Pb4u-Gnet Achieves Resolution-Adaptive Garment Simulation, Improving Performance Beyond Training Distributions

Garment simulation underpins increasingly vital applications such as virtual try-on and realistic digital human modelling, but conventional physics-based techniques are often too computationally demanding for real-time use. Aoran Liu from The University of Sydney, alongside Kun Hu and Clinton Ansun Mo from The University of Tokyo, with Qiuxia Wu, Wenxiong Kang and Zhiyong Wang et al, present a novel approach to tackle this challenge, introducing Pb4U-GNet , a resolution-adaptive framework utilising graph networks. Their research addresses a critical limitation of existing GNN-based methods: poor performance when simulating higher-resolution garments not seen during training, by dynamically adjusting information processing based on mesh density and local geometry. This breakthrough enables strong generalisation across varying mesh resolutions, even when trained on low-resolution data, representing a significant step towards practical, high-fidelity garment simulation.

Resolution-Adaptive Garment Simulation with Pb4U-GNet enables realistic

Scientists have developed a groundbreaking resolution-adaptive framework, Propagation-before-Update Graph Network (Pb4U-GNet), to dramatically improve garment simulation in computer vision and graphics applications. This innovative approach tackles the computational expense of traditional physics-based methods, paving the way for real-time performance in areas like virtual try-on and digital human modelling. The research addresses a critical limitation of existing graph neural networks (GNNs), their poor ability to generalise across different mesh resolutions, often exhibiting significant performance drops when applied to higher-resolution meshes than those used during training. Pb4U-GNet achieves this by decoupling message propagation from feature updates, allowing for flexible and adaptable information aggregation.
The team achieved this breakthrough by identifying two key factors hindering cross-resolution generalisation: fixed message-passing depth and resolution-dependent vertex displacement magnitudes. Existing GNNs employ a static message-passing depth, failing to adjust to variations in mesh density, while the scale of vertex movement inherently changes with resolution in garment simulation. To overcome these challenges, Pb4U-GNet incorporates dynamic propagation depth control, intelligently adjusting the number of message-passing iterations based on the mesh resolution. This ensures consistent receptive fields across different levels of detail, preventing information loss in fine meshes and over-smoothing in coarse ones.

Furthermore, the study unveils a geometry-aware update scaling mechanism, which rescales predictions based on local mesh characteristics. This innovative technique accounts for the fact that deformation magnitude scales with mesh density, ensuring physically consistent predictions regardless of resolution. Extensive experiments demonstrate that Pb4U-GNet, even when trained solely on low-resolution meshes containing approximately 11,000 triangles, exhibits remarkable generalisability to significantly higher resolutions, producing stable and realistic simulation results without the need for retraining. This research establishes a new benchmark in neural garment simulation, offering a practical solution to the long-standing problem of cross-resolution performance. The work opens exciting possibilities for dynamic mesh adaptation, allowing applications to adjust resolution based on computational resources and specific requirements. By decoupling propagation and updates, and introducing resolution-aware control and scaling, Pb4U-GNet not only improves simulation accuracy but also unlocks the potential for real-time garment simulation in a wider range of applications, from virtual fashion to immersive virtual reality experiences.

Resolution-Adaptive Graph Networks for Garment Simulation offer realistic

Scientists developed Propagation-before-Update Graph Network (Pb4U-GNet) to address limitations in cross-resolution generalisation within garment simulation using graph neural networks. The research tackled the problem of performance degradation when applying existing GNNs to higher-resolution meshes than those used during training. This study identified that fixed message-passing depths and resolution-dependent vertex displacement magnitudes were key factors hindering effective simulation across varying mesh densities. To overcome these challenges, the team engineered a resolution-adaptive framework that separates message propagation from feature updates, allowing for flexible control over receptive fields.

Pb4U-GNet employs dynamic propagation depth control, adjusting the number of message-passing iterations based directly on the mesh resolution, a crucial innovation for adapting to varying mesh densities. Experiments involved training the network on low-resolution meshes containing approximately 11,000 triangles, then evaluating its performance on significantly higher-resolution meshes without retraining. The approach further incorporates geometry-aware update scaling, which rescales predicted accelerations according to local mesh characteristics, ensuring physically consistent predictions across resolutions. This scaling mechanism compensates for the inherent relationship between deformation magnitude and mesh density, preventing over-smoothing on coarse meshes and insufficient context on fine meshes.

