Pith. sign in

REVIEW 4 major objections 8 minor 132 references

Making attention intrinsic and triangulation-agnostic turns it into a strong default layer for learning on triangle meshes.

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 · grok-4.5

2026-07-31 04:59 UTC pith:YXWHNPOS

load-bearing objection Simple mass-weighted FEM attention on intrinsic backbones is a real, usable technique; the mechanism claims hold, while the broadest “beats all transformers” headline overreaches on soft baselines. the 4 major comments →

arxiv 2607.24954 v1 pith:YXWHNPOS submitted 2026-07-27 cs.GR cs.CV

Intrinsic and Triangulation-Agnostic Attention: A Simple and Powerful Approach for Learning on Meshes

Ashwath Shetty , Zihan Zhu , Soeren Pirk , Noam Aigerman This is my paper
classification cs.GR cs.CV
keywords triangle meshesattention mechanismintrinsic learningtriangulation-agnosticfinite elementsshape deformationdense correspondencegeometry processing
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.

Standard attention has powered progress in text, images, and point clouds, but on triangle meshes it has lagged behind more geometric networks. This paper argues the missing ingredients are intrinsicality—respecting the mesh’s surface metric and topology—and triangulation-agnosticism—treating vertex features as samples of continuous functions on the underlying surface. The authors build queries, keys, and values with an existing intrinsic mesh network, then discretize continuous attention integrals with standard finite-element mass weighting. The resulting layers plug into prior mesh backbones and beat both mesh architectures and large point-cloud transformers on high-frequency signal prediction, fine deformation (including fingers), dense full and partial correspondence, and heat-kernel descriptors. A sympathetic reader cares because a simple, principled fix to attention appears to unlock the same aggregation power that transformers brought elsewhere, without discarding the mesh.

Core claim

If attention’s queries, keys, and values are produced by an intrinsic triangulation-agnostic network and the attention integrals are discretized with lumped-mass FEM quadrature, the mechanism itself becomes intrinsic and triangulation-agnostic, and that alone is enough to reach or exceed state-of-the-art accuracy on several standard mesh learning tasks.

What carries the argument

Mass-weighted mesh attention: continuous softmax attention rewritten as surface integrals, then discretized so each vertex’s aggregated feature is a lumped-mass-weighted sum of values (self-attention) or ordinary attention when keys/values live off-mesh (cross-attention), with an O(h²+ε) approximation guarantee under smooth fields.

Load-bearing premise

The claim rests on the idea that simple mass-weighted quadrature of continuous attention, fed by an existing intrinsic backbone on single-component manifold meshes, is accurate enough that the measured gains really come from triangulation-agnostic attention rather than backbone capacity or training setup.

What would settle it

Train the same backbone with and without mass weighting on a held-out set of heavily re-triangulated meshes (coarsened, subdivided, variable density) for high-frequency eigenfunction regression; if mass weighting does not preserve accuracy while naive attention collapses, or if a carefully tuned point-cloud transformer matches the intrinsic results on identical splits, the central claim is weakened.

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

If this is right

  • Mesh learning pipelines can add self- and cross-attention without giving up discretization robustness or the surface metric.
  • Cross-attention becomes a stronger conditioner than concatenating pose or other signals into mesh MLPs, enabling finer articulated deformation.
  • Features from the high-frequency predictor support dense full and partial correspondence via simple nearest-neighbor matching, without a specialized matching network.
  • Older intrinsic backbones can be upgraded with these layers and still surpass newer attention-free mesh methods on deformation.
  • The same construction suggests mesh-native representation learning and multimodal conditioning (vision/language) via cross-attention.

Where Pith is reading between the lines

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

  • Because self-attention cost still grows with vertex count, practical scaling may require hierarchical or sparse intrinsic attention rather than dense mass-weighted softmax alone.
  • Extending the same continuous-integral view across multiple connected components could remove the single-manifold restriction without abandoning triangulation-agnosticism.
  • If mass weighting is the decisive inductive bias, lighter backbones than full PDE-based blocks might still deliver most of the gain once Q/K/V are intrinsic.
  • Correspondence results from a naïve feature matcher imply that stronger map regularizers on top of these features could push partial and non-humanoid matching further.

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 / 8 minor

Summary. The paper adapts attention to triangle meshes by enforcing two properties: intrinsicality and triangulation-agnosticism. Queries, keys, and values are produced by an intrinsic, discretization-agnostic backbone (PoissonNet or DiffusionNet) and treated as samples of continuous fields; attention is then defined as continuous softmax-weighted integration over the surface and discretized with standard lumped-mass FEM quadrature, yielding a mass-weighted softmax attention that plugs into memory-efficient kernels. Cross-attention to discrete conditioning signals is shown to be triangulation-agnostic by construction (O(ε) error, App. A); self-attention is shown to approximate its continuous counterpart to O(h²+ε) (App. B). Empirically the authors report gains over mesh and point-cloud baselines on high-frequency eigenfunction prediction (Table 2), single- and arbitrary-source deformation including finger articulation (Tables 3–4), dense full and partial correspondence (Table 5), and HKS prediction on Thingi10K (Table 6), with runtime/memory analysis in Table 7.

