pith. sign in

arxiv: 2606.27088 · v1 · pith:RGFR74TYnew · submitted 2026-06-25 · 💻 cs.CV

SubdivAR: Autoregressive Next-Scale Prediction for Neural Mesh Subdivision

Huipeng Guo , Zikai Song , Hang Long , Jielei Zhang , Wenbing Li , Junkai Lin , Tianhao Zhao , Jinshen Zhang
show 2 more authors
Tianle Guo Wei Yang
This is my paper

Pith reviewed 2026-06-26 04:59 UTC · model grok-4.3

classification 💻 cs.CV
keywords neural mesh subdivisionautoregressive predictionmesh autoregressive representationtopology-aware transformervertex offset regressionsubdivision datasetgeometric detail synthesis3D surface refinement
0
0 comments X

The pith

SubdivAR reformulates mesh subdivision as autoregressive next-scale prediction to recover fine details while preserving topology.

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

The paper tries to establish that modeling subdivision levels as an ordered sequence allows an autoregressive neural model to generate higher-resolution surfaces more accurately than local refinement approaches. A sympathetic reader would care because classical and prior neural methods often smooth out details or fail on complex shapes, limiting use in asset creation. The core move is to treat each finer level as the next item to predict from the coarser one, using a transformer that mixes global context with topology rules. This is enabled by a new dataset providing supervision at multiple levels. If the claim holds, subdivision becomes a predictable sequence task rather than a fixed local rule set.

Core claim

SubdivAR introduces Mesh Autoregressive Representation (MAR) that arranges meshes at different subdivision levels into an ordered scale sequence, reformulating the task as autoregressive next-scale prediction. A Hybrid Topology-Aware Transformer combines global semantic attention with topology-constrained local feature aggregation to regress vertex offsets at each refinement stage. This preserves the subdivision topology while recovering fine-grained geometric details. The approach is trained on the FII-40K dataset of nearly 40,000 high-quality meshes with multi-level supervision and is reported to reduce Hausdorff Distance by 18.8% and Chamfer Distance by 14.2% over baselines while showing

What carries the argument

Mesh Autoregressive Representation (MAR), which sequences meshes across subdivision levels to enable next-scale coordinate prediction.

If this is right

  • The subdivision topology remains fixed while vertex offsets add geometric detail at each scale.
  • Global semantic attention and local topology constraints operate together inside the Hybrid Topology-Aware Transformer.
  • Training uses explicit multi-level subdivision supervision from the FII-40K dataset.
  • Reported error reductions reach 18.8% on Hausdorff Distance and 14.2% on Chamfer Distance relative to prior methods.
  • Performance remains stable on complex open-surface geometries.

Where Pith is reading between the lines

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

  • The sequencing idea could transfer to related tasks such as point-cloud densification or hierarchical surface reconstruction.
  • Coarse editable meshes could serve as controllable starting points inside larger text-to-3D pipelines.
  • Performance may depend on whether new meshes follow subdivision rules similar to those used when curating FII-40K.
  • Autoregressive next-scale prediction might extend to time-varying or animated mesh sequences.

Load-bearing premise

The FII-40K dataset of nearly 40,000 meshes supplies training signals diverse enough for the autoregressive model to generalize to unseen meshes and hierarchies.

What would settle it

Run SubdivAR on a fresh collection of meshes drawn from sources or topologies absent from FII-40K and measure whether the reported reductions in Hausdorff and Chamfer distances still hold.

Figures

Figures reproduced from arXiv: 2606.27088 by Hang Long, Huipeng Guo, Jielei Zhang, Jinshen Zhang, Junkai Lin, Tianhao Zhao, Tianle Guo, Wei Yang, Wenbing Li, Zikai Song.

Figure 1
Figure 1. Figure 1: High-fidelity mesh subdivision achieved by SubdivAR. Given a coarse input mesh (left, green), our model produce refined surfaces with faithful semantic and geometric details. Abstract Mesh subdivision is a fundamental operation for converting coarse, editable meshes into high-resolution surfaces, with broad applications in digital asset creation. Clas￾sical rule-based schemes rely on fixed local refinement… view at source ↗
Figure 2
Figure 2. Figure 2: Sharp Feature Collapse. The shape distortion is attributed to the fail￾ure in preserving sharp features during reparameterization. Neural mesh subdivision is limited by the scarcity of paired coarse-fine mesh data. To address this issue, we con￾struct the FII-40K dataset, which contains approximately 40,000 meshes with multi-level subdivision supervision. We begin by following the pipeline of Neural Subdiv… view at source ↗
Figure 3
Figure 3. Figure 3: Failure Modes. Comparing the raw mesh (left) with the reparameterized result (right), we identify two primary failure modes: significant shape distortion with a loss of sharp features( [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The pipeline of our SubdivAR framework. After careful filtering and validation, we construct FII-40K, a curated dataset of high-quality coarse￾fine mesh pairs that provides reliable supervision for training neural mesh subdivision models. 3.2 Mesh Subdivision through Next Scale Prediction Given a datapoint from FII-40K, we construct a unified sequence X by concatenating vertices across all subdivision leve… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison We demonstrate the subdivision performance of different models on the same coarse meshes. Our model leverages global information to successfully recover fine details(highlighted in the boxes). While previous methods often misinterpret the subdivision structure, leading to over-smoothing or mesh artifacts, our approach effectively overcomes these issues. 4.1 Experimental Setup Dataset… view at source ↗
Figure 6
Figure 6. Figure 6: Ablation studies on different components. (a) Normal loss: The additional normal loss [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation on Mode. All mode achieves su￾perior reconstruction quality, effectively recovering the complex structure of the robot’s lateral auricle com￾pared to other configurations on 80k scale. Hierarchical Modes and Data Scaling. Table 3c investigates the synergy between the hierarchical receptive field and train￾ing data volume. The first_cross and previ￾ous_cross modes utilize the sampling-point self-at… view at source ↗
Figure 8
Figure 8. Figure 8: Metric Distribution We visualize the metric distributions for both successful and failed samples. Because failed samples are selected based on subpar performance in specific dimensions, their distributions inevitably overlap with those of the successful ones. Furthermore, the failed samples demonstrate a broader distribution with prominent long-tail effects. A.3 Filtering To further guarantee geometric fid… view at source ↗
Figure 9
Figure 9. Figure 9: Test on coarse mesh on generated by Tripop1.0 The primary failure mode for several baseline models, such as NS [22] and NMR [52], stems from their inherent inability to process meshes with open boundaries. To facilitate a visual comparison on such geometries, we manually pre￾processed these meshes by stitching all boundary vertices to a singular auxiliary point. Furthermore, our model yields encouraging pr… view at source ↗
Figure 10
Figure 10. Figure 10: More Qualitative comparison We demonstrate the subdivision performance of different models on the same coarse meshes. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Failure cases Our method fails to recover sufficient geometric details in these examples. The missing details are highlighted with red boxes [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: More results 18 [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: More results [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: More results 19 [PITH_FULL_IMAGE:figures/full_fig_p019_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: More results [PITH_FULL_IMAGE:figures/full_fig_p020_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: More results 20 [PITH_FULL_IMAGE:figures/full_fig_p020_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: More results NeurIPS Paper Checklist 1. Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: The main claims in the abstract and introduction are well aligned with the contributions of the paper. Specifically, the abstract clearly states that the paper introduces SubdivAR, a topology-aware mesh refine… view at source ↗
read the original abstract

Mesh subdivision is a fundamental operation for converting coarse, editable meshes into high-resolution surfaces, with broad applications in digital asset creation. Classical rule-based schemes rely on fixed local refinement rules and often produce over-smoothed surfaces. Recent neural subdivision methods improve detail synthesis, but remain constrained by local modeling and exhibit limited generalizability. We present SubdivAR, a neural mesh subdivision framework based on our proposed Mesh Autoregressive Representation (MAR). MAR arranges meshes at different subdivision levels into an ordered scale sequence, reformulating subdivision as autoregressive next-scale prediction. To support this formulation, we introduce a Hybrid Topology-Aware Transformer that combines global semantic attention with topology-constrained local feature aggregation. SubdivAR adopts a next-scale coordinate prediction paradigm, regressing vertex offsets at each refinement stage to preserve subdivision topology while recovering fine-grained geometric details. To enable reliable learning, we construct FII-40K, a curated dataset of nearly 40,000 high-quality meshes with multi-level subdivision supervision. Experiments show that SubdivAR outperforms state-of-the-art baselines, reducing Hausdorff Distance and Chamfer Distance by 18.8% and 14.2%, respectively, and demonstrates strong robustness on complex open-surface geometries.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The manuscript introduces SubdivAR, a neural mesh subdivision method based on Mesh Autoregressive Representation (MAR) that reformulates subdivision as autoregressive next-scale prediction. It employs a Hybrid Topology-Aware Transformer combining global semantic attention with topology-constrained local aggregation, adopts next-scale coordinate prediction for vertex offsets, and constructs the FII-40K dataset of nearly 40,000 meshes with multi-level subdivision supervision. Experiments claim SubdivAR outperforms state-of-the-art baselines by reducing Hausdorff Distance by 18.8% and Chamfer Distance by 14.2%, with robustness on complex open-surface geometries.

Significance. If the empirical results are validated with transparent dataset protocols and retrained baselines, the autoregressive next-scale formulation could meaningfully advance neural subdivision beyond local modeling constraints, offering improved detail recovery and generalizability. The large-scale FII-40K dataset with explicit multi-level supervision would also constitute a reusable resource for the community.

major comments (1)
  1. [Abstract] Abstract: The headline performance claims (18.8% Hausdorff Distance and 14.2% Chamfer Distance reductions) are obtained exclusively on the newly constructed FII-40K dataset. No information is supplied on curation criteria, category balance, train/test split construction, or confirmation that baselines were retrained on identical data and subdivision levels; without these details the attribution of gains to the MAR + Hybrid Transformer architecture cannot be verified and the central empirical claim remains load-bearing.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for greater transparency on the FII-40K dataset and experimental protocol. We address this point directly below and will incorporate the requested details in the revision.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The headline performance claims (18.8% Hausdorff Distance and 14.2% Chamfer Distance reductions) are obtained exclusively on the newly constructed FII-40K dataset. No information is supplied on curation criteria, category balance, train/test split construction, or confirmation that baselines were retrained on identical data and subdivision levels; without these details the attribution of gains to the MAR + Hybrid Transformer architecture cannot be verified and the central empirical claim remains load-bearing.

    Authors: We agree that the current manuscript does not provide sufficient detail on FII-40K construction and baseline retraining within the abstract (and, upon re-examination, the Experiments section also lacks explicit statements on these points). In the revised version we will add a dedicated subsection under Experiments that specifies: (i) curation criteria used to assemble the ~40K meshes, (ii) category distribution and balance, (iii) exact train/test split methodology, and (iv) confirmation that every baseline was retrained from scratch on the identical FII-40K data and multi-level subdivision targets. These additions will allow readers to directly attribute performance differences to the MAR formulation and Hybrid Topology-Aware Transformer. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical results on new dataset

full rationale

The paper's central claims consist of an architectural reformulation (MAR as autoregressive next-scale prediction) and reported empirical gains (18.8% Hausdorff, 14.2% Chamfer reduction) on the newly constructed FII-40K dataset. No equations, fitted parameters presented as predictions, or load-bearing self-citations appear in the provided text. The performance numbers are external benchmarks against baselines rather than quantities forced by construction from the inputs. This is a standard empirical ML contribution with no detectable reduction of the claimed results to the method's own definitions.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract supplies no explicit free parameters, axioms, or invented entities beyond the high-level claim that the new representation and transformer enable the reported gains; all modeling choices remain opaque.

