HYPERBONES Realtime Bone-driven Neural Garment Simulation with Hypernetwork Conditioning

SIGGRAPH ASIA 2026 Conference Track
Paper (arXiv) Supplementary Interactive Results BibTex

Overview

Problem

Per-vertex neural simulators such as MeshGraphNet (HOOD) can be physically accurate, but they are too expensive for real-time use. On the other hand, PCA-parameters (shape/pose) dependent models fail to capture accurate dynamics of loose garments.

Insight

Body shape and garment topology are fixed per scene, while pose changes every frame. Caching that fixed encoding once makes the per-frame inference only a motion-dependent prediction.

Efficient Approach

Unlike per-vertex feed-forward, deforming a garment via virtual-bone-guided LBS transforms reduces the network's output bandwidth requirements (per-bone prediction), while using a network embedded in UV-space to predict high-frequency wrinkles significantly reduces the inference time.

Loading 0%

Interactive Visualization

HyperBones achieves real-time simulation of loose garments with body-shape generalization.

Loading 0%

Method & Architecture

Virtual Bones + Physics-AwareHypernetwork + Convolutional-MLP

HyperBones Method and Architecture

Training

Modules [B] and [C] share a graph encoder. Module [B] processes a canonical Body–Garment–Bones graph once to produce a shape code, skinning-weight corrections, and UV features. Module [C], a HOOD-style neural integrator, predicts physically-grounded positions under self-supervised physics losses, supervising Module [A] via a cross-branch consistency loss. Physics gradients flow through the shared encoder, making cached features dynamics-aware.

Inference

Module [C] is discarded; Module [B]’s outputs are precomputed once per body/garment. Only Module [A] runs per frame. Bone-Net corrects LBS-driven virtual bones per frame using pose history, with identity injected via hypernetwork-style FiLM conditioning on the cached shape code. A UV-space Conv-MLP adds wrinkle detail. This disentangled architecture enables realtime performance, significantly faster than MeshGraphNet-style architectures.

Initialization (LBS)

Loading 0%

+ BoneNet

Loading 0%

+ ConvMLP

Loading 0%

Comparison

Loading 0%
GAPS | REALTIME
HOOD | NOT REALTIME
HyperBones | REALTIME

Physics Errors

We measure per-frame Strain, Bending, Inertia, and Gravity errors and comapre with other methods below.

Method Strain ↓ Bending ↓ Inertia ↓ Gravity ↓
SNUG [Santesteban et al.] 3.722 0.181 0.613 0.516
NCS [Bertiche et al.] 4.013 0.147 1.082 0.613
GAPS [Chen et al.] 3.713 0.155 0.557 0.584
HOOD [Grigorev et al.] 3.611 0.142 0.524 0.598
Module-C (Standalone) 3.613 0.141 0.523 0.599
HyperBones (w/o Module-C) 5.116 0.305 2.310 0.537
HyperBones (w/o Llap & Linterp) 3.620 0.149 0.530 0.592
HyperBones 3.608 0.138 0.521 0.596

(Low gravity error alone does not mean better dynamics; garments that settle easily trade it for inertia error.)

Inference Time (ms)

Method T-shirt Shirt Pants Skirt Dress Hoodie
GAPS [Chen et al.] 2.02 2.02 1.98 2.72 3.02 2.14
HOOD [Grigorev et al.] 27.45 26.72 25.22 36.67 49.52 36.61
HyperBones 1.07 1.07 1.05 1.16 1.31 1.14

#Note: Some garment names have been shortened for convenience (e.g. Hooded Tight Dress → Hoodie). Please refer to the paper for the original sample names.

Direct Self-Supervision vs. Temporal Integration

Loading 0%
Direct Physics Supervision
Temporal Integration

Effect of Bone Density

Loading 0%
Loading 0%

For simple garments, 32 bones lead to high collision errors and 64 bones yield acceptable (but imperfect results). For complex loose garments, fewer than 128 bones produce visually degraded deformations and low simulation accuracy.

Geometrical Errors

Edge ( εe ), Area ( εa ), and Collision ( εc ) errors on 'Shirt' and 'Dress' meshes.

Shirt Dress
#Bones εe εa εc εe εa εc
32 4.812 6.412 0.687 (±0.598) 8.234 11.876 4.123 (±2.187)
64 2.687 3.487 0.213 (±0.254) 5.432 7.298 2.187 (±1.412)
128 2.312 3.101 0.118 (±0.187) 2.287 3.076 0.121 (±0.193)

Additional Results

Loading 0%

BibTeX Citation 

@misc{srivastava2026hyperbonesrealtimebonedrivenneural,
  title         = {HyperBones: Realtime Bone-driven Neural Garment Simulation with Hypernetwork Conditioning}, 
  author        = {Astitva Srivastava and Hsiao-Yu Chen and Ryan Goldade and
                   Philipp Herholz and Zhongshi Jiang and Gene Wei-Chin Lin and
                   Lingchen Yang and Nikolaos Sarafianos and Tuur Stuyck and
                   Doug Roble and Avinash Sharma and Egor Larionov},
  year          = {2026},
  eprint        = {2605.20460},
  archivePrefix = {arXiv},
  primaryClass  = {cs.GR},
  url           = {https://arxiv.org/abs/2605.20460}, 
}