acceptodds
Under review as a conference paper at ICLR 2027

PowerSim: Differentiable Physics Simulation and Rendering with Power Diagrams

Abstract

We introduce PowerSim, a method to bring physically grounded, differentiable dynamics to PowerFoam's power diagram based 3D representation. PowerSim directly couples a pre-trained PowerFoam scene to the Material Point Method (MPM) by exploiting a natural alignment between the two: the geometric and appearance properties of each primitive correspond closely to the quantities MPM already tracks as an object deforms. Consequently, simulated motion can drive the scene’s geometry and appearance directly, without an auxiliary representation in between. Built on this framework, we enable a range of applications on real and synthetic scenes: (1) simulating a static scene under user interaction, (2) recovering spatially varying material fields, (3) compositing primitives from independently captured scenes into a single simulation-ready scene and (4) ray-tracing reflections that update consistently as the object deforms. Our results suggest that PowerSim excels over previous frameworks for physically grounded dynamics, while unlocking unique advantages—such as secondary ray lighting effects on dynamic scenes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D vision, Physics-based Simulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_kv8KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PowerSim couples PowerFoam’s bounded power-diagram representation with differentiable Material Point Method (MPM) simulation. It updates primitive positions, scalar radii, dipole orientations, appearance directions, and adjacency according to simulated motion. The resulting pipeline supports material estimation from deformation videos, object selection from multiview masks, scene composition, and dynamic reflections. Experiments compare dynamic reconstruction and material estimation against PAC-NeRF and PhysDreamer.

Critical strengths
  1. The framework provides a practical way to animate PowerFoam scenes while updating geometry, appearance, and adjacency together.

  2. Dynamic reconstruction results are encouraging: mean PSNR is 25.73 dB, compared with 22.06 dB for PAC-NeRF and 19.00 dB for PhysDreamer.

  3. The optimization-free primitive-selection procedure is useful, achieving 0.80 mean IoU compared with 0.82 for SemanticFoam, which requires semantic optimization.

Critical weaknesses
  1. Minor conceptual novelty. The central strategy—coupling rendering primitives to MPM and differentiating through simulation and rendering for material estimation—is already established by prior work discussed in the paper. The proposed position, rotation, scale, and appearance updates are reasonable adaptations, but the paper does not establish a sufficiently substantial methodological advance beyond that transfer.

  2. The claimed representational advantage is not adequately validated. Explicit cell boundaries and adjacency are potentially useful, but the paper does not demonstrate quantitatively that they yield more accurate material volumes, contact behavior, or continuum deformation. Particle masses are assigned using occupied grid voxels rather than exact cell volumes, further limiting the demonstrated connection between explicit geometry and physical accuracy.

  3. The evaluation does not isolate the source of improvement. Better dynamic PSNR may partly reflect differences in the underlying static representations. Without controls for initial reconstruction quality and quantitative component ablations, it is difficult to attribute the gains to the proposed coupling.

  4. Material estimation gains are limited in scope. Young’s modulus estimation improves on average, but Poisson’s ratio error remains tied with PhysDreamer at 0.16. The experiments do not establish accurate recovery of heterogeneous material fields, despite this being a highlighted capability.

Questions for authors
  1. Beyond applying established differentiable MPM coupling to PowerFoam, what is the central technical insight that could not be obtained through a straightforward adaptation of existing pipelines?

  2. Can the authors quantify how explicit power-cell boundaries improve physical accuracy, particularly for volume evolution, contact, and large deformation?

  3. How much of the dynamic reconstruction improvement remains after accounting for differences in static reconstruction quality?

  4. Can the authors provide quantitative ablations of radius scaling, dipole rotation, appearance rotation, and adjacency rebuilding?

5 comments
AuthorsOpenReview ↗

Response to kv8K (1/3): Conceptual Novelty and Representational Advantage

We thank the reviewer for recognizing that PowerSim provides “a practical way to animate PowerFoam scenes while updating geometry, appearance, and adjacency together.” We appreciate reviewer’s assessment that the dynamic reconstruction results are “encouraging” and that the “optimization-free primitive-selection procedure is useful.”. We provide detailed responses to feedback as follows:

1/ Conceptual Novelty. We agree that coupling rendering primitives to MPM and differentiating through simulation and rendering were introduced by prior work. However, extending this paradigm to PowerFoam is non-trivial. Unlike Gaussians, whose covariance can be transformed directly by the deformation gradient, a PowerFoam primitive is a power cell coupled to its neighbours, with a radius, an oriented surface plane and appearance directions. Each requires a dedicated update rule to follow the MPM deformation, and our ablations (Fig. 7 and the quantitative ablation in 5/ below) show that each one contributes.

