Pith. sign in

REVIEW 5 major objections 4 minor 50 references

Self-supervised Learning of Latent Space Dynamics

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A neural integrator trained by minimizing a full-space physics energy can replace implicit Euler entirely in latent space, yielding stable CPU real-time simulation of rods, shells, and solids.

desk verdict Solid latent-space integrator paper; the autoencoder-fidelity premise is unquantified, which is the main thing a referee should push on. read the letter →

arxiv 2507.07440 v1 pith:XFRQT4KZ submitted 2025-07-10 cs.GR

Yue Li , Gene Wei-Chin Lin , Egor Larionov , Aljaz Bozic , Doug Roble , Ladislav Kavan , Stelian Coros , Bernhard Thomaszewski
show 2 more authors
Tuur Stuyck Hsiao-yu Chen
This is my paper · ORCID
classification cs.GR
keywords latentspaceintegrationself-supervisedlearningsubspacesimulationimplicitEulerdeformableobjectsreal-timeautoencoderphysics-informedloss
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 a small neural network can serve as the time integrator for elastic rods, shells, and solids entirely inside a learned latent space, with no full-space computation at inference. The training signal is the same variational implicit-Euler incremental potential used by full-space simulation, evaluated by decoding the predicted latents; the paper argues this self-supervised loss is what gives the integrator stability and generalization to unseen boundary conditions. The payoff is practical: CPU inference at 0.272–0.720 ms per frame, orders of magnitude faster than a full-space implicit-Euler solver, which would make physically plausible deformable dynamics feasible on VR headsets and mobile devices. A sympathetic reader would take the central claim to be that a physics-aware loss plus targeted data augmentation can replace explicit numerical integration in a reduced space without sacrificing plausibility.

What carries the argument

The load-bearing object is the latent-space integrator, an MLP that predicts the next latent code directly from the previous two latent codes and three steps of boundary-condition parameters, paired with an autoencoder $D \circ E$ that maps full-space states to latents. The physics enters through the loss: exactly the objective minimized in full-space implicit-Euler simulation, $E_{\text{total}}(\mathbf{x}_t(\mathbf{z}_t(\mathbf{q};\theta_I)), \mathbf{x}_{t-1}, \mathbf{x}_{t-2})$, with positions obtained by decoding the predicted latent. Two training-time mechanisms carry the stability argument: noise injected into the previous two latents (10% of the batch standard deviation) samples off-equilibrium states so the integrator learns to recover, and per-sample energy normalization by the average velocity magnitude $|\mathbf{v}_{t-1}|$ balances gradient contributions across different energy regimes. The autoencoder's first encoder layer and last decoder layer are initialized with PCA bases to suppress high-frequency reconstruction artifacts, because reconstruction error would otherwise introduce spurious forces into the physics loss.

What would settle it

To settle whether the self-supervised signal is faithful, decode the trained latents on a held-out sequence and compare the gradient of the incremental potential at the decoded state with the gradient at the matching ground-truth full-space state; if the discrepancy is large enough to reorder candidate next states, the training signal is biased. A direct version of the same test is to retrain with a deliberately degraded autoencoder (for example, stopping early so reconstruction error rises) and check whether the integrator's rollouts drift from the boundary conditions, oscillate around steady states, or lose kinetic energy faster than shown in the paper's limitation figure.

Watch

Extended reading notes

Core claim

The central claim is that the map from previous latent states and boundary-condition history to the next latent state can be learned as an MLP integrator $I(z_{t-1}, z_{t-2}, p_t, p_{t-1}, p_{t-2}; \theta_I)$, trained by minimizing the full-space incremental potential $E_{\text{total}} = E_{\text{inertial}} + E_{\text{elastic}} + E_{\text{external}} + E_{\text{bc}}$ on decoded latents. During training the decoder is used to evaluate the physics loss; during inference the decoder and all full-space energies are bypassed entirely. The paper reports that adding uniform noise scaled to 10% of the batch standard deviation to the input latents stabilizes autoregressive rollouts for thousands of frames, and that normalizing per-sample loss by the average velocity of the previous two states fixes a data-imbalance problem. Evaluated on hair-like rods, cloth shells, and volumetric solids, the method produces stable, plausible motion at per-frame CPU costs of 0.27–0.72 ms, roughly three orders of magnitude below the reference solver, while generalizing to out-of-distribution boundary velocities.

Load-bearing premise

The load-bearing premise is that the autoencoder reconstructs full-space states with high enough fidelity that the full-space incremental potential evaluated on decoded latents is a faithful training signal; the paper itself says larger reconstruction errors introduce spurious forces that hinder self-supervised training.

