{"id":"d75c3eb1-7b1f-4b76-8ded-62e5e7f333af","arxiv_id":"2506.20936","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"PhysRig animates articulated 3D objects by simulating them as deformable soft bodies driven by an embedded skeleton, and learns the material and motion parameters with a differentiable physics simulator.","lead":"PhysRig replaces the standard lightweight math of character animation with a soft-body physics simulation, so joints deform more naturally instead of collapsing or shrinking. It learns both material softness and skeleton motion parameters, which could improve animated characters, 4D content creation, and pose transfer.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The empirical claim of superiority over LBS rests on ground-truth data generated by PhysRig's own MPM simulator (Sec. 4, A.1), making the benchmark in-sample: low CD and high UR show the optimizer fits its own forward model, not generalizable realism.","rationale":"The framework is plausible: it uses standard MPM, a reasonable material-prototype parameterization, and a legitimate alternating optimization scheme. On its own terms, fitting deformations generated by the same differentiable simulator is a valid test of parameter recovery. However, the paper's central claim is stronger: it claims realistic articulated object modeling and consistent superiority over LBS. The only quantitative evidence for that claim is the self-generated benchmark. Because the same simulator is both the data generator and the inverse model, the evaluation cannot distinguish correct physics from overfitting to the solver's discretization and parameterization. The reader's weakest assumption identifies exactly this circularity, and I agree with it. No additional internal inconsistency is needed; the missing independent benchmark is the decisive gap. If the authors supplied an independent-simulator or real-capture validation, the paper would be substantially stronger, but as presented the central empirical claim is not supported.","tokens_in":13598,"tokens_out":6976,"duration_ms":87452,"concrete_test":"Use the same 17 objects and 120 motion settings, but generate ground-truth sequences with an independent continuum solver—for example, a finite-element elastodynamics code or a different MLS-MPM implementation with distinct grid resolution and substep counts—and rerun the Table 1 protocol without touching the optimization code. If PhysRig still shows near order-of-magnitude lower CD than LBS-3 and high UR, the central claim survives; if the advantage collapses, the reported benchmark is in-sample and the generalization claim must be withdrawn. A complementary check is to run the same protocol on one object with real 4D captures (e.g., AMASS) to test realism beyond any synthetic simulator.","verdict_should_be":"REJECT","load_bearing_attack":"Table 1's quantitative claims are load-bearing, and they depend on ground-truth sequences produced by PhysRig's own forward simulator. Sec. 4 states 'we generate a large amount of synthetic data using PhysRig', and Sec. A.1 confirms the dataset is built by simulating the same MPM formulation with the same grid and substep settings used for optimization (Sec. A.2). Thus each ground-truth target X'_gt is exactly F(X, E_gt, ν_gt, v_gt, Δt) from Eq. 1, and the optimizer minimizes Chamfer distance between its own simulation and that target using the same F. The inverse problem is in-sample: PhysRig is recovering parameters of the exact code that produced the data, while LBS is not in the generative model class. The consistent CD/UR margins in Table 1 and the ablations in Table 2 therefore do not establish that PhysRig produces more realistic skinning for real or independently simulated articulated objects. The user study inherits the same confound: participants compare PhysRig-generated animations against LBS animations sampled from the same self-generated benchmark, so it measures preference within PhysRig's own deformation distribution, not external realism. Sec. 4.3's pose-transfer demonstration is qualitative and adds no independent quantitative support.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"PhysRig is a differentiable physics-based skinning and rigging framework that replaces linear blend skinning with a volumetric soft-body simulation. The object is discretized into a volumetric point cloud, skeletal motion is encoded through driving-point velocities, and a Material Point Method simulator with a fixed-corotated hyperelastic model (Eqs. 4–9) produces deformed positions X' = F(X, E, ν, v, Δt) (Eq. 1). Given an observed motion sequence, the method solves an inverse-skinning problem (Sec. 3.1.4) that alternates between optimizing material prototypes (per-prototype Young's modulus and Poisson's ratio, Sec. 3.2) and per-frame driving-point velocities. For evaluation, the paper builds a synthetic dataset of 17 objects and 120 cases from Objaverse, The Amazing Animals Zoo, and MixaMo, with ground truth generated by the same MPM simulator (Secs. 4 and A.1), and reports user-study (UR) and Chamfer-distance (CD) comparisons against three LBS baselines (Table 1), component ablations (Table 2), a user study (Sec. A.4), and a pose-transfer demonstration (Sec. 4.3). The paper claims that PhysRig consistently outperforms LBS and produces more realistic and physically plausible deformations.","tokens_in":13919,"tokens_out":20061,"duration_ms":188137,"significance":"If its claims were established, PhysRig would be a genuinely useful contribution to articulated object modeling: it provides a differentiable inverse-skinning formulation that jointly recovers material properties and skeletal driving velocities through MPM, and the material-prototype parameterization (Sec. 3.2) is a compact, smooth way to represent spatially varying elasticity. The forward model is a standard MPM with the fixed-corotated energy, the alternating material/velocity optimization is described in enough detail to reproduce (Sec. A.2), and the authors ship a project page and a released dataset. These are real strengths. However, the central empirical claim — consistent superiority over LBS and more realistic deformations — is not supported by the presented evidence, because the benchmark ground truth is generated by the very same MPM formulation that the inverse model inverts, at the same 100^3 grid and 100-substep resolution (Secs. A.1 and A.2). The CD/UR margins in Tables 1 and 2 and the user-study preferences (Sec. A.4) are thus in-sample by construction and do not measure realism on independent data.","major_comments":[{"comment":"The evaluation is in-sample and does not establish the paper's central claim. Sec. 4 states that the authors 'generate a large amount of synthetic data using PhysRig', and Sec. A.1 confirms that each ground-truth sequence is produced by simulating the same MPM formulation with the same 100^3 grid and the same 100-substep, 4e-4 s time-step settings used in optimization (Sec. A.2). Each target mesh is therefore exactly F(X, E_gt, ν_gt, v_gt, Δt) from Eq. 1, and the optimizer in Sec. 3.1.4 minimizes a Chamfer loss between its own output of the same F and that target. Because LBS is not in this generative model class, the consistent CD and UR margins in Table 1 and the ablations in Table 2 show only that PhysRig can invert its own forward model; they do not show that PhysRig produces more realistic skinning for real or independently simulated objects. The user study (Sec. A.4) inherits the same confound because participants compare videos drawn from PhysRig's own deformation distribution. The paper itself defers real-world validation to future work (Sec. 5: 'integrating real-world priors'), which makes the absence of any independent test more, not less, consequential. To support the headline claim, the authors must evaluate on real captured deformation data (e.g., 4D scans) or at least on ground truth produced by an independently implemented simulator with different discretization, and they should restrict Sec. 4.1's claim that the method 'consistently outperforms all baselines' to what the evidence can support.","section":"Secs. 4, A.1, A.2"},{"comment":"The quantitative comparisons are reported without any measure of variability. Table 1 gives averages over the motion sequences of each object but no standard deviations, and the appendix's optimization procedure involves random initialization of prototypes and velocities (Sec. A.2), so repeated runs will produce a spread of CD and UR values. Without error bars or significance tests, the claimed margins over LBS-3 (e.g., Angelfish CD 0.021 vs 0.209, or Shark CD 0.016 vs 0.031) cannot be assessed, and the user-study scores in Table 1 have no uncertainty or inter-rater agreement statistics. Report per-sequence distributions and repeated-optimization statistics for the main comparisons.","section":"Sec. 4.1, Table 1"},{"comment":"There is a mismatch between the method's stated parameters and what the evaluation actually tests. The method is described as learning a per-prototype Poisson's ratio ν ∈ R^P (Sec. 3), and the inverse-skinning objective is said to recover E, ν, and velocities (Sec. 3.1.4), but Sec. A.1 states that the dataset assumes a homogeneous Poisson's ratio for all objects. Consequently the experiments never test recovery of spatially varying ν, and Fig. 5 cannot validate that component. In addition, Fig. 5 is purely qualitative: no quantitative error metric between learned and ground-truth material properties (e.g., relative error in E and ν) is reported anywhere, so the material-recovery claim is not numerically supported.","section":"Sec. 3.2 vs Sec. A.1"},{"comment":"The pose-transfer experiment is qualitative only. It consists of a single figure (Fig. 3) with no quantitative metric, no comparison against an LBS-based pose-transfer baseline, and no evaluation across a set of motions or