Pith. sign in

REVIEW 4 major objections 5 minor 47 references

Meshy T2 generates a mesh from an image in about 6 seconds by modeling one latent token per vertex and decoding geometry and topology in a single parallel pass.

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 ↗

T0 review · deepseek-v4-flash

2026-08-03 00:43 UTC pith:O7ZKTHSJ

load-bearing objection A serious method contribution with a central evaluation gap: the 'preserves artist-authored topology' claim is only tested against the authors' own pipeline's output — still deserves peer review. the 4 major comments →

arxiv 2607.28675 v1 pith:O7ZKTHSJ submitted 2026-07-28 cs.GR cs.CV

Meshy T2: Fast Native Mesh Generation with Flow Matching

Jiale Xu , Rendong Liang , Yuhao Long , Siyuan Shen , Zangyueyang Xian , Zeyi Xu , Yuanming Hu This is my paper
classification cs.GR cs.CV
keywords mesh generationflow matchingmesh VAEimage-to-3Dtopology generationper-vertex latent tokensretopologyface budget control
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.

Meshy T2 is a two-stage flow-matching system that turns a single image into a compact, artist-style mesh with a median latency of 6 seconds. The core claim is that a mesh is better modeled as an unordered set of continuous per-vertex latent tokens than as a serialized token sequence, because then all vertices, edges, and face orientations can be decoded in one parallel pass instead of being predicted one token at a time. The paper argues this vertex-set mesh VAE is nearly lossless: coordinates stay continuous, coincident vertices are not welded, and the artist-authored connectivity is preserved exactly, so the decoded mesh keeps high-precision detail and naturally separates into connected components. That design yields three capabilities — interactive speed, face-count control through a requested vertex budget, and direct support for multi-part assets — and the reported experiments show state-of-the-art geometric fidelity with a 3-second median retopology time.

Core claim

The paper claims that a mesh can be encoded as a set of continuous per-vertex latent tokens, with edge connectivity and face winding order decoded jointly from those tokens, and that this representation makes it possible to generate meshes by flow matching instead of autoregressive token prediction. The vertex-set mesh VAE avoids vertex quantization and vertex welding, so geometric precision and artist-authored topology are preserved. Generation proceeds as a two-stage cascade: a voxel-flow model first produces a coarse occupancy scaffold from the image, then a mesh-flow model populates the scaffold with per-vertex latent tokens under a requested vertex budget. The result is an end-to-end im

What carries the argument

The load-bearing mechanism is the vertex-set mesh VAE: one continuous latent token per vertex, with the decoder predicting vertex positions, an undirected edge set via a spacetime adjacency score, and per-vertex halfedge successor permutations that encode oriented faces exactly on manifold meshes (with a NULL element for open boundaries). A second mechanism is the assignment of spatial positions to unordered latent tokens through optimal transport onto a fixed low-discrepancy point set, which removes permutation ambiguity and lets the transformer use 3D rotary position encodings.

Load-bearing premise

The exact-preservation claim assumes the input meshes are manifold and consistently oriented, because the halfedge successor representation is defined only for manifold meshes; non-manifold meshes are repaired by splitting edges and vertices, and the paper does not report how often this rewrites the artist-authored topology it claims to preserve.

What would settle it

Take a large corpus of real artist meshes, apply the paper's repair procedure, and count how many meshes change topology; then compare the VAE's decoded connectivity to the original on repaired and unrepaired subsets. If a substantial fraction of assets are non-manifold and repair changes their component or edge structure, the 'preserved exactly' claim fails on that subset.

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

If this is right

  • Image-to-mesh generation completes in a median of 6 seconds, making interactive asset creation practical.
  • The requested vertex budget controls final face count (about 2× vertices for closed meshes) without post-hoc simplification.
  • Multi-part assets emerge from the generated vertex–edge graph, avoiding separate part-wise generation or stitching.
  • Geometry and topology are produced in a single parallel decoding pass, avoiding sequential error accumulation of token-by-token methods.
  • Retopology of a dense mesh can be done in about 3 seconds with low non-manifold edge counts.

