Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Auto-Regressive Surface Cutting

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read SeamGPT turns 3D mesh cutting into a next-token prediction task, beating standard UV unwrappers on distortion.

desk verdict A genuinely new autoregressive formulation for surface cutting, but the SOTA UV claim is undercut by an unfair comparison and a metric that rewards fragmentation. read the letter →

arxiv 2506.18017 v1 pith:LXQNYJM4 submitted 2025-06-22 cs.GR cs.AIcs.CV

Yang Li , Victor Cheung , Xinhai Liu , Yuguang Chen , Zhongjin Luo , Biwen Lei , Haohan Weng , Zibo Zhao
show 3 more authors
Jingwei Huang Zhuo Chen Chunchao Guo
This is my paper
classification cs.GRcs.AIcs.CV
keywords surfacecuttingUVunwrappingautoregressivegenerationseamprediction3Dmeshsegmentationconformaldistortionartist-styleseamspointcloudconditioning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

SeamGPT claims that the long-standing geometry problem of where to cut a 3D mesh for UV unwrapping can be re-cast as a language-modeling problem: the cutting seams become a sequence of 3D line segments, and a GPT-style transformer predicts them one quantized coordinate at a time, conditioned on a point cloud sampled from mesh vertices and edges. The paper trains this model on roughly 560K artist-created meshes whose UV-island boundaries are treated as expert ground-truth seams, so the model learns 'artist style' cutting that yields semantically coherent, low-distortion atlases. On the Flatten Anything and Toys4K benchmarks, the method reports the lowest average conformal distortion (13.04 vs 18.37 for XAtlas), and a user study with 20 professional artists rates its boundary quality and editability highest. It also shows that the generated seams can be used as clean boundary cues to improve PartField's 3D part segmentation. If true, the paper's significance is that a single learned sequence model can replace hand-tuned geometry optimization for a core graphics operation, while making the output align with semantic part boundaries.

What carries the argument

The central mechanism is the seam-as-token sequence. Each seam is a list of 3D line segments, each line segment is its two endpoint vertices, and each vertex is three quantized coordinates, so the whole seam is a flat sequence of $6N_s$ tokens that a causal Transformer decodes auto-regressively. The order is canonicalized by sorting vertices in yzx order and then edges lexicographically, which makes the same geometric seam always appear as the same token string. Shape conditioning comes from a point cloud encoder that samples only from mesh vertices and edges (rather than the surface uniformly), which the paper shows is what keeps predicted seams aligned with mesh topology and improves convergence. The hourglass decoder (borrowed from mesh generation) lets coordinates, vertices, and edges be processed at three causal levels of abstraction.

What would settle it

Compare SeamGPT's average conformal distortion on the benchmarks against the same unwrapping pipeline fed randomly chosen seams of matched length; if random seams achieve similar distortion, then the reported numbers do not establish semantic quality. Also, ask professional artists to label whether a random sample of the training meshes' UV boundaries are intentional cuts or automatic unwrapping artifacts.

Watch

Extended reading notes

Core claim

Formulating surface cutting as next-token prediction over quantized 3D seam coordinates. Given a mesh, SeamGPT samples 61,440 points on vertices and edges, encodes them into a shape latent, then an hourglass causal transformer predicts the seam as an ordered list of line segments, with coordinates quantized to 1024 bins per axis and sorted by a fixed yzx ordering so the sequence has a canonical form. The model is trained with cross-entropy on seam tokens plus a KL regularizer on the shape embedding, and a length embedding lets the user control cutting granularity. The paper reports that this single autoregressive model outperforms XAtlas, Nuvo, FAM, and a learned edge-classification baseline on average conformal distortion on both benchmarks, and that feeding its seams to PartField's patch-voting procedure yields cleaner part boundaries, with a failure case shown when the seam does not follow a semantic boundary.

Load-bearing premise

The load-bearing premise is that UV-island boundaries extracted from the filtered 560K artist meshes are reliable expert seams rather than leftovers of automatic unwrappers, since the benchmark distortion numbers measure flattening quality and cannot detect whether the model's cuts are semantically coherent.

