Pith. sign in

REVIEW 3 major objections 6 minor 44 references

ClothTransformer: Unified Latent-Space Transformers for Scalable Cloth Simulation

T0 review · 3 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Cloth simulation can be reformulated as autoregressive sequence modeling in a learned latent space, and a single ClothTransformer model reduces vertex error severalfold across body-driven, robotic, and free-fall scenarios.

desk verdict A credible unified latent-space cloth simulator whose headline 4–9× advantage is overstated by an unfair baseline protocol; worth engaging but needs revision. read the letter →

arxiv 2605.27852 v5 pith:BP6EXXN4 submitted 2026-05-27 cs.GR cs.CV

Yu Zhang , Yidi Shao , Wenqi Ouyang , Yushi Lan , Zhexin Liang , Chengrui Wu , Xudong Xu , Xingang Pan This is my paper · ORCID
classification cs.GRcs.CV
keywords clothsimulationlatent-spacelearningautoregressivetransformerscontinuouscollisiondetectionneuralphysicsgarmentanimationroboticmanipulationmesh-resolutiongeneralization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that cloth simulation, traditionally tied to mesh-specific solvers or single-scenario neural models, can be recast as a unified autoregressive prediction problem in a learned latent space. If correct, one trained model would handle diverse settings—garments on animated bodies, cloth grasped by robotic grippers, and cloth falling onto arbitrary rigid objects—without per-scenario tuning. The key promise is that the latent bottleneck makes temporal dynamics cost independent of mesh resolution, so high-fidelity meshes could be simulated at interactive speeds. The authors also claim a differentiable continuous-collision-detection loss, enabled by a newly built penetration-free dataset, suppresses the tunneling and self-penetration artifacts that plague discrete-check neural simulators.

What carries the argument

The load-bearing mechanism is the latent bottleneck: a cross-attention layer with a fixed number of learnable query tokens compresses the variable-sized cloth and collision mesh into K latent tokens, and a temporal Transformer with block-causal masking evolves these latents autoregressively. Decoding uses rest-shape vertex embeddings as queries against the predicted latents, followed by a GNN refinement. This makes the core temporal computation O(K²) and independent of mesh resolution. The second mechanism is the CCD module, which detects collisions by finding roots of cubic polynomials along inter-frame trajectories, uses a differentiable detect-then-regress loss to push penetrating vertice

What would settle it

Generate a held-out cloth sequence with deliberately dense high-frequency folds (for example, a crumpled sheet at 40k–100k vertices), keep the latent token count at 1024, and measure mean vertex error along a 100-frame rollout; if error grows steeply with fold density or rollout length beyond what a per-vertex model suffers, the latent bottleneck is discarding exactly the spatial information the dynamics need.

Watch

Extended reading notes

Core claim

The paper claims that a single Transformer, with no scenario-specific priors, can learn shared latent dynamics across three physically distinct cloth-simulation settings and achieve approximately 4–9× lower mean vertex error than prior state-of-the-art learning-based methods in each setting. The architecture compresses arbitrary-resolution cloth and collision geometry into a fixed-size set of latent tokens via cross-attention, evolves those tokens through a temporal Transformer, and reconstructs vertex positions through a rest-shape-conditioned decoder. Because the temporal model operates only on the fixed token set, inference cost scales with the number of latent tokens rather than vertex c

Load-bearing premise

The central claim stands or falls on whether a fixed set of 1024 learned latent tokens can retain enough spatial detail to reconstruct arbitrary high-resolution cloth meshes without error accumulating over long autoregressive rollouts.

Editorial extensions

If this is right

  • A single jointly trained model can generalize across body-driven garments, robotic manipulation, and free-fall collisions without per-scenario fine-tuning.
  • Temporal inference cost stops scaling with vertex count: the paper reports roughly 22 ms per frame at 5k vertices versus 275 ms at 40k vertices, while the strongest graph baseline grows from 130 ms to 472 ms over the same range.
  • The differentiable CCD loss combined with CCD post-processing removes self-penetration artifacts that DCD-based training leaves behind, including tunneling events within a single time step.
  • The model retains accuracy when evaluated at meshes roughly 11× the training resolution, outperforming all baselines in mean vertex error at 40k vertices.
  • The paper argues that the accuracy gap stems from the architecture itself, since even a graph baseline trained in its native self-supervised regime on a single scenario reaches 16.95 cm versus 6.92 cm for the proposed unified model on the Human Garment test set.

Reading between the lines

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

  • Editorial inference: the resolution-independent latent bottleneck should extend naturally to adaptive meshes or emerging topology such as tearing, provided the encoder and decoder can re-tokenize dynamically; the paper does not claim this, but the architecture makes it a plausible next step.
  • Editorial inference: the requirement of penetration-free ground truth sets a high data-generation bar; any future CCD-supervised simulator will need similarly strict solvers or a way to relax the penalty when training data contains residual intersections.
  • Editorial inference: the headline 4–9× advantage is measured against baselines retrained in a supervised mode; the paper's own appendix shows the gap shrinks to about 2.5× when the strongest graph baseline is trained natively on a single scenario, so practitioners should read the headline number in that protocol context.
  • Editorial inference: a natural stress test is to push mesh resolution beyond 40k vertices and increase wrinkle density while keeping the latent token count fixed; if error degrades sharply, the fixed bottleneck is the binding constraint on fidelity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes ClothTransformer, a Transformer-based framework that reformulates cloth simulation as autoregressive sequence modeling in a learned latent space. A spatial encoder compresses cloth vertices and collision triangles into a fixed set of latent tokens via cross-attention; a temporal Transformer evolves these latents; a spatial decoder reconstructs vertex positions. A single model is trained jointly on three scenarios (body-driven garments, robotic manipulation, free-fall collisions) using a newly generated GIPC-based penetration-free dataset, with an optional differentiable CCD loss and inference-time CCD post-processing. The main claims are: (1) one unified model outperforms prior state-of-the-art neural cloth simulators by roughly 4–9× in mean vertex error; (2) the latent formulation makes temporal dynamics computation independent of mesh resolution; and (3) the CCD module suppresses tunneling and self-intersections.

Significance. If the claims hold, this is a useful advance for learning-based cloth simulation: it is among the first attempts to use a single unified Transformer across qualitatively different cloth scenarios, it provides a fixed-size latent bottleneck that decouples the core dynamics from mesh complexity, and it shows a practical way to integrate continuous collision detection into neural training. The paper also ships a large penetration-free dataset, which is a valuable community resource. The architecture itself is coherent, and the ablations on latent compression, spatial GNN, and CCD components are informative. However, the main quantitative claim of 4–9× superiority is currently not supported as stated, because the baseline comparison in Table 2 relies on a supervised adaptation of a method whose native self-supervised regime gives much better numbers, as the paper itself reports in Appendix I.