pith-pipeline@v0.9.1-grok · 5776 in / 1153 out tokens · 22323 ms · 2026-06-26T04:59:45.032148+00:00 · methodology

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

58 extracted references · 5 canonical work pages

  1. [1]

    Recursively generated b-spline surfaces on arbitrary topo- logical meshes

    Edwin Catmull and James Clark. Recursively generated b-spline surfaces on arbitrary topo- logical meshes. InSeminal graphics: pioneering efforts that shaped the field, pages 183–188. 1998

    1998

  2. [2]

    Shapenet: An information-rich 3d model repository.arXiv preprint arXiv:1512.03012, 2015

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zekun Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository.arXiv preprint arXiv:1512.03012, 2015

    Pith/arXiv arXiv 2015

  3. [3]

    Triangular mesh surface subdivision based on graph neural network.Applied Sciences, 14(23), 2024

    Guojun Chen and Rongji Wang. Triangular mesh surface subdivision based on graph neural network.Applied Sciences, 14(23), 2024. ISSN 2076-3417. doi: 10.3390/app142311378

    work page doi:10.3390/app142311378 2024

  4. [4]

    Activating more pixels in image super-resolution transformer

    Xiangyu Chen, Xintao Wang, Jiantao Wen, Yu Qiao Zhou, Chao Dong Xiang, and Li Cheng. Activating more pixels in image super-resolution transformer. InCVPR, 2023

    2023

  5. [5]

    In: Proceedings of the ACM Special Interest Group on Computer Graphics and Interactive Techniques (SIGGRAPH) (August 2023).https://doi

    Yun-Chun Chen, Vladimir Kim, Noam Aigerman, and Alec Jacobson. Neural progressive meshes. SIGGRAPH ’23, New York, NY , USA, 2023. Association for Computing Machinery. ISBN 9798400701597. doi: 10.1145/3588432.3591531

    work page doi:10.1145/3588432.3591531 2023

  6. [6]

    Metro: measuring error on simplified surfaces.Computer Graphics Forum, 17(2):167–174, 1998

    Paolo Cignoni, Claudio Rocchini, and Roberto Scopigno. Metro: measuring error on simplified surfaces.Computer Graphics Forum, 17(2):167–174, 1998

    1998

  7. [7]

    Objaverse: A universe of annotated 3d objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142–13153, 2023

    2023

  8. [8]

    Behavior of recursive division surfaces near extraordinary points.Computer-Aided Design, 10(6):356–360, 1978

    Daniel Doo and Malcolm Sabin. Behavior of recursive division surfaces near extraordinary points.Computer-Aided Design, 10(6):356–360, 1978

    1978

  9. [9]

    A point set generation network for 3d object reconstruction from a single image

    Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 605–613, 2017

    2017

  10. [10]

    3d-future: 3d furniture shape with texture.International Journal of Computer Vision, 129 (12):3313–3337, 2021

    Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture.International Journal of Computer Vision, 129 (12):3313–3337, 2021

    2021

  11. [11]

    Surface simplification using quadric error metrics

    Michael Garland and Paul S Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 209–216, 1997

    1997

  12. [12]

    Spiralnet++: A fast and highly efficient mesh convolution operator

    Shunwang Gong et al. Spiralnet++: A fast and highly efficient mesh convolution operator. In ICCVW, 2019

    2019

  13. [13]

    Meshcnn: a network with an edge.ACM Transactions on Graphics (ToG), 38(4):1–12, 2019

    Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. Meshcnn: a network with an edge.ACM Transactions on Graphics (ToG), 38(4):1–12, 2019

