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

TL;DR

Pb4U-GNet tackles cross-resolution generalisation in neural garment simulation by decoupling message propagation from updates and by introducing resolution-aware propagation distance $D$ and per-vertex update scaling $s_i$, with $K = \left\lfloor D/\bar{L} \right\rfloor$ and $D = K_{\text{base}} \cdot \overline{L}_{\text{base}}$. It learns under physics-based self-supervision with six losses, on a garment-body graph that excludes absolute positions to promote scale-invariance. Empirical results on the VTO dataset show strong generalisation from low-resolution training to unseen higher resolutions and unseen garments, outperforming state-of-the-art graph-based simulators in stability and physical plausibility across resolutions. This approach enables real-time, multi-resolution garment simulation for applications like virtual try-on and digital humans without retraining on high-resolution meshes.

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.

Figures (5)

• Figure 1: Sample results of Pb4U-GNet, a resolution-adaptive garment simulation framework based on graph neural networks. Trained on low resolution meshes with $\sim$11K triangles, Pb4U-GNet generalises effectively to significantly higher resolutions, producing stable and realistic simulation results without retraining.
• Figure 2: Illustration of the proposed proposed Pb4U-GNet, which decouples message propagation with a propagation-before-update scheme. With a resolution-aware propagation control and a resolution-aware update scaling design, it enables resolution-adaptive garment simulation.
• Figure 3: Rendered simulation results on high-resolution garment meshes. Baseline methods often struggle to preserve realistic fabric stretch, leading to noticeable over-stretched artefacts. Baseline methods also fail to preserve realistic wrinkle details.
• Figure 4: Stretch loss vs. time. The plots illustrate the temporal evolution of log stretching energy for each method on the test sequence 07_02. Our method consistently maintains the lowest stretch energy across the simulations, demonstrating better physics validity.
• Figure 5: Physics loss vs. number of message propagation steps. Garments with different mesh resolutions require varying numbers of propagation steps to achieve stable and accurate simulation results.