Editorial extensions

If this is right

  • If the claim holds, deformable-object dynamics can run on CPUs at 30 FPS with per-frame costs around 0.27–0.72 ms, a speedup of roughly two to three orders of magnitude over full-space implicit Euler, which is enough for VR headsets and mobile platforms.
  • Under the same claim, the self-supervised incremental-potential loss prevents the per-step error accumulation that makes supervised latent-integrators trained with L2 loss diverge within about 100 frames, while the physics-based network stays stable for thousands of frames.
  • Because inference never evaluates full-space energies or derivatives, runtime is essentially an MLP forward pass (plus decoding when full-space positions are needed), so speed holds even for meshes with tens of thousands of degrees of freedom.
  • The same recipe applies across rods, shells, and solids, with only the elastic energy in the loss changing, so the method is a single framework rather than per-object tuning.
  • Training noise and energy normalization are necessary ingredients: without them, long rollouts drift from boundary conditions, oscillate at steady states, or bias low-energy configurations.

Reading between the lines

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

  • Beyond the paper, the 'decode during training only' pattern is a general recipe: any differentiable full-space simulator with a well-defined energy could supply the loss for a latent integrator, so contact, fracture, or fluid energies could be swapped in without changing the inference network.
  • Beyond the paper, the reported kinetic-energy shortfall relative to ground truth suggests a systematic damping bias inherited from the implicit-Euler objective; a natural test is to add a kinetic-energy or symplecticity penalty to the loss and measure whether energy conservation improves.
  • Beyond the paper, the fixed $\Delta t = 1/30$ s step and single-geometry latent subspaces are the next bottlenecks; conditioning the integrator on $\Delta t$ and using a geometry-agnostic latent representation would be direct extensions the paper leaves open.
  • Beyond the paper, the batch energy normalization points to a general principle for physics-based self-supervised learning: energy scales can vary by orders of magnitude within one batch, so a per-sample velocity normalization is a minimal way to keep low-energy configurations from being ignored.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper presents a subspace simulation framework for elastic rods, shells, and solids in which an autoencoder maps full-space configurations to a low-dimensional latent space and an MLP integrator predicts the next latent state autoregressively. The integrator is trained in a self-supervised manner by minimizing the full-space variational implicit Euler incremental potential evaluated on decoded latent states, rather than by regressing to ground-truth latents. Two auxiliary techniques are proposed: adding noise to integrator inputs during training to improve rollout stability, and a batch-wise energy normalization to handle data imbalance. Experiments cover five examples (two hair/rod cases, one cloth case, and two volumetric solid cases) and report CPU inference times between 0.272 ms and 0.720 ms per frame, with claims of stable long-horizon rollouts and generalization to unseen boundary conditions. Comparisons include full-space offline simulation, a supervised learning baseline, and a partial comparison to prior latent-space integrators.

Significance. If validated, the method would be a practically useful contribution: it offers a way to run nonlinear elastodynamics entirely in latent space at CPU real-time rates, avoiding full-space energy and derivative evaluations during inference. The self-supervised formulation is conceptually appealing because the training signal is the physical incremental potential rather than a ground-truth latent target, and the noise-augmentation and normalization ideas are simple and plausible. The paper also provides detailed timing tables and a clear implementation description. However, the evidence supporting the central claims is currently mostly qualitative: there are no quantitative reconstruction errors, energy errors, rollout errors, or error bars, and the test/train split is not cleanly held out for two of the five examples. The stress-test concern about decoder fidelity is therefore well placed: since Eq. (4) supervises the integrator through the decoded full-space energy, unquantified reconstruction distortion could train the integrator against spurious forces. The speedup claim is also overstated in one place.

