Pith. sign in

REVIEW 2 major objections 4 minor 45 references

Pretrained generative image models can turn any 3D object into a compact set of semantic superquadric primitives without any 3D training.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Prompted generative image models plus classical superquadric fitting yield category-agnostic 3D primitive abstractions with the lowest Chamfer distance on HumanPrim and Toys4K using 5–9 parts.

T0 review reviewed 2026-07-11 challenge →

load-bearing objection Solid training-free harness: generative paint-masks + classical superquadric fit give lowest CD with 5–9 parts; bottleneck study is honest, but headline numbers ride a closed generator. the 2 major comments →

arxiv 2607.05568 v1 pith:P6WN3IGW submitted 2026-07-06 cs.CV cs.AI

Harnessing Generative Image Models for Training-Free Primitive Shape Abstraction

Gregor Kobsik , Tim Elsner , Leif Kobbelt This is my paper
classification cs.CV cs.AI
keywords shape abstractionsuperquadricsgenerative image modelstraining-freepart segmentation3D primitivesvision-language modelsmulti-view reprojection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Good 3D shape abstraction needs two things: a semantic sense of which parts belong together, and a geometric fit of a simple primitive to each part. Learning-based methods usually supply both only inside their training categories and often only for upright objects; pure geometric optimizers ignore meaning and can split a chair leg arbitrarily. This paper shows that large generative image models already know how to name and paint object parts in ordinary images. The authors render four views of any mesh or point cloud, ask a vision-language model to list the semantic parts and assign colors, prompt a generative model to paint those colors onto the views, vote the labels back onto the 3D geometry, clean the clusters, and fit one superquadric per part by classical multi-start optimization. The pipeline has no learned 3D parameters, works across categories and orientations, and on two public benchmarks records the lowest surface error while using only five to nine primitives per object. A controlled swap of generated masks for human part labels raises accuracy further, proving that segmentation quality—not the fitter—is the current ceiling and that the method will inherit future gains in generative models for free.

Core claim

A training-free harness that elicits multi-view color-coded part masks from pretrained generative image models, reprojects them to 3D, and fits one superquadric per part yields category-agnostic, orientation-invariant shape abstractions that achieve the lowest Chamfer distance among evaluated methods on HumanPrim and Toys4K while averaging only 5–9 primitives per object; part segmentation, not primitive fitting, is the accuracy bottleneck.

What carries the argument

The five-step training-free harness: multi-view rendering, vision-language analysis that fixes a part-to-color map, generative painting of a consistent color-coded mask, per-pixel voting plus color-restricted spatial clustering, and parallel multi-start Chamfer optimization of a superquadric (with optional taper and bend) to each resulting part cloud.

Load-bearing premise

The method assumes a prompted generative image model can paint multi-view part masks that stay color-consistent across opposing viewpoints and accurate enough that simple voting and clustering recover clean semantic 3D part clouds.

What would settle it

Hold the superquadric fitter fixed and replace the generative masks with human-annotated part labels (e.g., PartNet hierarchies split into connected components). If Chamfer distance and volumetric IoU do not improve at matched primitive count, the claim that segmentation—not fitting—is the bottleneck is false. Separately, if the full pipeline no longer records the lowest Chamfer distance on HumanPrim and Toys4K under the same evaluation protocol, the main quantitative claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Abstraction accuracy will rise automatically as generative image models improve, without any 3D retraining.
  • Category-agnostic and orientation-invariant primitive abstraction becomes available for in-the-wild 3D data that previously required specialized training sets.
  • Compact semantic decompositions of 5–9 superquadrics can feed robotics, collision, simulation, and editing pipelines without dense over-segmentation.
  • The same render–analyze–paint–lift pattern can supply semantic parts for other 3D tasks that lack 3D part supervision.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Majority-vote ensembles over a few independent generative runs would likely dampen non-determinism in masks and raise run-to-run stability at modest extra cost.
  • Adaptive view selection that maximizes unobserved surface area should close much of the remaining volumetric gap on thin or sheet-like parts that fixed four-view layouts miss.
  • Closing an agentic loop in which the foundation model proposes primitive placements, scores fit quality, and revises the segmentation would couple semantic judgment directly to geometric error.
  • The same harness could replace category-specific 3D segmenters in CAD reverse-engineering and multi-object scene parsing where only images or meshes are available.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper proposes a training-free pipeline for semantic 3D shape abstraction into superquadrics. It renders four multi-view images of an input object, prompts a VLM to produce a structured part-to-color analysis, uses a generative image model to paint consistent color-coded segmentation masks, reprojects labels via per-pixel voting, extracts clean part point clouds by color-restricted spatial clustering, and fits one superquadric per part by parallel multi-start L-BFGS minimization of bidirectional Chamfer distance (Eq. 1). On HumanPrim and Toys4K the method reports the lowest Chamfer distances (0.079 and 0.093) among the evaluated baselines while using only 5–9 primitives on average and near-unit overlap rates; a PartNet ground-truth segmentation swap (Table 2) with the fitter held fixed isolates part segmentation quality as the accuracy bottleneck. The approach is presented as category-agnostic and orientation-invariant because it contains no learned 3D parameters.

