applications like shape editing, assembly planning, and semantic understanding. By ensuring non-overlapping, full coverage partitions guided by user-provided prompts, the method enables more reliable and controllable 3D content generation and analysis.

Key Findings

• The model produces complete, non-overlapping partitions by jointly modeling all parts in a shared latent space.
• Using 3D point prompts allows users to specify desired parts, enabling controllable decomposition from images or meshes.
• The approach achieves superior part quality and significantly improved part compatibility compared to prior methods.