Editorial extensions

If this is right

  • SeamGPT can be used as a drop-in pre-processing step before any standard UV unwrapper, replacing hand-placed seams with predicted artist-style cuts and lowering average conformal distortion below XAtlas on both manifold and non-manifold meshes.
  • Controlling a single length embedding changes cutting granularity from a few large islands to many small patches, giving artists a dial that currently requires manual seam-editing tools.
  • Because the model outputs seams as ordered line segments rather than per-face labels, the predictions can be snapped to mesh vertices and cut via geodesic paths, yielding clean boundaries that directly support 3D part segmentation when combined with a part-feature method such as PartField.
  • The same autoregressive formulation can in principle be applied to any surface-cutting task beyond UV mapping, including digital fabrication and mesh decomposition, since it only requires a mesh and a seam ground truth.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported distortion gains may partly come from having fewer, longer islands rather than from semantic alignment; a useful test would be to compare SeamGPT's seams against seams optimized purely for distortion, and measure semantic coherence separately (e.g., part segmentation accuracy) rather than flattening energy alone.
  • Because training data is filtered by removing 'overlapping UVs' and 'arbitrarily placed seams', the model may silently inherit whatever unwrapping convention dominates the source datasets; one could test domain robustness by fine-tuning on a small set of hand-annotated seams and measuring how much distortion and boundary quality change.
  • The same token-level formulation should transfer to curve outputs beyond straight line segments (e.g., splines), which the paper's limitation section hints at; if the decoder can emit control points instead of endpoint coordinates, the model could produce smoother and more compact seams.
  • A practical extension would be to make SeamGPT interactive: instead of generating a full seam in one pass, let an artist click anchor points and have the model complete the cut, which would more directly mimic the professional workflow the paper claims to imitate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes SeamGPT, an autoregressive transformer that generates cutting seams for 3D meshes by formulating surface cutting as next-token prediction over quantized 3D line-segment coordinates. The model is conditioned on point clouds sampled from mesh vertices and edges, trained on seam annotations extracted from UV-island boundaries of supposedly artist-created meshes, and evaluated on UV-unwrapping distortion (FAM and Toys4K benchmarks), a small user study, and a qualitative part-segmentation enhancement. The central claims are state-of-the-art UV-unwrapping performance with semantically meaningful, artist-style seams.

Significance. If the claims are substantiated, the paper offers a novel and potentially useful formulation: casting surface cutting as a sequence-generation problem, with a large-scale learned prior over seam placement. The idea of using autoregressive transformers for mesh cutting is original and could benefit UV unwrapping, texture editing, and part segmentation. However, the current evidence is insufficient to support the strong claims: the main quantitative comparisons are not apples-to-apples, the supervision assumption on 'artist seams' is unvalidated, and the semantic-coherence claim lacks any direct quantitative support. The paper's machine-checked aspects are limited; no code or trained model is released, and the ablations are mostly qualitative.

major comments (5)
  1. [Tables 1 and 2; Section 5] The comparison is not apples-to-apples. SeamGPT's predicted seams are flattened using Blender's Minimum Stretch algorithm, while XAtlas, Nuvo, and FAM are evaluated through their native pipelines. On 6 of the 13 FAM meshes (Ogre, Armadillo, Nefertiti, Dragon, Homer, Cow), XAtlas achieves lower average conformal distortion than SeamGPT, so the overall average gain is driven by a few large wins. The paper does not report the seam-length ratio R or chart counts for any baseline, so the distortion metric, which rewards fragmentation, is not controlled. Please evaluate all methods with identical flattening post-processing and report chart counts or seam budgets alongside distortion.
  2. [Section 5, Eq. (4); Section 6 (seam-length control)] The evaluation metric is average conformal energy, which can be arbitrarily lowered by adding more cuts; a mesh cut along every edge would have near-zero distortion. The paper introduces R (seam segment count over vertex count) as an empirical operating point for SeamGPT but does not control this quantity for baselines. Consequently, the lower average distortion in Tables 1 and 2 does not, by itself, demonstrate semantically meaningful or artist-style cuts. Please report additional metrics (e.g., number of charts, total seam length, alignment with semantic boundaries) and perform a controlled comparison where all methods are given a comparable seam budget.
  3. [Section 3 (Data Preparation)] The training supervision is not validated. The paper assumes that UV-island boundaries in the filtered Objaverse/Objaverse-XL/3D-FUTURE meshes are artist-created seams with semantic intent. The filtering is described in one sentence ('removed problematic UV unwrappings such as those with overlapping UVs or arbitrarily placed seams lacking semantic meaning'), with no analysis of how many meshes were removed or why the remaining seams are artist-defined rather than auto-generated by a UV unwrapper. If a large fraction of Objaverse UVs are produced by automatic tools, the model learns auto-unwrapper artifacts, and the benchmark distortion numbers do not reveal this. Please provide evidence of annotation quality (e.g., manual inspection statistics, inter-annotator agreement, or a comparison against a model trained on explicitly hand-authored seam annotations).
  4. [Section 7 and Figures 7, 9] The part-segmentation claim is supported only by qualitative examples, including an acknowledged failure case (Figure 9, last row). The text states that the approach 'significantly improves the accuracy of part decomposition,' but no quantitative segmentation metric (e.g., mIoU, boundary F-score, or part-consistent accuracy) is reported, and there is no comparison with other boundary-refinement methods. Please add quantitative evaluation on a standard part-segmentation benchmark and report performance with and without SeamGPT guidance.
  5. [Section 6 (Ablation Study)] The ablations are entirely qualitative. Figures 4 and 5 show visual comparisons for point-cloud vs. uniform sampling, point-cloud vs. mesh encoder, and pointer networks, but no quantitative distortion or seam-alignment numbers are given. The claim that edge/vertex sampling 'significantly improves model convergence' is not supported by any learning-curve or convergence data. Please provide quantitative ablation results, ideally with the same evaluation protocol as the main experiments.
minor comments (6)
  1. [Eq. (1)] The point-level expansion of the seam sequence contains a typo: the last two terms are both written as pNh_t; the second one should be pNs_t (or pNs_h, pNs_t).
  2. [Section 5 (Training Strategy) vs. Appendix A.2] The training hyperparameters are inconsistent: the main text states a 2,000-step warm-up, total batch size 128, and 64 Nvidia H20 GPUs, while Appendix A.2 states a batch size of 2 and a warm-up of one step. Please reconcile these numbers so the training setup is reproducible.
  3. [References and notation] PartField is cited as [23] in Section 7 but as [53] in Section 1; reference [53] is NerVE, not PartField. The name 'Xatalas' should be 'XAtlas'. Please correct these citation and spelling errors.
  4. [Tables 1 and 2] No error bars, standard deviations, or number of repeated runs are reported for any quantitative result. Given the small benchmark sizes (13 + 8 meshes) and the stochastic nature of autoregressive sampling, the claimed improvement over XAtlas on average distortion should be accompanied by variance estimates.
  5. [Section 5, User study] The user study (Table 3) involves only 20 participants and 3 test meshes, with no statistical significance testing reported. Please report confidence intervals or a significance test for the differences in Boundary quality and Editability.
  6. [Section 6] The phrase 'Does Pointer networks works?' should be corrected, and the caption of Figure 6 should be a complete sentence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SeamGPT is an empirical supervised model with an external distortion metric; no prediction reduces to its training input by construction.

full rationale

The paper's central chain is: extract seams from UV island boundaries in filtered meshes, train an autoregressive transformer to predict quantized seam coordinates conditioned on a point cloud, then evaluate the resulting UV unwrapping with an external conformal distortion metric. No equation in the paper defines the predicted seams in terms of the reported distortion, and no fitted parameter is renamed as a prediction. The seam-length ratio R in [0.1, 0.35] is an operating-point choice, not a parameter fit to the benchmark metric. The point-cloud encoder from [40] and the hourglass decoder from [15] are reused components; citing them is standard architectural reuse and does not force the reported numbers. The claim that seams are 'artist-style' relies on the unverified assumption that UV island boundaries in the filtered data reflect artist intent, but that is a supervision-validity concern, not a circular derivation: the evaluation metric is independent of that assumption. The part-segmentation combination is a post-hoc vote over SeamGPT seams and PartField labels, and the paper even includes a failure case. Thus the derivation is self-contained against external benchmarks, and no circular step can be exhibited from the paper's own equations or citations.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper's contribution is a learned model, so the main 'free parameters' are design choices such as quantization, point budget, and seam-length range. The most consequential assumption is that UV-island boundaries in the curated training set are expert ground truth; the evaluation metric also ignores semantic quality, which the user study only partially covers.

free parameters (4)
  • quantization_bins = 1024 (10-bit)
    Chosen as a precision versus tractability tradeoff; directly limits seam coordinate resolution. Section 4.1.
  • seam_length_ratio_R = [0.1, 0.35]
    Empirically determined valid range for seam segment count relative to vertex count; used to control cutting granularity. Section 6.
  • point_count = 61,440 (30,720 per vertices/edges)
    Sampling budget chosen for shape conditioning; affects the model's ability to capture mesh topology. Section 4.2.
  • max_sequence_length = 27,000 tokens
    Truncation length for training stability; affects whether long seams are fully generated. Appendix A.2.
assumptions (4)
  • domain assumption UV-island boundaries in the filtered training meshes correspond to semantically meaningful artist seams.
    Ground-truth seam extraction in Section 3 treats all island boundaries as cutting seams after filtering; if the filter misses arbitrary seams, the supervision is noisy and the learned model inherits that noise.
  • domain assumption Cutting seams can be adequately represented as sequences of straight line segments between quantized vertex positions in a fixed yzx order.
    Section 4.1 assumes any seam can be encoded by this discrete representation; curved seams or non-vertex-aligned cuts are not representable, which may limit applicability to smooth organic shapes.
  • domain assumption Point clouds sampled on vertices and edges, condensed by the pretrained encoder from [40], retain enough shape information for the decoder to predict valid seams.
    Used in Section 4.2; if the encoder loses local edge structure, seams will misalign with the mesh and create extra faces when cut.
  • domain assumption The conformal energy metric (Eq. 4) is the right scalar for comparing UV unwrapping quality across methods.
    Section A.1; the metric only measures per-face distortion and ignores chart count, packing efficiency, or semantic coherence, even though the paper's motivation emphasizes semantic quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Auto-Regressive Surface Cutting." pith.science (2026). https://pith.science/paper/LXQNYJM4

@misc{pith2026250618017,
  author       = {Pith},
  title        = {Pith review of: Auto-Regressive Surface Cutting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LXQNYJM4}},
  note         = {Machine review of arXiv:2506.18017}
}
read the original abstract

Surface cutting is a fundamental task in computer graphics, with applications in UV parameterization, texture mapping, and mesh decomposition. However, existing methods often produce technically valid but overly fragmented atlases that lack semantic coherence. We introduce SeamGPT, an auto-regressive model that generates cutting seams by mimicking professional workflows. Our key technical innovation lies in formulating surface cutting as a next token prediction task: sample point clouds on mesh vertices and edges, encode them as shape conditions, and employ a GPT-style transformer to sequentially predict seam segments with quantized 3D coordinates. Our approach achieves exceptional performance on UV unwrapping benchmarks containing both manifold and non-manifold meshes, including artist-created, and 3D-scanned models. In addition, it enhances existing 3D segmentation tools by providing clean boundaries for part decomposition.

Figures

Figures reproduced from arXiv: 2506.18017 by the authors.

Figure 1
Figure 1. SeamGPT generates surfaces cutting seams, facilitating UV flatten and part decomposition. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. SeamGPT architecture: Point cloud encoder extracts shape context; Causal transformer [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative UV flatten results on FAM benchmark ( Nefertiti, Cow, and Fandisk). [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Ablation of point sampling strategy. Point cloud sampling strategy. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Ablation study of encoder and decoder. Mesh encoder vs Point cloud encoder. An alter￾native approach for generating shape embeddings employs mesh encoders, as demonstrated by Zhou et al. [52]. We implemented an encoder combining graph convolutions (operating on both ve…
Figure 6
Figure 6. Figure 6: Seam length control and diversity. We can control the cutting granularity by adjusting seam [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Qualitative part segmentation results. Our method enhances existing part segmentation [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Seam prediction and UV unwrapping results. The left column displays the mesh with [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: More part segmentation results. Last row shows the [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Hunyuan3D-Buffalo 1.0: A Unified Multimodal Model for Scalable 3D Generation, Understanding, and Editing

    cs.CV 2026-08 conditional novelty 6.0 of 10

    A unified 3D multimodal model combines understanding, text-to-3D generation, instruction-guided editing, and part generation in one architecture, trained on an 87M-sample corpus, with claimed state-of-the-art results.