Where Pith is reading between the lines

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

  • The same near-lossless vertex-set representation could serve as a general mesh codec for compression, editing, and level-of-detail generation, not just image-conditioned creation.
  • The optimal-transport positional encoding for unordered token sets may transfer to other set-generation problems, such as point clouds or graphs, where permutation ambiguity slows training.
  • The padding-token count-control mechanism could be extended to control other attributes, such as part boundaries or material regions, by adding analogous condition channels.
  • Because the representation is defined only on manifold meshes, real-world non-manifold asset corpora may require topology-changing repair; the paper's exactness claim is conditional on repair being rare and harmless, which is not measured.

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

4 major / 5 minor

Summary. The paper proposes Meshy T2, a two-stage flow-matching framework for fast image-to-mesh generation. The first stage synthesizes a coarse 64^3 occupancy scaffold via an image-conditioned voxel flow; the second generates a per-vertex latent set with a mesh flow, conditioned on the image, the scaffold, and a requested vertex budget. The core representation is a vertex-set mesh VAE (extending SpaceMesh) that encodes each vertex as one continuous latent token and decodes vertex positions, undirected edges, and oriented face winding order from the latent set. The authors claim the representation is nearly lossless: coordinates are not quantized, coincident vertices are not welded, and artist-authored topology is preserved exactly. They report state-of-the-art geometric fidelity on a retopology benchmark (CD 0.020, HD 0.044, NC 0.860) and image-to-mesh generation within a median of 6 seconds, an order of magnitude faster than autoregressive baselines.

Significance. If the claims hold, Meshy T2 would be a meaningful advance: it combines a continuous, per-vertex latent representation with a parallel flow-based decoder, addressing the sequential bottlenecks and error accumulation of autoregressive mesh generators. The technical contributions are well motivated and detailed, particularly the use of Sinkhorn-normalized permutation matrices for halfedge successor prediction, the NULL-element extension for open boundaries, and the optimal-transport assignment of Sobol positional encodings to unordered latent tokens. The ablation in Table 1 provides a controlled comparison that clearly isolates the benefit of the OT assignment. The paper also makes explicit that the F=2V relation is Euler's formula and uses it only as a heuristic for face-count control. However, the empirical evaluation does not currently substantiate the headline claims: all geometric ground truth is generated by the authors' own Meshy 6 pipeline on a private, uncharacterized 115-asset benchmark, and no external artist-authored meshes are used to test the 'preserved exactly' claim. The three advertised capabilities—interactive speed, face-count control, and multi-part connectivity—a