major comments (3)
  1. [Abstract, §5.2, Table 2 vs Appendix I (Table 6)] The headline "approximately 4–9× lower error than prior state-of-the-art" is not supported as stated. On Human Garment, the SOTA GNN baseline achieves 59.13 cm MVE under the supervised multi-scenario adaptation in Table 2, but the same backbone trained in its native self-supervised regime on a single scenario reaches 16.95 cm MVE (Appendix I, Table 6). Against that number, the advantage of ClothTransformer (6.92 cm) is about 2.5×, not ~9×. The text in §5.2 asserts that this comparison "confirms that the gap in Table 2 stems from the SOTA GNN architecture itself, not the supervised setup," but the reported numbers show the opposite: the supervised adaptation changes the baseline by a factor of ~3.5. The comparison confounds supervision with single-scenario vs unified training, and no single-scenario supervised baseline is provided. For the two other scenarios, no native-regime baselines a
  2. [Tables 2, 5, 6; §5.2 and §5.3] All quantitative results are single-run point estimates with no error bars, standard deviations, or significance tests. Given that the SOTA GNN baseline varies by a factor of ~3.5 depending on training protocol (16.95 vs 59.13 cm MVE on the same scenario), the factor-level claims in the abstract and §5.2 cannot be distinguished from protocol or seed variation. The paper should report at least a few independent seeds (or equivalent variance estimates) for the headline comparisons and the key ablations. If full runs are too expensive, the paper should say so explicitly and temper the claims accordingly.
  3. [§5.4 and Appendix E] The scalability contribution is stated as "temporal dynamics computation independent of mesh resolution," which is true for the core Transformer, but the end-to-end inference time still grows linearly with mesh size (22.24 ms at 5k vertices to 275.27 ms at 40k vertices) and the reported timings exclude CCD post-processing (Appendix E adds ~30 ms, making the full pipeline ~52 ms at 5k and ~305 ms at 40k). The paper should be careful not to imply that total inference cost is mesh-resolution-independent; this distinction is important for readers assessing practical scalability.
minor comments (6)
  1. [Table 3 and Table 4 captions] These ablations use a smaller network configuration (hidden dimension 256, 6 layers) and only 50k training steps, as disclosed only in Appendix F. The main-text captions should state this directly; otherwise readers may interpret the numbers as full-scale model performance.
  2. [Table 2 caption] The caption says "Ours (CCD Loss)" and "Ours," but Table 2 reports raw predictions without post-processing. Please state explicitly in the caption that no CCD post-processing is applied in this table, to avoid confusion with Figure 4 where post-processing is applied uniformly.
  3. [§5.2] The sentence "up to ~16× lower than the SOTA GNN on Diverse Object Collision" derives from the supervised baseline in Table 2. If the headline comparison is revised, this sentence should be updated to be consistent with the revised baseline protocol.
  4. [Metric definitions] MVE is first used in Table 2 but defined only in Appendix G. A one-sentence definition in the main text would improve readability.
  5. [Figure 3] The labels t_c and t_safe are useful but not defined in the caption. A brief explanation would help.
  6. [Conclusion, Limitations] The Limitations paragraph honestly notes that material properties are implicit and topological changes are not handled. This is fine, but the abstract's strong "physical plausibility" phrasing should be aligned with these limitations.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the paper is an empirical system comparison, not a derivation; the 4–9x claim is a protocol-sensitive overstatement rather than a circular construction.

full rationale

ClothTransformer is an empirical systems paper: its central claims are measured errors on held-out sequences, not a derived chain of equations. The mapping in Eq. 1 is a supervised regressor, losses in Eqs. 4–7 are training objectives, and the reported MVE (Eq. 9) is an independent test metric; nothing is fitted to the test labels and then renamed a prediction. No self-definitional, ansatz-smuggled, or uniqueness-imported circular step appears. The most relevant issue is the baseline-comparison protocol, not circularity. Section 5.2 states that the Appendix I comparison 'confirm[s] that the gap in Table 2 stems from the SOTA GNN architecture itself, not the supervised setup,' but Appendix I/Table 6 reports the same backbone in its native self-supervised single-scenario regime at 16.95 cm MVE on Human Garment, versus 59.13 cm in Table 2 after supervised adaptation. The supervised setup changes the baseline by ~3.5x, so the architectural-causality conclusion is internally contradicted by the paper's own numbers, and the headline 'approximately 4–9x lower error' (Abstract, §1, §5.2, Conclusion) is overstated for that scenario. This is an evaluation-protocol / correctness risk, not a circular derivation. The only self-referential element is that the LayersNet baseline [35] is co-authored by present author Yidi Shao, but LayersNet is an external comparison point rather than a load-bearing justification of the method, so this does not constitute circularity. The dataset is generated by the authors using GIPC, which is standard practice for supervised neural simulation and does not make the evaluation circular. Overall, the central claim has independent empirical content; no step reduces by construction to its own inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

ClothTransformer introduces no new physical entities. Its unstated assumptions are the accuracy of the GIPC-generated ground truth, the availability of future collider geometry, the sufficiency of the chosen cloth model, and the adequacy of fixed-size latent compression across resolutions. The main unquantified free parameters are the latent count, the loss weights, and the CCD margin.