Central Insight. In Gaussian representation, each primitive must absorb the full deformation gradient, which stretches kernels into needle-like artifacts under large deformation. In PowerFoam, a cell's shape is also determined by its neighbours. We exploit this to split the deformation: the anisotropic part (stretch and shear) is carried collectively by the cell structure as MPM moves the sites, while each primitive only takes the rotation and uniform scale from the polar decomposition F=RSF=RS. This split is what keeps the cells well-formed under extreme deformation, and it is not available to existing per-primitive pipelines.

More importantly, PowerSim is a complete framework: foam selection and editing to obtain simulation particles, material estimation from video, and physics simulation with rasterized or ray-traced rendering. We also argue that its representation, PowerFoam, is well suited to physics simulation, thanks to three key properties supported by our experiments:

  • Non-overlapping cells: separated material stays sharp under extreme deformation instead of smearing across the gap (Fig. 2, against PhysGaussian, GaussianFluent and Vid2Sim).
  • Well-regularized cells: large stretch, twist and bend do not produce the needle-like artifacts of extremely elongated Gaussians artifacts (Fig. 2, and Fig. 7).
  • Unified rasterization and ray tracing: the simulated scene can be rendered with secondary effects such as dynamic mirror reflections (Fig.1, Fig. 8).

2/ Representational advantage. We thank the reviewer for this insightful question. PowerSim uses the same MPM solver, constitutive models and colliders as the baselines, so the physical behaviour (deformation, contact and mass) is determined by MPM rather than by the cell geometry. Our claimed advantage is that PowerFoam represents and renders the simulated deformation more faithfully, especially under large deformation (as shown in Fig. 2 and Fig. 7); we do not claim more accurate physics.

  • Material volumes and mass. By default, we assign particle volumes with the standard MPM grid rule, so that mass remains uniformly distributed regardless of how primitives were sampled during reconstruction. Because the cells are explicit, exact power-cell volumes are also available and can be used directly for simulation. We compare the two choices on the GSO benchmark, keeping everything else fixed, and they give comparable results (25.73 vs. 25.79 PSNR, 0.930 vs. 0.930 SSIM; per-object table in our response to Reviewer XVdH, 1a).
  • Volume evolution. Our radius scale got updated following MPM deformation gradient decomposition. Under uniform expansion or compression, the rendered cells follow the simulated volume closely; under general deformation (e.g. large deformation), they follow it through the motion of neighbouring sites. In a controlled test, the occupied-volume error is at most 0.5% under uniform dilation and 0.22% under simple shear (see our response to Reviewer XVdH, 5).
  • Contact. Contact is resolved on the MPM grid with the same collider as the baselines. The explicit cells do not change contact physics; they provide a sharp, non-overlapping visual boundary at the contact surface (as shown in multi-object simulation of Fig. 8).
  • Large deformation. Non-overlapping, well-regularized cells keep separated material sharp under extreme stretching, whereas Gaussian-based methods smear it into elongated kernels (Fig. 2, against PhysGaussian, GaussianFluent and Vid2Sim).
AuthorsOpenReview ↗

Response to kv8K (2/3): Isolation of static reconstruction quality and material estimation gains

3/ Isolation of static reconstruction quality. We thank the reviewer for suggesting this insightful comparison which further reveals interesting observation and strengthens our PowerSim result. As shown below, we separate reconstruction quality into a static score (first frame, before any simulation) and a dynamic score (averaged over simulated frames after first floor contact, where the object deforms).

MethodStatic PSNR / SSIM (first frame)Dynamic PSNR / SSIM (frames after floor contact)
PAC-NeRF26.32 / 0.93817.72 / 0.898
PhysDreamer27.26 / 0.95218.62 / 0.906
PowerSim (Ours)26.75 / 0.94323.71 / 0.915

Interestingly, the three methods start from comparable static quality, within 1 dB of each other. PowerSim is 0.51 dB below PhysDreamer and only 0.43 dB above PAC-NeRF on the static frame, yet it is 5.09 dB and 5.99 dB ahead of them, respectively, on the deforming frames. The gain therefore cannot be explained by first-frame reconstruction quality; it comes from how each representation follows the simulated deformation.

