Pith. sign in

REVIEW 5 major objections 4 minor 1 cited by

Hybrid Neural-MPM for Interactive Fluid Simulations in Real-Time

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A hybrid that alternates between a learned graph-network simulator and a material-point-method fallback makes real-time interactive 2D/3D fluid simulation practical while keeping rollout error low and letting users steer flow with sketches.

desk verdict Sensible hybrid simulator with a real-but-modest contribution; the fallback idea deserves a look, but the real-time numbers and control evaluation need tightening before the claims carry weight. read the letter →

arxiv 2505.18926 v1 pith:5MVM53GK submitted 2025-05-25 cs.LG physics.flu-dyn

Jingxuan Xu , Hong Huang , Chuhang Zou , Manolis Savva , Yunchao Wei , Wuyang Chen This is my paper · ORCID
classification cs.LGphysics.flu-dyn
keywords fluidsimulationmaterialpointmethodgraphneuralnetworkphysicsreal-timediffusionmodelinteractivecontrolhybridsolver
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

The paper tries to establish that real-time interactive fluid simulation does not have to choose between learned speed and numerical accuracy. Its proposed system runs a graph-neural-network physics model at deliberately coarse spatial and temporal resolution, which makes each step cheap, and monitors a cheap statistic, the cosine similarity of particle accelerations over the last ten steps, to decide when to fall back to a full material-point-method solve. On seven 2D and 3D scenarios covering water, sand, ramps, and rigid obstacles, the hybrid is claimed to cut latency by 11-29% relative to MPM while keeping rollout error below that of the neural-only model. A second component, a diffusion-based controller trained on force fields recovered by reversing simulations, lets a user steer fluid motion by drawing freehand arrows or oval shapes. If these claims hold, interactive graphics, virtual-reality, and design applications get a practical way to combine plausibility, speed, and controllability.

What carries the argument

The load-bearing mechanism is the fallback safeguard in the hybrid solver. A cheap fluid-complexity statistic, the average cosine similarity of per-particle accelerations over a $\delta_t=10$ step window, is monitored during neural-physics rollouts; the paper reports a Spearman correlation of $-0.3902$ between this statistic and simulation error in the Water 2D scenario, and uses it to trigger an MPM update when the similarity crosses the tuned threshold $r_c=0.8$. The second mechanism is the reversed-simulation data generator: given a forward MPM trajectory, Equation (2), $\ddot p_t = ((p_{t-1}-p_t)-\dot p_t\,\Delta t)/(\Delta t)^2 - g$, solves the external accelerations that would reverse the motion, producing ground-truth force fields for training the diffusion controller. Together these two components convert accuracy fallback and user control into a single real-time pipeline.

What would settle it

Go through held-out trajectories from WaterRamps, Sand 3D, and Water-Sand and compare each trigger decision with the neural model's actual rollout error: if many steps with similarity above 0.8 are followed by large error, or similarity below 0.8 coincide with accurate predictions, the safeguard is not measuring what the hybrid needs.

Watch

Extended reading notes

Core claim

The central claim is that the neural simulator and the numerical solver are complementary in time, not competitors. The paper's hybrid solver rolls out a GNN-based neural physics model trained at a reduced particle count and a doubled time step; because it predicts per-particle accelerations, the rollout is fast. A safeguard then compares, over windows of ten steps, the cosine similarity between successive per-particle acceleration vectors. When that similarity falls below the threshold $r_c=0.8$, meaning the flow has become too complex for the learned model, the system switches to the MPM solver for the next update. This single mechanism is said to suppress the long-run error accumulation of the neural model while preserving most of the speed gain: on Water 2D the hybrid finishes 1000 steps in 676.4 ms versus 1931.1 ms for the full neural physics, with a lower final grid RMSE. Separately, the paper claims a diffusion-based Fluid ControlNet, trained on force fields solved by reversing an MPM trajectory, generates external force fields from user sketches and aligns fluid particles with the sketched motion or shape better than a constant-force baseline.

Load-bearing premise

The whole speed-accuracy trade-off rests on one threshold: a cosine-similarity value of 0.8 over a ten-step window, tuned on a single water scene, is assumed to mark the moment when the learned simulator starts making errors in every other material and scenario.