major comments (5)
  1. [Sec. 4.6, Eq. (4)] The core training signal is the full-space incremental potential evaluated on decoded latents, and Sec. 4.6 states that 'high-fidelity reconstruction is crucial' because larger errors introduce spurious forces. Yet the paper reports no reconstruction error, no energy error, and no rollout error against ground truth anywhere; Tables 1 and 2 contain only timings. Please report (i) reconstruction error as mean/max vertex position error normalized by a characteristic length, (ii) the relative energy error |E_total(D(E(x))) - E_total(x)| / |E_total(x)| on training and test states, and (iii) per-frame rollout position error of the latent integrator versus the full-space reference for all five examples, with mean and standard deviation over at least three random seeds. Without these numbers, the claim that the self-supervised loss learns the correct physics is not verified.
  2. [Sec. 4.3] For the armadillo and bunny ears examples, the paper states 'We train on the initial 60% of the sequence ... and test on the complete sequence.' This means the test evaluation includes the training frames, so the claimed generalization to unseen motions is not measured on a held-out set. Please evaluate on the held-out 40% only and report separate metrics for interpolation (within the training distribution of boundary conditions) and extrapolation (outside that distribution).
  3. [Sec. 4.4, Table 2] The comparison with the latent-space integrators of Fulton et al. and Shen et al. does not actually run those methods. It times a single Jacobian-vector product for one vertex and compares that to the full proposed pipeline. This is not an end-to-end comparison and does not support the conclusion that the method is faster than prior latent-space integrators, since optimization-based baselines require repeated energy and derivative evaluations. Please report end-to-end timings of the actual baseline methods on the same examples and hardware, or clearly relabel the table as a component-level timing breakdown rather than a method comparison.
  4. [Sec. 4.5, Figs. 8-11] The ablation studies for supervised versus self-supervised learning, training noise, and data balancing are presented only through selected video frames. Please add quantitative metrics for each ablation: rollout position error versus ground truth, energy drift or kinetic-energy profile, and boundary-condition error, with error bars over multiple seeds. In addition, the supervised baseline in Fig. 8 uses different network capacities across rows; please include a matched-capacity comparison so the stability difference is not confounded by model size.
  5. [Sec. 4.4, Table 1] The text states that the method 'improves the performance by 3 orders of magnitude.' The speedups computed from Table 1 are approximately 372x for the hair examples, 280x for the armadillo, 264x for bunny ears, and 69x for cloth, i.e., between roughly 1.8 and 2.6 orders of magnitude. Please correct this claim to match the reported timings.
minor comments (4)
  1. [Eq. (6)] The normalization divides the loss by |v_{t-1}|, which has units of velocity, not energy. Please clarify the rationale for this choice and report sensitivity to the normalization; also note that the average velocity can be near zero for symmetric motions, so a small regularizer may be needed to avoid division by zero.
  2. [Tables 1 and 2] Timings are reported as single numbers from a workstation with an Nvidia 3080 and an AMD Threadripper CPU. Please specify the exact CPU model and thread count, state whether the timings are averaged over many frames, and report variance, since CPU real-time performance is a central claim.
  3. [Sec. 5, Fig. 12] The admitted kinetic-energy loss relative to ground truth is shown only as a qualitative side-by-side. Please add a quantitative plot of kinetic energy (and ideally total energy) versus time for the prediction and the reference, with error bars across seeds.
  4. [Sec. 4.1] The rod translation test is described as starting with a 'fast and out-of-distribution velocity' that is then decreased by an order of magnitude. Please state explicitly how this test velocity relates to the 10-120 m/s training range, since the current description makes it unclear whether the test begins outside or inside the training distribution.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the physics-energy training signal is external to the network and the integrator is not regressing its own target.

full rationale

The derivation is self-contained and does not reduce to its own inputs. The integrator is trained by minimizing the full-space incremental potential E_total (Eq. 4) evaluated on decoded latents, with physical parameters taken from real-world values; it is not trained to regress ground-truth latents, so the predicted next state is not a fitted rename of the data. The same E_total also generates the offline simulation data, but this is a consistent physical model rather than a circularity: at inference the network is evaluated without access to E_total or an optimizer, and the comparison to offline simulation is an approximation test. Self-citations to Martin et al. 2011 (variational implicit Euler), Fulton et al. 2019 (PCA initialization), and Li et al. 2024 (open-source simulation framework) are standard methodological references and are not load-bearing uniqueness claims. Sec. 4.6 asserts low autoencoder reconstruction error without reporting quantitative numbers, and Sec. 5 admits kinetic-energy loss relative to ground truth; these are missing-evidence and limitation statements, not circular reductions. No equation or fitted parameter is exhibited that makes a claimed prediction equivalent to an input by construction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The method's contributions are algorithmic: a learned latent integrator, a physics-based loss, and two training heuristics. The listed free parameters are hand-set or architectural choices, none fitted to achieve the target result in a circular way.