Reference graph

Works this paper leans on

53 extracted references · 41 canonical work pages · cited by 1 Pith paper

  1. [1]

    Achiam, S

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 10 Figure 8: Seam prediction and UV unwrapping results. The left column displays the mesh with predicted seams overlaid, the middle column shows the predicted seam lin...

    arXiv 2023

  2. [2]

    Bazazian, J

    D. Bazazian, J. R. Casas, and J. Ruiz-Hidalgo. Fast and robust edge extraction in unorganized point clouds. In 2015 international conference on digital image computing: techniques and applications (DICTA), pages 1–8. IEEE, 2015

    work page 2015

  3. [3]

    Bazazian and M

    D. Bazazian and M. E. Parés. Edc-net: Edge detection capsule network for 3d point clouds. Applied Sciences, 11(4):1833, 2021

    work page 2021

  4. [4]

    N. A. Carr, J. Hoberock, K. Crane, and J. C. Hart. Rectangular multi-chart geometry images. In Symposium on geometry processing, pages 181–190, 2006

    work page 2006

  5. [5]

    J. I. Daniels, L. K. Ha, T. Ochotta, and C. T. Silva. Robust smooth feature extraction from point clouds. In IEEE International Conference on Shape Modeling and Applications 2007 (SMI’07) , pages 123–136. IEEE, 2007

    work page 2007

  6. [6]

    Deitke, R

    M. Deitke, R. Liu, M. Wallingford, H. Ngo, O. Michel, A. Kusupati, A. Fan, C. Laforte, V . V oleti, S. Y . Gadre, E. VanderBilt, A. Kembhavi, C. V ondrick, G. Gkioxari, K. Ehsani, L. Schmidt, and A. Farhadi. Objaverse-xl: A universe of 10m+ 3d objects. arXiv preprint arXiv:2307.05663, 2023

    arXiv 2023

  7. [7]

    Deitke, D

    M. Deitke, D. Schwenk, J. Salvador, L. Weihs, O. Michel, E. VanderBilt, L. Schmidt, K. Ehsani, A. Kembhavi, and A. Farhadi. Objaverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022

    arXiv 2022

  8. [8]

    H. Fu, R. Jia, L. Gao, M. Gong, B. Zhao, S. Maybank, and D. Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, pages 1–25, 2021

    work page 2021

Show all 53 references
  1. [9]

    G. Gori, A. Sheffer, N. Vining, E. Rosales, N. Carr, and T. Ju. Flowrep: Descriptive curve networks for free-form design shapes. ACM Transactions on Graphics (TOG) , 36(4):1–14, 2017

    2017

  2. [10]

    Groueix, M

    T. Groueix, M. Fisher, V . G. Kim, B. C. Russell, and M. Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216–224, 2018

    2018

  3. [11]

    X. Gu, S. J. Gortler, and H. Hoppe. Geometry images. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques , pages 355–361, 2002

    2002

  4. [12]

    Gumhold, X

    S. Gumhold, X. Wang, R. S. MacLeod, et al. Feature extraction from point clouds. In IMR, pages 293–305, 2001

    2001

  5. [13]

    Hackel, J

    T. Hackel, J. D. Wegner, and K. Schindler. Contour detection in unstructured 3d point clouds. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 1610–1618, 2016

    2016

  6. [14]

    J. Han, J. Liu, Y . Jiang, B. Yan, Y . Zhang, Z. Yuan, B. Peng, and X. Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis, 2024

    2024

  7. [15]

    Z. Hao, D. W. Romero, T.-Y . Lin, and M.-Y . Liu. Meshtron: High-fidelity, artist-like 3d mesh generation at scale. arXiv preprint arXiv:2412.09548, 2024

    2024 arXiv

  8. [16]

    Himeur, T

    C.-E. Himeur, T. Lejemble, T. Pellegrini, M. Paulin, L. Barthe, and N. Mellado. Pcednet: A lightweight neural network for fast and interactive edge detection in 3d point clouds. ACM Transactions on Graphics (TOG), 41(1):1–21, 2021

