← ArchivePaper2026
HyperBones: Realtime Bone-driven Neural Garment Simulation with Hypernetwork Conditioning
Astitva Srivastava, Hsiao-Yu Chen, Ryan Goldade, Philipp Herholz, Zhongshi Jiang, Gene Wei-Chin Lin, Lingchen Yang, Nikolaos Sarafianos, Tuur Stuyck, Doug Roble, Avinash Sharma, Egor Larionov
A reduced space neural dynamics simulator with coarse and fine hypernetwork conditioned components predicts fast, physically plausible bone driven garment motion.
How to read this
- Category
- research paper, real time neural garment simulation via a reduced space bone driven representation
- Contributions
- Proposes a reduced space neural dynamics simulator split into an independent coarse level, virtual bone driven, and fine level, UV space convolutional wrinkle, component, so identity specific computation is precomputed once and only a lightweight network runs per frame
- Introduces a self supervised training scheme where the compact bone driven network is jointly trained alongside a heavier MeshGraphNet based per vertex simulator following HOOD, which acts as a teacher through a cross branch consistency loss without requiring precomputed simulation data
- Conditions both stages on a precomputed global body garment encoding via FiLM modulation, letting physics gradients from the teacher branch propagate into that cached identity representation during training
- Reports over 300 FPS on a commodity GPU with roughly a 25 times speedup over comparable state of the art neural garment simulators at comparable visual quality, generalizing across motions and body shapes for a fixed set of garments
- Context
- This paper extends the virtual bones idea from Pan et al's bone based garment deformation work and the self supervised, data free physics training approach pioneered by SNUG and extended by HOOD, which it directly uses as the teacher branch during training. It positions itself against both classical linear blend skinning, fast but poor on loose fitting garments, and full physics based or per vertex neural simulators like HOOD, accurate but too slow for interactive use, decoupling those two regimes into a fast runtime path trained by a slow, physically grounded one. builds_on is empty in the archive dataset even though the lineage to HOOD and SNUG is explicit inside the paper.
- Correctness
- Rests on quantitative experiments reporting roughly a 25 times performance improvement over state of the art neural garment simulation at comparable visual quality, plus an architecture where the HOOD based temporal integration branch acts purely as a physics teacher during training and is discarded at inference. As an arXiv preprint from an industry lab it is not yet peer reviewed, and the method is explicitly scoped to a fixed set of garments per trained model, so generalization to arbitrary unseen garment types at inference is not demonstrated.
- Clarity
- Technically dense, written for readers already comfortable with neural cloth simulation literature, HOOD, SNUG, TailorNet, and standard graph neural network and FiLM conditioning terminology, not an easy entry point for someone new to the space.
- How to read it
- Five minute pass: read the abstract and Figure 1 for the 25 times speedup claim and the coarse bone plus fine wrinkle map split. Second pass: read the introduction's three contributions and Section 3's description of the two stage architecture and the cross branch consistency loss, to see exactly what is discarded at inference versus kept. Third pass matters for anyone building real time garment simulation for games or interactive avatars, since the virtual bone placement via Farthest Point Sampling and the SMPL weight assignment borrowed from Pan et al are directly reusable rigging techniques independent of the neural training details.
This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →