Skip to content

Pb4U-GNet: Resolution-Adaptive Garment Simulation via Propagation-before-Update Graph Network

Conference: AAAI 2026 arXiv: 2601.15110 Code: github.com/adam-lau709/PB4U-GNet Area: 3D Vision Keywords: Garment Simulation, Graph Neural Network, Cross-Resolution Generalization, Message Propagation, Resolution Adaptation

TL;DR

This paper proposes Pb4U-GNet, which decouples message propagation from feature update (Propagation-before-Update) and incorporates resolution-aware propagation depth control and update scaling mechanisms, enabling garment simulation models trained solely on low-resolution meshes to generalize to high-resolution meshes.

Background & Motivation

Garment simulation is a core technology for applications such as virtual try-on and digital human modeling. Traditional physics-based methods (e.g., mass-spring systems) incur prohibitive computational costs, particularly on high-resolution meshes that require repeated constraint solving. Graph neural networks (GNNs) have emerged as effective acceleration alternatives, yet existing GNN methods suffer from severe cross-resolution generalization failures—performance degrades sharply on meshes outside the training resolution, especially at higher resolutions.

The authors identify two fundamental causes of cross-resolution failure:

Fixed message propagation depth: Existing GNNs use a fixed number of propagation layers, restricting each vertex to perceiving neighbors within a predefined hop count. On fine meshes, fixed depth leads to insufficient receptive field coverage; on coarse meshes, it causes over-smoothing.

Resolution-dependent displacement magnitude: In higher-resolution meshes, the same global motion is distributed across more vertices, resulting in smaller per-vertex displacements. This intrinsic resolution dependency causes models trained at low resolution to overestimate displacements at high resolution.

These two issues reveal a core contradiction: training directly on high-resolution meshes is computationally prohibitive, yet models trained at low resolution fail to generalize. This is the central challenge this paper addresses.

Method

Overall Architecture

The core innovation of Pb4U-GNet is decoupling message propagation from feature update. In conventional GNNs, each layer simultaneously performs message aggregation and feature update. Pb4U-GNet instead conducts \(K\) steps of pure message propagation to expand the receptive field, followed by a single unified feature update. This decoupled design allows the propagation depth \(K\) to be flexibly adjusted according to resolution without affecting the update frequency.

The overall pipeline is: input current mesh state \(\mathbf{X}_t\) → vertex/edge encoders → \(K\)-step message propagation → feature update → 15-layer MeshGraphNet refinement → vertex decoder predicting acceleration → resolution-aware scaling → forward Euler integration to obtain the next-timestep positions.

Key Designs

1. Propagation-before-Update (PbU): Decoupled Propagation and Update

Core Idea: Pure message aggregation is first performed to accumulate neighborhood information, followed by a unified feature update.

During the propagation phase, each vertex maintains an aggregated feature vector \(\mathbf{h}_{t,i}\), initialized to the vertex embedding \(\mathbf{v}_{t,i}\). At each step \(k\), neighborhood information is aggregated via a learnable message function \(f_m(\cdot)\) (implemented as an MLP):

\[\tilde{\mathbf{h}}^k_{t,i} = \text{LayerNorm}\left(\sum_{j \in \mathcal{N}(i)} f_m(\mathbf{h}^{k-1}_{t,i}, \mathbf{h}^{k-1}_{t,j}, \mathbf{e}_{t,ij})\right)\]

New and historical information are then fused via decay accumulation:

\[\mathbf{h}^k_{t,i} = \gamma \cdot \mathbf{h}^{k-1}_{t,i} + \tilde{\mathbf{h}}^k_{t,i}\]

where \(\gamma\) is a decay factor controlling the influence of historical messages. After \(K\) propagation steps, an update function \(f_u\) (MLP) integrates the original embedding with the accumulated features:

\[\mathbf{v}'_{t,i} = f_u(\mathbf{v}_{t,i}, \mathbf{h}^K_{t,i})\]

Design Motivation: Decoupling ensures that receptive field size is determined solely by the propagation step count \(K\), independent of update frequency, enabling flexible adaptation to different resolutions.

2. Resolution-Aware Propagation Control

Core Idea: Dynamically adjust propagation steps \(K\) based on mesh density to maintain a consistent physical propagation distance.

The effective physical propagation distance is defined as \(D = K_{\text{base}} \times \bar{L}_{\text{base}}\), where \(K_{\text{base}}\) is the number of propagation steps at the reference resolution and \(\bar{L}_{\text{base}}\) is the mean edge length at the reference resolution. For meshes at any resolution:

\[K = \lfloor D \times \bar{L}^{-1} \rfloor\]

As resolution increases (i.e., \(\bar{L}\) decreases), \(K\) scales proportionally to maintain consistent physical receptive field coverage.

Design Motivation: Grounded in the physical intuition that elastic wave propagation distance in cloth is independent of mesh discretization; thus, the same physical propagation distance should be maintained across different resolutions.

3. Resolution-Aware Update Scaling

Core Idea: Scale predicted accelerations according to the local geometric scale of each vertex.

Based on the geometric similarity principle from continuum mechanics (displacement fields scale linearly with element size), a per-vertex scaling factor is computed as:

\[\mathbf{s}_i = \frac{1}{|\mathcal{N}(i)|} \sum_{j \in \mathcal{N}(i)} l_{ij}\]

i.e., the mean length of all edges connected to vertex \(i\) in the rest state. The final acceleration is \(\mathbf{A}_{g,t} = \mathbf{S} \odot \tilde{\mathbf{A}}_{g,t}\).

Design Motivation: In higher-resolution meshes, each vertex represents a smaller area and mass; thus, per-vertex acceleration should be smaller for the same global deformation. This scaling restores physically consistent displacement magnitudes.

Loss & Training

The model is trained in a fully self-supervised manner without requiring ground-truth simulation data. Six physics-based loss terms are employed:

  • Stretch loss \(\mathcal{L}_{\text{stretch}}\): measures stretch/compression energy based on the St. Venant-Kirchhoff model
  • Bending loss \(\mathcal{L}_{\text{bending}}\): penalizes curvature changes between adjacent faces
  • Collision loss \(\mathcal{L}_{\text{collision}}\): quantifies garment-body interpenetration depth
  • Gravity loss \(\mathcal{L}_{\text{gravity}}\): encourages natural draping
  • Friction loss \(\mathcal{L}_{\text{friction}}\): penalizes tangential sliding at contact points
  • Inertia loss \(\mathcal{L}_{\text{inertia}}\): maintains temporal coherence

Total loss: \(\mathcal{L} = \mathcal{L}_{\text{stretch}} + \mathcal{L}_{\text{bending}} + \mathcal{L}_{\text{collision}} + \mathcal{L}_{\text{gravity}} + \mathcal{L}_{\text{friction}} + \mathcal{L}_{\text{inertia}}\)

Training details: training is conducted exclusively at the lowest resolution (11K triangles), with a 128-dimensional latent space. Both message propagation and update functions are 2-layer MLPs with 128 units, followed by 15 MeshGraphNet layers. Training runs for 100K iterations, taking approximately 36 hours on an RTX 4070 Ti.

Key Experimental Results

Main Results

Evaluated on the VTO dataset across four garment types (T-shirt, tank top, long-sleeve shirt, long skirt). All methods are trained only at Level 1 (11K) and tested at higher resolutions.

Resolution Metric Pb4U-GNet CCRAFT ESLR HOOD MGN
Lv.1 (11K) Total Loss -1.66E-02 4.24E-02 -2.56E-02 9.45E-03 4.70E-03
Lv.2 (18K) Total Loss 8.13E-03 1.10E-01 6.06E-02 2.49E-01 4.32E-01
Lv.3 (25K) Total Loss 6.34E-02 1.72E-01 1.73E-01 2.78E-01 1.44E+03
Lv.4 (38K) Total Loss 2.22E-01 2.82E-01 1.07E+05 2.57E+00 1.24E+06

At low resolution, all methods perform comparably; however, as resolution increases, competing methods degrade drastically (MGN reaches a total loss of \(1.24\times10^6\) at 38K), while Pb4U-GNet remains stable.

Ablation Study

Configuration Lv.1 (11K) Lv.3 (25K) Lv.4 (38K) Notes
Pb4U-GNet (full) -1.66E-02 6.34E-02 2.22E-01 Best
w/o propagation control -1.61E-03 1.08E+06 1.08E+09 Collapse at high resolution
w/o update scaling -5.78E-03 1.55E+13 7.34E+13 Severe collapse at high resolution
w/o both 4.70E-03 1.44E+03 1.24E+06 Degrades to baseline

Ablation results strongly demonstrate that both modules are indispensable for high-resolution generalization.

Key Findings

  1. Propagation depth scales with resolution: Figure 5 shows that, across different resolutions, physics loss converges only with more propagation steps for finer meshes.
  2. Adaptive computational efficiency: At low resolution, fewer propagation steps improve efficiency (50ms vs. MGN's 46.4ms); at high resolution, more steps ensure accuracy (196.4ms, yet with far superior total loss compared to baselines).
  3. Generalization to unseen garment types: Tested on a fitted dress and a cardigan, the total loss of 0.155 substantially outperforms CCRAFT's 0.264.

Highlights & Insights

  1. Rigorous problem analysis: Two independent factors underlying cross-resolution failure (receptive field and displacement scaling) are clearly identified, each addressed by a targeted solution.
  2. Elegant design: The Propagation-before-Update decoupling is simple yet effective, achieving resolution adaptation naturally through architectural design.
  3. Physical consistency: Both propagation distance control and update scaling carry clear physical interpretations (elastic wave propagation distance and the geometric similarity principle from continuum mechanics).
  4. Fully self-supervised: No expensive physics simulation data is required as supervision.

Limitations & Future Work

  1. The current approach validates only the fixed-\(\gamma\) decay accumulation scheme; more flexible multi-hop information fusion strategies warrant exploration.
  2. The resolution scaling relies on a simple linear relationship based on mean edge length, which may be insufficiently precise for non-uniform meshes.
  3. The method currently depends on predefined world-space distance thresholds for garment-body interaction handling, potentially limiting performance in extreme collision scenarios.
  4. Temporal adaptation (e.g., fast motions may require different propagation strategies) has not been investigated.
  • Relationship to HOOD (Grigorev 2023): HOOD employs hierarchical graph structures for long-range interaction but does not address cross-resolution generalization. The decoupling approach proposed here is more direct.
  • Limitations of super-resolution methods: Methods such as Zhang & Li 2024 rely on fixed-resolution coarse simulations; the proposed method performs simulation directly at arbitrary resolution, offering greater flexibility.
  • The concept of decoupling propagation and update may offer broader inspiration for graph network tasks requiring cross-scale generalization.

Rating

  • Novelty: ⭐⭐⭐⭐ — The decoupled propagation-update design is novel; both resolution-aware modules are grounded in solid physical motivation.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Comprehensive multi-resolution quantitative comparisons, unseen garment generalization, ablation studies, and efficiency analysis.
  • Writing Quality: ⭐⭐⭐⭐ — Problem analysis is clear, method motivation is well-articulated, and the structure is logical.
  • Value: ⭐⭐⭐⭐ — Addresses a critical bottleneck in practical deployment (train at low resolution, deploy at high resolution).