PhysDrape: Learning Explicit Forces and Collision Constraints for Physically Realistic Garment Draping

TL;DR

PhysDrape addresses the collision handling and physical fidelity gap in garment draping by coupling a Physics-Informed Graph Neural Network with an explicit two-stage solver. The workflow starts from a coarse Linear Blend Skinning initialization, then predicts residual deformations via a physics-aware graph, followed by a Learnable Force Solver to dissipate residual forces and a Differentiable Projection to enforce strict non-penetration against the body. The method leverages Saint Venant–Kirchhoff energy components ($E_{strain}$, $E_{bend}$, $E_{grav}$, $E_{coll}$) to drive self-supervised learning and ensures differentiability throughout the pipeline. Experimental results on CLOTH3D demonstrate state-of-the-art physical fidelity with negligible interpenetration (B2G ≈ 0.05%) and lower strain energy, while maintaining real-time inference. The work highlights a practical path toward physically grounded neural garment draping without relying on soft penalties, enabling robust, controllable, and real-time deployment.

Abstract

Deep learning-based garment draping has emerged as a promising alternative to traditional Physics-Based Simulation (PBS), yet robust collision handling remains a critical bottleneck. Most existing methods enforce physical validity through soft penalties, creating an intrinsic trade-off between geometric feasibility and physical plausibility: penalizing collisions often distorts mesh structure, while preserving shape leads to interpenetration. To resolve this conflict, we present PhysDrape, a hybrid neural-physical solver for physically realistic garment draping driven by explicit forces and constraints. Unlike soft-constrained frameworks, PhysDrape integrates neural inference with explicit geometric solvers in a fully differentiable pipeline. Specifically, we propose a Physics-Informed Graph Neural Network conditioned on a physics-enriched graph -- encoding material parameters and body proximity -- to predict residual displacements. Crucially, we integrate a differentiable two-stage solver: first, a learnable Force Solver iteratively resolves unbalanced forces derived from the Saint Venant-Kirchhoff (StVK) model to ensure quasi-static equilibrium; second, a Differentiable Projection strictly enforces collision constraints against the body surface. This differentiable design guarantees physical validity through explicit constraints, while enabling end-to-end learning to optimize the network for physically consistent predictions. Extensive experiments demonstrate that PhysDrape achieves state-of-the-art performance, ensuring negligible interpenetration with significantly lower strain energy compared to existing baselines, achieving superior physical fidelity and robustness in real-time.

Figures (6)

• Figure 1: Our method effectively resolves complex collisions and dissipates residual force, generating geometrically consistent and physically relaxed garments, whereas the baseline suffers from severe interpenetration (highlighted in red boxes) and high residual force (red regions in the force heatmap).
• Figure 2: Overview of PhysDrape. The pipeline begins with a coarse initialization generated via Linear Blend Skinning on the SMPL body. A Physics-Informed GNN then predicts residual displacements to recover geometric details. To guarantee physical validity, the prediction is refined by a two-stage differentiable module: a Learnable Force Solver that iteratively resolves unbalanced forces from StVK deformation, and a Differentiable Projection Layer that strictly enforces collision constraints.
• Figure 3: Visual comparison of collision handling on unseen body poses. The baseline method (top row) fails to resolve complex contacts, resulting in visible interpenetrations in complex regions such as the armpits and shoulders (highlighted in red). In contrast, PhysDrape (bottom row) strictly enforces geometric constraints, generating smooth and collision-free garments that robustly adapt to the body shape.
• Figure 4: Visualization of physical relaxation and force dissipation. We map residual forces to a heatmap where red denotes high stress. While DrapeNet exhibits widespread unnatural stretching, our solver effectively dissipates these forces from $T=3$ to $T=15$, reaching a stable and physically consistent equilibrium.
• Figure 5: Convergence analysis of the learnable solver. We plot Total Energy ($E_{total}$, blue) and Residual Force ($F_{res}$, orange) over iterations $T$. The sharp decrease and subsequent stabilization demonstrate that our solver efficiently drives the mesh to a physically consistent equilibrium.