{"id":"260bd823-8dd5-445d-aa66-87d5e184e3d8","arxiv_id":"2501.01393","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"EUNet learns a cloth's energy behavior from a single piece of fabric and uses it to animate diverse garments through energy optimization, beating garment-wise supervised baselines in synthetic tests.","lead":"This paper trains a neural network, EUNet, to learn the energy behavior of cloth, including stretching, bending, and damping, from a short simulated video of a single falling square of fabric. The learned energy is then used to animate T-shirts, jumpsuits, and dresses, reducing the need for large garment-specific datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Scalar energy supervision is too weak to identify the per-edge energy whose gradients drive animation; the Dress results suggest off-manifold failure.","rationale":"The reader's conditional verdict is justified, but I would place the load-bearing weight on identifiability rather than on the exactness of Eq. 5. Even under an ideal, energy-conserving simulator with no numerical damping, Eq. 6 provides only a scalar constraint per frame on a sum of ~1.4k edge-wise energy units, and Eq. 16 is a sparse inequality; this is insufficient to determine the per-edge energy gradients that LΦ uses. The Dress row in Table 4 is the observable symptom. The proposed recovery test with a known ground-truth model would settle the question: if EUNet cannot recover forces on held-out deformation modes even from clean synthetic data, the animation gains must be attributed to the optimization integrator and collision terms rather than to captured constitutive behavior. Because the paper otherwise provides coherent methodology, reproducible code, and competitive quantitative results, I do not recommend changing the reader's CONDITIONAL verdict; the condition should include the gradient-recovery test.","tokens_in":14704,"tokens_out":11363,"duration_ms":116647,"concrete_test":"Run a synthetic recovery experiment with known ground-truth energy: generate cloth trajectories with a fixed, known hyperelastic + bending + Rayleigh-damping model (e.g., StVK membrane with known Lamé constants plus a standard discrete bending energy) using a standard implicit solver. Train EUNet exactly as in §3.2 (Eqs. 8 and 16) on these trajectories. Then evaluate on deformation modes outside the cloth training set—pure shear of a patch, cylindrical bending of a flat sheet, uniaxial stretch of a skirt-shaped strip—by comparing EUNet's predicted force −∂Φ/∂x against the ground-truth force from the known model at the same deformed states. If the relative L2 force error exceeds ~20% or the energy is not recovered up to an additive constant on these probes, the scalar supervision plus contrastive loss does not identify the constitutive law, and the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that EUNet 'directly captures constitutive behaviors' from a cloth trajectory. What the garment animation actually consumes is the force -∂Φ/∂x through LΦ in Eq. 18, not the scalar energy itself. The only direct supervision available is the system-wide scalar ΔV in Eq. 6, derived from the energy balance Eq. 5. Since Φ is a sum of ~1.4k edge-wise MLP units (Eqs. 2-3), matching one scalar per frame (Eq. 8) leaves the per-edge energy massively underdetermined: many different ϕ_p, ϕ_d produce the same sum on the training manifold of a pinned square cloth. The vertex-wise contrastive loss (Eq. 16) adds only a one-sided inequality for 10% of vertices individually disturbed, so it does not fix the magnitudes, cross-terms, or off-manifold behavior of the energy gradients. This is not merely a theoretical worry: in Table 4, MGN-S+EUNet on Dress has 156.16±87.03 mm error, much worse than MGN-S+PHYS (103.72±62.30) and even the supervised LayersNet (108.06±69.74), while on T-shirt and Jumpsuit it wins. That pattern is consistent with an energy function that fits the cloth's deformation manifold but extrapolates poorly to a different garment topology/deformation mode. The paper validates only energy values (Table 1, Fig. 3), not energy gradients against known forces, so the constitutive-law claim is not directly supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a disentangled scheme for garment animation. In the first stage, an Energy Unit network (EUNet) is trained on trajectories of a single pinned square cloth to predict per-edge potential and dissipation energies, supervised only by per-frame scalar changes of the total mechanical energy (Eqs. (5)-(8)), plus a vertex-wise contrastive loss that encourages the observed states to be local energy minima (Eqs. (12)-(16)). In the second stage, the pre-trained EUNet is inserted as the energy