Editorial extensions

If this is right

  • Real-time interactive tools, such as games, VR, and live design previews, can run the neural model for most frames and invoke the exact MPM solver only during chaotic intervals, lowering average frame cost without sacrificing long-horizon stability.
  • Artists can specify fluid behavior with freehand arrows for motion direction and oval outlines for target regions, and the diffusion controller turns those sketches into force fields automatically, removing trial-and-error force tuning.
  • Because the neural physics component is trained once at low resolution, the reported latency gains compound as scenes grow, since the expensive MPM solver is used only under the trigger.
  • The same hybrid pattern, a fast learned predictor with a cheap confidence-triggered fallback to an exact solver, is claimed to transfer across the seven benchmark scenarios covering different materials, obstacle interactions, and 2D/3D domains.

Reading between the lines

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

  • The trigger statistic is a proxy for prediction confidence; replacing it with a learned uncertainty estimator or ensemble disagreement could tighten the reported trade-off, since the current proxy's Spearman correlation of $-0.39$ leaves room for misordered decisions.
  • The reversed-simulation strategy for generating control data is solver-agnostic: the same acceleration fields could train controllers for SPH, FLIP, or Eulerian solvers, which would let the control pipeline outlive the MPM choice.
  • A practical deployment would likely need per-scenario calibration of $r_c$ rather than a fixed 0.8, or a threshold that adapts over rollout time, to cover scenes outside the tuning set.
  • Extending the controller beyond the fixed 100-step window, the paper's stated limitation, is a natural next test: a diffusion model that decides when to stop applying force could handle harder or easier sketches than the fixed-length control.
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 proposes a hybrid fluid simulation system that combines a graph-neural-network (GNN) based neural physics simulator, operating at reduced spatial and temporal resolutions, with a Material Point Method (MPM) fallback triggered by a fluid-complexity metric. It further introduces a diffusion-based Fluid ControlNet, trained via a reverse simulation strategy, to generate external force fields from user sketches for interactive control. Experiments on 2D/3D water, sand, and mixed scenarios report improved error-latency trade-offs relative to neural-physics-only and MPM-only baselines, and control results that visually and quantitatively align with user sketches.

Significance. The hybrid fallback architecture is a practical and timely idea that could make neural-physics simulators more robust in deployed settings, and the reverse-simulation data-generation pipeline for control is creative and potentially reusable. The authors transparently report training costs and promise public release of code, models, and data. If the load-bearing concerns about latency accounting and control-direction consistency are resolved, the system would be a useful step toward real-time interactive fluid simulation. However, as presented, the central quantitative claims are not fully supported by the reported measurements.