major comments (4)
  1. [Sec. 3, Tables 2-3] The ground-truth geometry for both the retopology and image-to-mesh tasks is generated by the authors' own Meshy 6 pipeline on a private, curated 115-asset benchmark. The paper reports no curation criteria, no independent asset source, no training/test overlap statistics, and no external benchmark. Since Meshy T2 is developed by the same group, this creates an in-domain evaluation risk: the model may be specialized to Meshy 6's mesh distribution rather than to artist-authored meshes. The headline claims of 'state-of-the-art geometric fidelity' and 'artist-authored topology preserved exactly' are therefore not empirically established. The authors should report dataset provenance and overlap, and rerun the evaluation on external artist-authored meshes (e.g., public retopology datasets or manually retopologized assets).
  2. [Sec. 2.1, Eq. (3)] The face representation is lossless only for manifold meshes. The paper states that non-manifold meshes are 'repaired in advance by splitting the offending edges and vertices,' but it never reports how many training or test meshes required repair, what fraction of edges/vertices were split, or how this alters the connectivity that the VAE is claimed to preserve exactly. Concretely, if a non-manifold mesh is repaired, the decoded mesh does not preserve the original artist-authored connectivity; the 'exactly' claim is therefore conditioned on an unreported dataset assumption. The paper's own Sec. 4 concession about 'topological robustness on highly irregular source meshes' further limits the generality. Please report repair statistics and evaluate on unrepaired meshes, or explicitly restrict the scope of the losslessness claim.
  3. [Secs. 2.3 and 3] Face-count control is advertised as one of the three practical capabilities, but no experiment measures it. The vertex-count conditioning mechanism with pad tokens should produce meshes in the range [N/(1+p), N], yet no results show achieved vertex or face counts versus requested budgets, the spread across assets, or the effect on geometric fidelity. A simple quantitative evaluation (e.g., a plot of requested vs. achieved face count, with mean and variance, across at least two budgets) is needed to substantiate this capability rather than relying on the mechanism's design.
  4. [Tables 1-3] All headline metrics are single-run point estimates with no confidence intervals, multiple seeds, or statistical tests. On a 115-asset benchmark, differences such as BPT's 95.7% vs. 100% success rate or the FD gaps in Table 3 may not be significant. Additionally, the latency claims ('median of 6 seconds') are reported without specifying GPU hardware, software versions, or whether timing includes all preprocessing and decoding stages. Reporting per-sample distributions, seeds, and hardware details is necessary to support the speed and robustness claims.
minor comments (5)
  1. [Sec. 2.3] The notation for the vertex count condition is confusing: N is first introduced as the requested latent slot count, then the condition is set to the 'padded total token count'; clarify the relationship between N, the padding ratio p, and the final vertex count.
  2. [Fig. 6] The caption says the shaded band is a 'per-step min–max envelope over the same 10-step window'; it would help to define the windowing operation and state whether the envelope is over steps or over runs.
  3. [Sec. 3.2] The statement that absolute non-manifold counts for FastMesh and MeshSilksong 'are not directly comparable' because of dense outputs is reasonable, but the text should make clear which table rows are directly comparable and how success rates account for the 20-minute limit.
  4. [References] Reference [1] is a textbook; it should be cited at the point of the Euler formula in Sec. 2.3 (the current citation [1] appears in the text and is fine, but the sentence 'F=2V-4≈2V' would benefit from an explicit reference to the book's relevant chapter).
  5. [Sec. 3.3] The DINOv2 FD values are much larger than the Inception FD values; a brief note explaining the feature-space scale would help readers interpret the magnitudes.

Circularity Check

0 steps flagged

No circular derivation: the paper's predictions follow from trained flow models, an explicitly labeled Euler identity, and external prior work (SpaceMesh, TRELLIS), not from its own fitted constants.

full rationale

Walking the claimed derivation chain, none of the load-bearing steps reduces to its own inputs. The vertex budget relation F = 2V − 4 ≈ 2V is quoted as Euler's formula from an external textbook reference and used only to translate a vertex-count condition into an expected face budget; it is a mathematical identity, not a fitted parameter. The edge logit of Eq. (1) and the halfedge successor mapping of Eq. (3) are representation definitions credited to SpaceMesh [28], an external prior work with disjoint authorship, so invoking them is not a self-citation or an imported uniqueness theorem. The VAE losses, flow-matching objectives, and Sobol-OT positional encoding are all trained quantities; the reported retopology and image-to-mesh metrics measure the trained generators against held-out geometry, so they are genuine evaluations rather than renamed training targets. The one notable concern is evaluation provenance: the benchmark ground truth is stated to be 'generated by Meshy 6 and uniformly decimated to a 100k-triangle ground truth' (Sec. 3). That is a validity or generalization risk, since the target distribution may overlap with the authors' product, but it is not a circular reduction: the paper does not fit a parameter to the benchmark and then report that same fit as a prediction, and no equation in the paper makes a claimed result equal to an input by construction. The manifold-only representation caveat ('This representation is well defined only for manifold meshes... non-manifold meshes in the dataset are repaired in advance') is a stated limitation of the lossless-topology claim, not a circular step. Under the rule that circularity requires exhibiting a specific reduction, no such step exists here, so the appropriate score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The system's load is carried by a borrowed topology representation (SpaceMesh), a borrowed coarse-to-fine scaffolding strategy (TRELLIS), standard flow matching, and an unreported proprietary training corpus. The genuinely new machinery (Sobol OT position assignment, existence channel, NULL extension) is architectural rather than physical, so the ledger is dominated by domain assumptions about manifoldness and benchmark construction rather than by free fits.

