Keyframe‐Guided Self‐Supervised Learning Framework for Physics‐Based Garment Animation

ABSTRACT

Generating realistic and controllable 3D garment animations across multiple clothing categories under limited supervision remains a fundamental challenge in virtual clothing design and animation. To tackle this, we propose a physics‐aware, keyframe‐guided self‐supervised framework for multi‐category garment animation. First, a graph neural network (GNN) is employed to extract structural features from both human skeletal motion sequences and garment templates, and dynamically associate them to capture non‐rigid deformations in response to human motion. A physics‐constrained transformer module is then introduced to enforce physical plausibility through a multi‐objective loss that includes elasticity, bending, collision, and inertial consistency terms. Furthermore, we design a latent‐space‐based decoupling mechanism to encode garment simulation parameters, enabling fine‐grained parameter control by mapping keyframes to simulation attributes via radial basis function (RBF) interpolation. Lastly, an interpolated attention‐based sequence calibration module is proposed to inject keyframe‐guided topology features into the decoding process, enhancing the temporal continuity and topological consistency of the generated garment sequences. Experimental results demonstrate that our method produces physically realistic, dynamically stable, and highly detailed garment animations even under weak supervision, with strong generalization capabilities across diverse garment types.