Significance. If the claims hold, this is a useful and likely influential contribution to geometry processing: it closes a genuine gap (no prior intrinsic, triangulation-agnostic attention for meshes) with a minimal, easily adoptable modification. Notable strengths: formal discretization-error proofs (O(ε) cross-attention, O(h²+ε) self-attention) in Appendices A–B; well-designed internal controls (Table 1 mass-weighting ablation on coarsened/subdivided/variable-density meshes; Tables 3–4 backbone and component ablations with identical losses, including the hand-vertex loss, applied to baselines); a striking retroactive result (DiffusionNet+attention beating 2025 PoissonNet); and zero-shot transfer to external correspondence benchmarks (FAUST/SCAPE/SH19/DT4D) at 0.2 s per sample versus ~2 min for DiffuMatch. The limitation on the headline SOTA comparison (major comment 1) and missing variance reduce confidence in the breadth, not the core mechanism.

major comments (4)
  1. [§4.1.2, Table 2, §B.5.2] Table 2 is the sole quantitative support for the headline claim of 'exceeding point cloud transformers', but by the paper's own account (§B.5.2) two of the three transformer baselines did not converge: HodgeFormer 'fails to converge to satisfactory results' under both LR settings tried, and the parameter-matched PTV3 (5M) 'fails to converge even on the original discretization'. The only converged transformer baseline is PTV3 at 66x the parameters, trailing by 1.4 dB (33.8 vs 35.2) with no variance estimate. The comparison against PoissonNet (same backbone ±attention, 29.2→35.2 dB) is clean and convincing, but the broader SOTA claim needs repair: either (a) obtain converged HodgeFormer and parameter-matched PTV3 numbers (e.g., with author-provided configs or LR/schedule search on the regression task), or (b) scope the claim to what is demonstrated. As written, the abstract/intro overstate
  2. [§4, Tables 1–6; §B.3–B.5] Every table reports a single run, with 'the results of the best-performing model' (§B.3.1/B.4.1/B.5.1) and no seed variance. Several conclusions rest on margins that could be within run-to-run noise: Table 2 (1.4 dB over PTV3-46M), Table 5 (1.24 vs 1.7 on FAUST; 2.83 vs 3.4 on SH19), Table 6 (1.8 vs 2.8). Please report mean±std over at least 3 seeds for the headline tables, and clarify the checkpoint-selection protocol (best on which split — is the validation set also the test set?). This is cheap to add and materially affects the strength of the claims.
  3. [§3.4, Appendix B (Prop. B.1, Eq. 13)] The O(h²+ε) bound assumes (i) q,k,v ∈ C²(Ω), (ii) bounded keys/values, (iii) the quadrature estimate Eq. (13) at O(h²), and (iv) ε (the backbone's own discretization error) small. Learned features need not be C²; ε is asserted, not controlled; and Eq. (13) at O(h²) for lumped-mass vertex quadrature holds under mesh-quality assumptions (e.g., Delaunay-type triangulations) that are not stated — for obtuse meshes lumped masses can even be non-positive (the 1e-8 epsilon in §B.1 hints at this). Please state the mesh-regularity assumptions explicitly and add a numerical convergence study: you already have coarsened/subdivided variants (Table 1), so reporting attention-output error versus h on a known signal would directly validate the claimed rate, which currently is supported only qualitatively (Fig. 2).
  4. [§4.3, Table 5, §B.6] The correspondence model is trained only on SMPL humans (eigenfunction prediction with rotation/face-deletion augmentation, §B.6.1) and then evaluated on human-centric benchmarks (FAUST, SCAPE, SH19, DT4D) against DiffuMatch, which is category-agnostic. The 4-of-5 win is impressive as zero-shot transfer, but the domain asymmetry should be stated alongside Table 5, and the failure mode (DT4D-Intra, attributed to non-humanoid outliers) arguably illustrates the training-domain restriction rather than an anomaly. Also, baseline numbers are imported from DiffuMatch's Table 1: please confirm identical test splits and evaluation code for every imported entry (§B.6.2 says 'nearly all' reproduced — specify which entries differ).
minor comments (8)
  1. [§1] §1 states the discretizations have 'linearly-bounded error', but the self-attention bound is O(h²+ε). Clarify that 'linear' refers to ε only.
  2. [various] Typos: 'awarness' (§4.1.2), 'discretiation' (Appendix B, first line), 'feature extract' (Fig. 10 caption), 'Franccois' (ref [36]), 'learning overtriangle mesheshave' (§1).
  3. [§3.1, §B.1] The softmax in Eqs. (2)/(5) omits the usual 1/√d scaling. Please state whether scaling is used in practice and whether the log-mass addition (§B.1) interacts with it.
  4. [§3.4, Eqs. (8)–(11)] Eq. (9) defines α without M in the numerator, and Eq. (10) then re-applies M; readers may misread α as the final attention weight. Consider presenting Eq. (11) directly as 'softmax with a log-mass bias, followed by a mass-weighted value aggregation', which matches the §B.1 implementation.
  5. [§4.5, Table 7] Table 7 shows a 10x slowdown from self-attention at 300k faces (1039 ms vs 101 ms). A sentence on practical mesh-size limits, and whether sparse/local attention is compatible with the mass-weighted formulation, would help practitioners.
  6. [§B.1] Only single-head attention is used (§B.1). A brief ablation on head count (or a sentence justifying one head) would be useful.
  7. [Fig. 2, §4.1.1] Fig. 2: state how ground truth on the re-triangulated mesh was obtained (interpolation via closest point? recomputed eigenfunctions?), since eigenfunctions are mesh-dependent and sign/order-ambiguous at higher frequencies.
  8. [§4 / Supplement] Please state whether code and trained models will be released; the method's simplicity makes it easy to reimplement, but the baseline training details (Pointcept configs, HodgeFormer settings) matter for reproducibility of Table 2.