free parameters (3)
  • Pad ratio p for count conditioning = not reported
    Sec. 2.3 relaxes exact-count control into range control with 'pad tokens randomly appended... up to a ratio p'; the advertised face-budget guarantee [N/(1+p), N] depends on p, whose value is never given.
  • Maximum vertex count N_max = not reported
    The count condition is a Fourier embedding of N/N_max (Sec. 2.3); N_max determines the conditioning scale and is never stated.
  • VAE loss weights w_v=100, w_e=w_f=10 and KL weight 1e-4 = 100 / 10 / 10 / 1e-4
    Hand-chosen balancing weights (Sec. 2.1 implementation details) that shape reconstruction quality and therefore the ceiling for downstream generation.
axioms (5)
  • standard math Closed-manifold triangle meshes satisfy F = 2V - 4
    Sec. 2.3 uses Euler's formula with 3F = 2E to justify vertex-budget face control, cited to [1]; the boundary correction for the open meshes that the NULL extension explicitly supports is not accounted for.
  • domain assumption Training and benchmark meshes are, or are repaired to be, manifold and consistently orientable
    Sec. 2.1: 'This representation is well defined only for manifold meshes... non-manifold meshes in the dataset are repaired in advance by splitting.' The fraction of repaired meshes is unreported, yet the near-lossless claim rests on it.
  • domain assumption The SpaceMesh adjacency representation (Eq. 1 spacetime logit plus halfedge successor permutations) is a correct, near-lossless topology encoding
    Borrowed from [28] without re-derivation; the paper extends it with a NULL element but inherits the four-dimensional edge-embedding and permutation-fan machinery. If that representation fails on thin or boundary-heavy geometry, the 'preserved exactly' claim fails with it.
  • standard math Flow matching / rectified flow provides a valid generative model over per-vertex latent sets
    Used as the generative backbone (Secs. 2.2-2.3) with velocity prediction; standard theory from [17,18], no re-derivation.
  • domain assumption DINOv3 image features and the pretrained Voxel VAE latent space are sufficient conditions for the two stages
    The image condition is a frozen DINOv3 [30] and the voxel scaffold lives in a Voxel VAE latent (following TRELLIS [40]); the quality of both is taken as given, and stage-I errors propagate into stage-II.

pith-pipeline@v1.3.0-alltime-deepseek · 13996 in / 24777 out tokens · 229339 ms · 2026-08-03T00:43:42.701886+00:00 · methodology

0 comments
read the original abstract

Polygonal meshes are the standard surface representation of modern 3D pipelines, and generating high-quality meshes with artist-style topology is essential for film, gaming, and interactive 3D applications. Mainstream approaches serialize a mesh into a token sequence and decode it autoregressively, which is slow at inference and sensitive to error accumulation, making them impractical for interactive asset creation. We present Meshy T2, a fast native mesh generation framework built on flow matching. At its core is a vertex-set mesh VAE that encodes a mesh into one continuous latent token per vertex and decodes vertices, edge connectivity, and face winding order in a single pass, preserving high-precision geometry and artist-authored topology without vertex quantization or welding. Generation proceeds as a coarse-to-fine cascade of two flow-matching models: an image-conditioned voxel flow first sketches the overall shape as a coarse occupancy scaffold, and a mesh flow then populates the scaffold with per-vertex latent tokens, conditioned on the image, the scaffold, and a requested vertex budget. This design delivers three practical capabilities: interactive generation speed through parallel flow-based synthesis; effective face-count control through the requested vertex budget; and native support for multi-part assets, whose components emerge directly from the generated connectivity. In our experiments, Meshy T2 achieves state-of-the-art geometric fidelity and completes end-to-end image-to-mesh generation within a median of 6 seconds, over an order of magnitude faster than autoregressive baselines. Code and weights will be available at https://github.com/meshy-dev/meshy-t2.

