pith. sign in

arxiv: 2605.10586 · v1 · submitted 2026-05-11 · 💻 cs.CV

CausalGS: Learning Physical Causality of 3D Dynamic Scenes with Gaussian Representations

Nengbo Lu , Minghua Pan This is my paper

Pith reviewed 2026-05-12 03:30 UTC · model grok-4.3

classification 💻 cs.CV
keywords causal dynamics3D Gaussian splattinginverse physics inferencemulti-view videofuture predictionphysical causalitydynamic scenes
0
0 comments X

The pith

CausalGS decouples initial velocities from material properties to learn physical causality in 3D scenes from video alone.

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

The paper introduces a method to extract physical understanding from multi-view videos of moving 3D scenes. It does this by first inferring the starting motion of everything in the scene and the material traits that dictate how things respond to forces. These two pieces of information then drive a physics simulator that regularizes the learning, allowing the system to predict how the scene will evolve far into the future. A sympathetic reader would care because this approach avoids needing hand-crafted rules or perfect 3D models upfront, potentially letting AI systems discover cause and effect in the physical world just by watching.

Core claim

CausalGS learns the causal dynamics of complex dynamic 3D scenes solely from multi-view videos by using an inverse physics inference module to decouple the initial velocity field representing kinematics from the intrinsic material properties governing dynamics, and then employs a differentiable physics simulator to guide learning in a physics-regularized manner, achieving superior long-term future frame extrapolation and novel view interpolation without human annotations or strong priors.

What carries the argument

The inverse physics inference module, which jointly infers the initial velocity field for the scene's kinematics and the intrinsic material properties for its dynamics from multi-view video observations.

If this is right

  • The model outperforms prior methods on long-term future frame extrapolation in dynamic 3D scenes.
  • It maintains strong performance in novel view interpolation tasks.
  • The approach enables learning complex interactions between physical properties and causal relationships purely from visual data.
  • Physics regularization through the differentiable simulator improves the physical plausibility of predictions.

Where Pith is reading between the lines

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

  • If the decoupling works reliably, similar inverse inference could be applied to other data types like single-view video or point clouds.
  • This separation of kinematics and dynamics might help in creating more generalizable simulators for robotics planning.
  • Success here suggests that visual observations contain sufficient signals to recover underlying physical parameters in many everyday scenes.

Load-bearing premise

The video observations must contain enough information to accurately and uniquely separate the initial velocity field from the material properties without needing extra constraints or perfect geometry.

What would settle it

A test where the predicted future frames deviate significantly from ground truth in scenarios with known but hidden physical parameters, such as objects with different elasticity colliding.

Figures

Figures reproduced from arXiv: 2605.10586 by Minghua Pan, Nengbo Lu.

Figure 1
Figure 1. Figure 1: From a video stream of a real-world dynamic scene, CausalGS learns the interactions of physical properties within [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overall pipeline of our model begins with multi-view image sequences of a scene from video, from which an [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of our method against other models[ [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Each 3D Gaussian primitive is modeled as a La [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of rendering results against other models[7, 16, 18, 19, 32, 49] on the Dynamic Indoor Scene [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results of our method for unsupervised motion segmentation. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
read the original abstract

Learning a physical model from video data that can comprehend physical laws and predict the future trajectories of objects is a formidable challenge in artificial intelligence. Prior approaches either leverage various Partial Differential Equations (PDEs) as soft constraints in the form of PINN losses, or integrate physics simulators into neural networks; however, they often rely on strong priors or high-quality geometry reconstruction. In this paper, we propose CausalGS, a framework that learns the causal dynamics of complex dynamic 3D scenes solely from multi-view videos, while dispensing with the reliance on explicit priors. At its core is an inverse physics inference module that decouples the complex dynamics problem from the video into the joint inference of two factors: the initial velocity field representing the scene's kinematics, and the intrinsic material properties governing its dynamics. This inferred physical information is then utilized within a differentiable physics simulator to guide the learning process in a physics-regularized manner. Extensive experiments demonstrate that CausalGS surpasses the state-of-the-art on the highly challenging task of long-term future frame extrapolation, while also exhibiting advanced performance in novel view interpolation. Crucially, our work shows that, without any human annotation, the model is able to learn the complex interactions between multiple physical properties and understand the causal relationships driving the scene's dynamic evolution, solely from visual observations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The paper introduces CausalGS, a framework that uses 3D Gaussian representations to learn physical causality in dynamic scenes solely from multi-view videos. Its core component is an inverse physics inference module that decouples the initial velocity field (kinematics) from intrinsic material properties (dynamics); these are then fed into a differentiable physics simulator to provide physics-regularized supervision during training. The authors claim that the resulting model achieves state-of-the-art long-term future-frame extrapolation, competitive novel-view interpolation, and the ability to discover causal physical interactions without human annotations or strong geometric priors.