free parameters (6)
  • Latent dimensionality per example = 4 for rods and cloth, 12 for armadillo, 8 for bunny ears
    Chosen by hand; sets the subspace size and runtime. Reported in Tables 1 and 2.
  • Training noise scale = 10% of per-batch standard deviation, uniform distribution
    Ad hoc choice in the Data Augmentation paragraph of Sec 3.3; determines how far perturbed states are from equilibrium and is critical for autoregressive stability.
  • Energy normalization divisor = |v_{t-1}|, the average vertex velocity
    Introduced without derivation in Eq 6 to balance energy scales; changes the loss landscape and is central to the reported stability.
  • Boundary penalty weight w_bc = 1e5
    Fixed penalty weight for cloth and solid boundary conditions (Sec 3.3); a hand-set constant.
  • Network architectures = e.g., decoder 512,512,512,50 and integrator 512,512,512 for hair rotation
    Architecture sizes in Table 2 are design choices affecting capacity and runtime.
  • Training hyperparameters = Learning rate 1e-4, autoencoder 20k epochs, integrator 10k epochs, batch size 500
    Standard choices stated in Sec 3.3; not fitted to data but set by hand.
assumptions (4)
  • domain assumption The full-space implicit Euler incremental potential (Eq 1) correctly models the target dynamics.
    This potential generates the training data and is the loss (Eq 4); if it misrepresents the physics, the learned integrator inherits the error.
  • domain assumption The autoencoder latent space preserves enough information for the energy of decoded states to be an accurate training signal.
    Invoked in Sec 3.2 and Sec 4.6; reconstruction error must be small to avoid spurious forces.
  • domain assumption The next latent depends only on the previous two latents and boundary history, a second-order Markov assumption.
    The integrator input in Eq 3 assumes this reduced state is sufficient for implicit Euler.
  • domain assumption Backpropagation through the decoder into the latent integrator converges to a useful minimizer despite a non-convex energy landscape.
    Standard deep-learning assumption; the paper does not analyze the loss landscape.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-supervised Learning of Latent Space Dynamics." pith.science (2026). https://pith.science/paper/XFRQT4KZ

@misc{pith2026250707440,
  author       = {Pith},
  title        = {Pith review of: Self-supervised Learning of Latent Space Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XFRQT4KZ}},
  note         = {Machine review of arXiv:2507.07440}
}
read the original abstract

Modeling the dynamic behavior of deformable objects is crucial for creating realistic digital worlds. While conventional simulations produce high-quality motions, their computational costs are often prohibitive. Subspace simulation techniques address this challenge by restricting deformations to a lower-dimensional space, improving performance while maintaining visually compelling results. However, even subspace methods struggle to meet the stringent performance demands of portable devices such as virtual reality headsets and mobile platforms. To overcome this limitation, we introduce a novel subspace simulation framework powered by a neural latent-space integrator. Our approach leverages self-supervised learning to enhance inference stability and generalization. By operating entirely within latent space, our method eliminates the need for full-space computations, resulting in a highly efficient method well-suited for deployment on portable devices. We demonstrate the effectiveness of our approach on challenging examples involving rods, shells, and solids, showcasing its versatility and potential for widespread adoption.

Figures

Figures reproduced from arXiv: 2507.07440 by the authors.

Figure 1
Figure 1. Real-time Simulation: We propose a novel latent space integrator that allows for robust and stable CPU-real-time (0.489 ms per frame) autoregressive inference of dynamic motions. Boundary conditions are shown as red spheres and the colored squares indicate low-dimensional latents. dynamics induced by interactions between digital humans and virtual objects. However, these sim￾ulations are computationally expensive, a… view at source ↗
Figure 2
Figure 2. Method Overview: Our approach encodes full-space simulation data into low-dimensional latent vectors. We then learn an integrator in latent space in a self-supervised setting by minimizing the incremental potential in full space. During testing, as indicated by the blue dashed region, our integrator does not require any full-space computation and therefore runs in real-time on CPU. Proc. ACM Comput. Graph. Interact.… view at source ↗
Figure 3
Figure 3. Accelerating and Decelerating Rods: Thanks to our self-supervised learning strategy, the rods can adapt to different velocities and accelerations unseen during training and produce plausible simulations. 4.2 Shell Simulation We now turn to cloth simulation by applying our approach to simulate a discrete elastic shell by incorporating the StVK stretching energy and hinge-based bending energy under the influence of gr… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Rods Under Rotation Boundary Conditions: Our method produces plausible dynamics when adapting to changes in different boundary conditions, i.e. rods swing much higher at large rotation speeds (third row) but do not accumulate large momentum when the rotation speed is s…
Figure 5
Figure 5. Figure 5: Moving Cloth: Two pinned corners of a piece of cloth are being moved along a linear trajectory. Our method is able to generate stable simulation outcomes for a longer time horizon, despite being trained on only half of the sequence [PITH_FULL_IMAGE:figures/full_fig_p0…
Figure 6
Figure 6. Figure 6: Swinging Armadillo: Our method simulates the elastic deformation of an upside-down armadillo under gravity, with the prescribed motion of the boundary vertices shown in red. The simulation accurately captures the nonlinear behavior of the dynamic motion and demonstrate…
Figure 7
Figure 7. Figure 7: Bunny Ears: We simulate the elastic deformation of a pair of bunny ears attached to a rigid bunny head, with the prescribed motion of the head shown in green. The simulation captures the dynamic movement of the ears under gravity and predicts their motion on unseen seq…
Figure 8
Figure 8. Figure 8: Autoregressive Inference: We compare our approach with a supervised learning strategy. Supervised learning lacks stability when it comes to long rollouts due to per-step error accumulation. While increasing network capacity mitigates this issue to some extent (row 1-3)…
Figure 9
Figure 9. Figure 9: Data Balancing: Without balancing data between different system energies, the training process tends to favor high energy states due to the correspondingly large gradients. Using our balanced loss function improves the prediction, especially on low energy states. Mikló…
Figure 10
Figure 10. Figure 10: Steady State Solution: A cantilever beam clamped from one end is deformed under gravity. After training on the initial 100 frames of the simulation sequence, we continue the inference of the network for an additional 200 frames. The last frames of the inference are sh…
Figure 11
Figure 11. Figure 11: Training Noise: Without training noise, the network predictions gradually deviate from the boundary conditions and generate implausible states containing significantly large deformations. With training noise, our network predictions satisfy boundary conditions at all …
Figure 12
Figure 12. Figure 12: Limitation: our prediction (left), albeit stable, exhibits kinetic energy loss compared to the ground truth data (right). Edilson De Aguiar, Leonid Sigal, Adrien Treuille, and Jessica K Hodgins. 2010. Stable spaces for real-time clothing. ACM Transactions on Graphics …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 46 canonical work pages