Significance. If the empirical claims hold, the work supplies a concrete demonstration that large-scale 2D generative models can be harnessed, without any 3D fine-tuning or task-specific training, to supply the semantic decomposition that classical geometric optimizers have historically lacked. The explicit upper-bound experiment (Table 2) and the run-to-run variance quantification (Table 3) are methodological strengths that make the central claim falsifiable and show that future gains in generative models will transfer automatically. Compact, low-overlap, semantically coherent abstractions are directly useful for robotics, simulation and editing; the training-free framing therefore constitutes a useful alternative paradigm to both category-specific learned abstractors and purely geometric multi-primitive optimizers.

major comments (2)
  1. [§3.1–3.2, Table 1, Fig. 7, §4.3–4.5] The headline CD numbers in Table 1 (0.079 / 0.093) and the claim that the pipeline is immediately usable rest on the multi-view mask quality of the proprietary generator NanoBanana 2. Section 3.1–3.2 and the instance-ablation (Fig. 8) already document that left/right identities cannot be tracked and that type-level coloring is required; residual failures (inconsistent masks, small-cluster loss, color bleed) appear in §4.5 / Fig. 5. The PartNet GT swap (Table 2) correctly isolates segmentation as the bottleneck, yet Fig. 7 shows that most open-source generators fail the same cross-view consistency test and no open model is shown to recover comparable end-to-end CD. Without either an open-model baseline that matches the reported numbers or released intermediate masks, the load-bearing premise cannot be independently audited or stress-tested.
  2. [§4.2, Table 1] The four fixed opposing viewpoints are acknowledged in §4.2 to leave undersides of thin structures (wings, seats, tabletops) unobserved, producing primitives that are too thin and lowering IoU relative to EMS on Toys4K. Because the paper’s own upper-bound study shows that cleaner segmentation alone raises IoU by 11–17 points, an ablation that replaces the fixed camera set with even a simple adaptive or six-view schedule is needed to confirm that the residual IoU gap is not an artifact of the chosen capture geometry rather than of the generative stage.
minor comments (4)
  1. [§4.1] The precise version string and API parameters of NanoBanana 2 used for all reported numbers should be stated in §4.1 so that future readers can attempt approximate reproduction once the model is publicly versioned.
  2. [§3.3, Eq. (1)] In Eq. (1) the surface-area weights w_s are introduced without a short derivation or reference; a one-sentence pointer to the parametric sampling density of superquadrics would help readers unfamiliar with the classical literature.
  3. [Figs. 2–3] Figure 2 and Figure 3 annotate primitive counts in the lower-right corner, but the font size is small enough that the numbers become hard to read once the figures are scaled for print; enlarging them would improve clarity.
  4. [Appendix, Fig. 7] The appendix model-comparison figure (Fig. 7) lists prices but does not report any quantitative mask-consistency metric (e.g., cross-view color agreement or boundary F-score); even a simple aggregate score would make the qualitative ranking more transparent.

