acceptodds
Under review as a conference paper at ICLR 2027

Parts by Boundaries: Native Part-Aware 3D Generation with Seg-Voxel Representation

15 views

Abstract

Recent 3D generative models produce high-quality assets but typically treat objects as monolithic shapes. Existing part-aware approaches address this limitation through external segmentation, predefined layouts, or additional planning, leaving shape generation and part discovery largely decoupled. In this work, we propose a native part-aware 3D generation framework that models inter-part separation instead of explicit part assignments. Our key observation is that recovering part structure does not require explicit part assignments over the entire shape; it is sufficient to identify the interfaces where contacting parts should be separated. We therefore cast part decomposition as a local boundary prediction task, replacing open-ended part assignments with simple binary separation cues. To support this formulation, we introduce Seg-Voxel, a sparse voxel representation built upon the voxel design of TRELLIS.2 that jointly encodes geometry and inter-part boundaries. We further develop a native 3D generative pipeline that synthesizes Seg-Voxel to jointly generate object geometry and the separation structure required to partition it into individual components. The predicted boundaries are refined to ensure reliable geometric separation, then reassigned to adjacent components to recover coherent part assets. Experiments demonstrate that our method outperforms existing part-aware approaches in geometric quality, decomposition accuracy, and efficiency, enabling high-quality native part-aware generation without external segmentation or complex post-generation planning.

Primary area:
generative models
Keywords:
3D Generation, Part Generation

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_R91AOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Seg-Voxel, a representation and generation framework that jointly produces object geometry and part-separation boundaries from a single image. Built on TRELLIS.2, it adds a binary boundary attribute to the sparse voxel representation and fine-tunes the VAE and flow model to generate geometry and boundaries together. The decoded boundaries are completed using graph-based processing, temporarily removed to extract connected components, and reassigned to recover part surfaces. Experiments on HY3D-Bench evaluate reconstruction, part generation, and the two boundary-processing stages, reporting competitive whole-object geometry, improved part-level scores, and a runtime of 20.68 seconds per object.

Critical strengths
  • Representing part structure through local separation boundaries is a useful way to incorporate part discovery into native 3D generation. It avoids explicitly predicting global part identities or requiring an external segmentation model.

  • The experiments support the practical value of the approach: it preserves whole-object geometric quality while improving the reported part-level scores and achieving a low inference time. The reconstruction results and boundary-processing ablations provide useful evidence for the main design choices.

Critical weaknesses
  • The part-level metrics check whether each ground-truth part has a similar generated part, but do not directly check whether the output contains extra fragments.

  • The method treats disconnected pieces as different parts, even if they belong to the same annotated part. It also allows parts to share boundary faces. The paper should show the parts pulled apart to clarify whether their contact surfaces are complete.

  • The paper shows examples with different levels of complexity, but reports only average scores over 300 test objects. It does not give separate scores for objects with many parts, small or thin parts, or complicated contacts.

  • The paper reports 20.68 seconds per object, but does not give separate times for generation, boundary completion, connected-component extraction, and boundary assignment.

Questions for authors

The following questions correspond, in order, to the critical weaknesses above.

  1. Could you check whether the generated outputs contain extra fragments that are not reflected in the best-match scores? Is there an additional metric that could measure this?

  2. How do you handle an annotated part made of disconnected pieces? Please show the generated parts pulled apart so we can check whether their contact surfaces are complete and whether they share faces.

  3. Could you report separate scores for objects with many parts, small or thin parts, and complicated contacts? Please also show failure cases from the final outputs.

  4. How much of the 20.68 seconds is spent on generation, boundary completion, connected-component extraction, and boundary assignment?

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

The paper introduces a method that finetunes TRELLIS.2 for part-aware shape generation. It additionally encodes segmentation information in the voxel based on existing segmentation annotations, and finetunes the VAE as well as the latent generator to autoencode and generate the shapes with the segmentation flags. The segmentation flags, once generated, are used to obtain the segmentation defined on the surface of the shapes. They do this via a training-free post-processing step where any open curves are closed up to obtain the. final segmentation on the generated shape. The method beats state of the art part-level shape generation works in standard metrics and the qualitative results look promising.

