HyperBones: Realtime Bone-driven Neural Garment Simulation with Hypernetwork Conditioning

On September 7, 2026, arXiv released the HyperBones framework. This research proposes a real-time bone-driven neural clothing simulation method. The system achieves high-performance dynamic simulations by decoupling identity computation from shape conditioning, consisting of two independent levels of components: at the coarse level, lightweight neural networks are used to modify linear mixed skin (LBS) predictions; at the fine level, convolutional MLPs are used in the UV space to restore wrinkle details. This method employs a physics-based self-supervised training paradigm, allowing training without an offline simulator. Experiments show that compared to state-of-the-art autoregressive neural simulators, its speed is increased by more than 30 times, achieving approximately 1 millisecond per frame of interactive reasoning on consumer-grade GPUs, and demonstrating generalization capabilities across various movements and unseen human body shapes.