Point2Part: Unified 3D Partitioning From Point Prompts

Gallery

Given one point prompt per desired part, our model produces the parts as a complete partition of the whole.

generated parts drag to orbit · scroll to zoom · right-drag to pan

Pull apart

from a mesh

from one image

TL;DR/ Part decomposition as a partition of the whole

We formulate part decomposition as a joint partitioning of the entire shape, where parts are exclusive and exhaustive by construction, while enabling users to specify desired parts through point prompts.

Exclusive: Distinct parts have no volume overlap.
Exhaustive: Their union recovers the entire object.

Motivation/ Parts may leave gaps or overlap

Existing methods do not necessarily form a partition: parts may leave gaps (outlined) or overlap (red).

The same robot three ways, in one shared colouring: every part takes the colour of the ground-truth part it covers. On the left, a baseline decomposition assembled part by part from two published methods, three parts from CubePart and nine from X-Part, with the regions neither of them recovers left empty: first that decomposition, then a grey copy with the volume its parts claim twice painted red. On the right, our decomposition and our grey copy, which has no red on it.

Controllable/ Point prompts make decomposition directly controllable

A 3D mesh admits valid decompositions at different granularities, set by the prompts.

Results/ Best part quality on all three tasks

We lead on all part-quality metrics and improve geometric compatibility by an order of magnitude.

Single image → generated closed parts
Method	part CD ↓	part F1@.05 ↑	pen % ↓	watertight % ↑
Ours	0.0534	0.686	0.01	100.0
OmniPart	0.0643	0.610	0.96	94.7
PartPacker	0.0811	0.540	1.25	86.0
PartCrafter	0.1275	0.300	2.45	70.5

Mesh → generated closed parts
Method	part CD ↓	part F1@.05 ↑	pen % ↓	watertight % ↑
Ours	0.0273	0.844	0.06	100.0
X-Part	0.0453	0.730	3.05	90.8
CubePart	0.0471	0.740	2.09	100.0
HoloPart + P3-SAM	0.0529	0.710	1.01	32.6

Mesh → per-face labels
Method	mIoU ↑	part CD ↓	part F1@.05 ↑	sec ↓
Ours	69.80	0.0210	0.870	0.3
PartField	69.10	0.0531	0.725	1.1
P3-SAM	54.66	0.0433	0.748	19.1
S²AM3D	50.43	0.0316	0.805	0.3
PartSAM	50.31	0.0324	0.792	36.6
Point-SAM	39.12	0.0537	0.687	1.2
Ours, 4 points per part	74.84	0.0135	0.918	0.3

Citation

Cite this work

The preprint is on arXiv as arXiv:2609.38180.

@article{tsui2026point2part,
  title   = {Point2Part: Unified 3D Partitioning from Point Prompts},
  author  = {Tsui, Hao-Tang and Tuan, Yu-Rou and Ma, Xiaoxuan and
             Ugrinovic, Nicol{\'a}s and Shiratori, Takaaki and Kitani, Kris},
  journal = {arXiv preprint arXiv:2609.38180},
  year    = {2026}
}