The system delivers iterative message passing, enabling each vertex to exchange state information with its neighbours and update its latent feature before predicting its next position. Extensive experiments demonstrated that Pb4U-GNet, even when trained solely on low-resolution meshes, exhibits strong generalisability to diverse mesh resolutions, effectively addressing a fundamental challenge in neural garment simulation. This method achieves stable and realistic simulation results, paving the way for real-time applications requiring dynamic mesh resolution adjustments, such as virtual try-on and digital human modelling, without the computational expense of traditional physics-based methods.

Low-resolution training yields high-resolution garment simulation

Scientists have developed Propagation-before-Update Graph Network (Pb4U-GNet), a resolution-adaptive framework for garment simulation that overcomes limitations in cross-resolution generalisation. The research addresses a critical challenge in computer vision and graphics where conventional physics-based methods are computationally expensive, particularly for time-sensitive applications like virtual try-on and digital human modelling. Experiments demonstrate that Pb4U-GNet, trained solely on low-resolution meshes containing approximately 11,000 triangles, effectively generalises to significantly higher resolutions, producing stable and realistic simulation results without the need for retraining. This breakthrough delivers a substantial improvement in efficiency and adaptability for garment simulation.

The team measured performance by focusing on the ability of the network to accurately predict garment dynamics across varying mesh densities. Results demonstrate that Pb4U-GNet dynamically adjusts message-passing iterations based on mesh resolution, maintaining consistent receptive fields, a crucial factor for accurate simulation. Specifically, the dynamic propagation depth control ensures that the network considers an appropriate level of context, preventing over-smoothing on coarse meshes and insufficient information on fine meshes. Measurements confirm that this adaptive approach significantly improves the accuracy of predicted vertex positions compared to methods employing fixed message-passing depths.

Furthermore, scientists devised a resolution-aware update scaling strategy that rescales predicted accelerations based on local geometric scale. This ensures physically consistent deformation across meshes of different densities, addressing the issue where deformation magnitude scales with mesh density. Data shows that this scaling normalisation prevents physically inconsistent predictions on unseen resolutions, leading to more realistic and stable simulations. The breakthrough delivers a novel decoupling of message propagation from feature updates, allowing for flexible control of receptive fields and physically consistent deformation.

Extensive evaluations reveal that Pb4U-GNet significantly outperforms existing methods in cross-resolution generalisation. The work introduces three key contributions: a novel Propagation-before-Update Graph Network, a resolution-aware propagation control strategy, and a resolution-aware scaling update strategy. These advancements enable garment simulation capable of generalising to unseen garment resolutions even when trained solely on low-resolution meshes, opening up possibilities for real-time applications and dynamic mesh adaptation. The research paves the way for more efficient and versatile garment simulation in diverse computational environments.

Pb4U-GNet resolves variable mesh garment simulation

Scientists have developed a new framework, Propagation-before-Update (Pb4U-GNet), to improve garment simulation within computer graphics and vision applications. Traditional physics-based methods are often too slow for real-time use, and while Graph Neural Networks (GNNs) offer speed advantages, they typically struggle when applied to meshes of varying resolutions. This research addresses this limitation by introducing a resolution-adaptive approach that separates message propagation from feature updates within the GNN. Pb4U-GNet incorporates dynamic propagation depth control, which adjusts the number of message-passing iterations based on the mesh’s density, and geometry-aware update scaling, which calibrates predictions according to local mesh characteristics.

Experiments demonstrate that this framework, even when trained on low-resolution meshes, successfully generalises to higher resolutions, significantly improving simulation accuracy. Ablation studies confirmed the importance of both the propagation control and update scaling mechanisms for robust performance across different mesh levels. The significance of this work lies in its ability to overcome a fundamental challenge in garment simulation, the lack of cross-resolution generalisation in GNNs. By enabling accurate simulation on a wider range of mesh resolutions, Pb4U-GNet broadens the applicability of this technology to areas like virtual try-on and digital human modelling. The authors acknowledge that the framework’s performance is dependent on the quality of the initial low-resolution training data. Future research could explore extending the framework to handle more complex garment behaviours or integrating it with other simulation techniques to further enhance realism and efficiency.