major comments (5)
  1. [§B.1, §4.2] The reported neural-physics latency does not include the per-step cost of constructing the neighbor graph, even though the paper states that edges are dynamically updated every step (Appendix A.2: "edges in the graph are still dynamically updated by comparing the current particle-wise distances to R"). The aggregation is approximated as a matrix product A·o, which assumes the adjacency matrix is already available; building A requires a neighbor search that GNS-style simulators carry out at every rollout step. Since the claimed latency advantage over MPM is only 11–29% (Section 4.2, Figure 10), a per-step graph-construction cost of even a fraction of a millisecond could erase the margin. The p2g/g2p transfer cost when the MPM fallback fires is also not stated. The real-time claim is therefore not supported by the measurements as presented.
  2. [§3.2.2, §4.3] The training force fields are solved to reverse a forward trajectory ("restore positions ... reversely, from X_{Tctr} to X_1", Eq. (2)), while the user sketch is generated from the forward trajectory X_1...X_{Tctr}: arrows connect centroids at t=1 and t=T, and ovals mark the region at t=T. The baseline in §4.3 also computes a force by "moving particles from X_T to X_1". This is a directional inconsistency: a model trained to output reverse forces would, if applied forward, push particles away from the sketch target. Please clarify how the network learns to produce forward forces from reverse-force supervision, or state explicitly that forces are applied in reverse time at inference. If Eq. (2) is intended to compute forward accelerations, the sign of the velocity term appears incorrect.
  3. [§3.1.2, Table 1, Figure 5] The fallback threshold rc=0.8 is selected on Water 2D (Table 1), and Figure 5 reports a Spearman correlation of only -0.39 between the fluid-complexity metric and simulation error on that same scenario. The paper then applies rc=0.8 to all scenarios in Figure 10 without reporting per-scenario trigger frequency, sensitivity to rc, or the metric-error correlation on the other domains. A weak and potentially non-transferable trigger could either invoke MPM too often (destroying the latency benefit) or too rarely (accumulating error). Please provide evidence that the metric and threshold generalize, or adapt the threshold per scenario.
  4. [§4.3] The paper claims "interactive fluid control" and "real-time interactive applications," but no latency measurements are reported for the Fluid ControlNet or the interactive control loop. Since §3.2.3 states that whenever a user provides a control sketch the system falls back to MPM, the interactive phase may not meet real-time constraints. Please report control-loop latency (including sketch encoding, diffusion sampling, and MPM simulation) or revise the interactive claim accordingly.
  5. [§4.3, Table 3] The control "ground truth" is generated by the same reversed-simulation pipeline used to train the Fluid ControlNet. The evaluation therefore measures consistency with the training-data generation process, not an independently defined notion of successful control. An independent baseline (e.g., animator-specified targets or a different control-solver method) would strengthen the claim. As presented, the control numbers support internal self-consistency but not external validity of the control mechanism.
minor comments (4)
  1. [Abstract] The phrase "(-11~29% latency)" is ambiguous; clarify whether these values are latency reductions, absolute latencies, or a range, and ensure consistency with the numbers in Section 4.2 (e.g., the 29.8% reduction reported for Water-Sand 2D).
  2. [Conclusion] The word "revserve" in the conclusion should be "reverse."
  3. [§3.2.2] The paragraph says "we have two highly nontrivial sub-tasks" but then lists three steps; rephrase to avoid the numerical inconsistency.
  4. [§3.2.2 and §B.4] The symbol for the control horizon is written as both Tctl and Tctr; unify the notation.

Circularity Check

1 steps flagged · score 6.0 of 10

Hybrid RMSE gains are forced by the MPM-fallback definition; latency and control results remain independent.

  1. self definitional [Section 3.1.2, Eq. (1); Table 1; Figure 7]
    "To fuse the strengths of both approaches, we make our simulator hybrid. We primarily leverage neural physics for fast updates, but incorporate a safeguard mechanism to fall back to MPM in challenging scenarios and to empirically ensure simulation quality: Xt+1 = Neural Physics Update if update is 'good'; Fallback to MPM Update otherwise. ... when increasing our threshold rc (i.e. MPM will be more frequently triggered), the simulation fidelity will be corrected by MPM (RMSE ˜m is improved)."

    Eq. (1) defines the hybrid output at every triggered step as the MPM update, and MPM is the same solver used to generate the ground-truth trajectories against which grid RMSE is measured. Therefore the reported decrease in hybrid RMSE relative to neural physics (Table 1, Figure 7) is an arithmetic consequence of the fallback definition, not an independent empirical prediction: triggered steps contribute approximately zero error by construction. The genuinely empirical content is the trigger's ability to trade off latency against how often MPM is invoked; the error-improvement half of the headline claim reduces to the method's own definition.

full rationale

The only clear circularity is the hybrid solver's error improvement, which is entailed by Eq. (1) because the fallback literally substitutes the ground-truth MPM solver. This makes the RMSE reduction a construction artifact rather than a learned prediction. The paper's latency evidence, the fluid-complexity trigger, and the diffusion-based control results are independent of that reduction, so the circularity is partial. The skeptic's point about per-step graph-construction and MPM handoff costs in §B.1 is a measurement/correctness risk, not a circularity, and does not affect this score. There is no load-bearing self-citation: the GNS and Taichi references are external, and no uniqueness theorem from the authors is invoked. The control evaluation is a self-consistency test on data generated by the same reversed-simulation pipeline, but that is standard supervised evaluation on held-out trajectories rather than a definitional equivalence. Overall score 6 reflects one central claim component that reduces by construction while the latency and control contributions retain independent content.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The central claim depends on several tuned hyperparameters (rp, rt, rc, δt, λ, β, Tctr), a domain assumption that MPM is ground truth, and an ad hoc trigger metric with weak empirical support. No new physical entities are introduced.