free parameters (3)
  • N_latents = 1024
    Number of latent tokens; ablation Table 3 shows a trade-off between compression and accuracy. Chosen by hand, not derived.
  • CCD safety margin ϵ = not reported
    Appears in tsafe = max(0, tc − ϵ) in Eq. (2); value is never specified, yet it determines the strength of the CCD loss and post-processing behavior.
  • Loss weights λ_mse, λ_contact, λ_ccd = not reported
    Weights in Eqs. (6)–(7) are never assigned numerical values, so the exact training objective is under-specified.
assumptions (4)
  • domain assumption GIPC-simulated data with the stated material parameters is physically accurate and intersection-free ground truth.
    Dataset §4 uses GIPC [15]; all learned errors are measured against this simulation, so any systematic bias in the solver or parameters is inherited by the model.
  • domain assumption The Baraff-Witkin cloth model with the chosen parameters suffices for the three scenarios.
    Appendix C fixes Es=1e6 Pa, Eb=1e5 Pa, ν=0.49, G=5e6 Pa, density 200 g/m²; no comparison to real fabric measurements is provided.
  • domain assumption Collider geometry at the target frame C_{t+1} is available at inference time.
    Eq. (1) conditions on the lookahead collision mesh; this is valid for known animated trajectories or static objects, but not for fully interactive unknown obstacles.
  • domain assumption A fixed-size latent token set retains sufficient information across unseen meshes and resolutions.
    Central to the scalability claim; ablation Table 5 shows MVE degrades to 34.91 cm at 20k vertices and 57.32 cm at 40k vertices, indicating the compression is not lossless at higher resolutions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ClothTransformer: Unified Latent-Space Transformers for Scalable Cloth Simulation." pith.science (2026). https://pith.science/paper/BP6EXXN4

@misc{pith2026260527852,
  author       = {Pith},
  title        = {Pith review of: ClothTransformer: Unified Latent-Space Transformers for Scalable Cloth Simulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BP6EXXN4}},
  note         = {Machine review of arXiv:2605.27852}
}
abstract

Unified and scalable Transformers have recently achieved remarkable success in modeling diverse phenomena traditionally associated with computer graphics, such as 3D visual effects, rendering processes, and motion in videos. In this work, we take a step further by investigating whether modern Transformer techniques can tackle the challenging task of cloth simulation. To this end, we present ClothTransformer, a framework that reformulates cloth simulation as autoregressive sequence modeling in a learned latent space. Existing neural cloth simulators are largely specialized to single scenarios, intrinsically coupled to the mesh discretization, and lack robust collision handling. Our approach addresses these limitations through three contributions: (1) a unified Transformer architecture that handles diverse scenarios -- body-driven garments, robotic manipulation, and free-fall collisions -- under a single model and achieves approximately $4$--$9{\times}$ lower error than prior state-of-the-art methods across all scenarios; (2) a scalable latent-space formulation that compresses arbitrary-resolution meshes into a fixed-size set of latent tokens, making temporal dynamics computation independent of mesh resolution; and (3) a diverse-scenario high-fidelity penetration-free dataset of ${\sim}$493.4k frames spanning all three settings, which enables a differentiable Continuous Collision Detection (CCD) module to suppress penetration artifacts. Project Page: https://yucrazing.github.io/clothtransformer/

Figures

Figures reproduced from arXiv: 2605.27852 by the authors.