Circularity Check

0 steps flagged

No circularity: empirical pipeline and external-benchmark comparisons are self-contained; minor author self-citation of F2C is only a baseline, not load-bearing.

full rationale

The paper presents a training-free engineering pipeline (multi-view render o VLM analysis o generative mask painting o reprojection + color-restricted clustering o multi-start L-BFGS superquadric fitting minimizing bidirectional Chamfer, Eq. 1) whose accuracy claims are purely empirical: lowest CD on the external HumanPrim and Toys4K benchmarks versus independent baselines (PrimAny, EMS, SuperDec) plus the authors’ own prior F2C. The ground-truth PartNet ablation (Table 2) holds the identical fitter fixed and substitutes human labels, correctly isolating segmentation quality as the bottleneck rather than defining the result by construction. Superquadric formalism and the Chamfer objective are standard (Barr, Solina & Bajcsy, etc.); no parameter is fitted to a data subset and then re-presented as a prediction, no uniqueness theorem is imported from the authors, and no ansatz is smuggled via self-citation. The single self-citation of F2C [KHH*26] appears only as one of four evaluated baselines and does not underwrite the central claim. Consequently the derivation chain contains no circular step.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central empirical claim rests on standard geometry (superquadrics, Chamfer, multi-view projection), domain assumptions about foundation-model part understanding, and several hand-chosen pipeline hyperparameters. No new physical entities are postulated; the 'training-free harness' is an engineering composition. Free parameters affect clustering cleanliness and fit quality but are not fitted to claim the CD ranking post hoc in an opaque way—the paper reports them as design choices.

free parameters (5)
  • number_of_views
    Fixed to four opposing perspective views; coverage of thin/underside surfaces and IoU depend on this heuristic (§3.1, §4.2).
  • clustering_radius_r_and_min_cluster_size
    Color-restricted flood-fill radius and minimum cluster size discard noise but can remove thin legs; only lightly tuned (§3.2, §4.5).
  • Chamfer_backward_weight_lambda
    Balances surface coverage vs. primitive over-extension in Eq. 1; affects fitted shape without being learned from a labeled abstraction set.
  • multi_start_candidate_set
    Nine candidates from three shape types × three axis orientations from ApproxMVBB; choice of discrete starts shapes which local minima are reachable (§3.3).
  • generative_model_and_prompt_palette
    Results use NanoBanana 2 selected after qualitative comparison of eight models (Fig. 7); mask quality and thus CD/IoU depend on this external choice.
axioms (4)
  • domain assumption Scaled generative image models encode category-agnostic part-level visual knowledge that can be elicited by prompting without task-specific fine-tuning.
    Stated in abstract/intro and grounded by citation to Gabeur et al.; the pipeline's semantic stage stands or falls on this.
  • domain assumption Superquadric surfaces with optional taper/bend (Barr/Solina formalism) are an adequate primitive family for compact part abstraction.
    Adopted from classical literature and prior abstraction work; appendix restates F and parametric form.
  • ad hoc to paper Assigning one color per semantic part type (not per instance) yields reprojectable multi-view labels.
    §3.1 and instance-segmentation ablation (Fig. 8) justify this design after observing left/right identity failures.
  • domain assumption Bidirectional Chamfer distance on unit-cube-normalized parts is a sufficient objective for selecting among multi-start superquadric fits.
    §3.3 Eq. 1; standard surface metric, not volume-supervised.

reviewed 2026-07-11 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Harnessing Generative Image Models for Training-Free Primitive Shape Abstraction." pith.science (2026). https://pith.science/paper/P6WN3IGW