free parameters (7)
  • spatial downsampling ratio rp = 1/1.75
    Chosen by ablation on Water 2D (Figure 6b) to balance grid RMSE and latency; applied to all scenarios.
  • temporal downsampling ratio rt = 2
    Chosen by ablation on Water 2D (Figure 6a); combined with rp in spatiotemporal reduction.
  • fallback threshold rc = 0.8
    Tuned on Water 2D via Table 1 and Figure 6d; triggers MPM fallback when cosine similarity of accelerations exceeds threshold.
  • window size δt for fluid complexity = 10
    Default window for cosine similarity computation; no ablation reported.
  • smoothness coefficient λ = 0.1
    Hand-chosen in Equation 3 to enforce smoothness of target accelerations for control training.
  • smoothness exponent β = 2
    Hand-chosen in Equation 3.
  • control trajectory length Tctr = 100
    Fixed number of MPM steps for control; acknowledged as a limitation in Section 7.
assumptions (3)
  • domain assumption MPM is a faithful model of the target fluid dynamics and serves as ground truth for all evaluation.
    Used throughout as the reference for grid RMSE; the paper does not validate MPM against real physical measurements.
  • domain assumption Grid-level RMSE (normalized mass distribution) is a valid proxy for simulation fidelity.
    Defined in Section 3.1.1 as the evaluation metric; no user study or perceptual validation is provided.
  • ad hoc to paper Cosine similarity of per-particle accelerations over a window correlates with neural physics error, so it can be used as a fallback trigger.
    Introduced in Section 3.1.2 with Spearman correlation -0.3902; the threshold rc is tuned on this metric.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hybrid Neural-MPM for Interactive Fluid Simulations in Real-Time." pith.science (2026). https://pith.science/paper/5MVM53GK

@misc{pith2026250518926,
  author       = {Pith},
  title        = {Pith review of: Hybrid Neural-MPM for Interactive Fluid Simulations in Real-Time},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5MVM53GK}},
  note         = {Machine review of arXiv:2505.18926}
}
read the original abstract

We propose a neural physics system for real-time, interactive fluid simulations. Traditional physics-based methods, while accurate, are computationally intensive and suffer from latency issues. Recent machine-learning methods reduce computational costs while preserving fidelity; yet most still fail to satisfy the latency constraints for real-time use and lack support for interactive applications. To bridge this gap, we introduce a novel hybrid method that integrates numerical simulation, neural physics, and generative control. Our neural physics jointly pursues low-latency simulation and high physical fidelity by employing a fallback safeguard to classical numerical solvers. Furthermore, we develop a diffusion-based controller that is trained using a reverse modeling strategy to generate external dynamic force fields for fluid manipulation. Our system demonstrates robust performance across diverse 2D/3D scenarios, material types, and obstacle interactions, achieving real-time simulations at high frame rates (11~29% latency) while enabling fluid control guided by user-friendly freehand sketches. We present a significant step towards practical, controllable, and physically plausible fluid simulations for real-time interactive applications. We promise to release both models and data upon acceptance.

Figures

Figures reproduced from arXiv: 2505.18926 by the authors.

Figure 1
Figure 1. We target real-time, interactive fluid simulations. Our hybrid solver integrates a numerical simulator [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. GNN as our neural physics simulator. GNN Design. We first build the initial graph G(0) by assigning a node to each particle and connect￾ing particles as edges within a fixed “connectivity radius” R. The edge embeddings are learned from relative positional displacement and the magni￾tude ri,j = [(pi − pj ), ∥pi − pj∥]. Our neural physics consists of a stack of L = 10 GNN lay￾ers. The decoder predicts the per-particle… view at source ↗
Figure 3
Figure 3. Method Overview. To achieve real-time simulations, we cut latency by learning neural physics at a coarse spatiotemporal resolution, while safeguarding fidelity by automatically falling back to an MPM solver when complex fluid phenomena arise (Section 3.1). For interactive control, we train a diffusion-based generative model that infers external force fields directly from user sketches (Section 3.2). 3.1 Hybrid Real-… view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Our neural physics acceler￾ates simulations by learning and in￾ferring at low spatial (Nl num. par￾ticles) and temporal (∆t time steps) resolutions, with downsampling ra￾tios as rp, rt. To accelerate the simulation, we train our neural physics at low spatiotemporal res…
Figure 5
Figure 5. Figure 5: Negative correlation between “cosine similarity of particle accelera￾tions over frames” vs. “simulation errors of neural physics”. Scenario: Water 2D. Spearman correlation: -0.3902. To fuse the strengths of both approaches, we make our sim￾ulator hybrid. We primarily l…
Figure 6
Figure 6. Figure 6: Ablation studies of the trade-off between grid-level [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 7
Figure 7. Figure 7: Error trajectories during simulation (Water 2D). [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: We prepare our training data for generative control [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 9
Figure 9. Figure 9: Architecture design of our Fluid ControlNet. [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]
Figure 10
Figure 10. Figure 10: Trade-off between simulation error (grid [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]
Figure 11
Figure 11. Figure 11: Visualization of generative fluid control. Rows from top to bottom: Water (2D), Sand (2D), Water [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]
Figure 12
Figure 12. Figure 12: Complete results: hybrid simulation + fluid control. We start the simulation with our neural physics, [PITH_FULL_IMAGE:figures/full_fig_p009_12.png]
Figure 13
Figure 13. Figure 13: Step-wise correlations of ground-truth accelerations for fluid control. Left: before enforcing [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]
Figure 14
Figure 14. Figure 14: Visualizations of fluid simulations by different methods, over different scenarios. From left to right: [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: More visualization of generative fluid control. From top to bottom: Water2D, Sand2D, Water3D, and [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Hybrid Lagrangian-Eulerian Model for Lagrangian Fluid Simulation

    cs.CE 2026-08 conditional novelty 6.0 of 10

    A hybrid Lagrangian-Eulerian graph neural simulator with adaptive downsampling and cross-attention achieves state-of-the-art accuracy and rollout stability on particle-based fluid benchmarks.

Reference graph

Works this paper leans on

43 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [1]

    Layer normalization

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016

    arXiv 2016

  2. [2]

    Cfd simulation of pedestrian-level wind conditions around buildings: Past achievements and prospects

    Bert Blocken and Ted Stathopoulos. Cfd simulation of pedestrian-level wind conditions around buildings: Past achievements and prospects. Journal of Wind Engineering and Industrial Aerodynamics, 121:138–145, 2013

    work page 2013

  3. [3]

    Flip: A method for adaptively zoned, particle-in-cell calculations of fluid flows in two dimensions.Journal of Computational physics, 65(2):314–343, 1986

    Jeremiah U Brackbill and Hans M Ruppel. Flip: A method for adaptively zoned, particle-in-cell calculations of fluid flows in two dimensions.Journal of Computational physics, 65(2):314–343, 1986

    work page 1986

  4. [4]

    Message passing neural pde solvers

    Johannes Brandstetter, Daniel Worrall, and Max Welling. Message passing neural pde solvers. arXiv preprint arXiv:2202.03376, 2022

    arXiv 2022

  5. [5]

    Fluid simulation for computer graphics

    Robert Bridson. Fluid simulation for computer graphics. AK Peters/CRC Press, 2015

    2015

  6. [6]

    Fast approximate kNN graph construction for high dimensional data via recursive lanczos bisection

    Jie Chen, Haw-ren Fang, and Yousef Saad. Fast approximate kNN graph construction for high dimensional data via recursive lanczos bisection. Journal of Machine Learning Research, 10(Sep):1989–2012, 2009

    work page 1989

  7. [7]

    Learning meaningful controls for fluids

    Mengyu Chu, Nils Thuerey, Hans-Peter Seidel, Christian Theobalt, and Rhaleb Zayer. Learning meaningful controls for fluids. ACM Transactions on Graphics (TOG), 40(4):1–13, 2021

    work page 2021

  8. [8]