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

47 extracted references · 12 linked inside Pith

  1. [1]

    A K Peters/CRC Press, 2010

    Mario Botsch, Leif Kobbelt, Mark Pauly, Pierre Alliez, and Bruno Lévy.Polygon Mesh Processing. A K Peters/CRC Press, 2010

    2010

  2. [2]

    Meshxl: Neural coordinate field for generative 3d founda- tion models.Advances in Neural Information Processing Systems, 37:97141–97166, 2024

    Sijin Chen, Xin Chen, Anqi Pang, Xianfang Zeng, Wei Cheng, Yijun Fu, Fukun Yin, Zhibin Wang, Jingyi Yu, Gang Yu, et al. Meshxl: Neural coordinate field for generative 3d founda- tion models.Advances in Neural Information Processing Systems, 37:97141–97166, 2024

    2024

  3. [3]

    Meshanything: Artist-created mesh generation with autoregressive transformers

    Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Jiaxiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang Yu, Guosheng Lin, and Chi Zhang. Meshanything: Artist-created mesh generation with autoregressive transformers. InInternational Conference on Learning Representations, 2025

    2025

  4. [4]

    Meshanything v2: Artist-created mesh generation with adjacent mesh tokenization

    Yiwen Chen, Yikai Wang, Yihao Luo, Zhengyi Wang, Zilong Chen, Jun Zhu, Chi Zhang, and Guosheng Lin. Meshanything v2: Artist-created mesh generation with adjacent mesh tokenization. InProceedings of the IEEE/CVF International Conference on Computer Vision, 2025

    2025

  5. [5]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in Neural Information Processing Systems, volume 26, 2013

    2013

  6. [6]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. InInternational Conference on Machine Learning, 2024. 14

    2024

  7. [7]

    Romero, Tsung-Yi Lin, and Ming-Yu Liu

    Zekun Hao, David W. Romero, Tsung-Yi Lin, and Ming-Yu Liu. Meshtron: High-fidelity, artist-like 3d mesh generation at scale.arXiv preprint arXiv:2412.09548, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Meshcraft: Exploring efficient and controllable mesh generation with flow-based dits.arXiv preprint arXiv:2503.23022, 2025

    Xianglong He, Junyi Chen, Di Huang, Zexiang Liu, Xiaoshui Huang, Wanli Ouyang, Chun Yuan, and Yangguang Li. Meshcraft: Exploring efficient and controllable mesh generation with flow-based dits.arXiv preprint arXiv:2503.23022, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochre- iter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, volume 30, 2017

    2017

  10. [10]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Pith/arXiv arXiv 2022

  11. [11]

    Lrm: Large reconstruction model for single image to 3d.arXiv preprint arXiv:2311.04400, 2023

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d.arXiv preprint arXiv:2311.04400, 2023

    Pith/arXiv arXiv 2023

  12. [12]

    Fastmesh: Efficient artistic mesh generation via component decoupling

    Jeonghwan Kim, Yushi Lan, Armando Fortes, Yongwei Chen, and Xingang Pan. Fastmesh: Efficient artistic mesh generation via component decoupling. InInternational Conference on 3D Vision (3DV), 2026

    2026

  13. [13]

    Triflow: Generating artist-like 3d mesh topology via nearest-vertex vector fields

    Haoxuan Li, Ziya Erkoç, Daniele Sirigatti, Vladislav Rosov, Lei Li, Angela Dai, and Matthias Nießner. Triflow: Generating artist-like 3d mesh topology via nearest-vertex vector fields. arXiv preprint arXiv:2606.20131, 2026

    Pith/arXiv arXiv 2026

  14. [14]

    Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model.arXiv preprint arXiv:2311.06214, 2023

    Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model.arXiv preprint arXiv:2311.06214, 2023

    Pith/arXiv arXiv 2023

  15. [15]

    Meshflow: Efficient artistic mesh generation via meshvae and flow-based diffusion transformer

