CMU's Point2Part splits 3D models with a single click

Anyone working with 3D models knows the pain of taking things apart. Slicing a solid asset into articulated parts usually means hours of manual retopology, mesh booleans, and fixing ugly geometry overlaps.

Researchers from Carnegie Mellon University unveiled Point2Part, an AI system that decomposes 3D shapes or single images into clean, usable parts with simple point clicks.

No gaps, no overlap: the joint partition trick

Existing part-generation models usually generate each piece independently. That almost always creates a mess: parts either intersect and cause mesh clipping, or leave visible cracks along the seams.

Point2Part takes a completely different approach. It treats shape decomposition as a joint partitioning problem over the entire object space.

By design, every generated piece is both mutually exclusive and collectively exhaustive. When you put the parts back together, they fit like precision puzzle pieces.

In single-image tests, Point2Part achieved a 100% watertight mesh rate with a mesh penetration rate of just 0.01%, compared to 2.45% for PartCrafter and 0.96% for OmniPart.

When splitting existing 3D meshes, it improved geometric compatibility by an entire order of magnitude, slashing mesh penetration from 3.05% down to 0.06%.

Point, click, and explode

The user workflow is dead simple: you place one point prompt onto each component you want to isolate.

You can control granularity on the fly. Drop one point for broad body segments, or add more points to isolate smaller mechanical pieces.

The framework handles three primary pipelines: single image to closed parts, mesh to closed parts, and mesh face segmentation.

Speed is another big win. For mesh face labeling, Point2Part runs in just 0.3 seconds per model, hitting a 69.80 mIoU. By comparison, PartSAM takes 36.6 seconds.

With four prompt points per part, its accuracy jumps to 74.84 mIoU. Once separated, users can animate components outwards to create instant, textbook-ready exploded diagrams.

Why it matters

For game artists, CAD designers, and 3D printing hobbyists, Point2Part tackles a massive practical bottleneck.

Generative 3D has gotten very good at spitting out static, monolithic meshes. But a solid lump of polygons is practically useless if you need articulated limbs, opening car doors, or multi-material fabrication.

From an industry standpoint, moving 3D AI from visual novelties to functional, simulation-ready assets requires semantic part hierarchies. CMU's formulation is mathematically elegant because it enforces geometric consistency directly inside the architecture rather than relying on messy post-processing heuristics.

That said, temper your expectations slightly. For now, CMU has only released the project page and preprint paper (arXiv:2609.38180). While a GitHub repository is up, the code and model weights are promised for a future release. We will have to wait and see how robustly it handles messy, non-watertight real-world scans before declaring manual 3D part splitting completely obsolete.