    2021

  9. [17]

    Kharevych, B

    L. Kharevych, B. Springborn, and P. Schröder. Discrete conformal mappings via circle patterns. ACM Transactions on Graphics (TOG), 25(2):412–438, 2006

    2006

  10. [18]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, et al. Segment anything. InProceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023

    2023

  11. [19]

    B. Lévy, S. Petitjean, N. Ray, and J. Maillot. Least squares conformal maps for automatic texture atlas generation. In Seminal Graphics Papers: Pushing the Boundaries, V olume 2, pages 193–202. Arxiv, 2023. 13

    2023

  12. [20]

    M. Li, D. M. Kaufman, V . G. Kim, J. Solomon, and A. Sheffer. Optcuts: Joint optimization of surface cuts and parameterization. ACM transactions on graphics (TOG), 37(6):1–13, 2018

    2018

  13. [21]

    Y . Lin, C. Wang, J. Cheng, B. Chen, F. Jia, Z. Chen, and J. Li. Line segment extraction for large scale unorganized point clouds. ISPRS Journal of Photogrammetry and Remote Sensing , 102:172–183, 2015

    2015

  14. [22]

    A. Liu, C. Lin, Y . Liu, X. Long, Z. Dou, H.-X. Guo, P. Luo, and W. Wang. Part123: part-aware 3d reconstruction from a single-view image. In ACM SIGGRAPH 2024 Conference Papers , pages 1–12, 2024

    2024

  15. [23]

    M. Liu, M. A. Uy, D. Xiang, H. Su, S. Fidler, N. Sharp, and J. Gao. Partfield: Learning 3d feature fields for part segmentation and beyond, 2025

    2025

  16. [24]

    Lucquin, S

    V . Lucquin, S. Deguy, and T. Boubekeur. Seamcut: Interactive mesh segmentation for parame- terization. In ACM SIGGRAPH 2017 Technical Briefs, 2017

    2017

  17. [25]

    X. Ma, Y . Liu, T. Gao, Q. Huang, and H. Huang. Clr-wire: Towards continuous latent representations for 3d curve wireframe generation. arXiv preprint arXiv:2504.19174, 2025

    2025 arXiv

  18. [26]

    Mitani and H

    J. Mitani and H. Suzuki. Making papercraft toys from meshes using strip-based approximate unfolding. ACM transactions on graphics (TOG), 23(3):259–263, 2004

    2004

  19. [27]

    C. Nash, Y . Ganin, S. A. Eslami, and P. Battaglia. Polygen: An autoregressive generative model of 3d meshes. In International conference on machine learning , pages 7220–7229. PMLR, 2020

    2020

  20. [28]

    K. Park, U. Sinha, J. T. Barron, S. Bouaziz, D. B. Goldman, S. M. Seitz, and R. Martin-Brualla. Nerfies: Deformable neural radiance fields. ICCV, 2021

    2021

  21. [29]

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660, 2017

    2017

  22. [30]

    P. V . Sander, Z. J. Wood, S. Gortler, J. Snyder, and H. Hoppe. Multi-chart geometry images. arxiv, 2003

    2003

  23. [31]

    Sharp and K

    N. Sharp and K. Crane. Variational surface cutting. ACM Trans. Graph., 37(4), 2018

    2018

  24. [32]

    Sheffer and J

    A. Sheffer and J. C. Hart. Seamster: inconspicuous low-distortion texture seam layout. In IEEE Visualization, 2002. VIS 2002., pages 291–298. IEEE, 2002

    2002

  25. [33]

    Siddiqui, A

    Y . Siddiqui, A. Alliegro, A. Artemov, T. Tommasi, D. Sirigatti, V . Rosov, A. Dai, and M. Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. arXiv preprint arXiv:2311.15475, 2023

    2023 arXiv

  26. [34]

    Soliman, D

    Y . Soliman, D. Slepˇcev, and K. Crane. Optimal cone singularities for conformal flattening. ACM Transactions on Graphics (TOG), 37(4):1–17, 2018

    2018

  27. [35]

    Sorkine, D

    O. Sorkine, D. Cohen-Or, R. Goldenthal, and D. Lischinski. Bounded-distortion piecewise mesh parameterization. In IEEE Visualization, 2002. VIS 2002. , pages 355–362. IEEE, 2002