@misc{pith2026260705568,
  author       = {Pith},
  title        = {Pith review of: Harnessing Generative Image Models for Training-Free Primitive Shape Abstraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P6WN3IGW}},
  note         = {Machine review of arXiv:2607.05568}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Representing 3D shapes as compact sets of geometric primitives is fundamental to robotics, simulation, and scene understanding. Generative image models trained at scale have recently emerged as generalist visual learners that can identify and segment object parts directly in the image domain, across arbitrary categories and without task-specific training. Adapting such models to downstream tasks typically requires fine-tuning; we ask whether their pretrained capability can instead be harnessed directly, without any training, and answer affirmatively with a training-free harness. Our pipeline renders multi-view images of a 3D object, uses a vision-language model to analyze its semantic parts, prompts a generative image model to paint a color-coded part segmentation mask, reprojects it onto the geometry, and fits a superquadric primitive to each part via parameter optimization. The approach contains no learned parameters: it is category-agnostic and orientation-invariant, properties that previous learning-based models struggled with. Its accuracy ceiling rises with future generative-model improvements, which we confirm with a ground-truth segmentation study showing that part segmentation, not primitive fitting, is the current accuracy bottleneck. On HumanPrim and Toys4K, our method achieves the lowest Chamfer distance among all evaluated methods, using 5--9 primitives per object on average.

Figures

Figures reproduced from arXiv: 2607.05568 by Gregor Kobsik, Leif Kobbelt, Tim Elsner.

Figure 1
Figure 1. Figure 1: Pipeline overview. Given any 3D object, our framework extracts semantic part primitives in five steps: (1) Multi-View Render: We generate four perspective views and prompt a vision-language model (VLM) to identify the object’s semantic parts. (2) Generative Segmentation: A generative image model paints a color-coded segmentation mask based on that analysis. (3) 3D Reprojection: Pixel labels are projected b… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison on HumanPrim. Across di￾verse unseen categories, our method produces compact, semanti￾cally structured abstractions without category-specific training [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on Toys4K. Our method pro￾duces compact, semantically meaningful abstractions across a wide range of everyday objects — from simple primitives (apple, bottle) to articulated animals and complex geometries (bicycle, tree). introduced by Wang et al. [WCH∗ 25], quantifies how much the fit￾ted primitives overlap each other: OR = ∑x ∑θ Mθ(x) ∑x S θ Mθ(x) (2) where Mθ(x) = 1 if point x lie… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative determinism examples. Segmentation masks for a chair (top) and an insect (bottom) across five runs. Multi-View Segmentation Abstraction → → → → [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Failure cases. Each row shows a shape where the pipeline produces a poor result. Top: the generative model produces an inconsistent segmentation across views, although it does not af￾fect our pipeline. Bottom: color-restricted clustering wrongfully re￾moves the chair leg (green), which is absent from the final abstrac￾tion. 4.5. Failure Cases and Limitations Three failure modes are most prominent ( [PITH_… view at source ↗
Figure 6
Figure 6. Figure 6: VLM analysis output and pipeline results for two representative objects. For each object, the box shows the structured JSON returned by Call 1 (Analysis), listing the semantic parts, their exact descriptions, and assigned colors. Below: the four-view input, the generative-model segmentation mask (Call 2), and the final superquadric abstraction [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of image-generation VLMs on part segmentation. Color-coded segmentation masks for two hand￾picked objects (rows within each block), grouped into open-source and proprietary models. All models receive the identical input and the same prompt (Section ). The price above each column is the cost per edited image (USD), taken from the Artificial Analysis website [Art26]. Among open-source … view at source ↗
Figure 8
Figure 8. Figure 8: Instance segmentation ablation. An explicit instance￾segmentation prompt assigns a distinct color to each part instance (16 colors for the chair shown above). The resulting masks (top left) reveal that the generative image model cannot reliably track instance identities across views: left/right and front/back relation￾ships are especially prone to being swapped or lost. We therefore assign a single color p… view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison on ShapeNet. Our method recovers semantically meaningful part structure using a low number of primi￾tives. EMS produces geometrically arbitrary splits or fails to produce an abstraction at all; PrimitiveAnything achieves finer surface coverage only by using 10x more primitives. F2C and SuperDec were trained on this dataset; their in-distribution results are comparable to ours. flat o… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    : GPT-4 technical report

    Achiam J., Adler S., Agarwal S., Ahmad L., Akkaya I., et al. : GPT-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    Pith/arXiv arXiv 2023

  2. [2]

    : Iterative superquadric recomposition of 3D objects

    Alaniz S., Krawez M., Akata Z. : Iterative superquadric recomposition of 3D objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)