4/ Material estimation gains. We agree that the gain is concentrated in Young's modulus, and we argue this reflects what the benchmark can measure:

  • Poisson's ratio is weakly identifiable from visual dynamics. In a drop, EE governs the dominant visible response (compression and rebound at impact), while ν\nu only affects the subtle lateral bulging, which barely changes the rendered images. The ν\nu column therefore does not separate methods: the tie at 0.16 reflects that neither method can reliably recover ν\nu from this kind of motion
  • Heterogeneous material fields. To the best of our knowledge, there is no official benchmark provides ground-truth spatially varying materials. We instead demonstrate it qualitatively on real captures: the recovered field reproduces the observed motion of the carnation (Fig. 4).
AuthorsOpenReview ↗

Response to kv8K (3/3): Quantitative ablations of components

5/ Quantitative ablations of components. We thank reviewer for suggesting this analysis, which could help clearly indicate the effectiveness of each component update. Below, we provide more quantitative ablations of each component and we already updated this in the revised manuscript. For the ablation, we rerun the forward simulation process in the GSO benchmark by keeping each component at its rest value during simulation, while everything else is unchanged. As in Table 1 of the main paper, these scores are averaged over all 24 frames of the sequence, including the static first frame.

VariantPSNRSSIM
w/o dipole rotation24.030.909
w/o radius scaling24.160.903
w/o appearance rotation23.870.907
w/o adjacency rebuilding24.220.917
Full25.730.930

The full update rule achieves the best result, and removing any single component lowers PSNR by 1.5–1.9 dB, which indicates that each update contributes to a faithful simulation result.

Reviewer_gMYdOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors put forward a method that uses powerfoam as the representation of 3D scene apperance and geometry, then implement the rgb-supervised differentiable simulation similar to strategy of PhysDreamer. The main contribution lies in the coupling of the powerfoam representation and MPM simulation, especially during deformation. Such design enables high-fidelity dynamic reconstruction and material estimation, as well as well performance in several downstream application like scene editing, semantic selection, dynamic ray-traced reflection. The paper has done sound comparisons and ablation experiments to prove the effectiveness.

Critical strengths
  1. The work presents a cohesive and well-developed framework, integrating simulation, geometry and appearance updates, and differentiable rendering into a complete pipeline. Several pratical tasks along with sufficient comparisons have proven its practical usability. Diverse results shown in the paper and webpage looks great.
Critical weaknesses
  1. The evaluation doesn't include Vid2Sim despite using its benchmark. The PSNR and SSIM reported in the Vid2Sim paper are much higher than those reported for PowerSim.
Questions for authors
  1. see the weakness, the authors should clarify why the results from Vid2Sim are not included.
  2. Authors should report the compute cost and time for each part of the pipeline.
  3. Since the primitive update retains only local rotation and isotropic volume scaling from the deformation gradient, could some simulated stretching or shearing be inadequately represented in the rendered geometry?
  4. As shown in figure 2, when the bread has been stretched into 2 parts, there are clear pointcloud-liked artifacts between them. Can authors clarify what cause this? Despite the limited novelty, I will raise my score if the authors adequately address the concerns above, given the compelling results presented.
6 comments
AuthorsOpenReview ↗

Response to gMYd (1/2): Vid2Sim inclusion and compute cost and time report.

We thank reviewer's constructive feedback and positive assessment of PowerSim as a cohesive, well-developed framework with practical applications, thorough comparisons, and diverse results. We are also glad they found the results in the paper and on the project page compelling. We now address the reviewer's question as follows:

1/ Vid2Sim benchmark inclusion. Tables 1–2 compare PAC-NeRF, PhysDreamer and PowerSim because all three use the same MPM simulator and differ only in how each representation (NeRF, Gaussian splats, PowerFoam) consumes the MPM output to drive its rendering primitives, which isolates the effect of the representation. Vid2Sim instead uses a reduced-order simulator based on Linear Blend Skinning, so its results reflect both the solver and the representation. We agree that Vid2Sim should be reported, and we have added it to the appendix as a reference row next to Tables 1–2, noting that part of the difference may come from its different simulator.