Significance. If the decoupling is shown to be identifiable and the quantitative gains are reproducible, the work would represent a meaningful step toward causal, physics-aware neural rendering. The combination of Gaussian splatting with a differentiable simulator is efficient and avoids explicit mesh reconstruction, which could benefit downstream applications in robotics and simulation. The absence of human annotations is a notable strength if the causal factors prove robust.

major comments (3)
  1. [§3.2] §3.2 (Inverse Physics Inference Module): The central claim that the module uniquely recovers a pair (initial velocity field, material parameters) from multi-view video alone is load-bearing for all downstream causality assertions, yet the manuscript provides neither an identifiability argument nor a uniqueness regularizer. Because the forward map from (velocity, material) to observed trajectories is many-to-one (e.g., a stiffer modulus compensated by a different initial velocity distribution can yield indistinguishable short-term motion), the inferred quantities may be non-unique fits rather than causally valid factors.
  2. [§5] §5 (Experiments): The abstract asserts superiority on long-term extrapolation and the discovery of causal relationships, but the reported results lack error bars, dataset statistics, ablation studies that isolate the inverse module, and verification that the recovered parameters remain stable under small perturbations of the input views. Without these, the quantitative claims cannot be assessed and the circularity risk (simulator parameters guiding the very inference they later validate) remains unaddressed.
  3. [§4] §4 (Physics-Regularized Training): The training loop uses the differentiable simulator to supervise the inferred velocity and material fields. It is unclear whether the final predictions are independent of the simulator’s internal parameters or whether they reduce to fitting within the same closed loop; an explicit statement of which simulator parameters are held fixed versus learned would clarify this.
minor comments (3)
  1. [§3] Notation for the velocity field and material tensor is introduced without a clear table of symbols; readers must cross-reference multiple equations to reconstruct the full variable list.
  2. [Figure 3] Figure 3 (qualitative extrapolation results) would benefit from side-by-side ground-truth frames and error heat-maps to make the claimed long-term stability visually verifiable.
  3. [§2] The related-work section omits recent differentiable-simulator papers that also operate on implicit representations; adding these citations would strengthen the positioning.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment point by point below, providing clarifications where possible and committing to revisions that strengthen the presentation of our results and methodology.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (Inverse Physics Inference Module): The central claim that the module uniquely recovers a pair (initial velocity field, material parameters) from multi-view video alone is load-bearing for all downstream causality assertions, yet the manuscript provides neither an identifiability argument nor a uniqueness regularizer. Because the forward map from (velocity, material) to observed trajectories is many-to-one (e.g., a stiffer modulus compensated by a different initial velocity distribution can yield indistinguishable short-term motion), the inferred quantities may be non-unique fits rather than causally valid factors.

    Authors: We agree that a formal identifiability argument would bolster the causal claims. The current manuscript does not provide one, as establishing uniqueness for this inverse problem under general conditions is mathematically involved. In practice, the multi-view video observations over multiple time steps, combined with the differentiable simulator's physics constraints, provide sufficient regularization to yield parameters that produce accurate long-term rollouts. We will revise §3.2 to explicitly discuss the many-to-one nature of the forward map as a limitation, describe how the long-horizon prediction objective and multi-view consistency help mitigate ambiguities in our setting, and note the absence of an explicit uniqueness regularizer. revision: partial

  2. Referee: [§5] §5 (Experiments): The abstract asserts superiority on long-term extrapolation and the discovery of causal relationships, but the reported results lack error bars, dataset statistics, ablation studies that isolate the inverse module, and verification that the recovered parameters remain stable under small perturbations of the input views. Without these, the quantitative claims cannot be assessed and the circularity risk (simulator parameters guiding the very inference they later validate) remains unaddressed.

