acceptodds
Under review as a conference paper at ICLR 2027

Unified 3D Partitioning from Point Prompts

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. We instead formulate part decomposition as a joint partitioning of the entire shape, where the predicted parts are non-overlapping and jointly recover the entire shape. Our key insight is that part decomposition should consider all desired parts jointly, rather than modeling each part independently. To this end, we develop a promptable model for 3D part decomposition from images or meshes. Users can specify desired parts through 3D point prompts for controllable decomposition. Given one point prompt per desired part, our model produces the corresponding parts as a complete partition of the entire shape. We build on a pretrained 3D generation model and first obtain a shape latent from either an input image or mesh. We then introduce a prompt encoder that maps each 3D point prompt to a part token while attending to the shape latent. To decode the desired parts, we propose a novel part decoder jointly scoring the entire shape against all part tokens in a coarse-to-fine manner, assigning every position within the shape volume to exactly one part. We perform part decomposition in this shared shape latent space, enabling a unified model for image-to-part generation, mesh-to-part generation, and part segmentation. Our method outperforms existing works on all part-quality metrics across all three tasks, and improves compatibility among parts by an order of magnitude over previous SOTA methods. Code and models will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Part Decomposition, 3D Part Generation, Mesh Segmentation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_CrCNOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a method for partitioning 3D objects using point prompts. It uses the shape representation of a pretrained 3D generative model, jointly assigns interior points to the prompted parts, and extracts closed part meshes with little interpenetration from the same signed distance field. The method supports segmentation of existing meshes and part decomposition of image-generated shapes. It is evaluated through baseline comparisons and ablations on part generation and mesh segmentation.

Critical strengths
  1. The method uses point prompts to extract closed parts from a single whole shape, with very little interpenetration. This addresses a common problem with generating parts independently and has practical value.
  2. The experiments show good part quality and segmentation performance. For mesh-conditioned generation, [email protected] is 84.4, compared with 73.9 for the best baseline on this metric. On PartNeXt, mIoU is 57.76, compared with 54.89 for the best reported baseline. The method and training objectives are clearly specified. The experiments are well organized and include ablations, a different backbone, and an additional dataset; the method, equations, and figures are easy to follow.
Critical weaknesses

The joint prompt segmentation design is very close to Instruct-Particulate [1]. That work already lets part tokens attend to the whole shape and then to one another, while query points attend first to the shape and then to all part tokens. Final labels are assigned by argmax. Each part or query update uses the same attention order as Eqs. (4) and (5), though the layer schedule and scoring head differ. This work is not cited, and the related-work section describes earlier promptable methods as answering each prompt independently.

The paper’s ablation also shows a large reduction in interpenetration when independent assignments are replaced by argmax, with a much smaller change in part quality. The remaining contributions in interior volume partitioning and mesh extraction are not enough to convince me to recommend acceptance.

[1] Ruining Li et al. Instruct-Particulate: Scaling Feed-Forward 3D Object Articulation with Kinematic Control. arXiv:2606.14699, 2026. Paper.

Questions for authors
  1. Compared with Section 4.2 of Instruct-Particulate (arXiv:2606.14699), this paper completes part encoding before query decoding in Eqs. (4) and (5) and uses a different scoring head. Which problems do these differences address, and where do the existing results show their effects? Apart from interior volume partitioning and closed-mesh extraction, what is new in the joint prompt segmentation component?
  2. I am curious what would happen if a simple step to remove overlap were applied to CubePart or X-Part outputs. Only process regions where parts intersect: assign each location to the part with the nearest surface, choosing among the parts that already contain that location, and leave everything else unchanged. This keeps the union of the parts unchanged. Could this reduce interpenetration to a level close to this method while preserving part quality? Could you report pCD, pF1, and pen% before and after this step? This would affect how much weight I give to the low penetration rate when judging the contribution.

I would consider raising my score after a convincing response, particularly to Question 1.

2 comments

We sincerely thank the reviewer for recognizing the practical value, good performance, and clarity of our approach. We address the questions below.