term L_Phi in an optimization-based (incremental-potential) loss, Eq. (18), and drives MeshGraphNet and hierarchical HOOD-style simulators to animate Cloth3D garments of T-shirt, jumpsuit, and dress types without garment-wise supervision. Experiments report lower Euclidean errors and collision rates than garment-wise supervised baselines (MGN, LayersNet) and analytical-model baselines (MGN-S+PHYS, HOOD) on T-shirts and jumpsuits, with a clear degradation on dresses for the MGN-S+EUNet variant. The central claim is that EUNet 'directly captures constitutive behaviors' of the observed cloth without analytical physics priors or a differentiable simulator.","tokens_in":15013,"tokens_out":33262,"duration_ms":315137,"significance":"If the result holds, the paper addresses a genuine data-scarcity problem in learning-based garment animation, and the specific idea of supervising a per-edge energy function by scalar energy-balance residuals, without analytical model priors or differentiable simulators, is a novel and instructive design. The disentanglement of material behavior from garment topology is conceptually appealing, and the empirical gains on T-shirt and jumpsuit over strong self-supervised physics baselines are substantial. The paper ships its code, provides ablations of the dissipation and contrastive terms, reports runtime comparisons, and its cross-garment predictions are falsifiable. However, the constitutive-law claim is currently established only at the level of scalar energies, not of the force field that actually drives the animation; and because both training and test data are synthetic cloth simulations from the same material categories, the evaluation measures transfer within a family of simulators rather than independent physical validation.","major_comments":[{"comment":"The central claim that EUNet 'directly captures constitutive behaviors' requires that the learned force field -∂Φ/∂x be the correct material force, because the animation is driven by L_Phi = Phi in Eq. (18), whose gradient provides the forces on the mesh. The only direct supervision is the system-wide scalar Delta-V of Eq. (6), one number per frame constraining a sum of roughly 1.4k edge-wise energy units (Eqs. (2)-(3)); this leaves the per-edge energies underdetermined on the training manifold, and the vertex-wise contrastive loss (Eq. (16)) is a one-sided inequality applied to only 10% of the vertices, which does not fix the magnitudes or cross-terms of the per-edge force contributions. The paper validates energy values (Table 1) and shows a single-edge energy map (Fig. 3), but never compares predicted forces ∂Phi/∂x with reference forces from the simulator on held-out states. The Dress result in Table 4 (MGN-S+EUNet: 156.16±87.03 mm, worse than MGN-S+PHYS at 103.72±62.30 and LayersNet at 108.06±69.74) is consistent with an energy function that fits the pinned-cloth deformation manifold but extrapolates poorly to a new topology and deformation mode. I request (i) a held-out force/gradient validation against the Blender internal forces, and (ii) a controlled identifiability check in which the training data are generated from an analytic energy and the recovery of the force field is measured.","section":"Section 3.2-3.3, Eqs. (2)-(8), (16), (18); Tables 2/4"},{"comment":"The derivation of the vertex-wise contrastive loss appears to evaluate the incremental potential at the wrong state. In the backward-Euler/incremental-potential framework, the observed state X^t is the minimizer of the previous step's potential E^{t-1}(x) = Σ_i K(x_i, x_hat^{t-1}_i) + V_g + Phi, whereas Eq. (11) defines E with x_hat^t_i = x^t_i + h v^t_i, whose minimizer is the next state X^{t+1}. Consequently Eq. (12), E(Phi; X^t + ΔX) - E(Phi; X^t) > 0, is not a consequence of the trajectory being a backward-Euler solution: the first-order variation at X^t is Σ_i [-(m_i/h) v^t_i - m_i g + ∂Phi/∂x_i]·δx_i, which does not vanish, and perturbations along the motion or along gravity decrease E. As written, the loss in Eq. (16) pushes ∂Phi/∂x_i toward m_i g + (m_i/h) v^t_i on the sampled vertices, i.e., it injects a spurious velocity-dependent force into the learned constitutive law, and with λ = 10^6 this bias is potentially large. The derivation should be repaired (perturb the extrapolated state x_hat^t or use the previous step's incremental potential E^{t-1}), and the sensitivity of the animation results to Lcon should be reported.","section":"Section 3.2, Eqs. (9)-(15)"},{"comment":"The supervision target in Eqs. (5)-(6) assumes an exact discrete energy identity at the observed frame rate h, with velocities estimated by finite differences and with all energy loss between frames absorbed by the learned dissipation unit