    2019

  14. [14]

    Piecewise smooth surface reconstruction

    Hugues Hoppe, Tony DeRose, Tom Duchamp, Mark Halstead, Hubert Jin, John McDonald, Jean Schweitzer, and Werner Stuetzle. Piecewise smooth surface reconstruction. InProceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 295–302, 1994

    1994

  15. [15]

    Subdivision-based mesh convolution networks.ACM Transactions on Graphics (TOG), 41(3):1–16, 2022

    Shi-Min Hu, Zheng-Ning Liu, Meng-Hao Guo, Jun-Xiong Cai, Jiahui Huang, Tai-Jiang Mu, and Ralph R Martin. Subdivision-based mesh convolution networks.ACM Transactions on Graphics (TOG), 41(3):1–16, 2022

    2022

  16. [16]

    √ 3-subdivision

    Leif Kobbelt. √ 3-subdivision. InProceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 103–112, 2000

    2000

  17. [17]

    Armesh: Autoregressive mesh generation via next-level-of-detail prediction, 2025

    Jiabao Lei, Kewei Shi, Zhihao Liang, and Kui Jia. Armesh: Autoregressive mesh generation via next-level-of-detail prediction, 2025. URLhttps://arxiv.org/abs/2509.20824. 10

    arXiv 2025

  18. [18]

    Deeper insights into graph convolutional networks for semi-supervised learning, 2018

    Qimai Li, Zhichao Han, and Xiao-Ming Wu. Deeper insights into graph convolutional networks for semi-supervised learning, 2018. URLhttps://arxiv.org/abs/1801.07606

    Pith/arXiv arXiv 2018

  19. [19]

    Triposg: High-fidelity 3d shape synthesis using large-scale rectified flow models.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Yangguang Li, Zi-Xin Zou, Zexiang Liu, Dehu Wang, Yuan Liang, Zhipeng Yu, Xingchao Liu, Yuan-Chen Guo, Ding Liang, Wanli Ouyang, et al. Triposg: High-fidelity 3d shape synthesis using large-scale rectified flow models.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  20. [20]

    Swinir: Image restoration using swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 1833–1844, October 2021

    2021

  21. [21]

    Meshripple: Structured autoregressive generation of artist-meshes.arXiv preprint arXiv:2512.07514, 2025

    Junkai Lin, Hang Long, Huipeng Guo, Jielei Zhang, JiaYi Yang, Tianle Guo, Yang Yang, Jianwen Li, Wenxiao Zhang, Matthias Nießner, et al. Meshripple: Structured autoregressive generation of artist-meshes.arXiv preprint arXiv:2512.07514, 2025

    arXiv 2025

  22. [22]

    Neural subdivision.arXiv preprint arXiv:2005.01819, 2020

    Hsueh-Ti Derek Liu, Vladimir G Kim, Siddhartha Chaudhuri, Noam Aigerman, and Alec Jacobson. Neural subdivision.arXiv preprint arXiv:2005.01819, 2020

    arXiv 2005

  23. [23]

    Implicit neural distance optimization for mesh neural subdivision

    Ke Liu, Ning Ma, Zhihua Wang, Jingjun Gu, Jiajun Bu, and Haishuai Wang. Implicit neural distance optimization for mesh neural subdivision. In2023 IEEE International Conference on Multimedia and Expo (ICME), pages 2039–2044, 2023. doi: 10.1109/ICME55011.2023.00349

    work page doi:10.1109/icme55011.2023.00349 2039

  24. [24]

    Smooth subdivision surfaces based on triangles

    Charles Loop. Smooth subdivision surfaces based on triangles. 1987

    1987

  25. [25]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InInternational Conference on Learning Representations, 2017

    2017

  26. [26]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. URL https://arxiv.org/abs/2003.08934

    arXiv 2020

  27. [27]

