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

AI summary

Overview

Research area: Computer vision and graphics — neural (graph-neural-network) garment and cloth simulation, with a specific focus on cross-resolution generalisation.

Technical level: Advanced. The paper assumes familiarity with graph neural networks, message passing, mesh representations, autoregressive physics simulation, and physics-based loss terms.

Scope: The paper proposes Pb4U-GNet, a graph-network simulator that decouples message propagation from feature updates and adds two resolution-aware mechanisms so a model trained only on low-resolution garments (11K triangles) remains stable on much denser meshes up to 38K triangles.

What This Paper Is About

Physics-based cloth simulation is accurate but too slow for time-sensitive use, and existing graph neural network accelerators break down when the mesh resolution changes from what they were trained on. The authors trace this failure to two causes: a fixed message-passing depth that gives vertices the wrong amount of spatial context at different mesh densities, and the fact that per-vertex displacement magnitudes shrink as meshes get finer because the same global deformation is spread over more vertices. Pb4U-GNet addresses both so that a single model trained on coarse meshes can simulate garments at arbitrary resolutions without retraining.

Key Contributions

1. Pb4U-GNet architecture. A Propagation-before-Update Graph Network that separates iterative message propagation from feature updates, allowing the receptive field to be controlled independently of how often vertex features are refreshed.
2. Resolution-aware propagation control. A strategy that sets the number of message-passing steps from the mesh's mean edge length so the physical propagation distance stays constant across discretisations: given a base distance D = K_base × L̄_base, the step count is K = ⌊D × L̄⁻¹⌋.
3. Resolution-aware update scaling. A per-vertex scaling factor computed as the average length of a vertex's connected rest-state edges, applied to the predicted accelerations so that displacement magnitudes stay physically consistent across mesh densities.
4. Empirical demonstration of cross-resolution generalisation. Trained only on the lowest resolution (11K triangles) of the VTO dataset, the model is evaluated on 18K, 25K and 38K meshes and on unseen garment types, outperforming four graph-based baselines at higher resolutions.

Main Findings

• Baselines diverge at high resolution. On the held-out motion sequences, total physics loss at Level 4 (38K triangles) was 1.24E+06 for MGN, 2.57E+00 for HOOD, 1.07E+05 for ESLR and 2.82E-01 for CCRAFT, versus 2.22E-01 for Pb4U-GNet. Stretch loss was the clearest failure mode for the baselines.
• Comparable at the training resolution. At Level 1 (11K), Pb4U-GNet produced a total loss of -1.66E-02, alongside MGN (4.70E-03), HOOD (9.45E-03), ESLR (-2.56E-02) and CCRAFT (4.24E-02). The authors describe performance at this level as comparable to existing approaches.
• Progressive advantage as resolution rises. Total loss for Pb4U-GNet was 8.13E-03 at Level 2 (18K) and 6.34E-02 at Level 3 (25K), compared with 4.32E-01 and 1.44E+03 for MGN at those levels.
• Stable stretching energy over time. On test sequence 07_02 at the 38K setting, MGN and ESLR showed exploding stretch energy for the dress and long-sleeve garments, while Pb4U-GNet maintained the lowest stretching energy across all garments.
• Generalisation to unseen garments. On a form-fitting dress and a cardigan at the highest resolution, Pb4U-GNet reached a total loss of 1.55E-01 versus 2.64E-01 (CCRAFT), 1.39E+00 (ESLR), 1.74E+00 (HOOD) and 7.16E+04 (MGN).
• More propagation steps are needed on finer meshes. A study using a dress template at 12K, 25K and 48K resolutions found that physics loss generally decreases with more propagation steps and that higher-resolution meshes require more steps to converge.
• Adaptive depth saves compute at low resolution, spends it at high resolution. At 12K, Pb4U-GNet ran at 50.0 ms versus 46.4 ms (MGN), 50.8 ms (HOOD), 54.6 ms (ESLR) and 97.2 ms (CCRAFT). At 48K it ran at 196.4 ms with a physics loss of -2.42E-01, while CCRAFT took 761,499.3 ms and MGN took 144.5 ms with a loss of 5.75E+03.
• Both components are necessary. In the ablation, removing propagation control gave total losses of -1.61E-03 (11K), 1.08E+06 (25K) and 1.08E+09 (38K); removing update scaling gave -5.78E-03, 1.55E+13 and 7.34E+13; removing both gave 4.70E-03, 1.44E+03 and 1.24E+06, against -1.66E-02, 6.34E-02 and 2.22E-01 for the full model.
• Qualitative behaviour. Rendered comparisons at 38K showed MGN with severe distortions and topological artefacts such as tearing and collapsing, HOOD and ESLR with slipping and misalignment around the shoulders, and CCRAFT with overstretched garments lacking fine wrinkle detail. Pb4U-GNet's outputs are described as closely matching physics-simulated reference results.

Methodology in Plain English

The garment and body are represented as a graph: garment vertices connect through mesh topology, and extra "world edges" link garment vertices to body vertices that are within a distance threshold. Each vertex carries features such as velocity, mass, surface normals, material parameters and a garment-or-body indicator, but absolute positions are deliberately excluded so the model is translation-invariant. Edge features are kept relative — current and rest-state direction vectors plus the ratio of current to rest length — rather than absolute lengths, which helps the model transfer between meshes with different edge densities.