of Eq. (3). Blender's integrator, the pinned-vertex constraints, and numerical damping inject energy effects that are not necessarily representable by this dissipation model, so the learned Phi_d (and, through Eq. (8), Phi_p) is entangled with solver artifacts. The transfer to Cloth3D then rests on the assertion in Section 4.2 that garments in Cloth3D share the same constitutive behavior as the observed cloth because both are labeled silk, leather, denim, and cotton; no evidence is given that the Blender material parameters and the Cloth3D stiffness attributes match, and the test-time attribute vector a is not shown to be within the training distribution of a. The evaluation therefore measures transfer within a family of synthetic simulators. I recommend a cross-simulator transfer experiment (or a test on a measured cloth dataset), and at minimum a report of the actual material parameter values used in Blender and the a vectors used for the Cloth3D garments.","section":"Section 3.2, Eqs. (4)-(6); Sections 3.1 and 4.2"},{"comment":"The Dress column contradicts the paper's success narrative and is not discussed. MGN-S+EUNet achieves 156.16±87.03 mm on Dress, worse than the analytical-model baseline MGN-S+PHYS (103.72±62.30) and the garment-wise supervised LayersNet (108.06±69.74), and the overall standard deviation of MGN-S+EUNet (147.83 mm) exceeds its mean. The dress is the test case that most directly probes the 'topology-agnostic constitutive law' claim, since its deformation manifold (loose folds, self-contact) is farthest from the pinned square cloth used in training. The paper should analyze the causes of this degradation (out-of-distribution edge features, self-collision, or the distribution of a) and explain why MGN-H+EUNet does not degrade on Dress (92.83 mm) while MGN-S+EUNet does.","section":"Section 4.2, Tables 2 and 4"}],"minor_comments":[{"comment":"In-text table references are inconsistent: Section 4.2 says quantitative results are reported in 'Table 5', but the garment-error results are in Table 2; the appendix duplicates that table as Table 4, and the time-efficiency table in A.2.3 is also called Table 5. The numbering and all in-text references should be corrected.","section":"Section 4.2; Appendix A.2.2-A.2.3"},{"comment":"The contrastive-noise scale is not defined: the text says the standard deviations range 'from σ/10 to σ' but σ is never specified; since this noise determines the activation of Lcon, the paper should define σ (e.g., relative to the cloth's mean edge length) and report the sensitivity of the results to λ and σ.","section":"Section 4.1"},{"comment":"The material attribute vector a ∈ R^5 is never defined for the four training materials nor for the Cloth3D test garments; its construction, normalization, and train/test distribution should be specified because the transfer claim depends on the EUNet being evaluated at test-time a values inside the training distribution.","section":"Section 3.2, Eq. (2); Appendix A.1.2"},{"comment":"Table 1 reports mean squared-energy errors without units or a reference scale, and several entries have standard deviations far larger than their means (e.g., denim 201.63 ± 670.55), which indicates heavy-tailed errors; the claim of 'low errors' should be supported by relative errors, medians, or a baseline magnitude.","section":"Table 1"},{"comment":"The dissipation unit takes as input the squared norm of the edge-midpoint velocity, a purely translational quantity that also enters the kinetic-energy balance in Eq. (4); the paper should justify this as a Rayleigh-type dissipation and explain how the learned ϕ_d avoids double-counting rigid translational motion in the energy balance.","section":"Section 3.2, Eq. (3)"},{"comment":"The external-force term Lexternal is described in the appendix as a random constant force, but the Cloth3D animation protocol does not specify what forces f are applied there; please clarify whether the human-body motion enters through Lexternal or only through Lcollision.","section":"Appendix A.1.4 and Section 4.2"},{"comment":"There are several minor writing issues: 'ontain' in the Table 1 caption should be 'obtain', the model name is rendered both as 'St. Venant-Kirchhoff' and 'Saint Venant Kirchhoff', and the statement that EUNet is 'translation- and rotation-equivalent' should be phrased as input invariance of the energy (equivariant energy under rigid motions) for precision.