    Efficient k-nearest neighbor graph construction for generic similarity measures

    Wei Dong, Charikar Moses, and Kai Li. Efficient k-nearest neighbor graph construction for generic similarity measures. In Proceedings of the 20th International Conference on World Wide Web, pages 577–586, 2011

    work page 2011

Show all 43 references
  1. [9]

    Fluidnexus: 3d fluid reconstruction and prediction from a single video

    Yue Gao, Hong-Xing Yu, Bo Zhu, and Jiajun Wu. Fluidnexus: 3d fluid reconstruction and prediction from a single video. arXiv preprint arXiv:2503.04720, 2025

    2025 arXiv

  2. [10]

    Neurofluid: Fluid dynamics grounding with particle-driven neural radiance fields

    Shanyan Guan, Huayu Deng, Yunbo Wang, and Xiaokang Yang. Neurofluid: Fluid dynamics grounding with particle-driven neural radiance fields. In International conference on machine learning, pages 7919–7929. PMLR, 2022

    2022

  3. [11]

    Sparsectrl: Adding sparse controls to text-to-video diffusion models

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330–348. Springer, 2024

    2024

  4. [12]

    Cameractrl: Enabling camera control for text-to-video generation

    Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024

    2024 arXiv

  5. [13]

    Difftaichi: Differentiable programming for physical simulation

    Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Frédo Durand. Difftaichi: Differentiable programming for physical simulation. ICLR, 2020

    2020

  6. [14]

    Taichi: a language for high-performance computation on spatially sparse data structures

    Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand. Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transac- tions on Graphics (TOG), 38(6):201, 2019

    2019

  7. [15]

    Freeman, and Frédo Durand

    Yuanming Hu, Jiafeng Liu, Xuanda Yang, Mingkuan Xu, Ye Kuang, Weiwei Xu, Qiang Dai, William T. Freeman, and Frédo Durand. Quantaichi: A compiler for quantized simulations. ACM Transactions on Graphics (TOG), 40(4), 2021

    2021

  8. [16]

    Plasticinelab: A soft-body manipulation benchmark with differentiable physics

    Zhiao Huang, Yuanming Hu, Tao Du, Siyuan Zhou, Hao Su, Joshua B Tenenbaum, and Chuang Gan. Plasticinelab: A soft-body manipulation benchmark with differentiable physics. arXiv preprint arXiv:2104.03311, 2021

    2021 arXiv

  9. [17]

    The affine particle-in-cell method

    Chenfanfu Jiang, Craig Schroeder, Andrew Selle, Joseph Teran, and Alexey Stomakhin. The affine particle-in-cell method. ACM Transactions on Graphics (TOG), 34(4):1–10, 2015

    2015

  10. [18]

    Accelerating particle and fluid simulations with differentiable graph networks for solving forward and inverse problems

    Krishna Kumar and Yonjin Choi. Accelerating particle and fluid simulations with differentiable graph networks for solving forward and inverse problems. In Proceedings of the SC’23 Workshops of the International Conference on High Performance Computing, Network, Storage, and An...

    2023

  11. [19]

    Gns: A generalizable graph neural network-based simulator for particulate and fluid modeling

    Krishna Kumar and Joseph Vantassel. Gns: A generalizable graph neural network-based simulator for particulate and fluid modeling. arXiv preprint arXiv:2211.10228, 2022

    2022 arXiv

  12. [20]

    Mpmnet: A data-driven mpm framework for dynamic fluid-solid interaction

    Jin Li, Yang Gao, Ju Dai, Shuai Li, Aimin Hao, and Hong Qin. Mpmnet: A data-driven mpm framework for dynamic fluid-solid interaction. IEEE Transactions on Visualization and Computer Graphics, 2023

    2023

  13. [21]

    Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids

    Yunzhu Li, Jiajun Wu, Russ Tedrake, Joshua B Tenenbaum, and Antonio Torralba. Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. arXiv preprint arXiv:1810.01566, 2018

    2018 arXiv

  14. [22]

    Interactive localized liquid motion editing

    Zherong Pan, Jin Huang, Yiying Tong, Changxi Zheng, and Hujun Bao. Interactive localized liquid motion editing. ACM Transactions on Graphics (TOG), 32(6):1–10, 2013