    2023

  3. [3]

    https://artificialanalysis.ai/image/leaderboard/editing, 2026

    Artificial Analysis : Image editing leaderboard. https://artificialanalysis.ai/image/leaderboard/editing, 2026. Accessed: 2026-05-30

    2026

  4. [4]

    Barr A. H. : Superquadrics and angle-preserving transformations. IEEE Computer Graphics and Applications 1, 1 (1981), 11--23

    1981

  5. [5]

    : Efficiently approximating the minimum-volume bounding box of a point set in three dimensions

    Barequet G., Har-Peled S. : Efficiently approximating the minimum-volume bounding box of a point set in three dimensions. Journal of Algorithms 38, 1 (2001), 91--109

    2001

  6. [6]

    X., Funkhouser T., Guibas L., Hanrahan P., Huang Q., Li Z., Savarese S., Savva M., Song S., Su H., Xiao J., Yi L., Yu F

    Chang A. X., Funkhouser T., Guibas L., Hanrahan P., Huang Q., Li Z., Savarese S., Savva M., Song S., Su H., Xiao J., Yi L., Yu F. : ShapeNet : An information-rich 3D model repository. arXiv preprint arXiv:1512.03012 (2015)

    Pith/arXiv arXiv 2015

  7. [7]

    : SuperDec : 3D scene decomposition with superquadric primitives

    Fedele E., Sun B., Guibas L., Pollefeys M., Engelmann F. : SuperDec : 3D scene decomposition with superquadric primitives. In Proceedings of the IEEE/CVF International Conference on Computer Vision (2025)

    2025

  8. [8]

    : Image generators are generalist vision learners

    Gabeur V., et al. : Image generators are generalist vision learners. arXiv preprint (2026). DeepMind

    2026

  9. [9]

    M., et al

    Gemini Team , Anil R., Borgeaud S., Alayrac J.-B., Yu J., Soricut R., Schalkwyk J., Dai A. M., et al. : Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)

    Pith/arXiv arXiv 2023

  10. [10]

    : Residual primitive fitting of 3D shapes with SuperFrusta

    Ganeshan A., Gadelha M., Groueix T., Chen Z., Chaudhuri S., Kim V., Yifan W., Ritchie D. : Residual primitive fitting of 3D shapes with SuperFrusta . arXiv preprint arXiv:2512.09201 (2025)

    arXiv 2025

  11. [11]

    https://blog.google/innovation-and-ai/technology/ai/nano-banana-2/, 2025

    Google DeepMind : Nano Banana 2 . https://blog.google/innovation-and-ai/technology/ai/nano-banana-2/, 2025. Google Blog

    2025

  12. [12]

    : 3D part segmentation via geometric aggregation of 2D visual features

    Garosi M., Tedoldi R., Boscaini D., Mancini M., Sebe N., Poiesi F. : 3D part segmentation via geometric aggregation of 2D visual features. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2025)

    2025

  13. [13]

    : Self-supervised learning of hybrid part-aware 3D representations of 2D Gaussians and superquadrics

    Gao Z., Yi R., Huang Y., Chen W., Zhu C., Xu K. : Self-supervised learning of hybrid part-aware 3D representations of 2D Gaussians and superquadrics. In Proceedings of the IEEE/CVF International Conference on Computer Vision (2025)

    2025

  14. [14]

    : Segmentation and Recovery of Superquadrics

    Jakli c A., Leonardis A., Solina F. : Segmentation and Recovery of Superquadrics. Springer, 2000

    2000

  15. [15]

    : Learning fine-to-coarse cuboid shape abstraction

    Kobsik G., Henkel M., He Y., Czech V., Elsner T., Lim I., Kobbelt L. : Learning fine-to-coarse cuboid shape abstraction. In Proceedings of Eurographics (2026)

    2026

  16. [16]

    C., Lo W.-Y., Doll \'a r P., Girshick R

    Kirillov A., Mintun E., Ravi N., Mao H., Rolland C., Gustafson L., Xiao T., Whitehead S., Berg A. C., Lo W.-Y., Doll \'a r P., Girshick R. : Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023), pp. 3992--4003