Circularity Check

0 steps flagged

No significant circularity: empirical mesh-attention method with self-contained FEM discretization and external benchmark evaluation.

full rationale

The paper's load-bearing construction is a standard continuous-to-discrete step: Q/K/V are produced by an off-the-shelf intrinsic backbone, treated as samples of continuous fields, and attention is defined via surface integrals (Eqs. 5–7) then discretized by lumped-mass quadrature (Eqs. 8–11), with O(ε) and O(h²+ε) error bounds proved in Appendices A–B from elementary analysis/FEM. That chain does not define the output in terms of the target quantity, nor fit a parameter and relabel it as a prediction. Empirical claims (high-frequency signals, deformation, correspondence, HKS) are trained and scored on held-out poses, OOD shapes, Thingi10K, and external correspondence suites (FAUST, SCAPE, etc.), so results are not forced by normalization identities. Use of PoissonNet/DiffusionNet (including author-overlapping PoissonNet) is ordinary backbone choice and ablation control (Tables 3–4), not a self-citation uniqueness theorem or ansatz that makes the central claim true by construction. No circular steps meet the quote-and-reduce standard.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

Load-bearing content is mostly standard FEM/attention plus domain assumptions of intrinsic mesh learning. Free parameters are ordinary training hyperparameters and architectural widths, not physical constants fitted to prove a law. No new particles or forces; the ‘invented’ object is the mesh attention layer itself as an engineering construct.

free parameters (4)
  • Attention channel dims (cross 128, self 126) and single head = 128 / 126, 1 head
    Chosen architectural widths; performance depends on these capacity choices.
  • Training hyperparameters (lr=0.005, 200k iters, batch 16, hand loss λ=1) = Adam lr 0.005; λ_hand=1
    Hand-set optimization and loss weighting, including extra hand-vertex loss that directly targets finger articulation claims.
  • Number of predicted eigenfunctions / PE dimension (first 64) = 64
    Truncation level for high-frequency targets and positional encoding fed to deformation.
  • Lumped-mass epsilon 1e-8 before log-mass in softmax = 1e-8
    Ad hoc stabilizer for area-weighted attention implementation.
axioms (5)
  • domain assumption Mesh functions are treated as samples of continuous fields on a manifold approximated by the triangulation; triangulation-agnosticism means F_i ≈ F(x_i).
    Section 3.1; standard geometry-processing modeling assumption.
  • standard math Lumped-mass vertex quadrature approximates integrals of C² integrands with O(h²) error.
    Invoked in Appendix B Eq. (13) and Proposition B.1.
  • domain assumption Q/K/V produced by PoissonNet or DiffusionNet are sufficiently intrinsic and triangulation-agnostic when inputs are.
    Section 3.2; inherits claims from cited backbones.
  • ad hoc to paper Keys/values bounded and fields C² so attention discretization error bounds apply to learned features.
    Propositions A.1 and B.1; smoothness of trained networks is assumed, not verified.
  • domain assumption SMPL correspondence and transferred Laplacian eigenfunctions define a valid multi-mesh supervised signal for intrinsic learning.
    Section 4.1 training setup.
invented entities (1)
  • Mass-weighted intrinsic mesh self-attention (FEM quadrature form of continuous softmax attention) independent evidence
    purpose: Provide triangulation-agnostic feature aggregation on mesh vertices while remaining compatible with efficient attention kernels.
    Core proposed operator, Eq. (11); engineering construct rather than a physical entity.

pith-pipeline@v1.2.0-grok45-kimik3 · 28609 in / 3264 out tokens · 65734 ms · 2026-07-31T04:59:55.562496+00:00 · methodology

0 comments
read the original abstract

This work proposes an adaptation of the attention mechanism for triangle meshes. The core observation is that endowing the attention mechanism with critical properties for learning over meshes -- intrinsicality and triangulation-agnosticism -- enables it to attain state-of-the-art results over several learning-based tasks in geometry-processing. The above is achieved by modifying the attention mechanism from the bottom up based on simple principles from geometry-processing. Namely, the quantities used within attention -- queries, keys and values -- are created by an intrinsic, triangulation-agnostic network, and treated as discretizations of continuous functions. From that, we devise an appropriate attention mechanism that operates over triangle meshes through standard FEM discretization of the resulting integrals of the above functions. Surprisingly, as far as we know, this straightforward approach has not been utilized for learning over meshes. Experiments show our method exceeds current state of the art, including both mesh-based architectures as well as point cloud transformers. Namely, we show significant improvements on several common benchmarks and tasks -- predicting canonical high-frequency signals; predicting deformations; computing dense correspondences, both between full shapes and partial ones; and predicting feature descriptors.