    Authors: We acknowledge these omissions weaken the quantitative assessment. In the revised version we will add error bars from at least three independent runs with different random seeds, include full dataset statistics (number of scenes, frames, camera configurations), and insert new ablation studies that remove or replace the inverse physics inference module while keeping other components fixed. We will also report results on input-view perturbations to demonstrate stability of the recovered velocity and material fields. Regarding circularity, the simulator's internal parameters (time-step size, gravity, base material model coefficients) are held fixed throughout training and inference; only the scene-specific velocity field and material properties are optimized by the inference module. revision: yes

  3. Referee: [§4] §4 (Physics-Regularized Training): The training loop uses the differentiable simulator to supervise the inferred velocity and material fields. It is unclear whether the final predictions are independent of the simulator’s internal parameters or whether they reduce to fitting within the same closed loop; an explicit statement of which simulator parameters are held fixed versus learned would clarify this.

    Authors: We will expand §4 with a dedicated paragraph that explicitly lists the simulator parameters. Fixed parameters include the integration time step, collision restitution coefficients, and the underlying constitutive model constants (e.g., reference density and Poisson's ratio ranges drawn from standard physics tables). Learned quantities are strictly limited to the initial velocity field and the per-object material parameters (Young's modulus, damping coefficients) inferred by the inverse module. Predictions are generated by rolling out the simulator with these inferred values; the simulator itself is never optimized, breaking any closed-loop dependency on its internal settings. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper describes an inverse physics inference module that decouples initial velocity (kinematics) from material properties (dynamics) directly from multi-view video observations using Gaussian representations, then feeds the inferred factors into a differentiable simulator for physics-regularized training. The central results—long-term future frame extrapolation and causal understanding—are obtained by applying this learned model to held-out future frames and novel views. No quoted step in the abstract or described framework reduces the reported predictions or causal claims to the training inputs by construction (e.g., no self-definitional loop where the output is the fitted input renamed, no uniqueness theorem imported from self-citation, and no ansatz smuggled via prior work). The regularization loop is a standard supervised training mechanism whose extrapolation performance is evaluated separately on future data, leaving the derivation self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

Based solely on the abstract, the central claim rests on the unverified ability of video observations to determine unique velocity and material fields and on the accuracy of an unspecified differentiable physics simulator.

axioms (2)
  • domain assumption A differentiable physics simulator exists that can accurately evolve scene state given only the inferred initial velocity field and material properties.
    Invoked when the inferred quantities are fed into the simulator to regularize learning.
  • domain assumption Multi-view video alone contains sufficient information to uniquely recover the initial velocity field and intrinsic material properties without additional priors.
    Core premise of the inverse physics inference module.
invented entities (1)
  • Inverse physics inference module no independent evidence
    purpose: Decouples complex dynamics into initial velocity field and material properties from video input.
    New component introduced to enable physics-regularized learning without explicit priors.

pith-pipeline@v0.9.0 · 5530 in / 1488 out tokens · 51047 ms · 2026-05-12T03:30:12.385660+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

55 extracted references · 55 canonical work pages

  1. [1]

    Hugo Bertiche, Meysam Madadi, and Sergio Escalera. 2020. PBNS: physically based neural simulator for unsupervised garment pose space deformation

    work page 2020

  2. [2]

    Junhao Cai, Yuji Yang, Weihao Yuan, Yisheng He, Zilong Dong, Liefeng Bo, Hui Cheng, and Qifeng Chen. 2024. Gic: Gaussian-informed continuum for physical property identification and simulation.Advances in Neural Information Processing Systems37 (2024), 75035–75063. CausalGS: Learning Physical Causality of 3D Dynamic Scenes with Gaussian Representations ICM...

    work page 2024

  3. [3]

    Ang Cao and Justin Johnson. 2023. Hexplane: A fast representation for dynamic scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Los Alamitos, CA, USA, 130–141

    work page 2023

  4. [4]

    Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. 2022. Masked-attention mask transformer for universal image segmen- tation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. IEEE, Los Alamitos, CA, USA, 1290–1299

    work page 2022

  5. [5]

    Mengyu Chu, Lingjie Liu, Quan Zheng, Aleksandra Franz, Hans-Peter Seidel, Christian Theobalt, and Rhaleb Zayer. 2022. Physics informed neural fields for smoke reconstruction with sparse data.ACM Transactions on Graphics (ToG)41, 4 (2022), 1–14

    work page 2022

  6. [6]