objects. It therefore does not provide independent support for the claimed versatility of the framework, and the cross-species generalization statement in Sec. 4.3 is not backed by measured results.","section":"Sec. 4.3"}],"minor_comments":[{"comment":"The dataset description is internally inconsistent: Sec. A.1 reports 40 motion sequences giving 120 cases (40 original plus 80 material-configuration variants), while the caption of Table 1 says that the dataset 'consists of 17 diverse objects ... totaling 120 motion sequences'. Clarify whether the count is of sequences or of cases.","section":"Sec. A.1 vs Table 1"},{"comment":"The MPM update equations use inconsistent notation. The affine velocity-gradient matrix C_p in Eq. 7 is never defined, a separate velocity gradient ∇v^{t+1}_p is introduced in Eq. 9, and the driving-point update of Eq. 10 uses ∇v_i without defining a grid-node velocity gradient. In addition, the external force f_i appears inside the particle-to-grid sum in Eq. 7, whereas such forces are normally applied to the grid after the gather; please align the equations with a standard reference and define every symbol.","section":"Sec. 3.1.3, Eqs. 7–10"},{"comment":"Table 2 is hard to read: entries such as 'w/o Locating4.31 0.186' and 'Prototypes: 25- 0.147 - 0.229 - 0.023 2000' are missing separators, the convergence column mixes iteration counts with metric values, and the arrow in the 'Converge Iteration↓' header is ambiguous because CD is lower-better while UR is higher-better. Reformat the table and explain the symbols.","section":"Table 2"},{"comment":"The subsection heading 'Affinity-Based Seg via Spectral Clustering' contains a truncated word; spell out 'Segmentation'.","section":"Sec. 3.3.1"},{"comment":"Several bibliographic entries are incomplete: [1] ('Adobe. Mixamo') has no year or venue, and [31] lists an animation studio name in place of a conventional author field; complete these entries for consistency with the journal's style.","section":"References"},{"comment":"The ablation text says that per-point materials 'struggle to find the optimal solution'; since per-point is strictly the most expressive representation, the favorable prototype results demonstrate better optimization convergence rather than higher expressiveness, and the text should state this interpretation explicitly rather than treating the comparison as a test of representational power.","section":"Sec. 4.2, Table 2"}],"recommendation":"reject","confidential_remarks":"The decisive issue is the self-generated benchmark. The dataset is listed as a contribution in Sec. 1 and is publicly released, yet its ground truth is produced by the same MPM formulation, at the same grid and substep resolution, that the paper's own inverse model fits; therefore Tables 1 and 2 and the user study cannot support the paper's headline claims. The core idea — differentiable MPM-based inverse skinning with material prototypes — is plausible and could be publishable, but repairing the evaluation requires new data sources (real 4D captures or an independent simulator), quantitative material-recovery metrics, and error bars, i.e., a substantially different experimental section. On that basis I recommend rejection rather than major revision; a rewritten manuscript with an independent benchmark could reasonably be considered anew. I saw no problematic citation or novelty-disclosure issues: prior differentiable-MPM work is cited, and the material-prototype formulation appears to be new in this context."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"PhysRig is a serious attempt to replace LBS with a differentiable MPM-based skinning pipeline, and the material prototype parameterization is a genuinely useful idea. The forward physics is standard corotated MPM, and the contribution is the inverse skinning formulation: learning per-region material properties and driving-point velocities from observed motion. The paper does a good job of describing the optimization strategy and ablations; the material prototype count and initialization study are informative.\n\nThe soft spot is the evaluation. The ground-truth sequences are generated by PhysRig's own forward simulator (Sec. 4 and A.1), so the inverse model is recovering parameters from data in the same model family. That makes the Chamfer Distance and user-study comparisons with LBS in-sample by construction. The optimizer is essentially inverting its own generator, and LBS is not in that generative class, so the consistent margins in Table 1 do not establish external realism. The user study inherits the same bias: participants compare animations drawn from PhysRig's own deformation distribution. The pose-transfer section is qualitative only. None of this makes the method useless—if the goal is to show that the optimizer can fit its own simulator, it works—but the headline claim of consistently beating LBS is not supported by independent evidence.