Figures

Figures reproduced from arXiv: 2607.24954 by Ashwath Shetty, Noam Aigerman, Soeren Pirk, Zihan Zhu.

Figure 1
Figure 1. Figure 1: We propose an attention layer designed specifically for meshes, so that it is intrinsic and discretization agnostic. Using these layers we attain state-of-the-art results in several applications: predicting Fourier-like high dimensional embeddings of vertices; deforming characters with a level of granularity unattained before (note fingers); computing mesh-to-mesh dense correspondences ; and, predicting sh… view at source ↗
Figure 2
Figure 2. Figure 2: Triangulation-agnostic vs. naive attention. When faced with a mesh with different triangulation than that of the training set, our triangulation-agnostic attention still predicts signals that match the ground truth; naive attention predicts incorrect signals. differential operators, and define a simple, triangulation-agnostic attention mechanism through it. Learning on meshes. Many methods for learning ove… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison on high frequency signal prediction (Sec￾tion 4.1). Our method accurately predicts high-frequency functions. PoissonNet [54] makes incorrect predictions on intricate areas such as the fingers, while hodgeformer [63] and PTV3 [101] (both with 5x and 90x more parameters) are not triangulation-agnostic and thus make incorrect predictions on differently-triangulated meshes (bottom row) [PITH_FULL_I… view at source ↗
Figure 4
Figure 4. Figure 4: Prediction of high frequency intrinsic functions (Sec￾tion 4.1 for out-of-distribution shapes. Our model, trained solely on SMPL [67] meshes, makes highly-accurate predictions of high Frequency signals on shapes that are significantly different (prediction of same signal shown on all shapes). shapes [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Improvement in performance for new and old meth￾ods via our attention layers, shown on the deformation task. Incorporating our simple layers leads to an immediate improvement for an older method from several years ago, DiffusionNet [80] (DN), which outperforms the current state of the art, PoissonNet [54] (PN); note articulation of fingers. Incorporating our layers with Poisson￾Net itself leads to an even … view at source ↗
Figure 6
Figure 6. Figure 6: Articulations of Arbitrary Source Meshes. Our method can articulate arbitrary humanoids from diverse sources (artist-generated, 3D Scanned, AI-generated), without the use of rigs or training on these shapes. Each triplet shows the ground truth target position in green, our prediction, and the prediction of PoissonNet [54], which lacks the ability to predict correct fine-grained deformations (e.g., hands). … view at source ↗
Figure 7
Figure 7. Figure 7: Generalization of our deformation network to out-of-distribution shapes. Our method, trained solely on human SMPL [67] meshes, generalizes to other types of humanoids, and exhibits more accurate deformations than PoissonNet [54]. Target pose shown in green next to predicted deformations [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Computed correspondences between pairs of meshes. Our method produces dense, smooth, and continuous correspondences (visualized by transferring a texture from one mesh to the other using the correspondence map). Diffumatch [68] exhibits discontinuities and erroneous matchings. We further stress-test the accuracy of our predicted features by computing partial matches between a partial mesh and a full one. T… view at source ↗
Figure 9
Figure 9. Figure 9: Predicted heat kernel signature [88] features over general meshes. Our method reliably predicts intrinsic features on a dataset of general meshes, matching the ground truth significantly better than PoissonNet (PN) [54] [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison against foundational feature extractors on the correspondence task (Section 4.3). The features from our method provide significantly better correspondences than both a foundational point cloud feature extractors [113], as well as a 2D feature extract on meshes [92]. Our method is the only one that produces smooth, continuous matching. Method MSE (×10−5 ) ↓ PSNR ↑ PoissonNet 2.8 15 Ours 1.8 17 … view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

132 extracted references · 4 canonical work pages

  1. [1]

    Ahmed Abdelreheem, Abdelrahman Eldesokey, Maks Ovsjanikov, and Peter Wonka. 2023. Zero-Shot 3D Shape Correspondence. InSIGGRAPH Asia 2023 Con- ference Papers(Sydney, NSW, Australia)(SA ’23). Association for Computing Ma- chinery, New York, NY, USA, Article 59, 11 pages. doi:10.1145/3610548.3618228

    arXiv 2023

  2. [2]

    Noam Aigerman, Kunal Gupta, Vladimir G Kim, Siddhartha Chaudhuri, Jun Saito, and Thibault Groueix. 2022. Neural jacobian fields: Learning intrinsic mappings of arbitrary meshes.SIGGRAPH(2022)