  1. [11]

    InSIGGRAPH Asia 2023 Conference Papers

    LiCROM: Linear-Subspace Continuous Reduced Order Modeling with Neural Fields. InSIGGRAPH Asia 2023 Conference Papers. 1–12. Peter Yichen Chen, Maurizio M Chiaramonte, Eitan Grinspun, and Kevin Carlberg

    work page 2023

  2. [12]

    Model reduction for the material point method via an implicit neural representation of the deformation map.J. Comput. Phys.478 (2023), 111908. Proc. ACM Comput. Graph. Interact. Tech., Vol. 8, No. 4, Article

    work page 2023

  3. [13]

    Nuttapong Chentanez, Miles Macklin, Matthias Müller, Stefan Jeschke, and Tae-Yong Kim

    Crom: Continuous reduced-order modeling of pdes using implicit neural representations.arXiv preprint arXiv:2206.02607(2022). Nuttapong Chentanez, Miles Macklin, Matthias Müller, Stefan Jeschke, and Tae-Yong Kim

    arXiv 2022

  4. [14]

    Yinwei Du, Yue Li, Stelian Coros, and Bernhard Thomaszewski

    Stable spaces for real-time clothing.ACM Transactions on Graphics (ToG)29, 4 (2010), 1–9. Yinwei Du, Yue Li, Stelian Coros, and Bernhard Thomaszewski

    work page 2010

  5. [15]

    Lawson Fulton, Vismay Modi, David Duvenaud, David IW Levin, and Alec Jacobson

    Nonlinear compliant modes for large- deformation analysis of flexible structures.ACM Transactions on Graphics42, 2 (2022), 1–11. Lawson Fulton, Vismay Modi, David Duvenaud, David IW Levin, and Alec Jacobson

    work page 2022

  6. [16]

    InACM SIGGRAPH 2024 Conference Papers

    Contourcraft: Learning to resolve intersections in neural multi-garment simulations. InACM SIGGRAPH 2024 Conference Papers. 1–10. Artur Grigorev, Bernhard Thomaszewski, Michael J Black, and Otmar Hilliges

    work page 2024

  7. [18]

    Oshri Halimi, Tuur Stuyck, Donglai Xiang, Timur M Bagautdinov, He Wen, Ron Kimmel, Takaaki Shiratori, Chenglei Wu, Yaser Sheikh, and Fabian Prada

    Subspace clothing simulation using adaptive bases.ACM Transactions on Graphics (TOG)33, 4 (2014), 1–9. Oshri Halimi, Tuur Stuyck, Donglai Xiang, Timur M Bagautdinov, He Wen, Ron Kimmel, Takaaki Shiratori, Chenglei Wu, Yaser Sheikh, and Fabian Prada

    work page 2014

  8. [19]