Figure 1
Figure 1. ClothTransformer generalizes to unseen test cases across three diverse scenarios. Left two: Diverse Object Collision—cloth falling onto unseen rigid objects (sword, character). Middle two: Human Garment—unseen body, garment, and animation combinations (front-flip, dancing). Right two: Robotic Manipulation—unseen cloth meshes grasped and lifted by a robotic gripper. Abstract Unified and scalable Transformers have rec… view at source ↗
Figure 2
Figure 2. Overview of the proposed auto-regressive cloth simulation architecture. The framework consists of three main components: (1) A Spatial Encoder (left) that compresses the physical state of the history cloth mesh at frame T and the lookahead collision geometry at frame T + 1 into a compact set of latent tokens. (2) A Temporal Transformer (middle) that models the dynamics in the latent space, predicting the future late… view at source ↗
Figure 3
Figure 3. Qualitative comparison on unseen test sequences. Columns 1, 4: Diverse Object Collision (sword, stick). Columns 2, 5: Human Garment (running, front-flip). Columns 3, 6: Robotic Manipulation (static resting, grasping). uniformly to all methods so the visual comparison focuses on shape fidelity rather than residual penetration artifacts. Quantitative Comparison. Our method achieves the best MVE across all three scenar… view at source ↗
Figures from the paper (9 more)
Figure 3
Figure 3. Figure 3: DCD vs. CCD. DCD checks only at discrete time steps and can miss mid-step penetrations (left). CCD sweeps the entire trajectory between two consecutive frames to locate the exact collision time tc, then corrects the vertex to a safe position tsafe before the collision …
Figure 4
Figure 4. Figure 4: Effect of differentiable CCD. Close-up views of collision-prone regions. Top row (front flip, cloth on angel): cloth–collision penetrations that DCD loss cannot fully resolve; our CCD loss targets self-collision and does not improve these cases, but CCD post-processing…
Figure 4
Figure 4. Figure 4: Qualitative comparison on unseen test sequences. Columns 1, 4: Diverse Object Collision (sword, stick). Columns 2, 5: Human Garment (running, front-flip). Columns 3, 6: Robotic Manipulation (static resting, grasping). 5.3 Ablation Studies Impact of the CCD Module. Our …
Figure 5
Figure 5. Figure 5: DCD vs. CCD. DCD checks only at discrete time steps and can miss mid-step penetrations (left). CCD sweeps the entire trajectory between two consecutive frames to locate the exact collision time tc, then corrects the vertex to a safe position tsafe before the collision …
Figure 5
Figure 5. Figure 5: Effect of differentiable CCD. Close-up views of collision-prone regions. Top row (front flip, cloth on angel): cloth–collision penetrations that DCD loss cannot fully resolve; our CCD loss targets self-collision and does not improve these cases, but CCD post-processing…
Figure 6
Figure 6. Figure 6: Visual comparison of different latent compression rates. N=512 produces abnormal deformation artifacts due to excessive compression. The uncompressed variant (No Comp.) exhibits similar artifacts due to insufficient training convergence. Our default N=1024 achieves a g…
Figure 7
Figure 7. Figure 7: CCD vs. DCD-based self-collision handling on a challenging folded-cloth grasping scenario. ContourCraft [10] and our CCD loss both reduce self-collisions but leave residual artifacts. Our CCD post-processing fully resolves remaining intersections. I CCD vs. DCD-Based S…
Figure 8
Figure 8. Figure 8: Unified vs. specialized training on the Human Garment scenario. Each column shows a different frame. Top: SOTA GNN trained unsupervised on Human Garment only. Middle: SOTA GNN trained unsupervised on the unified dataset (all three scenarios). Bottom: our unified model.…
Figure 9
Figure 9. Figure 9: Additional qualitative comparisons. Each row shows a different scenario. From left to right: [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]

Review history (3 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 1 linked inside Pith

  1. [1]

    David Baraff and Andrew P. Witkin. Large steps in cloth simulation. In Steve Cunningham, Walt Bransford, and Michael F. Cohen, editors,Proceedings of the 25th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 1998, Orlando, FL, USA, July 19-24, 1998, pages 43–54. ACM, 1998. doi: 10.1145/280814.280821. URL https: //doi.org/10.1145...

    arXiv 1998

  2. [2]

    Deepsd: Automatic deep skinning and pose space deformation for 3d garment animation

    Hugo Bertiche, Meysam Madadi, Emilio Tylson, and Sergio Escalera. Deepsd: Automatic deep skinning and pose space deformation for 3d garment animation. In2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 5451–5460. IEEE, 2021. doi: 10.1109/ICCV48922.2021.00542. URL https://doi.org/10.1109...

    arXiv 2021

  3. [3]

    Neural cloth simulation.ACM Transac- tions on Graphics (TOG), 41(6):1–14, 2022

    Hugo Bertiche, Meysam Madadi, and Sergio Escalera. Neural cloth simulation.ACM Transac- tions on Graphics (TOG), 41(6):1–14, 2022

    2022

  4. [4]

    Robust treatment of collisions, contact and friction for cloth animation.ACM Trans

    Robert Bridson, Ronald Fedkiw, and John Anderson. Robust treatment of collisions, contact and friction for cloth animation.ACM Trans. Graph., 21(3):594–603, 2002. doi: 10.1145/ 566654.566623. URLhttps://doi.org/10.1145/566654.566623

    arXiv 2002

  5. [5]

    Efficient learning of mesh-based physical simulation with bi-stride multi-scale graph neural network

    Yadi Cao, Menglei Chai, Minchen Li, and Chenfanfu Jiang. Efficient learning of mesh-based physical simulation with bi-stride multi-scale graph neural network. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu...