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the stress-test concern about scalar energy supervision underdetermining the force field lands, and I have found an additional internal inconsistency in the contrastive-loss derivation (Eqs. (12)-(15) evaluate the incremental potential at the current state rather than at the extrapolated or next state). I nevertheless recommend major revision rather than rejection because the central idea is defensible and the empirical gains on T-shirt and jumpsuit are substantial; the fixes called for (force-level validation, corrected contrastive-loss derivation, and analysis of the Dress failure) are within the scope of the manuscript. I would also note that the contribution is demonstrated only on synthetic data of the same material categories; for a journal submission, a cross-simulator or real-measurement validation would materially raise the confidence in the constitutive-law claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zhi,\n\nQuick take on arXiv:2501.01393 (Shao, Loy, Dai). The core idea is genuinely new: instead of training a garment-specific model on thousands of frames, they fit a per-edge energy network (EUNet) to the trajectory of a single pinned square cloth, then use that learned energy as a constraint in an optimization-based animator for T-shirts, jumpsuits, and dresses. The edge-wise decomposition into potential and dissipation branches, trained from system-level energy differences plus a vertex-wise contrastive loss, is a real combination I haven't seen in the cited prior work.\n\nWhat's done well: the ablations are clean—removing the dissipation branch or the contrastive loss degrades energy prediction and gradient quality. Held-out garment results beat the garment-wise supervised baselines (MGN, LayersNet) on T-shirt and jumpsuit, and collision rates are much lower. The speed comparison shows EUNet is roughly on par with the analytic StVK/bending models. Code is promised. The appendix also honestly notes that self-collisions are excluded from training and that the edge-wise discretization is coarse.\n\nSoft spots, in decreasing order of importance. First, identifiability: the supervision is one scalar per frame (the change in total energy), while the network has about 1.4k edge-wise units. The contrastive loss only adds one-sided inequalities on 10% of vertices with small noise. Nothing in the training directly supervises the per-edge forces that actually drive animation. The Dress row in Table 4 (156 mm with MGN-S+EUNet, worse than the analytic PHYS baseline) looks like exactly the off-manifold failure you'd expect from an energy that fits the cloth's deformation manifold but extrapolates badly to a different topology. The paper validates energy values, not energy gradients against known forces, so the phrase \"directly captures constitutive behaviors\" is stronger than what's shown. Second, the test physics isn't independent: EUNet is trained on Blender cloth and evaluated on Cloth3D garments, both synthetic and using the same material categories. That's a fair sanity check but not evidence the learned law generalizes to real fabric. Third, the material attribute vector a is described as R^5 but never specified—what exactly is being conditioned on? Minor: the text references \"Table 5\" where the main table is Table 2, a harmless but sloppy pointer.\n\nOverall: the central claim is plausible but not fully demonstrated. The Dress failure and the identifiability gap need addressing—either by gradient-level supervision, more severe perturbations, or a real cloth dataset. I'd send this to reviewers; it's a solid contribution that needs revision, not a desk reject. I wouldn't cite it until the energy gradients are validated.\n\nCheers.","headline":"A genuinely new disentangled energy-learning scheme for cloth animation, with a clear identifiability gap and validation limited to synthetic data; worth refereeing but not ready as-is.","tokens_in":15493,"tokens_out":2228,"would_cite":false,"duration_ms":21308,"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":"A single falling cloth can teach EUNet to animate any garment of the same fabric.","keywords":["garment animation","constitutive laws","energy-based optimization","cloth simulation","data-efficient learning","graph neural network","EUNet"],"falsifier":"Train EUNet on cloth trajectories generated by a simulator with known analytical material parameters (for example, a classical linearly elastic stretch model with fixed stiffness coefficients), then compare EUNet's predicted stretching and bending energy gradients against the analytical energy gradients on the same states; if the gradients diverge by more than the simulator's numerical noise, the balance-equation supervision is not recovering the true constitutive law.","tokens_in":14499,"feed_emoji":"👗","tokens_out":9054,"duration_ms":78951,"temperature":0.7,"pith_summary":"The paper tries to show that the material behavior of a fabric can be learned from the trajectories of a single pinned