    Pattern-Based Cloth Registration and Sparse-View Animation.ACM Trans. Graph. 41, 6 (2022), 196–1. David Harmon and Denis Zorin

    work page 2022

Show all 50 references
  1. [21]

    Sagar Imambi, Kolla Bhanu Prakash, and GR Kanagachidambaresan

    Real-time Physically Guided Hair Interpolation.ACM Transactions on Graphics (TOG)43, 4 (2024), 1–11. Sagar Imambi, Kolla Bhanu Prakash, and GR Kanagachidambaresan

    2024

  2. [23]

    ACM Transactions on Graphics (ToG)31, 4 (2012), 1–10

    Fast automatic skinning transformations. ACM Transactions on Graphics (ToG)31, 4 (2012), 1–10. Theodore Kim and John Delaney

    2012

  3. [24]

    Egor Larionov, Ye Fan, and Dinesh K

    Subspace fluid re-simulation.ACM Transactions on Graphics (TOG)32, 4 (2013), 1–9. Egor Larionov, Ye Fan, and Dinesh K. Pai

    2013

  4. [25]

    Graph.40, 2, Article 15 (April 2021), 17 pages

    Frictional Contact on Smooth Elastic Solids.ACM Trans. Graph.40, 2, Article 15 (April 2021), 17 pages. https://doi.org/10.1145/3446663 Kookjin Lee and Kevin T Carlberg

    2021 doi

  5. [26]

    Model reduction of dynamical systems on nonlinear manifolds using deep convolutional autoencoders.J. Comput. Phys.404 (2020), 108973. Minchen Li, Zachary Ferguson, Teseo Schneider, Timothy R Langlois, Denis Zorin, Daniele Panozzo, Chenfanfu Jiang, and Danny M Kaufman

    2020

  6. [27]

    ACM Trans

    Incremental potential contact: intersection-and inversion-free, large-deformation dynamics. ACM Trans. Graph.39, 4 (2020),

    2020

  7. [28]

    InSIGGRAPH Asia 2023 Conference Papers

    Subspace-preconditioned gpu projective dynamics with contact for cloth simulation. InSIGGRAPH Asia 2023 Conference Papers. 1–12. Yue Li et al

    2023

  8. [29]

    WuKong. (2024). https://github.com/liyuesolo/Wukong2024. Yue Li, Marc Habermann, Bernhard Thomaszewski, Stelian Coros, Thabo Beeler, and Christian Theobalt

    2024

  9. [30]

    Qing Lyu, Menglei Chai, Xiang Chen, and Kun Zhou

    Accelerate Neural Subspace- Based Reduced-Order Solver of Deformable Simulation by Lipschitz Optimization.ACM Transactions on Graphics (TOG) 43, 6 (2024), 1–10. Qing Lyu, Menglei Chai, Xiang Chen, and Kun Zhou

    2024

  10. [31]

    Pingchuan Ma, Peter Yichen Chen, Bolei Deng, Joshua B Tenenbaum, Tao Du, Chuang Gan, and Wojciech Matusik

    Real-time hair simulation with neural interpolation.IEEE Transactions on Visualization and Computer Graphics28, 4 (2020), 1894–1905. Pingchuan Ma, Peter Yichen Chen, Bolei Deng, Joshua B Tenenbaum, Tao Du, Chuang Gan, and Wojciech Matusik

    2020

  11. [33]

    Jorge Nocedal and Stephen J Wright

    Simplicits: Mesh-Free, Geometry-Agnostic Elastic Simulation.ACM Transactions on Graphics (TOG)43, 4 (2024), 1–11. Jorge Nocedal and Stephen J Wright. 1999.Numerical optimization. Springer. Shaowu Pan, Steven L Brunton, and J Nathan Kutz

    2024

  12. [34]

    Alex Pentland and John Williams

    Neural implicit flow: a mesh-agnostic dimensionality reduction paradigm of spatio-temporal data.Journal of Machine Learning Research24, 41 (2023), 1–60. Alex Pentland and John Williams

    2023

  13. [35]

    InSIGGRAPH Asia 2023 Conference Papers

    Learning Contact Deformations with General Collider Descriptors. InSIGGRAPH Asia 2023 Conference Papers. 1–10. Cristian Romero, Dan Casas, Maurizio M Chiaramonte, and Miguel A Otaduy

    2023

  14. [36]

    ACM Transactions on Graphics (TOG)41, 4 (2022), 1–11

    Contact-centric deformation learning. ACM Transactions on Graphics (TOG)41, 4 (2022), 1–11. Cristian Romero, Dan Casas, Jesús Pérez, and Miguel Otaduy