    2022

  3. [3]

    Noam Aigerman, Roi Poranne, and Yaron Lipman. 2014. Lifted bijections for low distortion surface mappings.ACM Trans. Graph.33, 4, Article 69 (July 2014), 12 pages. doi:10.1145/2601097.2601158

    arXiv 2014

  4. [4]

    Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. 2005. SCAPE: Shape Completion and Animation of People. InSIGGRAPH

    2005

  5. [5]

    Souhaib Attaiki and Maks Ovsjanikov. 2023. Shape non-rigid kinematics (snk): A zero-shot method for non-rigid shape matching via unsupervised functional map regularized reconstruction.Advances in Neural Information Processing Systems 36 (2023), 70012–70032

    2023

  6. [6]

    Souhaib Attaiki, Gautam Pai, and Maks Ovsjanikov. 2021. Dpfm: Deep partial functional maps. In2021 International Conference on 3D Vision (3DV). IEEE, 175–185

    2021

  7. [7]

    Stephen W Bailey, Dalton Omens, Paul Dilorenzo, and James F O’Brien. 2020. Fast and deep facial deformations.ACM Transactions on Graphics (TOG)39, 4 (2020), 94–1

    2020

  8. [8]

    Bailey, Dave Otte, Paul Dilorenzo, and James F

    Stephen W. Bailey, Dave Otte, Paul Dilorenzo, and James F. O’Brien. 2018. Fast and Deep Deformation Approximations.ACM Transactions on Graphics37, 4 (Aug. 2018), 119:1–12. doi:10.1145/3197517.3201300 Presented at SIGGRAPH 2018, Los Angeles

    arXiv 2018

  9. [9]

    Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. 2016. Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image. InComputer Vision – ECCV 2016 (Lecture Notes in Computer Science). Springer International Publishing

    2016

  10. [10]

    Federica Bogo, Javier Romero, Matthew Loper, and Michael J. Black. 2014. FAUST: Dataset and evaluation for 3D mesh registration. InCVPR

    2014

  11. [11]

    Davide Boscaini, Jonathan Masci, Emanuele Rodolà, and Michael Bronstein. 2016. Learning shape correspondence with anisotropic convolutional neural networks. Advances in neural information processing systems29 (2016)

    2016

  12. [12]

    Michael M Bronstein, Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre Van- dergheynst. 2017. Geometric deep learning: going beyond euclidean data.IEEE Signal Processing Magazine34, 4 (2017), 18–42

    2017

  13. [13]

    Shuhao Cao. 2021. Choose a Transformer: Fourier or Galerkin. InAdvances in Neural Information Processing Systems (NeurIPS 2021), Vol. 34. arXiv:arXiv: 2105.14995 [cs.CL] https://openreview.net/forum?id=ssohLcmn4-r

    Pith/arXiv arXiv 2021

  14. [14]

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. 2021. Emerging Properties in Self-Supervised Vision Transformers. InProceedings of the International Conference on Computer Vision (ICCV)

    2021

  15. [15]

    Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. 2022. Cross-Attention of Disentangled Modalities for 3D Human Mesh Recovery with Transformers. In European Conference on Computer Vision (ECCV)

    2022

  16. [16]

    Pointcept Contributors. 2023. Pointcept: A Codebase for Point Cloud Perception Research. https://github.com/Pointcept/Pointcept

    2023

  17. [17]

    Fu, Stefano Ermon, Atri Rudra, and Christopher Ré

    Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness. InAdvances in Neural Information Processing Systems (NeurIPS)

    2022

  18. [18]

    Tri Dao and Albert Gu. 2024. Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality. InInternational Conference on Machine Learning (ICML)

    2024

  19. [19]

    Pim De Haan, Maurice Weiler, Taco Cohen, and Max Welling. 2020. Gauge equivariant mesh cnns: Anisotropic convolutions on geometric graphs.arXiv preprint arXiv:2003.05425(2020)

    Pith/arXiv arXiv 2020

  20. [20]

    Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 2019. Learning elementary structures for 3D shape generation and matching. InNeurips

    2019

  21. [21]

    Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. 2020. Deep geometric functional maps: Robust feature learning for shape correspondence. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 8592–8601

    2020

  22. [22]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale.ICLR(2021)

    2021

  23. [23]

    Marvin Eisenberger, Zorah Lähner, and Daniel Cremers. 2019. Smooth Shells: Multi-Scale Shape Registration with Functional Maps. doi:10.48550/arXiv.1905. 12512

    doi:10.48550/arxiv.1905 2019

  24. [24]

    Mohammad Farazi and Yalin Wang. 2024. A Recipe for Geometry-Aware 3D Mesh Transformers.2025 IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV)(2024), 3290–3300. https://api.semanticscholar.org/CorpusID: 273798014

    2024

  25. [25]

    Matthias Fey, Jan Eric Lenssen, Frank Weichert, and Heinrich Müller. 2018. Splinecnn: Fast geometric deep learning with continuous b-spline kernels. In Proceedings of the IEEE conference on computer vision and pattern recognition. 869–877