    2013

  15. [23]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    2023

  16. [24]

    Learning mesh- based simulation with graph networks

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter Battaglia. Learning mesh- based simulation with graph networks. In International conference on learning representations, 2020

    2020

  17. [25]

    Controlling liquids using meshes

    Karthik Raveendran, Nils Thuerey, Christopher J Wojtan, and Greg Turk. Controlling liquids using meshes. In Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation, 2012

    2012

  18. [26]

    A neural material point method for particle-based simulations

    Omer Rochman-Sharabi, Sacha Lewin, and Gilles Louppe. A neural material point method for particle-based simulations. 2024

    2024

  19. [27]

    Learning to simulate complex physics with graph networks

    Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. In International conference on machine learning, pages 8459–8468. PMLR, 2020

    2020

  20. [28]

    Particle-based liquid control using animation templates

    Arnaud Schoentgen, Pierre Poulin, Emmanuelle Darles, and Philippe Meseure. Particle-based liquid control using animation templates. In Computer Graphics Forum, volume 39, pages 79–88. Wiley Online Library, 2020

    2020

  21. [29]

    Interactive cfd simulations with virtual reality to support learning in mixing

    Serkan Solmaz and Tom Van Gerven. Interactive cfd simulations with virtual reality to support learning in mixing. Computers & Chemical Engineering, 156:107570, 2022

    2022

  22. [30]

    Stable fluids

    Jos Stam. Stable fluids. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 779–786. 2023

    2023

  23. [31]

    Visualizing large-scale and high- dimensional data

    Jian Tang, Jingzhou Liu, Ming Zhang, and Qiaozhu Mei. Visualizing large-scale and high- dimensional data. In Proceedings of the 25th International Conference on World Wide Web, pages 287–297, 2016

    2016

  24. [32]

    Neural sph: Improved neural modeling of lagrangian fluid dynamics

    Artur P Toshev, Jonas A Erbesdobler, Nikolaus A Adams, and Johannes Brandstetter. Neural sph: Improved neural modeling of lagrangian fluid dynamics. arXiv preprint arXiv:2402.06275, 2024

    2024 arXiv

  25. [33]

    From early virtual garment simulation to interactive fashion design

    Pascal V olino, Frederic Cordier, and Nadia Magnenat-Thalmann. From early virtual garment simulation to interactive fashion design. Computer-aided design, 37(6):593–608, 2005

    2005

  26. [34]

    Boximator: Generating rich and controllable motions for video synthesis

    Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis. arXiv preprint arXiv:2402.01566, 2024

    2024 arXiv

  27. [35]

    Motionctrl: A unified and flexible motion controller for video generation

    Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024

    2024

  28. [36]

    Draganything: Motion control for anything using entity representation

    Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331–348. Springer, 2024. 11

    2024

  29. [37]

    Camco: Camera-controllable 3d-consistent image-to-video generation

    Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vah- dat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024

    2024 arXiv

  30. [38]

    Interactive liquid splash modeling by user sketches

    Guowei Yan, Zhili Chen, Jimei Yang, and Huamin Wang. Interactive liquid splash modeling by user sketches. ACM Transactions on Graphics (TOG), 39(6):1–13, 2020

    2020

  31. [39]

    Direct-a-video: Customized video generation with user-directed camera movement and object motion

    Shiyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024

    2024

  32. [40]

    Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory

    Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023

    2023 arXiv

  33. [41]

    Slimmable neural networks

    Jiahui Yu, Linjie Yang, Ning Xu, Jianchao Yang, and Thomas Huang. Slimmable neural networks. arXiv preprint arXiv:1812.08928, 2018

    2018 arXiv

  34. [42]

    Tora: Trajectory-oriented diffusion transformer for video generation

    Zhenghao Zhang, Junchao Liao, Menghao Li, Zuozhuo Dai, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024

    2024 arXiv

  35. [43]

    Motiondirector: Motion customization of text-to-video diffusion models

    Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. In European Conference on Computer Vision, pages 273–290. Springer, 2024. 12 A Details of...

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.