    Siming Fan, Jingtan Piao, Chen Qian, Hongsheng Li, and Kwan-Yee Lin. 2023. Simulating fluids in real-world still images. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision. IEEE, Los Alamitos, CA, USA, 15922–15931

    work page 2023

  7. [7]

    Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. 2022. Fast dynamic radiance fields with time-aware neural voxels. InSIGGRAPH Asia 2022 Conference Papers. Association for Computing Machinery, New York, NY, USA, 1–9

    work page 2022

  8. [8]

    Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. 2023. K-planes: Explicit radiance fields in space, time, and appearance. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Los Alamitos, CA, USA, 12479–12488

    work page 2023

  9. [9]

    Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Schölkopf

    work page

  10. [10]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Los Alamitos, CA, USA, 21295–21304

    work page

  11. [11]

    Yue Gao, Hong-Xing Yu, Bo Zhu, and Jiajun Wu. 2025. FluidNexus: 3D fluid reconstruction and prediction from a single video. InProceedings of the Computer Vision and Pattern Recognition Conference. IEEE, Los Alamitos, CA, USA, 26091– 26101

    work page 2025

  12. [12]

    Michelle Guo, Matt Jen-Yuan Chiang, Igor Santesteban, Nikolaos Sarafianos, Hsiao-yu Chen, Oshri Halimi, Aljaž Božič, Shunsuke Saito, Jiajun Wu, C Karen Liu, et al . 2025. PGC: Physics-Based Gaussian Cloth from a Single Pose. In Proceedings of the Computer Vision and Pattern Recognition Conference. IEEE, Los Alamitos, CA, USA, 21215–21225

    work page 2025

  13. [13]

    Shuting He, Peilin Ji, Yitong Yang, Changshuo Wang, Jiayi Ji, Yinglin Wang, and Henghui Ding. 2025. A survey on 3d gaussian splatting applications: Segmenta- tion, editing, and generation

    work page 2025

  14. [14]

    Haorui Ji, Rong Wang, Tao Jun Lin, and Hongdong Li. 2025. JADE: Joint-aware Latent Diffusion for 3D Human Generative Modeling. In2025 International Con- ference on 3D Vision (3DV). IEEE, IEEE, Los Alamitos, CA, USA, 791–801

    work page 2025

  15. [15]

    Ali Kamali, Mohammad Sarabian, and Kaveh Laksari. 2023. Elasticity imaging using physics-informed neural networks: Spatial discovery of elastic modulus and Poisson’s ratio.Acta biomaterialia155 (2023), 400–409

    work page 2023

  16. [16]

    Deqi Li, Shi-Sheng Huang, Zhiyuan Lu, Xinran Duan, and Hua Huang. 2024. St-4dgs: Spatial-temporally consistent 4d gaussian splatting for efficient dynamic scene rendering. InACM SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, New York, NY, USA, 1–11

    work page 2024

  17. [17]

    Jinxi Li, Ziyang Song, and Bo Yang. 2023. Nvfi: Neural velocity fields for 3d physics learning from dynamic videos.Advances in Neural Information Processing Systems36 (2023), 34723–34751

    work page 2023

  18. [18]

    Jinxi Li, Ziyang Song, and Bo Yang. 2025. TRACE: Learning 3D Gaussian Physical Dynamics from Multi-view Videos. arXiv:2508.09811 [cs.CV] https://arxiv.org/ abs/2508.09811

    work page arXiv 2025

  19. [19]

    Jinxi Li, Ziyang Song, Siyuan Zhou, and Bo Yang. 2025. FreeGave: 3D Physics Learning from Dynamic Videos by Gaussian Velocity. InProceedings of the Com- puter Vision and Pattern Recognition Conference. IEEE, Los Alamitos, CA, USA, 12433–12443

    work page 2025

  20. [20]

    Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. 2021. Neural scene flow fields for space-time view synthesis of dynamic scenes. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. IEEE, Los Alamitos, CA, USA, 6498–6508

    work page 2021

  21. [21]

    Xingyu Lin, Zhiao Huang, Yunzhu Li, Joshua B Tenenbaum, David Held, and Chuang Gan. 2022. Diffskill: Skill abstraction from differentiable physics for deformable object manipulations with tools

    work page 2022

  22. [22]

    Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. 2024. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Los Alamitos, CA, USA, 21136–21145

    work page 2024

  23. [23]

    Yuchen Lin, Chenguo Lin, Jianjin Xu, and Yadong Mu. 2025. OmniphysGS: 3d constitutive gaussians for general physics-based dynamics generation

    work page 2025

  24. [24]

    Fangfu Liu, Hanyang Wang, Shunyu Yao, Shengjun Zhang, Jie Zhou, and Yueqi Duan. 2024. Physics3d: Learning physical properties of 3d gaussians via video diffusion

    work page 2024

  25. [25]

    Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. 2021. Neural actor: Neural free-view synthesis of human actors with pose control.ACM transactions on graphics (TOG)40, 6 (2021), 1–16

    work page 2021

  26. [26]

    Yufei Liu, Junshu Tang, Chu Zheng, Shijie Zhang, Jinkun Hao, Junwei Zhu, and Dongjin Huang. 2024. ClotheDreamer: Text-Guided Garment Generation with 3D Gaussians

    work page 2024

  27. [27]

    Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. 2024. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In2024 International Conference on 3D Vision (3DV). IEEE, IEEE, Los Alamitos, CA, USA, 800–809

    work page 2024

  28. [28]

    Pingchuan Ma, Peter Yichen Chen, Bolei Deng, Joshua B Tenenbaum, Tao Du, Chuang Gan, and Wojciech Matusik. 2023. Learning neural constitutive laws from motion observations for generalizable pde dynamics. InInternational Conference on Machine Learning. PMLR, PMLR, Honolulu, HI, USA, 23279–23300

    work page 2023

  29. [29]

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis.Commun. ACM65, 1 (2021), 99–106

    work page 2021

  30. [30]

    Seungtae Nam, Daniel Rho, Jong Hwan Ko, and Eunbyung Park. 2023. Mip-grid: Anti-aliased grid representations for neural radiance fields.Advances in Neural Information Processing Systems36 (2023), 2837–2849

    work page 2023

  31. [31]

    Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. 2021. Neural scene graphs for dynamic scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Los Alamitos, CA, USA, 2856– 2865

    work page 2021

  32. [32]

    2023.𝑝-Poisson surface reconstruction in curl-free flow from point clouds.Advances in Neural Information Processing Systems36 (2023), 60077–60098

    Yesom Park, Taekyung Lee, Jooyoung Hahn, and Myungjoo Kang. 2023.𝑝-Poisson surface reconstruction in curl-free flow from point clouds.Advances in Neural Information Processing Systems36 (2023), 60077–60098

    work page 2023

  33. [33]

    Albert Pumarola Peris, Enric Corona Puyane, Gerard Pons-Moll, Francesc Moreno- Noguer, et al. 2021. D-NeRF: Neural radiance fields for dynamic scenes

    work page 2021

  34. [34]

    Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. 2023. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Los Alamitos, CA, USA, 16632– 16642

    work page 2023

  35. [35]

    Siyuan Song and Hanxun Jin. 2024. Identifying constitutive parameters for complex hyperelastic materials using physics-informed neural networks.Soft Matter20, 30 (2024), 5915–5926

    work page 2024

  36. [36]

    Ziyang Song and Bo Yang. 2022. Ogc: Unsupervised 3d object segmentation from rigid dynamics of point clouds.Advances in Neural Information Processing Systems35 (2022), 30798–30812

    work page 2022

  37. [37]

    Ziyang Song and Bo Yang. 2024. Unsupervised 3d object segmentation of point clouds by geometry consistency.IEEE Transactions on Pattern Analysis and Machine Intelligence46, 12 (2024), 8459–8473

    work page 2024

  38. [38]

    Zhaoqi Su, Liangxiao Hu, Siyou Lin, Hongwen Zhang, Shengping Zhang, Justus Thies, and Yebin Liu. 2023. Caphy: Capturing physical properties for animat- able human avatars. InProceedings of the IEEE/CVF International Conference on Computer Vision. IEEE, Los Alamitos, CA, USA, 14150–14160

    work page 2023

  39. [39]

    Hengyi Wang, Jingwen Wang, and Lourdes Agapito. 2023. Co-slam: Joint coordi- nate and sparse parametric encodings for neural real-time slam. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Los Alamitos, CA, USA, 13293–13302

    work page 2023

  40. [40]