    2002

  28. [36]

    Springborn, P

    B. Springborn, P. Schröder, and U. Pinkall. Conformal equivalence of triangle meshes. In ACM SIGGRAPH 2008 papers, pages 1–11. arxiv, 2008

    2008

  29. [37]

    P. P. Srinivasan, S. J. Garbin, D. Verbin, J. T. Barron, and B. Mildenhall. Nuvo: Neural uv mapping for unruly 3d representations. In European Conference on Computer Vision, pages 18–34. Springer, 2024

    2024

  30. [38]

    Stojanov, A

    S. Stojanov, A. Thai, and J. M. Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. CVPR, 2021

    2021

  31. [39]

    C. Tang, P. Bo, J. Wallner, and H. Pottmann. Interactive design of developable surfaces.ACM Transactions on Graphics (TOG), 35(2):1–12, 2016. 14

    2016

  32. [40]

    T. H. Team. Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation, 2025

    2025

  33. [41]

    Vinyals, M

    O. Vinyals, M. Fortunato, and N. Jaitly. Pointer networks. Advances in neural information processing systems, 28, 2015

    2015

  34. [42]

    X. Wang, Y . Xu, K. Xu, A. Tagliasacchi, B. Zhou, A. Mahdavi-Amiri, and H. Zhang. Pie-net: Parametric inference of point cloud edges. Advances in neural information processing systems , 33:20167–20178, 2020

    2020

  35. [43]

    Y . Wang, Y . Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, and J. M. Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1–12, 2019

    2019

  36. [44]

    H. Weng, Z. Zhao, B. Lei, X. Yang, J. Liu, Z. Lai, Z. Chen, Y . Liu, J. Jiang, C. Guo, T. Zhang, S. Gao, and C. L. P. Chen. Scaling mesh generation via compressive tokenization. arXiv preprint arXiv:2411.07025, 2024

    2024 arXiv

  37. [45]

    Yamauchi, S

    H. Yamauchi, S. Gumhold, R. Zayer, and H.-P. Seidel. Mesh segmentation driven by gaussian curvature. The Visual Computer, 21:659–668, 2005

    2005

  38. [46]

    Yang, Y .-C

    Y . Yang, Y .-C. Guo, Y . Huang, Z.-X. Zou, Z. Yu, Y . Li, Y .-P. Cao, and X. Liu. Holopart: Generative 3d part amodal segmentation. arXiv preprint arXiv:2504.07943, 2025

    2025 arXiv

  39. [47]

    Y . Yang, Y . Huang, Y .-C. Guo, L. Lu, X. Wu, E. Y . Lam, Y .-P. Cao, and X. Liu. Sampart3d: Segment any part in 3d objects. arXiv preprint arXiv:2411.07184, 2024

    2024 arXiv

  40. [48]

    J. Young. xatlas - mesh parameterization / uv unwrapping library, 2024

    2024

  41. [49]

    L. Yu, X. Li, C.-W. Fu, D. Cohen-Or, and P.-A. Heng. Ec-net: an edge-aware point set consolidation network. In Proceedings of the European conference on computer vision (ECCV) , pages 386–402, 2018

    2018

  42. [50]

    Zhang, J

    Q. Zhang, J. Hou, W. Wang, and Y . He. Flatten anything: Unsupervised neural surface parameterization. arXiv preprint arXiv:2405.14633, 2024

    2024 arXiv

  43. [51]

    K. Zhou, J. Synder, B. Guo, and H.-Y . Shum. Iso-charts: stretch-driven mesh parameterization using spectral analysis. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH symposium on Geometry processing, pages 45–54, 2004

    2004

  44. [52]

    Y . Zhou, C. Wu, Z. Li, C. Cao, Y . Ye, J. Saragih, H. Li, and Y . Sheikh. Fully convolutional mesh autoencoder using efficient spatially varying kernels. Advances in neural information processing systems, 33:9251–9262, 2020

    2020

  45. [53]

    X. Zhu, D. Du, W. Chen, Z. Zhao, Y . Nie, and X. Han. Nerve: Neural volumetric edges for parametric curve extraction from point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13601–13610, 2023. 15

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.