    2018

  26. [26]

    Lawson Fulton, Vismay Modi, David Duvenaud, David I. W. Levin, and Alec Jacobson. 2019. Latent-space Dynamics for Reduced Deformable Simulation. Computer Graphics Forum(2019)

    2019

  27. [27]

    Lin Gao, Jie Yang, Tong Wu, Yu-Jie Yuan, Hongbo Fu, Yu-Kun Lai, and Hao Zhang. 2019. SDM-NET: Deep generative network for structured deformable mesh.ACM Transactions on Graphics (TOG)38, 6 (2019), 1–15

    2019

  28. [28]

    Kim, and Rana Hanocka

    William Gao, Noam Aigerman, Thibault Groueix, Vladimir G. Kim, and Rana Hanocka. 2023. TextDeformer: Geometry Manipulation using Text Guidance. SIGGRAPH (Conference track)(2023)

    2023

  29. [29]

    Oshri Halimi, Or Litany, Emanuele Rodola, Alex M Bronstein, and Ron Kimmel

  30. [30]

    Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. 2019. MeshCNN: a network with an edge.ACM Trans. Graph.38, 4, Article 90 (July 2019), 12 pages. doi:10.1145/3306346.3322959

    arXiv 2019

  31. [31]

    Wenchong He, Zhe Jiang, Chengming Zhang, and Arpan Man Sainju. 2020. CurvaNet: Geometric deep learning based on directional curvature for 3D shape analysis. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2214–2224

    2020

  32. [32]

    Daniel Holden, Jun Saito, and Taku Komura. 2015. Learning an Inverse Rig Mapping for Character Animation. InProceedings of the 14th ACM SIGGRAPH / Eurographics Symposium on Computer Animation(Los Angeles, California) (SCA ’15). Association for Computing Machinery, New York, NY, USA, 165–173. doi:10.1145/2786784.2786788

    arXiv 2015

  33. [33]

    Alec Jacobson, Ilya Baran, Jovan Popovic, and Olga Sorkine. 2011. Bounded biharmonic weights for real-time deformation.ACM Trans. Graph.30, 4 (2011), 78

    2011

  34. [34]

    Alec Jacobson, Zhigang Deng, Ladislav Kavan, and JP Lewis. 2014. Skinning: Real-time Shape Deformation. InACM SIGGRAPH 2014 Courses

    2014

  35. [35]

    Tao Ju, Scott Schaefer, and Joe Warren. 2005. Mean value coordinates for closed triangular meshes.ACM Siggraph 2005 Papers(2005), 561–566

    2005

  36. [36]

    Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and Franccois Fleuret

  37. [37]

    Ladislav Kavan, Steven Collins, Jiří Žára, and Carol O’Sullivan. 2008. Geometric skinning with approximate dual quaternion blending.ACM Transactions on Graphics (TOG)27, 4 (2008), 1–23

    2008

  38. [38]

    Kim, and Rana Hanocka

    Hyunwoo Kim, Itai Lang, Thibault Groueix, Noam Aigerman, Vladimir G. Kim, and Rana Hanocka. 2025. MeshUp: Multi-Target Mesh Deformation via Blended Score Distillation.3DV(2025)

    2025

  39. [39]

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. 2023. Neural Operator: 10 Learning Maps Between Function Spaces with Applications to PDEs.JMLR24, 1, Article 89 (2023), 97 pages

    2023

  40. [40]

    Alon Lahav and Ayellet Tal. 2020. Meshwalker: Deep mesh understanding by random walks.ACM Transactions on Graphics (TOG)39, 6 (2020), 1–13

    2020

  41. [41]

    Barak Lenz, Opher Lieber, Alan Arazi, Amir Bergman, Avshalom Manevich, Barak Peleg, Ben Aviram, Chen Almagor, Clara Fridman, Dan Padnos, Daniel Gissin, Daniel Jannai, Dor Muhlgay, Dor Zimberg, Edden M. Gerber, Elad Dolev, Eran Krakovsky, Erez Safahi, Erez Schwartz, Gal Cohen, Gal Shachaf, Haim Rozenblum, Hofit Bata, Ido Blass, Inbal Magar, Itay Dalmedigos...