For reference, Vid2Sim reports higher reconstruction scores on this benchmark (30.17 PSNR / 0.961 SSIM vs. 25.73 / 0.930 for PowerSim) and a lower Poisson's ratio error (0.06 vs. 0.16), while PowerSim achieves a lower Young's modulus error (0.44 vs. 0.51 in log⁡10E\log_{10} E) despite using a single camera view. Note that Vid2Sim jointly refines its skinning model and material parameters using multi-view videos. In contrast, our Table 1 results evaluate forward simulation with the benchmark's ground-truth material parameters, scored on the rendered frames from one camera, and our Table 2 estimates are fitted from that single view.

More importantly, we additionally show that Vid2Sim might work well for drop test, where deformation are moderate but fail to capture large deformation. To test this, we ran Vid2Sim's own simulator on the same bread-roll tearing example and just updated Fig. 2 to include Vid2Sim under this test (video results could also be found on our updated project website). As results suggest, Vid2Sim fails to capture extreme deformation under stretching and does not show seperation in this setup. Also, the original design of Vid2Sim limits to homogeneous material properties estimation (global scalar) while our PowerSim offers heterogeneous option (scalar field) which demonstrate another advantage of our PowerSim as compared to Vid2Sim.

2/ Compute cost and run time. We thank the reviewer for pointing this out. Below we provide the runtime and compute cost report and we also updated these information in the revised manuscript for more details. All timings are measured on a single NVIDIA A100 80 GB GPU.

a) Simulation Forward: On the GSO benchmark, we simulate 24 frames per object, matching the benchmark's reference video. Each object has 500k simulated particles with 2,500 MPM substeps per frame (runs at 13 s per frame), results in 5.2 min for the full sequence, with a peak GPU memory (VRAM) usage of about 28 GB. On real world scene such as carnation and garden scene, we only simulate a subset of its particles (which are relevant to moving parts of the scene), results in 195k/1.2M primitives for carnation and 229k/1.5M primitives for garden scene. We use 1,333 substeps per frame and simulate 80 frames for both scenes, where carnation scene finishes its run at 1.6s per frame (2.1 min in total) with about 6 GB of peak GPU memory, while garden scene runs at about 1.7s per frame (about 2.3 min in total) with about 7 GB.

b) Material Estimation/Optimization: On the GSO benchmark (Table 2), we skip the first optimisation stage, which recovers the initial velocity because the benchmark already specifies each object's initial state: the object is released and falls freely under gravity. This takes 11.9 min per object on average (median 6.3 min) on a single A100 GPU, for about 500k particles per object and 10 optimizaton iterations. On real captured videos such as the carnation scene, the initial velocity induced by the unobserved interaction is unknown, so we first optimise it from the first 3 frames. This first stage takes 7 min. We then optimise a spatially varying EE field over all 25 frames, which takes 53 min. This extra cost comes from additional number of primitives (1.2M PowerFoam primitives) as well as treating material properties as spatial field rather than global scalar as conducted in the GSO benchmark). The peak GPU memory for this stage is about 26 GB VRAM.

c) Foam Selection: Our dynamic primitive selection requires no optimisation: each primitive's label is obtained by a render-weighted vote of the 2D masks over the training views. For GSO benchmark, we do not apply foam selection to fairly compare with previous methods. On real world scene (e.g. carnation scene), our foam selection mechansim can take the multi-view 2D segmentation masks of the simulated object and produce selected primitives in about 1–3 min depending on the scene size with a peak GPU memory (VRAM) usage of about 5 GB.

AuthorsOpenReview ↗

Response to gMYd (2/2): Local update and Figure 2 clarification

3/ Local rotation and isotropic volume scaling. We thank the reviewer for this thoughtful question, which highlights a deliberate design choice at the core of PowerSim: each primitive stays isotropic and compact, while anisotropic deformation is carried by the evolving geometry of the power diagram. Under large deformation such as shearing and stretching, PowerSim represents them at the level of neighbouring cells rather than within individual primitives. Although each primitive retains only the local rotation and isotropic volume scaling from the deformation gradient, two mechanisms carry the anisotropic part of the deformation:

(i) Geometry. Power-cell faces are bisector planes between neighbouring sites. When MPM moves the sites apart along one axis, the cells elongate along that axis, so stretching and shearing are captured by the collective arrangement of cells rather than by any single primitive’s shape. Also, the dipole plane that carries the surface is attached to its site. The site's position already accounts for translation, so the dipole plane only needs the local rotation RR from the polar decomposition F=RSF = RS to follow the surface.