    Weiyu Li, Antoine Toisoul, Tom Monnier, Roman Shapovalov, Rakesh Ranjan, Ping Tan, and Andrea Vedaldi. Meshflow: Efficient artistic mesh generation via meshvae and flow-based diffusion transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5849–5858, 2026

    2026

  16. [16]

    Treemeshgpt: Artistic mesh generation with autoregressive tree sequencing

    Stefan Lionar, Jiabin Liang, and Gim Hee Lee. Treemeshgpt: Artistic mesh generation with autoregressive tree sequencing. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 26608–26617, 2025

    2025

  17. [17]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InInternational Conference on Learning Representations, 2023

    2023

  18. [18]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InInternational Conference on Learning Representations, 2023

    2023

  19. [19]

    Lato.2: Factorized 3d mesh generation with vertex and topology flow.arXiv preprint arXiv:2607.10623, 2026

    Hang Long, Tianhao Zhao, Junkai Lin, Youjia Zhang, Huipeng Guo, Rendong Liang, Jiale Xu, Jozef Hladký, Matthias Nießner, and Wei Yang. Lato.2: Factorized 3d mesh generation with vertex and topology flow.arXiv preprint arXiv:2607.10623, 2026

    Pith/arXiv arXiv 2026

  20. [20]

    Lorensen and Harvey E

    William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. InProceedings of the 14th Annual Conference on Computer Graphics and Interactive T echniques, pages 163–169, 1987. 15

    1987

  21. [21]

    Faithful contouring: Near- lossless 3d voxel representation free from iso-surface

    Yihao Luo, Xianglong He, Chuanyu Pan, Yiwen Chen, Jiaqi Wu, Yangguang Li, Wanli Ouyang, Yuanming Hu, Guang Yang, and ChoonHwai Yap. Faithful contouring: Near- lossless 3d voxel representation free from iso-surface. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14408–14418, 2026

    2026

  22. [22]

    Dinov2: Learning robust visual features without supervision.T ransactions on Machine Learning Research, 2024

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision.T ransactions on Machine Learning Research, 2024

    2024

  23. [23]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceed- ings of the IEEE/CVF International Conference on Computer Vision, pages 4195–4205, 2023

    2023

  24. [24]

    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

  25. [25]

    Dual marching cubes: Primal contouring of dual grids

    Scott Schaefer and Joe Warren. Dual marching cubes: Primal contouring of dual grids. Computer Graphics Forum, 24(2):195–201, 2005

    2005

  26. [26]

    Glu variants improve transformer.arXiv preprint arXiv:2002.05202, 2020

    Noam Shazeer. Glu variants improve transformer.arXiv preprint arXiv:2002.05202, 2020

    Pith/arXiv arXiv 2002

  27. [27]

    Flexible isosurface extraction for gradient-based mesh optimization.ACM T ransactions on Graphics, 42(4):1–16, 2023

    Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization.ACM T ransactions on Graphics, 42(4):1–16, 2023

    2023

  28. [28]

    Spacemesh: A continuous representation for learning manifold surface meshes

    Tianchang Shen, Zhaoshuo Li, Marc Law, Matan Atzmon, Sanja Fidler, James Lucas, Jun Gao, and Nicholas Sharp. Spacemesh: A continuous representation for learning manifold surface meshes. InSIGGRAPH Asia 2024 Conference Papers, 2024

    2024

  29. [29]

    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

  30. [30]

    Oriane Siméoni, Huy V Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, et al. Dinov3. arXiv preprint arXiv:2508.10104, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    On the distribution of points in a cube and the approximate evaluation of integrals.USSR Computational Mathematics and Mathematical Physics, 7(4):86–112, 1967

    Ilya M Sobol’. On the distribution of points in a cube and the approximate evaluation of integrals.USSR Computational Mathematics and Mathematical Physics, 7(4):86–112, 1967

    1967

  32. [32]

    Topology-preserved auto-regressive mesh generation in the manner of weaving silk

    Gaochao Song, Zibo Zhao, Haohan Weng, Jingbo Zeng, Rongfei Jia, and Shenghua Gao. Topology-preserved auto-regressive mesh generation in the manner of weaving silk. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  33. [33]

    Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    2024

  34. [34]

    Meshflow: Mesh generation with equivariant flow matching

    Qi Sun, Kiyohiro Nakayama, Jing Nathan Yan, Qixing Huang, Alexander Rush, Leonidas Guibas, Gordon Wetzstein, Jing Liao, and Guandao Yang. Meshflow: Mesh generation with equivariant flow matching. InACM SIGGRAPH 2026 Conference Papers, 2026

    2026

  35. [35]

    Rethinking the inception architecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2818–2826, 2016. 16

    2016

  36. [36]

    Fourier features let networks learn high frequency functions in low dimensional domains

    Matthew Tancik, Pratul P Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. InAdvances in Neural Information Processing Systems, volume 33, pages 7537–7547, 2020

    2020

  37. [37]

    Edgerunner: Auto-regressive auto-encoder for artistic mesh generation

    Jiaxiang Tang, Max Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang. Edgerunner: Auto-regressive auto-encoder for artistic mesh generation. InInternational Conference on Learning Representations, volume 2025, pages 35913–35934, 2025

    2025

  38. [38]

    Nexus: Native mesh generation with diffusion.arXiv preprint arXiv:2607.13563, 2026

    Hanxiao Wang, Ying-Tian Liu, Yuan-Chen Guo, Qi-Yuan Feng, Zi-Xin Zou, Ding Liang, Biao Zhang, and Yan-Pei Cao. Nexus: Native mesh generation with diffusion.arXiv preprint arXiv:2607.13563, 2026

    Pith/arXiv arXiv 2026

  39. [39]

    Haohan Weng, Zibo Zhao, Biwen Lei, Xianghui Yang, Jian Liu, Zeqiang Lai, Zhuo Chen, Yuhong Liu, Jie Jiang, Chunchao Guo, Tong Zhang, Shenghua Gao, and C. L. Philip Chen. Scaling mesh generation via compressive tokenization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11093–11103, 2025

    2025

  40. [40]

    Structured 3d latents for scalable and versatile 3d generation

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21469–21480, 2025

    2025

  41. [41]

    Native and compact structured latents for 3d generation

    Jianfeng Xiang, Xiaoxue Chen, Sicheng Xu, Ruicheng Wang, Zelong Lv, Yu Deng, Hongyuan Zhu, Yue Dong, Hao Zhao, Nicholas Jing Yuan, and Jiaolong Yang. Native and compact structured latents for 3d generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14419–14429, 2026

    2026

  42. [42]

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

    Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. In- stantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models.arXiv preprint arXiv:2404.07191, 2024

    Pith/arXiv arXiv 2024

  43. [43]

    Freesplatter: Pose-free gaussian splatting for sparse-view 3d reconstruction.arXiv preprint arXiv:2412.09573, 2024

    Jiale Xu, Shenghua Gao, and Ying Shan. Freesplatter: Pose-free gaussian splatting for sparse-view 3d reconstruction.arXiv preprint arXiv:2412.09573, 2024

    Pith/arXiv arXiv 2024

  44. [44]

    Meshweaver: Sparse-voxel-guided surface weaving for autoregressive mesh generation

    Jiale Xu, Wang Zhao, and Ying Shan. Meshweaver: Sparse-voxel-guided surface weaving for autoregressive mesh generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5912–5922, 2026

    2026

  45. [45]

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

    2023

  46. [46]

    Deepmesh: Auto-regressive artist-mesh creation with reinforcement learning

    Ruowen Zhao, Junliang Ye, Zhengyi Wang, Guangce Liu, Yiwen Chen, Yikai Wang, and Jun Zhu. Deepmesh: Auto-regressive artist-mesh creation with reinforcement learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10612–10623, 2025

    2025

  47. [47]

    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 Hladký, Matthias Nießner, and Wei Yang. Lato: 3d mesh flow matching with structured topology preserving latents.arXiv preprint arXiv:2603.06357, 2026. 17

    arXiv 2026