square cloth, and that this learned behavior is enough to animate arbitrary garments made of the same fabric. Instead of collecting large garment-wise datasets and supervising models to match each garment's dynamics, the authors disentangle the problem: first learn a constitutive law as an energy function (EUNet), then roll out garment motion by energy optimization constrained by that law. The payoff would be data-efficient garment animation that generalizes across topologies and materials, and that replaces hand-designed cloth models without needing a differentiable simulator. On benchmark garments made of the same materials, the paper reports that simulators constrained by EUNet achieve lower Euclidean errors and lower garment-human collision rates than garment-wise supervised baselines.","feed_headline":"One falling cloth teaches a network to animate any garment","feed_subtitle":"EUNet learns the fabric's energy from a pinned cloth, then uses it to drive T-shirts, jumpsuits, and dresses.","key_machinery":"The central object is the Energy Unit Network (EUNet), which predicts per-edge energy units whose sum over the mesh is the cloth's total energy. Each edge unit has a potential branch taking edge-length differences and relative vertex-normal angles (representing stretch and bending) and a dissipation branch taking edge velocities scaled by adjacent face area (representing air damping). The network is trained purely from trajectory data by matching the predicted energy change $\\Phi^{t+1}_p + \\Phi^{t+1}_d - \\Phi^t_p$ to the energy change extracted from a kinetic-plus-gravity balance equation, plus a contrastive loss derived from the backward-Euler principle that energy is minimized, so any vertex disturbance must raise total energy. Because the energy units are edge-wise and inputs are invariant to translation and rotation, the learned law is topology-independent and can be inserted directly into the incremental-potential optimization used for animation.","core_discovery":"EUNet directly captures constitutive behavior from observed cloth motion in the form of an energy function, with no analytical cloth model and no differentiable simulator as prior. The total energy of a cloth is decomposed into edge-wise potential units (functions of edge-length change and relative bending angles) and edge-wise dissipation units (functions of edge velocities, area-weighted). Training extracts the change in potential-plus-dissipation energy between adjacent frames from a balance equation of kinetic, gravitational, potential, and dissipated energy, and a vertex-wise contrastive loss enforces that random disturbances increase total energy, mirroring backward-Euler energy minimization. Once trained, the energy function is embedded as the material term in an optimization-based physics loss to animate garments of different topologies. The paper reports that this disentangled pipeline outperforms garment-wise supervised training and analytical-model baselines in Euclidean error and collision rate on benchmark garments.","pith_inferences":["The paper leaves implicit that the same edge-wise energy decomposition could be applied to other deformable materials, such as sand, metal, or snow, where an energy formulation is available; the contrastive disturbance loss would then act as a stabilizer for learned constitutive models.","A testable extension not run in the paper is cross-solver transfer: train EUNet on trajectories from one simulator and animate with a different integrator; if the learned energy law is truly material, the animated garments should stay accurate, and if not, the law has absorbed solver-specific damping.","Because the paper only evaluates on synthetic cloth, the strongest open question is whether the same training signal works on real captured trajectories; if it does, the disentangled scheme would connect video observation of cloth directly to garment animation.","The dissipation branch may not separate true air drag from numerical viscosity arising in the simulator; a probe would be to animate the same garment at different simulation time-step sizes and check that the learned material energy stays unchanged."],"forward_implications":["A pre-trained EUNet can animate unseen garments of the same material without any garment-level ground-truth data, since the same edge-wise energy law transfers across different meshes.","The learned energy function can replace analytical cloth models in a physics-loss simulator, and the paper reports lower errors and collision rates than either garment-wise supervised models or analytical-model baselines.","Because each material only needs a cloth trajectory, adding a new fabric to the animation pipeline requires collecting one simple