(ii) Appearance. Each primitive’s dipole frame and appearance directions rotate with the local rotation R from the polar decomposition F = RS, so surface detail follows the material as it deforms.

This design gives PowerSim a key advantage over PhysGaussian. Anisotropic 3DGS absorbs deformation by stretching individual primitives, which under significant deformation produces highly elongated primitives and visual artifacts (Fig. 2). PhysGaussian therefore requires an additional regularization to prevent overly skinny kernels (Sec. 3.8 of PhysGaussian). In contrast, PowerSim offloads anisotropic shape changes to the geometry of neighbouring cells, so the representation can deform significantly without elongating any primitive and without such regularization.

As a result, PowerSim produces realistic results across a range of materials, including under large deformation such as tearing a bread roll (Fig. 2) and twisting bread (Fig. 7), with minimal per-scene tuning. PhysGaussian, by contrast, shows visual artifacts without internal particle filling. These results are best viewed in the videos on our project webpage.

4/ Figure 2 clarification. In Fig.2, we apply two grips that prescribe the velocity of the material at both ends of the bread (specified within a small slab). The rest of the bread follows only through elastic forces. As the two halves pull apart, the material in the middle thins out, and a few small groups of particles lose contact with both sides. The specks correspond to small groups of simulated particles that detach from the thinning middle section. Because we omit gravity in this setup, so the detached fragments do not fall while the two ends continue to move apart.

Thank you for the detailed response and additional experiments. I agree that Vid2Sim's use of multi-view supervision makes the reported scores difficult to compare directly with. The runtime and memory breakdown is also clear. Regarding local rotation and isotropic scaling, the volume and normal errors reported in your response to Reviewer XVdH largely address my concern about how faithfully the rendered geometry captures the deformation. The clarification of Fig.2 is acceptable. Given the overall quality of the visual results and comprehensive quantitative comparisons, I will maintain my original recommendation of weak acceptance.

Reviewer_XVdHOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PowerSim couples PowerFoam power-diagram primitives to the Material Point Method (MPM) for differentiable simulation and rendering. It updates primitive positions, isotropic radii, dipole frames, appearance directions and adjacency, and fits spatial material fields from monocular deformation videos. Its main technical contribution is this coupling, building on PowerFoam's geometry, appearance and ray-tracing support. Render-weighted mask voting provides object segmentation for scene composition and simulation. Evaluation comprises a 12-object FEM-generated benchmark, qualitative deformation ablations and demonstrations of dynamic reflections, and segmentation on five Mip-NeRF 360 scenes.

Critical strengths
  1. The coupling of PowerFoam and MPM is concrete and useful: geometry, appearance and adjacency evolve within the captured representation, supporting deformation, composition and reflections. The Fig. 7 ablation reports cleaner boundaries and more consistent appearance when dipole frames and appearance directions are updated, supporting the practical value of the integration.

  2. The reported benchmark gains are meaningful, including both dynamic reconstruction and material estimation. The paper is easy to follow, with clear figures and demonstrations.

Critical weaknesses
  1. Consistency between PowerFoam geometry and the MPM state remains insufficiently validated. For example, rest volumes come from occupied-voxel particle counts rather than occupied power-cell geometry. The radius update matches the bounding-sphere volume factor without enforcing corresponding occupied-cell volume change. Rotation-only dipole frames omit general stretch-induced changes of base-plane normals. Fixed nonzero normal displacements prevent surface relief from scaling consistently under uniform dilation (Sec. 4.1; A.2). These raise concerns about whether the conversions are physically plausible under simulation.

  2. Material-estimation baselines are limited to PAC-NeRF and PhysDreamer (Tables 1–2). Pixie, which Sec. 2 describes as predicting material fields from visual features, should be included for comparison. Approaches like OmniPhysGS, NeRF2Physics, and VLMs using only images could provide complementary comparisons as well.

  3. Similar in the case of the segmentation comparison. Table 5 compares LabelGS, GaussianGrouping and SemanticFoam, but does not establish an advantage. Broader, applicable comparisons like FlashSplat, OmniSeg3D, or GaussianCut would be applicable segmentation baselines as well. There are also methods that do not require post-optimization but directly assign, like GaussianCut.