    2022

  15. [37]

    Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia

    Learning contact corrections for handle-based subspace dynamics.ACM Transactions on Graphics (TOG)40, 4 (2021), 1–12. Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia

    2021

  16. [38]

    InACM SIGGRAPH 2023 Conference Proceedings

    Data-Free Learning of Reduced-Order Kinematics. InACM SIGGRAPH 2023 Conference Proceedings. 1–9. Siyuan Shen, Yin Yang, Tianjia Shao, He Wang, Chenfanfu Jiang, Lei Lan, and Kun Zhou

    2023

  17. [39]

    Eftychios Sifakis and Jernej Barbic

    High-order differentiable autoencoder for nonlinear model reduction.ACM Transactions on Graphics40, 4 (2021). Eftychios Sifakis and Jernej Barbic. 2012a. FEM simulation of 3D deformable solids: a practitioner’s guide to theory, discretization and model reduction. InACM SIGGRAP...

    2021

  18. [40]

    InProceedings of the 2005 ACM SIGGRAPH/Eurographics symposium on Computer animation

    Robust quasistatic finite elements and flesh simulation. InProceedings of the 2005 ACM SIGGRAPH/Eurographics symposium on Computer animation. 181–190. Adrien Treuille, Andrew Lewis, and Zoran Popović

    2005

  19. [43]

    Cem Yuksel, Scott Schaefer, and John Keyser

    A unified approach for subspace simulation of deformable bodies in multiple domains.ACM Transactions on Graphics (TOG)34, 6 (2015), 1–9. Cem Yuksel, Scott Schaefer, and John Keyser

    2015

  20. [44]

    Jonas Zehnder, Yue Li, Stelian Coros, and Bernhard Thomaszewski

    Hair meshes.ACM Transactions on Graphics (TOG)28, 5 (2009), 1–7. Jonas Zehnder, Yue Li, Stelian Coros, and Bernhard Thomaszewski

    2009

  21. [45]

    Meng Zhang, Tuanfeng Y

    Ntopo: Mesh-free topology optimization using implicit neural representations.Advances in Neural Information Processing Systems34 (2021), 10368–10381. Meng Zhang, Tuanfeng Y. Wang, Duygu Ceylan, and Niloy J. Mitra

    2021

  22. [46]

    Graph.40, 6, Article 235 (Dec

    Dynamic neural garments.ACM Trans. Graph.40, 6, Article 235 (Dec. 2021), 15 pages. https://doi.org/10.1145/3478513.3480497 Mianlun Zheng and Jernej Barbic

    2021

  23. [47]

    Zhongtian Zheng, Tongtong Wang, Qijia Feng, Zherong Pan, Xifeng Gao, and Kui Wu

    Multi-Resolution Real-Time Deep Pose-Space Deformation.ACM Transactions on Graphics (TOG)43, 6 (2024), 1–11. Zhongtian Zheng, Tongtong Wang, Qijia Feng, Zherong Pan, Xifeng Gao, and Kui Wu

    2024

  24. [48]

    Zeshun Zong, Xuan Li, Minchen Li, Maurizio M Chiaramonte, Wojciech Matusik, Eitan Grinspun, Kevin Carlberg, Chenfanfu Jiang, and Peter Yichen Chen

    Proxy Asset Generation for Cloth Simulation in Games.ACM Transactions on Graphics (TOG)43, 4 (2024), 1–12. Zeshun Zong, Xuan Li, Minchen Li, Maurizio M Chiaramonte, Wojciech Matusik, Eitan Grinspun, Kevin Carlberg, Chenfanfu Jiang, and Peter Yichen Chen

    2024

  25. [49]

    InSIGGRAPH Asia 2023 Conference Papers

    Neural stress fields for reduced-order elastoplasticity and fracture. InSIGGRAPH Asia 2023 Conference Papers. 1–11. Proc. ACM Comput. Graph. Interact. Tech., Vol. 8, No. 4, Article

    2023

  26. [57]

    Publication date: August 2025

    2025

  27. [2003]

    InProceedings of the 2003 ACM SIGGRAPH/Eurographics symposium on Computer animation

    Discrete shells. InProceedings of the 2003 ACM SIGGRAPH/Eurographics symposium on Computer animation. Citeseer, 62–67. Fabian Hahn, Bernhard Thomaszewski, Stelian Coros, Robert W Sumner, Forrester Cole, Mark Meyer, Tony DeRose, and Markus Gross

    2003

  28. [2005]

    Venant-Kirchhoff deformable models.ACM transactions on graphics (TOG)24, 3 (2005), 982–990