    Geometric deep learning on graphs and manifolds using mixture model cnns

    Federico Monti et al. Geometric deep learning on graphs and manifolds using mixture model cnns. InCVPR, 2017

    2017

  28. [28]

    Deepsdf: Learning continuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165–174, 2019

    2019

  29. [29]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. InAdvances in neural information processing systems, 2019

    2019

  30. [30]

    Convolutional occupancy networks, 2020

    Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks, 2020. URLhttps://arxiv.org/abs/2003.04618

    arXiv 2020

  31. [31]

    Learning mesh-based simulation with graph networks

    Tobias Pfaff et al. Learning mesh-based simulation with graph networks. InICLR, 2021

    2021

  32. [32]

    Wandt, J

    Rolandos Alexandros Potamias, Stylianos Ploumpis, and Stefanos Zafeiriou. Neural mesh simplification. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18562–18571, 2022. doi: 10.1109/CVPR52688.2022.01803

    work page doi:10.1109/cvpr52688.2022.01803 2022

  33. [33]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660, 2017

    2017

  34. [34]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space.Advances in neural information processing systems, 30, 2017

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space.Advances in neural information processing systems, 30, 2017

    2017

  35. [35]

    Pu-gcn: Point cloud upsampling using graph convolutional networks

    Guocheng Qian, Abdulellah Abualshour, Guohao Li, Ali Thabet, and Bernard Ghanem. Pu-gcn: Point cloud upsampling using graph convolutional networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11683–11692, 2021. 11

    2021

  36. [36]

    Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis.Advances in Neural Information Processing Systems, 34:6087–6101, 2021

    Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis.Advances in Neural Information Processing Systems, 34:6087–6101, 2021

    2021

  37. [37]

    Meshgpt: Generating triangle meshes with decoder-only transformers

    Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19615–19625, 2024

    2024

  38. [38]

    Using shape to categorize: Low-shot learning with an explicit shape bias

    Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1798–1808, 2021

    2021

  39. [39]

    Structured 3d latents for scalable and versatile generation.arXiv:2412.01506, 2024

    Jiaming Sun et al. Structured 3d latents for scalable and versatile generation.arXiv:2412.01506, 2024

    Pith/arXiv arXiv 2024

  40. [40]

    Visual autoregressive mod- eling: Scalable image generation via next-scale prediction.arXiv preprint arXiv:2404.02905, 2024

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive mod- eling: Scalable image generation via next-scale prediction.arXiv preprint arXiv:2404.02905, 2024

    arXiv 2024

  41. [41]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InNeurIPS, 2017

    2017

  42. [42]

    Feastnet: Feature-steered graph convolutions for 3d shape analysis

    Nitika Verma et al. Feastnet: Feature-steered graph convolutions for 3d shape analysis. In CVPR, 2018

    2018

  43. [43]

    Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models.arXiv:2404.07191, 2024

    Jiale Xu et al. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models.arXiv:2404.07191, 2024

    Pith/arXiv arXiv 2024

  44. [44]

    Cwf: Consolidating weak features in high-quality mesh simplification.ACM Transactions on Graphics (TOG), 43(4), 2024

    Rui Xu, Longdu Liu, Ningna Wang, Shuangmin Chen, Shiqing Xin, Xiaohu Guo, Zichun Zhong, Taku Komura, Wenping Wang, and Changhe Tu. Cwf: Consolidating weak features in high-quality mesh simplification.ACM Transactions on Graphics (TOG), 43(4), 2024. ISSN 0730-0301. doi: 10.1145/3658159. URLhttps://doi.org/10.1145/3658159

    work page doi:10.1145/3658159 2024

  45. [45]

    Pu-net: Point cloud upsampling network

    Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-net: Point cloud upsampling network. InProceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  46. [46]

    3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models.ACM Transactions On Graphics (TOG), 42(4):1–16, 2023

    Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models.ACM Transactions On Graphics (TOG), 42(4):1–16, 2023