    2023

  6. [6]

    Chiaramonte, Kevin Carlberg, and Eitan Grinspun

    Yue Chang, Peter Yichen Chen, Zhecheng Wang, Maurizio M. Chiaramonte, Kevin Carlberg, and Eitan Grinspun. Licrom: Linear-subspace continuous reduced order modeling with neural fields. In June Kim, Ming C. Lin, and Bernd Bickel, editors,SIGGRAPH Asia 2023 Conference Papers, SA 2023, Sydney, NSW, Australia, December 12-15, 2023, pages 111:1–111:12. ACM, 202...

    arXiv 2023

  7. [7]

    Peter Yichen Chen, Jinxu Xiang, Dong Heon Cho, Yue Chang, G. A. Pershing, Henrique Teles Maia, Maurizio M. Chiaramonte, Kevin T. Carlberg, and Eitan Grinspun. CROM: continuous reduced-order modeling of pdes using implicit neural representations. InThe Eleventh Interna- tional Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023...

    2023

  8. [8]

    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

Show all 44 references
  1. [9]

    Black, and Otmar Hilliges

    Artur Grigorev, Michael J. Black, and Otmar Hilliges. HOOD: hierarchical graphs for generalized modelling of clothing dynamics. InIEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, 11 pages 16965–16974. IEEE, 2...

    2023

  2. [10]

    Black, Otmar Hilliges, and Bernhard Thomaszewski

    Artur Grigorev, Giorgio Becherini, Michael J. Black, Otmar Hilliges, and Bernhard Thomaszewski. Contourcraft: Learning to resolve intersections in neural multi-garment simula- tions. In Andres Burbano, Denis Zorin, and Wojciech Jarosz, editors,ACM SIGGRAPH 2024 Conference Pape...

    2024

  3. [11]

    Make-it- animatable: An efficient framework for authoring animation-ready 3d characters

    Zhiyang Guo, Jinxu Xiang, Kai Ma, Wengang Zhou, Houqiang Li, and Ran Zhang. Make-it- animatable: An efficient framework for authoring animation-ready 3d characters. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 10783–10792, 2025

    2025

  4. [12]

    Predicting physics in mesh- reduced space with temporal attention

    Xu Han, Han Gao, Tobias Pfaff, Jian-Xun Wang, and Liping Liu. Predicting physics in mesh- reduced space with temporal attention. InThe Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. URL https://ope...

    2022

  5. [13]

    From physically-based to learning-based in cloth simulation: evolution and future - a scoping review.Vis

    Jing He, Yuanjie Cao, TangSheng Guo, Wei Liang, Jin Huang, Qian Liu, Huaiyuan Yang, Sen Liu, and Ruhan He. From physically-based to learning-based in cloth simulation: evolution and future - a scoping review.Vis. Comput., 41(15):12711–12742, 2025. doi: 10.1007/ S00371-025-0418...

    2025 doi

  6. [14]

    Holzschuh, Qiang Liu, Georg Kohl, and Nils Thuerey

    Benjamin J. Holzschuh, Qiang Liu, Georg Kohl, and Nils Thuerey. Pde-transformer: Efficient and versatile transformers for physics simulations. InForty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025. OpenReview.net,

    2025

  7. [15]

    Chitalu, Huancheng Lin, and Taku Komura

    Kemeng Huang, Floyd M. Chitalu, Huancheng Lin, and Taku Komura. GIPC: fast and stable gauss-newton optimization of IPC barrier energy.ACM Trans. Graph., 43(2):23:1–23:18, 2024. doi: 10.1145/3643028. URLhttps://doi.org/10.1145/3643028

    2024 doi

  8. [16]

    Generating datasets of 3d garments with sewing patterns

    Maria Korosteleva and Sung-Hee Lee. Generating datasets of 3d garments with sewing patterns. arXiv preprint arXiv:2109.05633, 2021

    2021 arXiv

  9. [17]

    Langlois, Denis Zorin, Daniele Panozzo, Chenfanfu Jiang, and Danny M

    Minchen Li, Zachary Ferguson, Teseo Schneider, Timothy R. Langlois, Denis Zorin, Daniele Panozzo, Chenfanfu Jiang, and Danny M. Kaufman. Incremental potential contact: intersection- and inversion-free, large-deformation dynamics.ACM Trans. Graph., 39(4):49, 2020. doi: 10.1145/...