Questions for authors
  1. For Tables 1–2, which frames and viewpoints are used for fitting and evaluation, and how are baseline inputs, reconstruction quality, optimization budgets, density, scale and boundary/contact conditions matched? Does the recovered material field retain its advantage on a new loading or motion sequence with these assumptions fixed?

  2. Under controlled dilation and shear, how large are occupied-volume, surface-position and normal errors relative to the intended deformation, and how do these affect recovered material parameters?

  3. Would GaussianFluent be an applicable comparator for the large-deformation setting in Fig. 2? It seems to be better at modeling high dynamic simulations.

7 comments
AuthorsOpenReview ↗

Response to XVdH (1/4): Consistency between PowerFoam geometry and the MPM state

We thank reviewer for recognizing “the coupling of PowerFoam and MPM is concrete and useful,” and for highlighting the “cleaner boundaries and more consistent appearance” in the Fig. 7 ablation. We also appreciate their assessment that “the reported benchmark gains are meaningful, including both dynamic reconstruction and material estimation,” and that the paper is “easy to follow, with clear figures and demonstrations.” We now address their questions about geometric accuracy under deformation, the evaluation protocol and the additional baselines below.

1/ Consistency between PowerFoam geometry and MPM state. a) Rest Volume: Our rationale is to follow the standard MPM practice in assigning volume to particles: each occupied background-grid cell’s volume is divided equally among the particles inside it. This accounts for local particle density. Considering a region where PowerFoam places primitives more densely, each particle receives a proportionally smaller volume, so mass stays uniformly distributed regardless of how primitives were sampled during reconstruction.

Exact power-cell volumes are also available, since the cells are explicit, and we can also use them directly for simulation. We compare the two choices on the GSO benchmark, replacing the grid-based volume with the exact power-cell volume and keeping everything else fixed:

PSNR ↑

Volumebackpackbellblocksbuscreamelephantgrandfatherleatherlionmariosofaturtleavg
Grid volume23.6924.3929.5626.8425.8222.8022.0135.3325.1720.3124.7228.1125.73
Power-cell volume23.3824.7629.5726.5626.2522.9521.9335.9125.6920.3024.6727.5325.79

SSIM ↑

Volumebackpackbellblocksbuscreamelephantgrandfatherleatherlionmariosofaturtleavg
Grid volume0.8890.9440.9530.9380.9230.9080.8890.9780.9300.9410.9180.9530.930
Power-cell volume0.8850.9460.9550.9360.9260.9100.8890.9790.9340.9320.9170.9500.930

As shown, two versions are comparable in terms of PSNR and SSIM. This indicates that the grid-based assignment is already a good approximation of the exact cell volume, and that PowerSim is not tied to either choice; both are supported. An interesting direction for future work is to jointly optimize geometry and simulation-relevant quantities such as cell volumes, so that the cells are tuned for simulation as well as rendering.

b) Radius update and cell volume: We agree that the radius update matches the bounding-sphere volume factor and does not explicitly enforce the corresponding occupied-cell volume change. However, the occupied region is maintained implicitly by construction: it is the power cell clipped by the primitive's dipole plane, and both follow the simulation. The cell faces move with the neighbouring sites, and the dipole plane moves and rotates with the primitive's own site. We verify this quantitatively in 5/: under controlled deformation, the occupied-volume error is at most 0.5% for uniform dilation and at most 0.22% for simple shear up to γ=0.5\gamma = 0.5.

c) Rotation-only dipole frames: We agree that rotating the dipole frame by RR alone does not reproduce the exact stretch-induced change of each base-plane normal, but it is not the only mechanism through which the surface follows the deformation. Each dipole plane is attached to its primal site (Sections 3 and 4.1), so under stretch the sites move apart and carry their planes with them, and the overall surface follows the deformation through the motion of the sites; RR then orients each local plane to match the rotating material. We quantify the remaining error in 5/: the mean normal error is at most 0.2° under uniform dilation and 3.4° under simple shear up to γ=0.5\gamma = 0.5.

d) Fixed nonzero normal displacements: We really appreciate the reviewer for pointing this out, which comes from a typo in Eq.2 and our confusing explanation in Appendix A.2. In our implementation, the normal displacements are scaled by the primitive radius in the same way as the in-plane offsets. By "fixed" in Appendix A.2, we mean that the stored values si,ks_{i,k} and di,kd_{i,k} are not explicitly updated during simulation. Because they are defined relative to ri\mathbf{r}_i, updating the radius rescales them implicitly in world space, so the surface relief scales consistently under uniform dilation without a separate update. We have corrected Eq. 2 and clarified this in Appendix A.2 of the revised manuscript.