Critical strengths
  1. The idea of encoding an additional segmentation voxel for part-level shape generation is interesting and novel.
  2. The paper achieves state-of-the-art performance on part-level shape generation compared to many SOTA baselines. This demonstrates the effectiveness of their method.
  3. The method is a simple modification of the TRELLIS.2 pipeline that enables it to generate part-level. This shows that the methods is compatible with SOTA shape generator.
Critical weaknesses
  1. The Mesh to seg-voxel conversion seems to have some flaws. For example, if a small or narrow part lies entirely within the boundary-face band, removing that band would make that segment unrecoverable from the post-processing. How does the method handle this case, and how often does it occur on the training dataset?
  2. The post-processing algorithm could wrongly close a loop and form a segment that doesn't suppose to be there. For example, if the ground truth segmentation contains a region with a small bottleneck in the middle (imaging an 8 like region), then the post-processing algorithm would wrongly segment the two region into two separate regions. This is a algorithmic weakness rather than something that can be fixed with more training. Does this occur in practice and if so how much error does it introduce to the part segmentation. A discussion of these algorithmic failure cases would be good to have.
Questions for authors

See weaknesses

Reviewer_wm2pOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Seg-Voxel, a representation that extends TRELLIS.2 with binary inter-part boundary attributes. Instead of generating explicit part assignments, the method jointly generates geometry and separation boundaries through a fine-tuned VAE and flow model. Boundary completion and reassignment then recover individual parts through mesh connectivity. Experiments show improved part-level geometry metrics and inference efficiency over the evaluated baselines while preserving whole-object reconstruction quality.

Critical strengths

Reformulating part structure as local separation boundaries is a meaningful contribution at the representation level. It avoids explicit part identities and predefined part counts, integrating decomposition into an existing geometry generation framework with a small extension to its output attributes. I find this a useful direction for native compositional 3D generation. Despite the limitations in validation and implementation details below, the representation offers enough insight to support acceptance.

The experiments support the practical potential of this formulation. Seg-Voxel preserves the base VAE’s whole-object reconstruction quality at both 5123512^3 and 102431024^3, while the generation results improve part-level metrics over the evaluated baselines. The boundary-processing ablations also help establish why completion and reassignment are needed to turn predicted local boundaries into usable part decompositions.

Critical weaknesses
  1. The evaluation does not directly establish how reliably the VAE reconstructs separation boundaries, although this is central to the proposed representation. The balanced binary cross-entropy addresses class imbalance, but the paper reports geometry-based reconstruction metrics without boundary precision, recall, F1, or IoU. It is also unclear whether Table 1 includes boundary completion. Since small boundary gaps can merge otherwise distinct parts, the current results do not sufficiently separate the quality of the learned boundary representation from the contribution of post-processing.
  2. Important details of training and part recovery remain unclear. The paper gives fine-tuning schedules for both the Seg-VAE and Seg-SLAT flow model, but does not specify which parameters are updated or frozen. The boundary-completion procedure also lacks sufficient detail about pair selection, search neighborhoods, candidate scoring, and stopping conditions. Its robustness to entirely missed boundaries or spurious boundaries is not evaluated. Additionally, the 102431024^3 results establish high-resolution VAE reconstruction, while the resolution and refinement procedure used for end-to-end generation remain unspecified.
Questions for authors
  1. Could you report boundary-voxel precision, recall, F1, or IoU before boundary completion? Does Table 1 include the completion and reassignment stages? Reporting part merge/split errors before and after completion would help distinguish VAE reconstruction quality from post-processing effectiveness.

  2. Are all parameters of the Seg-VAE encoder/decoder and Seg-SLAT flow model fine-tuned?

  3. How does boundary completion handle entirely missed boundaries and incorrectly predicted boundaries, rather than local gaps in otherwise correct boundaries? Please clarify the main search parameters and stopping conditions.

  4. What output resolution is used for Table 2 and its reported runtime? Does end-to-end generation include high-resolution cascade refinement?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.