    2020

  10. [18]

    Wang, Timur Levent Kesdogan, Duygu Ceylan, and Olga Sorkine- Hornung

    Peizhuo Li, Tuanfeng Y . Wang, Timur Levent Kesdogan, Duygu Ceylan, and Olga Sorkine- Hornung. Neural garment dynamics via manifold-aware transformers.Comput. Graph. Forum, 43(2):i–iii, 2024. doi: 10.1111/CGF.15028. URLhttps://doi.org/10.1111/cgf.15028

    2024 doi

  11. [19]

    Swingar: Spectrum-inspired neural dynamic deformation for free-swinging garments.IEEE Trans

    Tianxing Li, Rui Shi, Qing Zhu, and Takashi Kanai. Swingar: Spectrum-inspired neural dynamic deformation for free-swinging garments.IEEE Trans. Vis. Comput. Graph., 30(10): 6913–6927, 2024. doi: 10.1109/TVCG.2023.3346055. URL https://doi.org/10.1109/ TVCG.2023.3346055

    2024

  12. [20]

    Gartrans: Transformer-based archi- tecture for dynamic and detailed garment deformation.Comput

    Tianxing Li, Zhi Qiao, Zihui Li, Rui Shi, and Qing Zhu. Gartrans: Transformer-based archi- tecture for dynamic and detailed garment deformation.Comput. Vis. Media, 11(6):1209–1226,

  13. [21]

    Spectrum-enhanced graph attention network for garment mesh deformation.IEEE Trans

    Tianxing Li, Rui Shi, Qing Zhu, Liguo Zhang, and Takashi Kanai. Spectrum-enhanced graph attention network for garment mesh deformation.IEEE Trans. Pattern Anal. Mach. Intell., 47(8):7153–7170, 2025. doi: 10.1109/TPAMI.2025.3570523. URL https://doi.org/10. 1109/TPAMI.2025.3570523

    2025

  14. [22]

    URL https://doi.org/10.26599/cvm.2025

    doi: 10.26599/CVM.2025.9450448. URL https://doi.org/10.26599/cvm.2025. 9450448

    2025

  15. [23]

    Meshgraphnetrp: Improving generalization of gnn-based cloth simulation

    Emmanuel Ian Libao, Myeongjin Lee, Sumin Kim, and Sung-Hee Lee. Meshgraphnetrp: Improving generalization of gnn-based cloth simulation. In Julien Pettré, Barbara Solenthaler, Rachel McDonnell, and Christopher Peters, editors,Proceedings of the 16th ACM SIGGRAPH Conference on M...

    2023

  16. [24]

    SENC: handling self-collision in neural cloth simulation

    Zhouyingcheng Liao, Sinan Wang, and Taku Komura. SENC: handling self-collision in neural cloth simulation. In Ales Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol, editors,Computer Vision - ECCV 2024 - 18th European Conference, Milan, 12 I...

    2024 doi

  17. [25]

    Smpl: A skinned multi-person linear model

    Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi-person linear model. InSeminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851–866. 2023

    2023

  18. [26]

    Zero-1-to-3: Zero-shot one image to 3d object

    Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl V ondrick. Zero-1-to-3: Zero-shot one image to 3d object. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2023

    2023

  19. [27]

    Tailornet: Predict- ing clothing in 3d as a function of human pose, shape and garment style

    Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. Tailornet: Predict- ing clothing in 3d as a function of human pose, shape and garment style. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages ...

    2020

  20. [28]

    Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to dress 3d people in generative clothing. In2020 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020,...