Instead of the usual scheme where each layer propagates messages and then immediately updates features, Pb4U-GNet runs K rounds of pure message accumulation (with LayerNorm and a decay factor γ controlling how much earlier messages matter), then applies a single learnable update that fuses the original vertex embedding with the accumulated messages. Because propagation and update are decoupled, K can be chosen freely per resolution, and a fixed stack of 15 MeshGraphNet blocks then refines the features.

K is not a hyperparameter search but a physical calculation: the authors calibrate an effective propagation distance at the lowest training resolution and divide it by the mesh's mean edge length, so finer meshes automatically get more hops. Separately, each vertex's predicted acceleration is multiplied by a scale factor equal to the average rest-state length of its incident edges, which corrects for the fact that vertices on fine meshes undergo smaller displacements under the same global motion.

Training is fully self-supervised, requiring no ground-truth simulation: the loss combines stretch (St. Venant–Kirchhoff), bending, collision, gravity, friction and inertia terms. Garment meshes are first posed with linear blend skinning to match an SMPL body pose, giving a coarse starting state that the network refines. Training used the VTO dataset with four garment types (T-shirt, tank top, long-sleeve shirt, long dress) at five mesh resolutions from 11K to 38K triangles; only the 11K resolution was used for training. Four motion sequences were held out for testing. The model was trained for 100,000 iterations, taking approximately 36 hours on a single NVIDIA RTX 4070 Ti GPU, with vertex and edge features encoded into a 128-dimensional latent space.

Why This Matters

Impact on research. The paper reframes cross-resolution failure not as a data-scarcity problem but as an architectural and geometric one: fixed receptive fields and resolution-dependent displacement magnitudes. Its propagation-before-update design and edge-length-based scaling are general ideas that could transfer to other mesh-based neural simulators (fluids, deformable solids, cloth) that face the same resolution-transfer problem.

Real-world applications:

• Virtual try-on and online apparel retail, where garments must be simulated at detail levels matched to a user's device.
• Virtual reality and digital human modelling, where a single asset needs to run on both headsets and high-end workstations.
• Game and film production, where level-of-detail systems switch mesh densities dynamically during a scene.
• Simulation-driven content pipelines that need fast, physically plausible draping without training a separate network per resolution.

Industry relevance. The training cost argument is significant: training directly on high-resolution meshes is described as computationally prohibitive, and the paper's approach trains only on 11K meshes and deploys at 38K and above. The latency table also shows the practical trade-off directly — the method spends more time than baselines only when higher resolution actually demands it, while a baseline such as CCRAFT reached 761,499.3 ms at 48K with a worse physics loss.

Future Directions

• Extension beyond garments. The propagation-depth and update-scaling ideas are presented as general resolution-adaptation mechanisms; testing them on other simulated systems (fluids, soft bodies, yarn-level cloth) is an open question.
• Pushing the resolution range further. The reported evaluation tops out at 38K triangles, with Figure 5 studying up to 48K; behaviour at substantially higher densities is not reported.
• Replacing hand-calibrated parameters. The base propagation distance D and the step count K_base are calibrated from the lowest training resolution, and the decay factor γ is a design choice; learning these adaptively is not explored.
• Ground-truth comparison. Evaluation relies on the same physics-based loss terms used for training, following prior work, since the method is self-supervised; comparison against reference physics simulation remains largely qualitative.

Target Audience

Researchers and graduate students working on neural physics simulation, graph neural networks for mesh data, or computer graphics for garments and digital humans. It is also relevant to graphics engineers in virtual try-on, gaming, VFX and VR who need one trained model to handle multiple mesh resolutions at different computational budgets. Readers without a background in message passing, mesh representations and physics-based loss functions will find the methodology section demanding.

Authors’ abstract

Garment simulation is fundamental to various applications in computer vision and graphics, from virtual try-on to digital human modelling. However, conventional physics-based methods remain computationally expensive, hindering their application in time-sensitive scenarios. While graph neural networks (GNNs) offer promising acceleration, existing approaches exhibit poor cross-resolution generalisation, demonstrating significant performance degradation on higher-resolution meshes beyond the training distribution. This stems from two key factors: (1) existing GNNs employ fixed message-passing depth that fails to adapt information aggregation to mesh density variation, and (2) vertex-wise displacement magnitudes are inherently resolution-dependent in garment simulation. To address these issues, we introduce Propagation-before-Update Graph Network (Pb4U-GNet), a resolution-adaptive framework that decouples message propagation from feature updates. Pb4U-GNet incorporates two key mechanisms: (1) dynamic propagation depth control, adjusting message-passing iterations based on mesh resolution, and (2) geometry-aware update scaling, which scales predictions according to local mesh characteristics. Extensive experiments show that even trained solely on low-resolution meshes, Pb4U-GNet exhibits strong generalisability across diverse mesh resolutions, addressing a fundamental challenge in neural garment simulation.