\n\nI would add that the paper does not flag this circularity as a limitation; the appendix describes the dataset generation but treats it as a neutral benchmark. That is the main thing to fix: either evaluate on real captured deformations or on independent simulators with different constitutive models, or at least temper the claim.\n\nWho is this for? Researchers in differentiable physics for animation and 4D generation will find the material prototype formulation and the driving-point initialization useful, even if the evaluation needs another pass. It deserves a serious referee—the method is new and the flaw is fixable—but a reviewer should push for an external benchmark before acceptance.","headline":"PhysRig is a plausible differentiable MPM-based skinning framework with a novel material prototype parameterization, but its headline claim of beating LBS rests on an in-sample benchmark generated by its own simulator.","tokens_in":14407,"tokens_out":1882,"would_cite":false,"duration_ms":19537,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Physics-based skinning recovers motion linear blends miss","keywords":["differentiable physics simulation","material point method","skinning","rigging","inverse skinning","material prototypes","linear blend skinning","pose transfer"],"falsifier":"Run the inverse-skinning procedure on motion capture of real soft tissue or on sequences produced by an independent finite-element solver; if the Chamfer-distance advantage over LBS shrinks or disappears on out-of-simulator data, the central claim of recovering realistic deformations fails.","tokens_in":13397,"feed_emoji":"🦴","tokens_out":3356,"duration_ms":37349,"temperature":0.7,"pith_summary":"This paper tries to establish that skinning an articulated character can be treated as an inverse physics problem: instead of blending bone transformations on the surface, embed the skeleton in a deformable volume, simulate the volume with a differentiable Material Point Method, and optimize the material parameters and driving-point velocities so the simulated surface matches observed motion. The stakes are practical: classic Linear Blend Skinning is simple but collapses joints, shrinks volume, and cannot represent soft tissue or flexible appendages, while physics-based skinning has been non-differentiable and hard to tune. PhysRig addresses the tuning problem by introducing material prototypes, a small vocabulary of learnable elastic properties, so the optimization space stays small and smooth. If the claim holds, animators, pose-transfer systems, and 4D generators can replace surface-weight blending with a physically consistent volume model that is still trainable by gradient descent.","feed_headline":"Physics-based skinning recovers motion linear blends miss","feed_subtitle":"PhysRig drives soft-body volumes with embedded skeleton points, learning materials and velocities from observed meshes.","key_machinery":"The engine is a differentiable Material Point Method (MPM) simulator with a Fixed Corotated hyperelastic constitutive model. Particles carry mass, velocity, deformation gradient, and volume; B-spline interpolation transfers them to an Eulerian background grid for stress-based force updates and back. The skeleton is injected through driving points, which write velocities into nearby grid nodes; material prototypes are Gaussian ellipsoids whose Mahalanobis-distance softmax weights blend Young's modulus and Poisson's ratio across the volume. An alternating scheme optimizes material parameters over whole sequences and driving-point velocities frame by frame.","core_discovery":"The central claim is that the deformation of an articulated object can be recovered by inverting a differentiable continuum-mechanics simulator. PhysRig represents the object as particles in a soft-body volume driven by driving points that carry skeletal motion, and it learns both the volume's elastic parameters (Young's modulus and Poisson's ratio, expressed through material prototypes) and the driving-point velocities by minimizing Chamfer distance to an observed mesh sequence. On the paper's synthetic benchmark, built from Objaverse, Mixamo, and Amazing Animals Zoo meshes with 120 motion-material cases, this inverse-skinning procedure yields larger user-study preference scores and lower Chamfer distances than three Linear Blend Skinning baselines, including one initialized with ground-truth skinning weights.","pith_inferences":["The head-to-head with LBS is evaluated on synthetic sequences produced by PhysRig's own forward simulator, so the reported margin likely reflects in-sample fitting; a fair test would use real captured deformations or an independent solver.","A natural testable extension is to pretrain material prototypes on categories and transfer them to unseen characters, which the paper only gestures at.","The driving-point formulation resembles reduced-order control; combining it with reduced coordinates could cut the 100^3 grid cost