AuthorsOpenReview ↗

Response to XVdH (2a/4): More material estimation baseline comparisons

2/ Material-estimation baselines. We thank the reviewer for the suggestion. Tables 1–2 compare against PAC-NeRF and PhysDreamer because this isolates how the representation affects simulation-based material estimation when the simulation solver is fixed. All three methods use the same MPM simulator, inputs, and optimisation budget, and differ only in how each representation (NeRF, Gaussian splats, PowerFoam) consumes the MPM output to drive its rendering primitives. This controlled setting isolates the effect of the primitive, which is the contribution of our paper.

Pixie, OmniPhysGS, NeRF2Physics, and VLMs address a complementary problem: they infer material properties from priors, either a learned prior over appearance (Pixie’s UNet over feature fields; NeRF2Physics’ language-model proposals fused with CLIP features) or a video-diffusion prior on how such an object should move (OmniPhysGS). None of them observes the object’s actual response to interaction. In contrast, PowerSim estimates EE and ν\nu solely from the object’s observed deformation in the captured video, without any learned material prior.

Nevertheless, as the reviewer suggests, we evaluate these methods following the Tab. 2 benchmark as a reference. Since several of them do not natively output EE and ν\nu, we adapted each as follows:

  • Pixie: We fit and voxelise an F3RM CLIP feature field from the renders, apply the pretrained UNet to predict material class, EE, and ν\nu per voxel, and take the median over occupied voxels.
  • OmniPhysGS: In the official implementation, OmniPhysGS is a material-type classifier rather than a material-value estimator. Every constitutive expert uses the same fixed E=2×106E = 2\times10^6 Pa and ν=0.4\nu = 0.4, and optimisation only learns per-particle logits that select among these experts. Following the official pipeline, the input is the object’s 3DGS and a text prompt (“<object description> falling ...”), and the rendered MPM simulation is supervised by a text-to-video diffusion model via score distillation. Note that, EE and ν\nu are not optimised, we use the same constant values for every object, as reflected in its row.
  • NeRF2Physics. We follow the official pipeline and extend its original properties (mass density, friction, hardness) to EE and ν\nu. We fit a Nerfacto field to the multi-view renders and fuse its CLIP features onto a point cloud of the object. A VLM (Qwen3-VL-8B) proposes five candidate materials with EE and ν\nu ranges from one canonical view, and CLIP-based kernel regression assigns each point a value from these candidates. The object-level estimate is the mean over points, so its EE reflects the language model’s notion of the material (e.g., wood or rubber) rather than the observed behaviour.
  • VLM. We query Qwen3-VL-8B-Instruct with a single rendered view of the object on a white background, using the prompt below. We repeat this independently for three views per object and average log⁡10E\log_{10} E and ν\nu over the three answers. The prompt that we use is
You will be given an image of an object. Estimate the effective Young's modulus (in Pa) and Poisson's ratio of the object as a whole, i.e. the values a physics simulator should use for the object's deformation. Answer only with JSON of the form {"description": description, "youngs_modulus_pa": number, "poisson_ratio": number}. Use plain numbers.

MAE of log₁₀E (lower is better)

CategoryMethodbackpackbellblocksbuscreamelephantgrandfatherleatherlionmariosofaturtleavg
Prior-basedPixie0.881.571.751.491.831.781.921.961.252.052.711.411.72
Prior-basedOmniPhysGS1.462.052.191.421.041.500.910.782.270.481.920.671.39
Prior-basedNeRF2Physics4.296.366.545.714.083.384.123.295.633.685.034.954.76
Prior-basedVLM query2.495.954.383.983.070.200.392.145.742.104.952.373.15
Observation-basedPAC-NeRF3.281.084.023.303.223.052.991.202.343.370.201.942.50
Observation-basedPhysDreamer0.101.070.740.400.280.540.430.331.011.180.680.440.60
Observation-basedPowerSim (ours)0.060.810.160.600.370.360.471.210.200.600.330.090.44
AuthorsOpenReview ↗

Response to XVdH (2b/4): More material estimation baseline comparisons

Absolute error of ν (lower is better)