  42. [42]

    Peizhuo Li, Kfir Aberman, Rana Hanocka, Libin Liu, Olga Sorkine-Hornung, and Baoquan Chen. 2021. Learning Skeletal Articulations with Neural Blend Shapes. ACM Transactions on Graphics (TOG)40, 4 (2021), 1

    2021

  43. [43]

    Xiao-Juan Li, Jie Yang, and Fang-Lue Zhang. 2022. Laplacian Mesh Transformer: Dual Attention and Topology Aware Network for 3D Mesh Classification and Segmentation. InComputer Vision – ECCV 2022: 17th European Conference, Tel A viv, Israel, October 23–27, 2022, Proceedings, Part XXIX(Tel Aviv, Israel). Springer- Verlag, Berlin, Heidelberg, 541–560. doi:10...

    work page doi:10.1007/978-3-031-19818-2_31 2022

  44. [44]

    Yiming Li, Zhiding Yu, Christopher Bongsoo Choy, Chaowei Xiao, José Manuel Álvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. 2023. VoxFormer: Sparse Voxel Transformer for Camera-Based 3D Semantic Scene Completion. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023), 9087–9098. https://api.semanticscholar.org/CorpusID:257102923

    2023

  45. [45]

    Yaqian Liang, Shanshan Zhao, Baosheng Yu, Jing Zhang, and Fazhi He. 2022. MeshMAE: Masked Autoencoders for 3D Mesh Data Analysis. InEuropean Conference on Computer Vision

    2022

  46. [46]

    Kevin Lin, Lijuan Wang, and Zicheng Liu. 2021. End-to-End Human Pose and Mesh Reconstruction with Transformers. InCVPR

    2021

  47. [47]

    Yaron Lipman, David Levin, and Daniel Cohen-Or. 2008. Green coordinates. ACM Transactions on Graphics (TOG)27, 3 (2008), 1–10

    2008

  48. [48]

    Lipman, O

    Y. Lipman, O. Sorkine, D. Cohen-Or, D. Levin, C. Rossi, and H.P. Seidel. 2004. Dif- ferential coordinates for interactive mesh editing. InProceedings Shape Modeling Applications, 2004.181–190

    2004

  49. [49]

    Or Litany, Tal Remez, Emanuele Rodola, Alex Bronstein, and Michael Bronstein

  50. [50]

    Bronstein, and Michael M

    Or Litany, Tal Remez, Emanuele Rodolà, Alexander M. Bronstein, and Michael M. Bronstein. 2017. Deep Functional Maps: Structured Prediction for Dense Shape Correspondence.2017 IEEE International Conference on Computer Vision (ICCV) (2017), 5660–5668. https://api.semanticscholar.org/CorpusID:4215682

    2017

  51. [51]

    Isabella Liu, Zhan Xu, Wang Yifan, Hao Tan, Zexiang Xu, Xiaolong Wang, Hao Su, and Zifan Shi. 2025. RigAnything: Template-Free Autoregressive Rigging for Diverse 3D Assets.arXiv preprint arXiv:2502.09615(2025)

    arXiv 2025

  52. [52]

    Minghua Liu, Chong Zeng, Xinyue Wei, Ruoxi Shi, Linghao Chen, Chao Xu, Mengqi Zhang, Zhaoning Wang, Xiaoshuai Zhang, Isabella Liu, Hongzhi Wu, and Hao Su. 2024. MeshFormer: High-Quality Mesh Generation with 3D-Guided Reconstruction Model.arXiv preprint arXiv:2408.10198(2024)

    Pith/arXiv arXiv 2024

  53. [53]

    Wenjie Liu, Hao Wu, Xin Qiu, Yingqi Fan, Yihang Zhang, Anhao Zhao, Yunpu Ma, and Xiaoyu Shen. 2026. ViCA: Efficient Multimodal LLMs with Vision-Only Cross-Attention.ArXivabs/2602.07574 (2026). https://api.semanticscholar.org/ CorpusID:285452725

    Pith/arXiv arXiv 2026

  54. [54]

    Arman Maesumi, Tanish Makadia, Thibault Groueix, Vladimir Kim, Daniel Ritchie, and Noam Aigerman. 2025. PoissonNet: A Local-Global Approach for Learning on Surfaces.ACM Trans. Graph.44, 6, Article 175 (Dec. 2025), 16 pages. doi:10.1145/3763298

    work page doi:10.1145/3763298 2025

  55. [55]

    Robin Magnet and Maks Ovsjanikov. 2024. Memory-Scalable and Simplified Functional Map Learning. 4041–4050. doi:10.1109/CVPR52733.2024.00387

    arXiv 2024

  56. [56]

    Jonathan Masci, Davide Boscaini, Michael Bronstein, and Pierre Vandergheynst

  57. [58]

    Mitchel, Vladimir G

    Thomas W. Mitchel, Vladimir G. Kim, and Michael Kazhdan. 2021. Field Convo- lutions for Surface CNNs. (October 2021), 10001–10011

    2021

  58. [59]

    Federico Monti, Davide Boscaini, Jonathan Masci, Emanuele Rodola, Jan Svo- boda, and Michael M Bronstein. 2017. Geometric deep learning on graphs and manifolds using mixture model cnns. InProceedings of the IEEE conference on computer vision and pattern recognition. 5115–5124

    2017

  59. [60]

    Luca Morreale, Noam Aigerman, Vladimir Kim, and Niloy Mitra. 2021. Neural Surface Maps. doi:10.1109/CVPR46437.2021.00461

    arXiv 2021

  60. [61]

    Luca Morreale, Noam Aigerman, Vladimir Kim, and Niloy Mitra. 2024. Neural Semantic Surface Maps.Computer Graphics Forum43 (04 2024). doi:10.1111/cgf. 15005

    doi:10.1111/cgf 2024

  61. [62]