    2020

  21. [29]

    Cristian Romero, Dan Casas, Jesús Pérez, and Miguel A. Otaduy. Learning contact corrections for handle-based subspace dynamics.ACM Trans. Graph., 40(4):131:1–131:12, 2021. doi: 10.1145/3450626.3459875. URLhttps://doi.org/10.1145/3450626.3459875

    2021

  22. [30]

    Battaglia

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W. Battaglia. Learning mesh-based simulation with graph networks. In9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. URLhttps://op...

    2021

  23. [31]

    Battaglia

    Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter W. Battaglia. Learning to simulate complex physics with graph networks. InProceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event,...

    2020

  24. [32]

    Chiaramonte, and Miguel A

    Cristian Romero, Dan Casas, Maurizio M. Chiaramonte, and Miguel A. Otaduy. Contact-centric deformation learning.ACM Trans. Graph., 41(4):70:1–70:11, 2022. doi: 10.1145/3528223. 3530182. URLhttps://doi.org/10.1145/3528223.3530182

    2022

  25. [33]

    Otaduy, and Dan Casas

    Igor Santesteban, Miguel A. Otaduy, and Dan Casas. SNUG: self-supervised neural dynamic garments. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 8130–8140. IEEE, 2022. doi: 10.1109/ CVPR52688.2022.0079...

    2022

  26. [34]

    Transformer with implicit edges for particle-based physics simulation

    Yidi Shao, Chen Change Loy, and Bo Dai. Transformer with implicit edges for particle-based physics simulation. In Shai Avidan, Gabriel J. Brostow, Moustapha Cissé, Giovanni Maria Farinella, and Tal Hassner, editors,Computer Vision - ECCV 2022 - 17th European Conference, Tel Av...

    2022 doi

  27. [35]

    Otaduy, and Dan Casas

    Igor Santesteban, Nils Thuerey, Miguel A. Otaduy, and Dan Casas. Self-supervised collision handling via generative 3d garment models for virtual try-on. InIEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 11763– 11773. Comp...

    2021

  28. [36]

    Wang, Duygu Ceylan, Xin Sun, and Dinesh Manocha

    Qingyang Tan, Yi Zhou, Tuanfeng Y . Wang, Duygu Ceylan, Xin Sun, and Dinesh Manocha. A repulsive force unit for garment collision handling in neural networks. In Shai Avidan, Gabriel J. Brostow, Moustapha Cissé, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Visi...

    2022

  29. [37]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  30. [38]

    Towards multi-layered 3d garments animation

    Yidi Shao, Chen Change Loy, and Bo Dai. Towards multi-layered 3d garments animation. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 14315–14324. IEEE, 2023. doi: 10.1109/ICCV51070.2023.01321. URL https://doi.org/10.1...

    2023

  31. [39]

    A fast & robust solution for cubic & higher-order polynomials

    Cem Yuksel. A fast & robust solution for cubic & higher-order polynomials. InACM SIGGRAPH 2022 Talks, pages 1–2. 2022

    2022

  32. [40]

    No Comp

    Chong Zeng, Yue Dong, Pieter Peers, Hongzhi Wu, and Xin Tong. Renderformer: Transformer- based neural rendering of triangle meshes with global illumination. In Ginger Alford, Hao (Richard) Zhang, and Adriana Schulz, editors,Proceedings of the Special Interest Group on Computer...

    2025

  33. [42]

    Fully convolutional graph neural networks for parametric virtual try-on.Comput

    Raquel Vidaurre, Igor Santesteban, Elena Garces, and Dan Casas. Fully convolutional graph neural networks for parametric virtual try-on.Comput. Graph. Forum, 39(8):145–156, 2020. doi: 10.1111/CGF.14109. URLhttps://doi.org/10.1111/cgf.14109

    2020 doi

  34. [467]

    doi: 10.1007/978-3-031-20062-5\_26

    Springer, 2022. doi: 10.1007/978-3-031-20062-5\_26. URL https://doi.org/10. 1007/978-3-031-20062-5_26

    2022 doi

  35. [2020]

    URLhttp://proceedings.mlr.press/v119/sanchez-gonzalez20a.html

  36. [2025]

    URLhttps://openreview.net/forum?id=3BaJMRaPSx

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.