    2023

  47. [47]

    Clay: A controllable large-scale generative model for creating high- quality 3d assets.ACM Trans

    Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high- quality 3d assets.ACM Trans. Graph., 43(4), July 2024. ISSN 0730-0301. doi: 10.1145/ 3658146

    2024

  48. [48]

    Torr, and Vladlen Koltun

    Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H.S. Torr, and Vladlen Koltun. Point transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16259–16268, October 2021

    2021

  49. [49]

    Lato: 3d mesh flow matching with structured topology preserving latents.arXiv preprint arXiv:2603.06357, 2026

    Tianhao Zhao, Youjia Zhang, Hang Long, Jinshen Zhang, Wenbing Li, Yang Yang, Gongbo Zhang, Jozef Hladk `y, Matthias Nießner, and Wei Yang. Lato: 3d mesh flow matching with structured topology preserving latents.arXiv preprint arXiv:2603.06357, 2026

    arXiv 2026

  50. [50]

    Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation.arXiv preprint arXiv:2501.12202, 2025

    Zibo Zhao, Zeqiang Lai, Qingxiang Lin, Yunfei Zhao, Haolin Liu, Shuhui Yang, Yifei Feng, Mingxin Yang, Sheng Zhang, Xianghui Yang, et al. Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation.arXiv preprint arXiv:2501.12202, 2025

    Pith/arXiv arXiv 2025

  51. [51]

    Pymesh-geometry processing library for python.GitHub repository, 2018

    Qingnan Zhou and Eitan Grinspun. Pymesh-geometry processing library for python.GitHub repository, 2018

    2018

  52. [52]

    Neural mesh refinement.Frontiers of Information Technology & Electronic Engineering, 26(5):695–712, 2025

    Zhiwei Zhu, Xiang Gao, Lu Yu, and Yiyi Liao. Neural mesh refinement.Frontiers of Information Technology & Electronic Engineering, 26(5):695–712, 2025. 12

    2025

  53. [53]

    Interpolating subdivision for meshes with arbitrary topology

    Denis Zorin, Peter Schröder, and Wim Sweldens. Interpolating subdivision for meshes with arbitrary topology. InProceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 189–192, 1996. A Dataset A.1 Data Collection We collect meshes from several public datasets and apply a unified preprocessing and filtering pipeline ...

    1996

  54. [54]

    Models >40,000 faces are simpli- fied via QEM (Garland and Heckbert[11])

    Decimation and Cleaning:Filter meshes with <2,500 faces. Models >40,000 faces are simpli- fied via QEM (Garland and Heckbert[11]). We per- form cleaning and repair non-manifold structures using standard mesh processing libraries (Zhou and Grinspun [51])

  55. [55]

    Orientation and Normalization:Face orienta- tions are corrected using view-dependent visibility methods(algorithm 1) and meshes are normalized into a[−1,1]canonical space

  56. [56]

    [22]) to a coarse resolution of 750–850 faces and re- parametrization

    Subdivision Generation:Cleaned meshes un- dergo stochastic decimation (Liu et al. [22]) to a coarse resolution of 750–850 faces and re- parametrization. Algorithm 1Mesh Normal Orientation Correction Require:verticesV, facesF Ensure:corrected mesh(V ′, F ′) 1:(F ′, C)←BFSOrient(F) 2:V ′ ←Normalize(V) 3:M←(V ′, F ′) 4:Sample cameras{c i}Nc i=1 on sphere 5:f...

    arXiv

  57. [57]

    Limitations

    and LATO [ 49] attempt to decode structured meshes from high-resolution voxel features. While promising, these methods currently face significant hurdles, including incomplete modeling, non-manifold holes, and the loss of fine-grained details, which hinder their practical adoption. Alternatively, the autoregressive approach, exemplified by AR-Mesh [ 17], ...

    arXiv

  58. [58]

    Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...