HOOD: Hierarchical Graphs for Generalized Modelling of Clothing Dynamics

1. Introduction

• Physics-based simulation can produce impressive results but is computationally expensive and sensitive to initial conditions, limiting its fit for real-time budgets.The cited state-of-the-art approaches are not designed for the strict computation budgets of real-time applications.Source: §1 · p. 2
• Skinning-based methods struggle with loose garments, while many learning-based methods are garment-specific and require retraining for each outfit.Skinning-based approaches can produce compelling results for tight-fitting garments such as shirts and sportswear.Source: §1 · p. 2
• HOOD uses graph neural networks to predict garment deformations and generalizes across garment structures, body shapes, and motions through local reasoning.The method reasons about the mapping between local deformations, forces, and accelerations.Source: §1 · p. 2
• Hierarchical message passing propagates stiff stretching waves on coarse graph levels while preserving fine detail such as folds and wrinkles.The authors report improved qualitative and quantitative predictions at equal computation budgets.Source: §1 · p. 2
• An incremental-potential loss for implicit time stepping enables fully unsupervised training without ground-truth annotations.The formulation learns multiscale dynamics, material effects, collision response, and frictional contact with the body.Source: §1 · p. 2
• A single trained network predicts motion for varied garments, generalizes to unseen garment types and shapes, and supports runtime material and topology changes.The reported capabilities include opening zippers and unbuttoning shirts.Source: §1 · p. 2

2. Related Work

• Physics-based clothing simulation can produce highly realistic, detailed results but is often computationally expensive.The field also studies mechanical modeling, material behavior, time integration, collision handling, and differentiable simulation.Source: §2 · p. 2
• Skinning-based learned models can run impressively fast, but struggle with loose garments and often require retraining when garment geometry changes.Some approaches target loose garments, but still rely on linear blend skinning and parametric body models, limiting dynamic motion generation.Source: §2 · p. 3
• Traditional mesh methods are typically fixed-topology, while implicit shape methods handle arbitrary topology at the cost of training data, pipeline compatibility, and rendering efficiency.The paper contrasts these methods with its graph formulation, which supports varied topologies using an efficient mesh representation.Source: §2 · p. 3
• PBNS trains static garment deformations with potential energy, while SNUG adds recurrent dynamic prediction but remains limited to tight-fitting and known garments [5].The paper says its approach avoids skinning and supports arbitrary garment types, including free-flowing clothing and changing topology.Source: §2 · p. 3
• MeshGraphNets generalize well, but default message passing propagates signals slowly and makes the required step count difficult to choose [40].For clothing, too few steps can cause excessive stretching and unnatural dynamics; HOOD addresses this with hierarchical multi-level propagation.Source: §2 · p. 3
• Earlier graph-pooling strategies expand propagation radius, while HOOD's coarsening lets the network learn transitions between levels without manually designed transfer operators [28] [16] [11].The cited concurrent work by Cao et al. [11] uses pre-computed coarsened geometries and handcrafted aggregation weights.Source: §2 · p. 3
• The authors identify their work as the first graph-based approach for dynamic garment motion on dressed virtual humans.They distinguish garment motion on dressed virtual humans from graph-based modeling of other physical systems, including fabric simulation.Source: §2 · p. 3

3. Method

• A single network predicts motion across garment types and shapes, generalizes to unseen clothing, and supports runtime changes to material parameters and topology.These capabilities combine garment-independent graph dynamics, hierarchical message passing, and self-supervised physics-based training.Source: §3 · p. 3

3.1. Background

• Garment and nearby body vertices form a graph whose node and edge features encode state, geometry, and physical properties; local material parameters also support multi-material garments.The network predicts per-vertex accelerations, which update garment velocities and positions.Source: §3.1 · p. 4

3.2. Hierarchical Message Passing

• A nested hierarchy accelerates signal propagation, with roughly equal computation yielding a 48-edge radius versus 15 edges for single-level message passing.Shared nodes transfer information implicitly between levels, while coarse levels handle fast stretching waves and finer levels preserve folds and wrinkles.Source: §3.2 · pp. 4–5

3.3. Garment Model

• A physics-based incremental potential combines stretching, bending, collision, inertia, and friction terms to model cloth mechanics and train without supervision.Minimizing the potential over next-step positions is equivalent to an implicit Euler update.Source: §3.3 · p. 5

3.4. Training

• Training samples body poses, shapes, garment templates, fit perturbations, and material parameters, using corresponding physical losses to teach the network their effects.The training set uses 52 AMASS pose sequences; garment meshes are initialized with skinning, which is not used at inference.Source: §3.4 · pp. 5–6
• Autoregressive training increases prediction length gradually to five steps, which the authors find sufficient for inference over hundreds of steps.Feature normalization uses statistics collected from linearly skinned garments because ground-truth data are unavailable.Source: §3.4 · p. 6

4.1. Comparison to State-of-the-Art Methods

• The perceptual study used 30 participants, eight AMASS pose sequences, and garments from each learned baseline’s training set.Four sequences were from the validation set used in and four were new, more dynamic sequences; SMPL shape parameters were sampled from U(−2, 2).Source: §4.1 · pp. 6–7
• Participants preferred HOOD over learned methods in a clear majority of cases, while preferences versus ARCSIM were nearly equal.Results came from a perceptual study comparing paired videos; ARCSIM was not compared on new sequences because it failed to resolve many self-collisions.Source: Figure 5 · p. 7, §4.1 · p. 7

4.2. Hierarchical Architecture

• HOOD’s hierarchical model outperformed Fine15 in inference speed and loss, while Fine48 slightly improved loss and gradient norm at significantly lower inference speed.The hierarchical model propagates across 48 edges per time step versus 15 for Fine15; Figure 7 illustrates stretching artifacts from Fine15’s limited propagation radius.Source: §4.2 · pp. 7–8, Table 1 · p. 7, Figure 7 · p. 7

4.3. Generalization

• HOOD generated results for modified and entirely new garments, including a manually extracted garment from a 3D scan.The model also supports garment resizing at inference time, material changes across a garment, and topology changes such as buttoning or unbuttoning.Source: §4.3 · p. 8, Figure 8 · p. 8, Figure 9 · p. 8

• HOOD predicts nodal accelerations from a graph-based garment representation, with hierarchical message passing accelerating stiff-wave propagation.The authors use an optimization-based physical-simulation formulation for supervision and report plausible motion across garment types, body shapes, unseen garments, and new pose sequences.Source: §5 · p. 8
• The method does not model garment–garment interactions, so results may exhibit self-penetrations.The authors identify self-supervised handling of garment self-collisions as future work because naïvely adding many collision edges would be prohibitively expensive.Source: §5 · p. 9
• Body motions faster than those seen during training may cause failures, while severe body self-intersections may produce erroneous body–garment correspondences.The authors describe continuous collision detection as a possible strategy for the motion-velocity limitation; automatic resolution of body self-penetrations is beyond this work’s scope.Source: §5 · p. 9