Question 1. We thank the reviewer and now discuss this concurrent work in Sec. 2. The two methods address fundamentally different problems: surface segmentation versus solid decomposition. Instruct-Particulate relies on PartField’s output, an existing part segmentation model, whereas ours decomposes the entire solid directly from whole-shape latents. Surface segmentation is a special case of our more general formulation (Sec. 3.3). While both methods share similar attention operations, our scoring head assigns spatial queries to jointly encoded part representations, enabling dense volumetric partitioning and closed-part generation. We further introduce consistency supervision to learn part-level representations from whole-shape features, whose effectiveness is demonstrated in Table 3(b).

Question 2. We thank the reviewer for suggesting this additional evaluation. We apply overlap-removal post-processing to X-Part and CubePart using Trimesh to reassign each location to the nearest part. Invalid X-Part meshes are first repaired with Trimesh, followed by Blender voxel remeshing when necessary.

MethodspCD[email protected][email protected]pen%latency
Ours2.7357.084.40.0624.4
XPart before4.5352.573.43.05147.2
XPart after4.5452.272.50.01156.7
CubePart before4.7151.572.52.0929.1
CubePart after4.7251.273.70.0035.4

Although post-processing reduces penetration to nearly zero, it does not improve part quality (pCD, pF1) and introduces additional latency. We discussed our residual penetration in Appendix B. Our method achieves substantially better part quality and faster inference without overlap-removal post-processing, demonstrating that its benefits extend beyond low penetration alone.

Thank you for the response. The Q2 experiment answers my question directly.

Regarding Q1, the revised Section 2 now discusses Instruct-Particulate. Given that it is concurrent work, and that this paper extends joint partitioning to interior volumes and closed-part generation, I no longer consider this a major concern and have raised my score to 3.

Reviewer_XiC5OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed a 3D part decomposition method. The main motivation is that part decomposition needs to consider all parts jointly rather than individually. The authors also assume each position only belongs to one part. The 3D decomposition system is built on a pretrained 3d generation model. The system takes user-provided 3D points as prompts. Part encoder transforms point prompt into part tokens by attending it to shape latent. Then the part decoder uses the part tokens and shape latent to predict part assignments and further obtain segmentation or part meshes. Experiments demonstrate strong segmentation performance compared with prior baselines.

Critical strengths
  • The insight that part decomposition needs to consider all parts jointly instead of modeling each part individually is interesting and well motivated. Joint assignment provides a principled way to make the parts form a coherent partition
  • The segmentation results are strong, both quantitatively and qualitatively.
Critical weaknesses
  • The quality of generated parts is not satisfactory.
    • Low whole geometry quality. The part decoder uses original SDF values to extract part meshes. Therefore, it is expected that the part mesh perfectly aligns with the whole geometry. However, as shown in Table 1 and Figure 3, the whole geometry quality is lower than previous baselines.
    • As shown in Figure 9, the generated part interfaces are not smooth and demonstrate obvious flaw.
  • The authors use inter-part penetration as an important metric for part generation. However, a small amount of penetration may be acceptable and could improve the appearance of part interfaces.
Questions for authors
  • How could whole-shape geometry quality be improved?
  • How could the quality of the generated part interfaces be improved?
4 comments

We thank the reviewer for the positive assessment and constructive feedback on geometric quality. We address the questions below.

Question 1. We apologize for the confusion. We'd like to clarify that whole-shape metrics are measured using the frozen, pretrained Hunyuan3D-2.1 backbone’s prediction (Fig. 2, ∂Ω\partial \Omega), rather than our method’s. Despite this reconstruction limitation, our method achieves superior part-level quality. Fine-tuning the HY3D model or adopting a more advanced backbone could further improve the quality of whole-shape geometry. We have revised Sec. 4.1 to make this clearer.

Question 2. The quality of part interfaces is affected by ambiguous boundary supervision in the training data, where overlapping ground-truth parts can introduce inconsistencies. Since inter-part penetration is common in existing public datasets, we already filter samples with severe penetration during training. While small overlaps may improve visual appearance, our goal is to produce geometrically compatible parts that form a valid partition of the whole shape. Further improvements could be achieved with graph-cut optimization and boundary-smoothness regularization.

Thanks to the authors for the detailed response. The segmentation results are impressive. However, if the paper claims to support part generation, the quality of the generated parts should be further improved. The method could simply use the ground-truth SDF values from the original mesh, rather than the Hunyuan 2.1 predictions, for reconstruction to improve the overall mesh surface quality. The key point is that the interface quality should be carefully addressed and further improved. Therefore, I would maintain my score.