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 2024. 4d gaussian splatting for real-time dynamic scene rendering. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. IEEE, Los Alamitos, CA, USA, 20310–20320

    work page 2024

  41. [41]

    Jiahao Wu, Rui Peng, Zhiyan Wang, Lu Xiao, Luyang Tang, Jinbo Yan, Kaiqiang Xiong, and Ronggang Wang. 2025. Swift4D: Adaptive divide-and-conquer Gauss- ian Splatting for compact and efficient reconstruction of dynamic scene

    work page 2025

  42. [42]

    Donglai Xiang, Timur Bagautdinov, Tuur Stuyck, Fabian Prada, Javier Romero, Weipeng Xu, Shunsuke Saito, Jingfan Guo, Breannan Smith, Takaaki Shiratori, et al. 2022. Dressing avatars: Deep photorealistic appearance for physically simulated clothing.ACM Transactions on Graphics (TOG)41, 6 (2022), 1–15

    work page 2022

  43. [43]

    Jun Xiang, Xuan Gao, Yudong Guo, and Juyong Zhang. 2024. Flashavatar: High- fidelity head avatar with efficient gaussian embedding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Los Alamitos, CA, USA, 1802–1812

    work page 2024

  44. [44]

    Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 2023. 3d-aware image generation using 2d diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision. IEEE, Los Alamitos, CA, USA, 2383–2393

    work page 2023

  45. [45]

    Haozhe Xie, Zhaoxi Chen, Fangzhou Hong, and Ziwei Liu. 2025. Generative Gaussian splatting for unbounded 3D city generation. InProceedings of the Com- puter Vision and Pattern Recognition Conference. IEEE, Los Alamitos, CA, USA, 6111–6120

    work page 2025

  46. [46]

    Tianyi Xie, Yiwei Zhao, Ying Jiang, and Chenfanfu Jiang. 2025. Physanimator: Physics-guided generative cartoon animation. InProceedings of the Computer Vision and Pattern Recognition Conference. IEEE, Los Alamitos, CA, USA, 10793– 10804

    work page 2025

  47. [47]

    Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chen- fanfu Jiang. 2024. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Los Alamitos, CA, USA, 4389–4398. ICMR ’26, June 16–July 3, 2026, Chicago, IL, USA Trovato et al

    work page 2024

  48. [48]

    Yiteng Xu, Peishan Cong, Yichen Yao, Runnan Chen, Yuenan Hou, Xinge Zhu, Xuming He, Jingyi Yu, and Yuexin Ma. 2023. Human-centric scene understand- ing for 3d large-scale scenarios. InProceedings of the IEEE/CVF International Conference on Computer Vision. IEEE, Los Alamitos, CA, USA, 20349–20359

    work page 2023

  49. [49]

    Gengshan Yang, Chaoyang Wang, N Dinesh Reddy, and Deva Ramanan. 2023. Reconstructing animatable categories from videos. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Los Alamitos, CA, USA, 16995–17005

    work page 2023

  50. [50]

    Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. 2024. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. IEEE, Los Alamitos, CA, USA, 20331–20341

    work page 2024

  51. [51]

    Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. 2020. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Los Alamitos, CA, USA, 5336–5345

    work page 2020

  52. [52]

    Hong-Xing Yu, Yang Zheng, Yuan Gao, Yitong Deng, Bo Zhu, and Jiajun Wu. 2023. Inferring hybrid neural fluid fields from videos.Advances in Neural Information Processing Systems36 (2023), 63595–63608

    work page 2023

  53. [53]

    Qi Zhao, Xingyu Ni, Ziyu Wang, Feng Cheng, Ziyan Yang, Lu Jiang, and Bohan Wang. 2025. Synthetic video enhances physical fidelity in video synthesis

    work page 2025

  54. [54]

    Yang Zheng, Qingqing Zhao, Guandao Yang, Wang Yifan, Donglai Xiang, Florian Dubost, Dmitry Lagun, Thabo Beeler, Federico Tombari, Leonidas Guibas, et al

    work page

  55. [55]

    InEuropean Conference on Computer Vision

    Physavatar: Learning the physics of dressed 3d avatars from visual obser- vations. InEuropean Conference on Computer Vision. Springer, Springer, Cham, Switzerland, 262–284

    work page