    Real-time subspace integration for St. Venant-Kirchhoff deformable models.ACM transactions on graphics (TOG)24, 3 (2005), 982–990. Jernej Barbič, Funshing Sin, and Eitan Grinspun

    2005

  29. [2006]

    Jiahong Wang, Yinwei Du, Stelian Coros, and Bernhard Thomaszewski

    Model reduction for real-time fluids.ACM Transactions on Graphics (TOG)25, 3 (2006), 826–834. Jiahong Wang, Yinwei Du, Stelian Coros, and Bernhard Thomaszewski

    2006

  30. [2008]

    InACM SIGGRAPH 2008 papers

    Discrete elastic rods. InACM SIGGRAPH 2008 papers. 1–12. Hugo Bertiche, Meysam Madadi, and Sergio Escalera

    2008

  31. [2009]

    In ACM SIGGRAPH 2009 papers

    Deformable object animation using reduced optimal control. In ACM SIGGRAPH 2009 papers. 1–9. Jernej Barbič and Doug L James

    2009

  32. [2010]

    Discrete viscous threads.ACM Transactions on graphics (TOG)29, 4 (2010), 1–10. Proc. ACM Comput. Graph. Interact. Tech., Vol. 8, No. 4, Article

    2010

  33. [2011]

    In ACM SIGGRAPH 2011 papers

    Example-based elastic materials. In ACM SIGGRAPH 2011 papers. 1–8. Vismay Modi, Nicholas Sharp, Or Perel, Shinjiro Sueda, and David IW Levin

    2011

  34. [2012]

    Otman Benchekroun, Jiayi Eris Zhang, Siddartha Chaudhuri, Eitan Grinspun, Yi Zhou, and Alec Jacobson

    Interactive editing of deformable simulations.ACM Transactions on Graphics (TOG)31, 4 (2012), 1–8. Otman Benchekroun, Jiayi Eris Zhang, Siddartha Chaudhuri, Eitan Grinspun, Yi Zhou, and Alec Jacobson

    2012

  35. [2013]

    Kris K Hauser, Chen Shen, and James F O’Brien

    Subspace integration with local deformations.ACM Transactions on Graphics (TOG) 32, 4 (2013), 1–10. Kris K Hauser, Chen Shen, and James F O’Brien

    2013

  36. [2014]

    Yue Chang, Peter Yichen Chen, Zhecheng Wang, Maurizio M Chiaramonte, Kevin Carlberg, and Eitan Grinspun

    A reduced model for interactive hairs.ACM Transactions on Graphics (TOG)33, 4 (2014), 1–11. Yue Chang, Peter Yichen Chen, Zhecheng Wang, Maurizio M Chiaramonte, Kevin Carlberg, and Eitan Grinspun

    2014

  37. [2015]

    ACM Transactions on Graphics (TOG)34, 4 (2015), 1–11

    Linear subspace design for real-time shape deformation. ACM Transactions on Graphics (TOG)34, 4 (2015), 1–11. Kui Wu and Cem Yuksel

    2015

  38. [2020]

    Hugo Bertiche, Meysam Madadi, and Sergio Escalera

    Pbns: Physically based neural simulator for unsupervised garment pose space deformation.arXiv preprint arXiv:2012.11310(2020). Hugo Bertiche, Meysam Madadi, and Sergio Escalera

    2020 arXiv

  39. [2021]

    PyTorch.Programming with TensorFlow: solution for edge computing applications(2021), 87–104. Proc. ACM Comput. Graph. Interact. Tech., Vol. 8, No. 4, Article

    2021

  40. [2022]

    Gaurav Bhokare, Eisen Montalvo, Elie Diaz, and Cem Yuksel

    Neural cloth simulation.ACM Transactions on Graphics (TOG) 41, 6 (2022), 1–14. Gaurav Bhokare, Eisen Montalvo, Elie Diaz, and Cem Yuksel

    2022

  41. [2023]

    Miklós Bergou, Basile Audoly, Etienne Vouga, Max Wardetzky, and Eitan Grinspun

    Fast Complementary Dynamics via Skinning Eigenmodes.ACM Transactions on Graphics (TOG)42, 4 (2023), 1–21. Miklós Bergou, Basile Audoly, Etienne Vouga, Max Wardetzky, and Eitan Grinspun

    2023

  42. [2024]

    InACM SIGGRAPH 2024 Conference Papers

    Real-Time Hair Rendering with Hair Meshes. InACM SIGGRAPH 2024 Conference Papers. 1–10. Menglei Chai, Changxi Zheng, and Kun Zhou

    2024

Pith tools

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