    2023

  17. [17]

    : PASTA : Controllable part-aware shape generation with autoregressive transformers

    Li J., et al. : PASTA : Controllable part-aware shape generation with autoregressive transformers. arXiv preprint arXiv:2407.13677 (2024)

    Pith/arXiv arXiv 2024

  18. [18]

    : P3-SAM : Native 3D part segmentation

    Li J., et al. : P3-SAM : Native 3D part segmentation. arXiv preprint arXiv:2509.06784 (2025)

    arXiv 2025

  19. [19]

    : PartField : Learning part field representations for generalizable 3D part segmentation

    Liu M., et al. : PartField : Learning part field representations for generalizable 3D part segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2025)

    2025

  20. [20]

    : SegviGen : Repurposing 3D generative model for part segmentation

    Li L., Feng H., Huang Z., Chen H., Nie W., Hou S., Fan K., Hu P., Wang S., Li B., Sheng L. : SegviGen : Repurposing 3D generative model for part segmentation. arXiv preprint arXiv:2603.16869 (2026)

    Pith/arXiv arXiv 2026

  21. [21]

    : Superquadrics for segmenting and modeling range data

    Leonardis A., Jakli c A., Solina F. : Superquadrics for segmenting and modeling range data. IEEE Transactions on Pattern Analysis and Machine Intelligence 19, 11 (1997), 1289--1295

    1997

  22. [22]

    C., Nocedal J

    Liu D. C., Nocedal J. : On the limited memory BFGS method for large scale optimization. Mathematical Programming 45, 1--3 (1989), 503--528

    1989

  23. [23]

    Liu W., Wu Y., Ruan S., Chirikjian G. S. : Robust and accurate superquadric recovery: A probabilistic approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022), pp. 2676--2685

    2022

  24. [24]

    Liu W., Wu Y., Ruan S., Chirikjian G. S. : Marching-primitives: Shape abstraction from signed distance function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023), pp. 8558--8567

    2023

  25. [25]

    : PartSLIP : Low-shot part segmentation for 3D point clouds via pretrained image-language models

    Liu M., Zhu Y., Cai H., Han S., Cai Z., Porikli F., Su H. : PartSLIP : Low-shot part segmentation for 3D point clouds via pretrained image-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023), pp. 21736--21746

    2023

  26. [26]

    A., Aubry M

    Monnier T., Austin J., Kanazawa A., Efros A. A., Aubry M. : Differentiable blocks world: Qualitative 3D decomposition by rendering primitives. In Advances in Neural Information Processing Systems (2023), vol. 36

    2023

  27. [27]

    X., Yi L., Tripathi S., Guibas L

    Mo K., Zhu S., Chang A. X., Yi L., Tripathi S., Guibas L. J., Su H. : PartNet : A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)

    2019

  28. [28]

    Pentland A. P. : Parts: Structured descriptions of shape. In Proceedings of the AAAI Conference on Artificial Intelligence (1986)

    1986

  29. [29]

    O., Geiger A

    Paschalidou D., Ulusoy A. O., Geiger A. : Superquadrics revisited: Learning 3D shape parsing beyond cuboids. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019), pp. 10344--10353

    2019

  30. [30]

    : Neural parts: Learning expressive deformable geometry with invertible neural networks

    Paschalidou D., Van Gool L., Geiger A. : Neural parts: Learning expressive deformable geometry with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021), pp. 3204--3215

    2021

  31. [31]

    : Exploiting the potential of GPT-4V for zero-shot 3D point cloud understanding

    Qi G., et al. : Exploiting the potential of GPT-4V for zero-shot 3D point cloud understanding. arXiv preprint arXiv:2401.07572 (2024)

    Pith/arXiv arXiv 2024

  32. [32]

    : Recovery of parametric models from range images: The case for superquadrics with global deformations

    Solina F., Bajcsy R. : Recovery of parametric models from range images: The case for superquadrics with global deformations. IEEE Transactions on Pattern Analysis and Machine Intelligence 12, 2 (1990), 131--147

    1990

  33. [33]

