Point2Part: Unified 3D Partitioning from Point Prompts

What it claims

Point2Part guarantees by construction that the parts it predicts are exclusive (no shared volume) and exhaustive (their union is the whole solid). The model never has to learn this property. The guarantee comes from a single argmax over all part scores at each interior point, so it holds for any model output. No suppression, flood filling or overlap penalty is used. The authors are from Carnegie Mellon University, and the paper was posted 29 September 2026. They evaluate on PartObjaverse-Tiny (Yang et al., 2024) and contrast with P3-SAM, which merges per-click masks with suppression and flood fill, and with part generators such as PartCrafter, PartPacker, X-Part, OmniPart, CubePart and HoloPart. In the authors' account, none of those generators constrains parts to stay out of one another.

How it works

The backbone is Hunyuan3D. A mesh encoder or the image-conditioned flow model maps the input to shape latents Z in one shared space, and the decomposition runs on Z. For image input, the decoded whole mesh is where the 3D prompts are placed.

flowchart LR
  I[Image] --> F[Hunyuan3D flow model]
  M[Mesh] --> E[Hunyuan3D mesh encoder]
  F --> Z[Shared shape latents Z]
  E --> Z
  P[Point prompts] --> PE[Prompt encoder]
  Z --> PE
  PE --> D[Part decoder: argmax per point]
  Z --> D
  D --> G[Generation: closed part meshes]
  D --> S[Segmentation: label per face]

The model has three parts:

• Prompt encoder. Each prompt has K points. Each point is positionally embedded and cross-attends to Z. The K tokens are then attention-pooled into one token per part, and the parts self-attend to each other.
• Part decoder. Each interior query point attends to Z and then to the part tokens. A dot product with each part token gives the score, so each token acts as a dynamic linear classifier in the style of Mask2Former. One pass over the queries decodes all N parts, and the paper says cost "grows only slightly with N".
• Closed part meshes. Part j's field keeps the whole-shape SDF where the argmax picks j and takes its absolute value elsewhere. Marching cubes on that field then gives one watertight mesh per part. Its outer faces coincide with the original surface, and its inner faces are the cut surfaces.

Boundaries are refined coarse-to-fine at higher resolution. In training, each prompt samples K from 1 to 4 surface points of the ground-truth part, and inference uses K=1 by default. The target label at each interior point is the nearest annotated part by signed distance, which resolves overlaps in the annotations. Training queries are concentrated near part boundaries. The loss combines classification, Dice and a consistency term on part tokens.

Evidence

The fetched text contains no result tables. It states only the abstract-level claims: the method beats SOTA on all part-quality metrics in all three tasks and improves part compatibility by an order of magnitude. The metrics, baselines and numbers behind these claims are not on the fetched page.

How to use it

Code is announced at https://henrytsui000.github.io/Point2Part. The page gives no install commands, weights, licence or hardware requirements. The method depends on pretrained Hunyuan3D components, so their licence terms apply.

Limits

The fetched text stops partway through the training objective, so ablations, quantitative results and any stated limitations were not available. Quality on image input depends on the mesh Hunyuan3D generates, because the partition covers that mesh and not the image. Primitive-based decomposition methods are excluded from comparison as not comparable under the paper's metrics.