cloth sequence instead of a large garment dataset.","Energy-constrained rollouts remain stable over long horizons, avoiding the error accumulation that the paper observes in supervised garment-wise baselines."],"supporting_citations":[{"why":"Supplies the backward-Euler incremental-potential minimization that grounds both the vertex-wise contrastive loss and the optimization-based animation scheme.","marker":"[9]"},{"why":"Supplies a hierarchical graph-network simulator and an energy-optimization rollout that serve as the strongest comparison baseline for EUNet-constrained animation.","marker":"[10]"},{"why":"Supplies the self-supervised physics-loss formulation that the animation loss is built on, and provides the analytical cloth model used as a physics-constrained baseline.","marker":"[26]"},{"why":"Supplies the graph-network simulator architecture used to test whether EUNet can constrain an existing learned simulator.","marker":"[24]"},{"why":"Supplies a garment-wise supervised baseline trained on large garment data, which EUNet-constrained animation is compared against.","marker":"[28]"},{"why":"Supplies benchmark garments and material categories used to evaluate the accuracy and plausibility of animated garments.","marker":"[2]"}],"fun_headline_variants":["One cloth drop trains energy model to animate any garment","Learn cloth energy from one falling piece, animate all clothes","Single cloth trajectory teaches network to drive diverse outfits","No physics priors: one cloth's motion yields general garment animation","EUNet: one cloth's energy captures physics, animates any garment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training signal assumes that every change in the system's energy between two consecutive frames is exactly the potential-plus-dissipation change the network learns, with velocities computed from frame positions; if the simulation software's numerical damping, collisions, or constraints quietly add or remove energy, the learned material law absorbs those solver artifacts instead of the true fabric behavior.","fun_headline_variants_meta":{"raw":{"variants":["One cloth drop trains energy model to animate any garment","Learn cloth energy from one falling piece, animate all clothes","Single cloth trajectory teaches network to drive diverse outfits","No physics priors: one cloth's motion yields general garment animation","EUNet: one cloth's energy captures physics, animates any garment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00024,"raw_usage":{"total_tokens":1556,"prompt_tokens":1018,"completion_tokens":538,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":454}},"tokens_in":634,"tokens_out":538,"duration_ms":5616,"temperature":1.0,"reasoning_tokens":454,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:28:36.872194+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train EUNet on cloth trajectories generated by a simulator with known analytical material parameters (for example, a classical linearly elastic stretch model with fixed stiffness coefficients), then compare EUNet's predicted stretching and bending energy gradients against the analytical energy gradients on the same states; if the gradients diverge by more than the simulator's numerical noise, the balance-equation supervision is not recovering the true constitutive law.","supporting_citations":[{"cited_title":"Gast, Craig A","cited_arxiv_id":null,"evidence_quote":"Supplies the backward-Euler incremental-potential minimization that grounds both the vertex-wise contrastive loss and the optimization-based animation scheme."},{"cited_title":"Black, and Otmar Hilliges","cited_arxiv_id":null,"evidence_quote":"Supplies a hierarchical graph-network simulator and an energy-optimization rollout that serve as the strongest comparison baseline for EUNet-constrained animation."},{"cited_title":"Otaduy, and Dan Casas","cited_arxiv_id":null,"evidence_quote":"Supplies the self-supervised physics-loss formulation that the animation loss is built on, and provides the analytical cloth model used as a physics-constrained baseline."},{"cited_title":"Battaglia","cited_arxiv_id":null,"evidence_quote":"Supplies the graph-network simulator architecture used to test whether EUNet can constrain an existing learned simulator."},{"cited_title":"Towards multi-layered 3d garments animation","cited_arxiv_id":null,"evidence_quote":"Supplies a garment-wise supervised baseline trained on large garment data, which EUNet-constrained animation is compared against."},{"cited_title":"CLOTH3D: Clothed 3D humans","cited_arxiv_id":null,"evidence_quote":"Supplies benchmark garments and material categories used to evaluate the accuracy and plausibility of animated garments."}],"review_version":1}