    Stojanov S., Thai A., Rehg J. M. : Using shape to categorize: Low-shot learning with an explicit shape bias

  34. [34]

    : Learning adaptive hierarchical cuboid abstractions of 3D shape collections

    Sun Q., Wang Z., Xu K., Cohen-Or D., Chen B. : Learning adaptive hierarchical cuboid abstractions of 3D shape collections. ACM Transactions on Graphics (SIGGRAPH Asia) 38, 6 (2019)

    2019

  35. [35]

    : Llm-primitives: Large language model for 3D reconstruction with primitives

    Tian K., Hu Z., Guan Y., Zhang J. : Llm-primitives: Large language model for 3D reconstruction with primitives. In ACM SIGGRAPH Asia 2025 Conference Papers (2025)

    2025

  36. [36]

    J., Efros A

    Tulsiani S., Su H., Guibas L. J., Efros A. A., Malik J. : Learning shape abstractions by assembling volumetric primitives. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2017), pp. 2635--2643

    2017

  37. [37]

    : Light-SQ : Structure-aware shape abstraction with superquadrics for generated meshes

    Wang Y., Chen W., Hu Z., Zhang R., Yin Y., Wu R., Luo K., Qian S., Ma Y., Li H., et al. : Light-SQ : Structure-aware shape abstraction with superquadrics for generated meshes. In ACM SIGGRAPH Asia 2025 Conference Papers (2025)

    2025

  38. [38]

    MeshSegmenter: Zero-Shot Mesh Semantic Segmentation via Texture Synthesis

    Yu Z., et al. : MeshSegmenter : Zero-shot mesh semantic segmentation via texture synthesis. arXiv preprint arXiv:2407.13675 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  39. [39]

    : Unsupervised learning of fine structure generation for 3D point clouds by 2D projection matching

    Yang B., Chen S. : Unsupervised learning of fine structure generation for 3D point clouds by 2D projection matching. ACM Transactions on Graphics 40, 4 (2021)

    2021

  40. [40]

    Y., Cao Y.-P., Liu X

    Yang Y., Huang Y., Guo Y.-C., Lu L., Wu X., Lam E. Y., Cao Y.-P., Liu X. : SAMPart3D : Segment any part in 3D objects. arXiv preprint arXiv:2411.07184 (2024)

    Pith/arXiv arXiv 2024

  41. [41]

    : PrimitiveAnything : Human-crafted 3D primitive assembly generation with auto-regressive transformer

    Ye J., He Y., Zhou Y., Zhu Y., Xiao K., Liu Y.-J., Yang W., Han X. : PrimitiveAnything : Human-crafted 3D primitive assembly generation with auto-regressive transformer. In ACM SIGGRAPH 2025 Conference Papers (2025)

    2025

  42. [42]

    A., Han J., Thomas R., Zhang H., Du Y., Chen H., Engelmann F., You S., Guibas L

    You Y., Uy M. A., Han J., Thomas R., Zhang H., Du Y., Chen H., Engelmann F., You S., Guibas L. J. : Img2 CAD : Reverse engineering 3D CAD models from images through VLM -assisted conditional factorization. In ACM SIGGRAPH Asia 2024 Conference Papers (2024)

    2024

  43. [43]

    : SweepNet : Unsupervised shape abstraction via neural sweeping

    Zhao M., et al. : SweepNet : Unsupervised shape abstraction via neural sweeping. In Proceedings of the European Conference on Computer Vision (ECCV) (2024)

    2024

  44. [44]

    : Point- SAM : Promptable 3D segmentation model for point clouds

    Zhou Y., Lyu J., et al. : Point- SAM : Promptable 3D segmentation model for point clouds. In Proceedings of the International Conference on Learning Representations (ICLR) (2025)

    2025

  45. [45]

    : 3D-PRNN : Generating shape primitives with recurrent neural networks

    Zou C., Yumer E., Yang J., Ceylan D., Hoiem D. : 3D-PRNN : Generating shape primitives with recurrent neural networks. In Proceedings of the IEEE International Conference on Computer Vision (2017), pp. 900--909

    2017

This paper was first reviewed by grok-4.5 on July 11, 2026.