We thank the reviewer for this valuable suggestion. Following the reviewer's recommendation, we incorporated ground-truth SDF values for mesh extraction, substantially improving both whole-shape quality:

MethodpCD↓[email protected]↑CD↓[email protected]↑
CubePart4.7172.51.2197.5
HoloPart5.2973.91.5495.0
X-Part4.5373.41.1997.3
Ours (HY3D SDF)2.7384.41.5694.1
Ours (GT SDF)2.2687.80.85100.0

With ground-truth SDF, our method now achieves the best results across all metrics. These results further demonstrate that the pretrained backbone's reconstruction quality was a major limiting factor and that our method can achieve stronger performance with higher-quality geometry. We hope this additional experiment helps clarify the source of the performance gap and addresses the reviewer's concern. We will include this option in the released codebase.

I appreciate the authors’ additional experiments. However, the quality of the interfaces between parts, which is my primary concern, still needs improvement. Therefore, I will maintain my score.

Reviewer_bUEdOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a unified framework for point-prompted 3D part decomposition. By combining a shared shape latent, a prompt encoder, and a joint part decoder, the method produces mutually exclusive and exhaustive partitions. It supports image-to-part generation, mesh-to-part generation, and part segmentation. Experiments on PartObjaverse-Tiny and additional segmentation experiments on PartNeXt demonstrate improvements in part quality and geometric compatibility.

Critical strengths
  1. The paper addresses inter-part overlap and incomplete decomposition through joint partitioning, providing a unified solution for three related tasks.
  2. The method achieves consistent improvements in part quality and substantially reduces inter-part penetration. The ablations and cross-dataset evaluation further support its effectiveness.
Critical weaknesses
  1. Limited methodological novelty and insufficient isolation of the core contribution. The framework largely combines established components. The exclusive and exhaustive partitioning properties follow directly from argmax assignment, which is widely used in segmentation.
  2. Although the framework supports variable numbers of part prompts, its effectiveness for substantially larger or mismatched part counts remains unclear. Moreover, every intended part requires explicit point prompts, limiting its applicability.
  3. Unclear significance of the unified framework. Although the unified framework achieves strong performance across three tasks, the specific benefits of sharing a common architecture, compared with separately trained task-specific variants under controlled settings, remain unclear.
  4. Limited improvement in whole-shape fidelity. For mesh inputs, the method achieves worse whole-shape CD and F1 than several baselines (Table 1), suggesting that part-level improvements come with some loss of overall geometric accuracy.
Questions for authors

Please see the weaknesses.

2 comments

We sincerely appreciate the reviewer’s recognition of our unified solution and the effectiveness of our model design. We respond to each point as follows.

Weakness 1. We would like to clarify that the novelty of our work lies not in the complexity of any individual operator, but in the overall problem formulation and framework design. Existing methods may produce overlapping parts or missing components, rather than a coherent decomposition of the whole shape. To address this limitation, we formulate 3D part decomposition as a joint partitioning problem and develop a prompt-conditioned framework that produces complete and mutually compatible parts by construction. Furthermore, our model components are carefully designed to work together, yielding SOTA part quality and efficient inference, while supporting all three tasks within a single model. The results demonstrate the effectiveness of our overall approach, rather than any isolated operator.

Weakness 2. Fig. 5 demonstrates strong performance at large part counts, as further visualized in Fig. 10. To reduce ambiguity in part decomposition and give users direct control over which parts they want to extract, we intentionally introduce point prompts. Point-based prompting has also been explored in established promptable 3D segmentation methods (eg, Point-SAM and S²AM3D).

Weakness 3. We clarify that our approach only trains a single model, rather than combining any task-specific training procedures, objectives, or adaptations. Training separate models for each task would repeat the same training process. At inference time, the same model supports all three tasks. This unified design eliminates the need to train separate models, improving the framework's usability and practical applicability.

Weakness 4. We clarify that whole-shape metrics are measured using the frozen, pretrained Hunyuan3D-2.1 backbone’s prediction (Fig. 2, ∂Ω\partial \Omega) rather than our method’s (as revised in Sec. 4.1). Despite starting with the least accurate whole-shape geometry, our method achieves the best part-level quality, demonstrating our approach’s effectiveness.

My concerns have been addressed and I raise my score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.