PhySkin: Physics-based Bone-driven Neural Garment Simulation

Abstract

Recent advances in digital avatar technology have enabled the generation of compelling virtual characters, but deploying these avatars on compute-constrained devices poses significant challenges for achieving realistic garment deformations. While physics-based simulations yield accurate results, they are computationally prohibitive for real-time applications. Conversely, linear blend skinning offers efficiency but fails to capture the complex dynamics of loose-fitting garments, resulting in unrealistic motion and visual artifacts. Neural methods have shown promise, yet they struggle to animate loose clothing plausibly under strict performance constraints. In this work, we present a novel approach for fast and physically plausible garment draping tailored for resource-constrained environments. Our method leverages a reduced-space quasi-static neural simulation, mapping the garment's full degrees of freedom to a set of bone handles that drive deformation. A neural deformation model is trained in a fully self-supervised manner, eliminating the need for costly simulation data. At runtime, a lightweight neural network modulates the handle deformations based on body shape and pose, enabling realistic garment behavior that respects physical properties such as gravity, fabric stretching, bending, and collision avoidance. Experimental results demonstrate that our method achieves physically plausible garment drapes while generalizing across diverse poses and body shapes, supporting zero-shot evaluation and mesh topology independence. Our method's runtime significantly outperforms past works, as it runs in microseconds per frame using single-threaded CPU inference, offering a practical solution for real-time avatar animation on low-compute devices.

Figures (5)

• Figure 1: We propose a garment draping model that runs in microseconds using single-threaded CPU compute, which handles garments across a range of poses and body shapes. Our deformation approach performs physics-based quasi-static simulation in reduced subspaces defined by node-based transformations, to which the garment is skinned. The model is trained in a self-supervised manner and supports topology and resolution-independent inference.
• Figure 2: Method overview. Given a garment $\mathcal{G}$ and pose $\theta$, our method first applies a sequence of two LBS computations to predict the location of the garment for the particular pose indicated by the wide arrows. Given the body geometry $\mathcal{B}$, we sample its surface with points and compute a modulation signal using a transformer network $\xi_\mathit{hyper}$ for $\mathcal{MLP}_{\theta}$. The Pose Modulator and Node Deformer networks work in conjunction to predict the best corrective for the LBS to predict the most optimal drape as measured by the physically based loss $\mathcal{L}_{\mathit{phys}}$.
• Figure 3: Comparison of Self-Supervised Training with Supervised Training. Our proposed self-supervised training scheme does not require additional simulation software and yet can achieve the same result as when using supervision. On the right, we map the Hausdorff distance between the two results indicating that they are indeed slightly different.
• Figure 4: Qualitative Drape Comparison. Our method is qualitatively compared against draping results from GAPS and HOOD on a variety of body shapes and poses. Here we demonstrate that our method produces comparable quality to both GAPS and HOOD, with slightly better overall collision handling, and all at a fraction of the cost.
• Figure 5: Resolution Agnostic Draping. We demonstrate that our method is able to produce physically plausible drapes for varying resolutions of the garment it was trained on. In contrast, HOOD fails to produce an adequate drape when applied to mesh resolutions that differ significantly from those it was trained on.