CategoryMethodbackpackbellblocksbuscreamelephantgrandfatherleatherlionmariosofaturtleavg
Prior-basedPixie0.010.160.020.010.060.110.020.020.010.140.080.050.06
Prior-basedOmniPhysGS (constant prediction)0.050.090.030.050.030.050.020.020.010.090.010.010.04
Prior-basedNeRF2Physics0.030.150.050.020.030.150.050.040.070.180.110.100.08
Prior-basedVLM query0.050.190.070.000.070.150.080.080.110.160.110.090.10
Observation-basedPAC-NeRF0.210.230.330.160.120.060.360.260.140.330.300.010.21
Observation-basedPhysDreamer0.260.010.100.140.140.040.330.110.060.440.050.290.16
Observation-basedPowerSim (ours)0.170.180.160.120.170.270.010.110.160.250.090.210.16

As result shown, PowerSim reduces the log₁₀ E error by roughly 3× over the best prior-based baseline. For ν, prior-based methods score lower. Note that, OmniPhysGS outputs a constant ν = 0.4 for every object yet achieves the lowest error, suggesting that ground-truth ν values in this benchmark cluster near typical defaults. Poisson’s ratio is usually weakly identifiable from observed motion, as it has a much smaller effect on the dynamics compared to E; PowerSim estimates it solely from the observed deformation without a prior, whereas the baselines benefit from priors aligned with this narrow range.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their time and their constructive, detailed feedback. We are encouraged that the reviewers found PowerSim to be a "cohesive, well-developed framework" with practical applications, thorough comparisons and diverse results (gMYd), that "the coupling of PowerFoam and MPM is concrete and useful" (XVdH), and that it offers "a practical way to animate PowerFoam scenes while updating geometry, appearance, and adjacency together" (kv8K). We also appreciate the reviewers' recognition that the benchmark gains are "meaningful, including both dynamic reconstruction and material estimation" (XVdH) and the dynamic reconstruction results "encouraging" (kv8K), that our "optimization-free primitive-selection procedure is useful" (kv8K), and that the paper is "easy to follow, with clear figures and demonstrations" (XVdH), with compelling results in the paper and on the project page (gMYd).

We have addressed all of the reviewers' concerns in the individual responses, and we have updated the submission PDF and its appendix to incorporate these changes. The main updates are summarized below.

New comparisons

  • Large deformation (Fig. 2): We have added GaussianFluent (XVdH) and Vid2Sim (gMYd) to the bread-roll tearing example, together with a citation to GaussianFluent in the related work.
  • GSO benchmark (Tables 1–2): We have added Vid2Sim as a reference row in the appendix, with a note on its different simulator (gMYd).
  • Material estimation: We have added prior-based baselines (Pixie, OmniPhysGS, NeRF2Physics and a VLM query) alongside the observation-based methods (XVdH).
  • 3D segmentation: We have added FlashSplat, GaussianCut and OmniSeg3D, with run times (XVdH).

New analyses

  • Geometric consistency: We provide a comparison of grid-based and exact power-cell rest volumes, and a controlled dilation and shear study reporting occupied-volume, surface and normal errors (XVdH).
  • Static vs. dynamic reconstruction: Static and dynamic reconstruction scores are reported separately, showing that our gain does not come from static reconstruction quality (kv8K).
  • Quantitative ablation: We show that each update rule (dipole rotation, radius scaling, appearance rotation, adjacency rebuilding) contributes to the final result (kv8K).
  • Compute cost: We report run time and peak GPU memory for forward simulation, material estimation and foam selection (gMYd).

Clarifications

  • Corrected a typo in Eq. 2 and clarified the normal-displacement update in Appendix A.2 (XVdH).
  • Added the full evaluation protocol for Tables 1–2: frames, viewpoints, inputs, optimization budget, density, scale and contact conditions (XVdH).
  • Clarified the design rationale of keeping each primitive isotropic while the power diagram carries the anisotropic deformation (kv8K, gMYd), and the simulation setup of Fig. 2 (gMYd).

Project website. We have added a Rebuttal section to the project website with further qualitative video examples, including the GaussianFluent and Vid2Sim comparisons under large deformation. We encourage the reviewers to view these results as videos.

We hope these responses and revisions resolve the reviewers' concerns, and we are happy to provide any further clarification during the discussion period.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.