    Sanjeev Muralikrishnan, Niladri Dutt, Siddhartha Chaudhuri, Noam Aigerman, Vladimir Kim, Matthew Fisher, and Niloy J Mitra. 2024. Temporal Residual Jacobians for Rig-Free Motion Transfer. (2024), 93–109

    2024

  62. [63]

    Akis Nousias and Stavros Nousias. 2025. HodgeFormer: Transformers for Learn- able Operators on Triangular Meshes through Data-Driven Hodge Matrices. arXiv preprint(2025)

    2025

  63. [64]

    Ahmed A A Osman, Timo Bolkart, and Michael J. Black. 2020. STAR: A Sparse Trained Articulated Human Body Regressor. InEuropean Conference on Computer Vision (ECCV). 598–613. https://star.is.tue.mpg.de

    2020

  64. [65]

    Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. 2012. Functional maps: a flexible representation of maps between shapes.ACM Transactions on Graphics (ToG)31, 4 (2012), 1–11

    2012

  65. [66]

    Mikhail Panine, Maxime Kirgo, and Maks Ovsjanikov. 2022. Non-Isometric Shape Matching via Functional Maps on Landmark-Adapted Bases. InComputer graphics forum, Vol. 41. Wiley Online Library, 394–417

    2022

  66. [67]

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. 2019. Expressive Body Capture: 3D Hands, Face, and Body from a Single Image. InProceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). 10975–10985

    2019

  67. [68]

    Emery Pierson, Lei Li, Angela Dai, and Maks Ovsjanikov. 2025. DiffuMatch: Category-Agnostic Spectral Diffusion Priors for Robust Non-rigid Shape Match- ing.arXiv preprint arXiv:2507.23715(2025)

    Pith/arXiv arXiv 2025

  68. [69]

    Adrien Poulenard and Maks Ovsjanikov. 2018. Multi-directional geodesic neural networks via equivariant convolution.ACM Trans. Graph.37, 6, Article 236 (Dec. 2018), 14 pages. doi:10.1145/3272127.3275102

    arXiv 2018

  69. [70]

    Markus Rabe and Charles Staats. 2021. Self-attention Does Not Need 𝑂(𝑛 2) Memory. doi:10.48550/arXiv.2112.05682

    Pith/arXiv arXiv doi:10.48550/arxiv.2112.05682 2021

  70. [71]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. InInternational Conference on Machine Learning. https://api.semanticscholar.org/CorpusID:231591445

    2021

  71. [72]

    Alec Radford and Karthik Narasimhan. 2018. Improving Language Understanding by Generative Pre-Training. https://api.semanticscholar.org/CorpusID:49313245

    2018

  72. [73]

    Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. 2018. Contin- uous and orientation-preserving correspondences via functional maps.ACM Transactions on Graphics (ToG)37, 6 (2018), 1–16

    2018

  73. [74]

    Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer

    Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer

  74. [75]

    Cristian Romero, Dan Casas, Jesus Perez, and Miguel A. Otaduy. 2021. Learning Contact Corrections for Handle-Based Subspace Dynamics.ACM Trans. on Graphics (Proc. of ACM SIGGRAPH)40, 4 (2021). http://gmrv.es/Publications/ 2021/RCPO21

    2021

  75. [76]

    Jean-Michel Roufosse, Abhishek Sharma, and Maks Ovsjanikov. 2019. Unsuper- vised deep learning for structured shape matching. InProceedings of the IEEE/CVF International Conference on Computer Vision. 1617–1627

    2019

  76. [77]

    Rusinkiewicz and M

    S. Rusinkiewicz and M. Levoy. 2001. Efficient variants of the ICP algorithm. In Proceedings Third International Conference on 3-D Digital Imaging and Modeling. 145–152. doi:10.1109/IM.2001.924423

    arXiv 2001

  77. [78]

    Patrick Schmidt, Janis Born, Marcel Campen, and Leif Kobbelt. 2019. Distortion- minimizing injective maps between surfaces.ACM Trans. Graph.38, 6, Article 156 (Nov. 2019), 15 pages. doi:10.1145/3355089.3356519

    arXiv 2019

  78. [79]

    John Schreiner, Arul Asirvatham, Emil Praun, and Hugues Hoppe. 2004. Inter- surface mapping.ACM Trans. Graph.23, 3 (Aug. 2004), 870–877. doi:10.1145/ 1015706.1015812

    arXiv 2004

  79. [80]

    Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. 2020. DiffusionNet: Discretization Agnostic Learning on Surfaces.ACM Transactions on Graphics (TOG)41 (2020), 1 – 16. https://api.semanticscholar.org/CorpusID: 233880804

    2020

  80. [81]

    Siyuan Shen, Yin Yang, Tianjia Shao, He Wang, Chenfanfu Jiang, Lei Lan, and Kun Zhou. 2021. High-order differentiable autoencoder for nonlinear model reduction.ACM Transactions on Graphics

    2021

Showing first 80 references.