and move the method toward real time."],"forward_implications":["Inverse skinning can recover not just surface motion but plausible internal elastic parameters for articulated objects.","Pose transfer no longer requires predicting skinning weights; driven volumes can adapt to novel skeletons and materials.","Soft regions such as trunks, ears, fatty tissue, and fins deform through stress propagation instead of linear blends, removing candy-wrapper and volume-loss artifacts.","The same volume representation accepts meshes, point clouds, and Gaussian splats, so it can slot into neural 4D reconstruction and generation pipelines.","Because the simulation is differentiable, material properties and motion can be optimized jointly with other losses in a deep learning loop."],"supporting_citations":[{"why":"Supplies the Material Point Method formulation used as the differentiable physics core of the framework.","marker":"[9]"},{"why":"Pinocchio provides the coarse skinning weights that initialize driving point placement and velocities.","marker":"[3]"},{"why":"RigNet serves as both an initialization baseline and a comparison method for LBS-based inverse skinning.","marker":"[37]"},{"why":"MagicPose4D inspires the pose-transfer application by extracting skeletons and bone angles from mesh sequences.","marker":"[42]"},{"why":"Objaverse is a source of the synthetic dataset meshes used to construct the evaluation benchmark.","marker":"[4]"},{"why":"The Amazing Animals Zoo dataset supplies additional animal meshes for the synthetic benchmark.","marker":"[31]"},{"why":"Mixamo provides humanoid character meshes and motion data used in the dataset and pose-transfer experiments.","marker":"[1]"},{"why":"The affine particle-in-cell method underlies the MPM discretization that enables differentiable simulation.","marker":"[12]"}],"fun_headline_variants":["PhysRig: differentiable physics beats linear blend skinning","Soft-body physics replaces LBS for realistic articulated motion","Learn material and velocity from meshes via differentiable physics","Physics-based skinning overcomes LBS artifacts and volume loss","Differentiable soft-body rigs driven by skeleton motion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that synthetic motion sequences generated by PhysRig's own forward simulator are a reliable test bed for skinning realism, so the comparison with LBS is measured on the same model that generates the ground truth.","fun_headline_variants_meta":{"raw":{"variants":["PhysRig: differentiable physics beats linear blend skinning","Soft-body physics replaces LBS for realistic articulated motion","Learn material and velocity from meshes via differentiable physics","Physics-based skinning overcomes LBS artifacts and volume loss","Differentiable soft-body rigs driven by skeleton motion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000708,"raw_usage":{"total_tokens":3190,"prompt_tokens":945,"completion_tokens":2245,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":2166}},"tokens_in":561,"tokens_out":2245,"duration_ms":20473,"temperature":1.0,"reasoning_tokens":2166,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:38:14.825992+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the inverse-skinning procedure on motion capture of real soft tissue or on sequences produced by an independent finite-element solver; if the Chamfer-distance advantage over LBS shrinks or disappears on out-of-simulator data, the central claim of recovering realistic deformations fails.","supporting_citations":[{"cited_title":"A moving least squares material point method with displacement discontinuity and two-way rigid body coupling.ACM Transactions on Graphics (TOG), 37(4):1–14, 2018","cited_arxiv_id":null,"evidence_quote":"Supplies the Material Point Method formulation used as the differentiable physics core of the framework."},{"cited_title":"Automatic rigging and anima- tion of 3d characters","cited_arxiv_id":null,"evidence_quote":"Pinocchio provides the coarse skinning weights that initialize driving point placement and velocities."},{"cited_title":"Objaverse-xl: A universe of 10m+ 3d objects.Advances in Neural Information Processing Systems, 36:35799–35813, 2023","cited_arxiv_id":null,"evidence_quote":"Objaverse is a source of the synthetic dataset meshes used to construct the evaluation benchmark."},{"cited_title":"The Amazing Ani- mals Zoo Dataset","cited_arxiv_id":null,"evidence_quote":"The Amazing Animals Zoo dataset supplies additional animal meshes for the synthetic benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Mixamo provides humanoid character meshes and motion data used in the dataset and pose-transfer experiments."},{"cited_title":"The affine particle-in-cell method.ACM Transactions on Graphics (TOG), 34(4):1–10,","cited_arxiv_id":null,"evidence_quote":"The affine particle-in-cell method underlies the MPM discretization that enables differentiable simulation."}],"review_version":1}