pith. sign in

arxiv: 2605.20460 · v3 · pith:JDTWSUUHnew · submitted 2026-05-19 · 💻 cs.GR · cs.CV

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
show 4 more authors
Tuur Stuyck Doug Roble Avinash Sharma Egor Larionov
This is my paper

Pith reviewed 2026-05-21 06:12 UTC · model grok-4.3

classification 💻 cs.GR cs.CV
keywords garment simulationneural dynamicsreal-time animationbone-driven modelinghypernetworkphysics supervisionwrinkle generation
0
0 comments X

The pith

Virtual bones drive a neural network to simulate realistic garment dynamics in real time.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to create a fast neural method for animating loose clothing that looks physically correct. It separates the simulation into a coarse level where virtual bones control the main garment shape through a lightweight neural net, and a fine level that adds wrinkles using a convolutional map. This setup allows the system to run very fast while handling different body shapes and motions for a set of fixed garments. The key is a training approach that uses physics-based supervision so no external simulator is needed during use. If successful, this could make high-quality clothing animation practical for real-time applications like video games and virtual reality.

Core claim

The authors propose a reduced-space neural dynamics simulator that uses a set of virtual bones integrated with a hypernetwork-conditioned neural network for coarse garment motion, followed by a trained convolutional neural map to recover fine-scale wrinkle details. By decoupling identity-specific aspects and employing a physics-supervision scheme during training, the method achieves physically plausible results without an external simulator at runtime, running at over 300 frames per second on a commodity GPU while generalizing to various motions and body shapes for a fixed set of garments.

What carries the argument

Hypernetwork-conditioned virtual bone drivers for coarse-level dynamics combined with a convolutional neural map for fine-scale details.

If this is right

  • Real-time performance at 300+ FPS enables use in interactive applications.
  • Generalization to different body shapes and motions supports diverse character animations.
  • Support for a fixed set of garments allows pre-training for specific clothing items.
  • Physics supervision removes dependency on external simulators during deployment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar bone-driven approaches might apply to simulating other soft body elements like hair or flesh.
  • Extending the hypernetwork to handle garment changes or tears could broaden applications.
  • Integration with full character animation pipelines would test end-to-end performance.

Load-bearing premise

That the physics-supervision during training produces dynamics that remain accurate and stable without an external simulator for guidance at runtime.

What would settle it

A direct comparison of the neural simulation outputs against a high-accuracy physics-based simulator for a sequence of complex, unseen motions and body shapes.

Figures

Figures reproduced from arXiv: 2605.20460 by Astitva Srivastava, Avinash Sharma, Doug Roble, Egor Larionov, Gene Wei-Chin Lin, Hsiao-Yu Chen, Lingchen Yang, Nikolaos Sarafianos, Philipp Herholz, Ryan Goldade, Tuur Stuyck, Zhongshi Jiang.

Figure 1
Figure 1. Figure 1: We propose an efficient neural garment simulation method, which handles garments across a range of motions and body shapes for a fixed set of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Method overview. [A] Pose-Conditioned Deformation: given a pose sequence, virtual bones are first transformed via LBS (Skeleton → Bones) using SMPL skinning weights, then corrected by Bone-Net conditioned on the shape code z via the Shape Modulator. The corrected bones drive garment vertices via LBS (Bones → Garment) with learned skin weight corrections Δw. The resulting posed position map is concatenated … view at source ↗
Figure 3
Figure 3. Figure 3: Output of the individual deformation stages in Module [A] for 3 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison: GAPS relies on a GRU-based motion state that ac￾cumulates drift over long sequences, leading to progressive deformation artifacts. Our method adopts HOOD’s autoregressive integration scheme, maintaining stable simulation quality regardless of sequence length but with much improved performance (see supplementary video). Method Tshirt Tshirt-Unzipped Pants Long-Skirt Dress Hooded-Tight-Dress GAPS… view at source ↗
Figure 5
Figure 5. Figure 5: Generalization to multiple body shapes [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation on Number of Bones: For most garments, 32 bones lead to interpenetration and 64 yield acceptable but imperfect results. For most garments, fewer than 128 bones produce visually degraded deformations (see supplementary video) [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation over Temporal Integration: Applying the physics loss directly to Module A’s output produces unstable dynamics (left), as sparse bone transforms cannot represent dense per-vertex inertial states. Supervis￾ing Module A through the per-vertex GNN oracle (Module C) yields stable, physically plausible results (right). comparable or superior simulation accuracy while operating at more than 25× the infer… view at source ↗
read the original abstract

Recent advances in garment simulation have brought high-quality results closer to real-time performance. Physics-based simulators can produce accurate motion, but remain too computationally expensive for interactive applications. In contrast, linear blend skinning is efficient, but cannot capture the complex dynamics of loose-fitting garments, often leading to unrealistic motion and visual artifacts. Neural methods offer a promising alternative, yet they still struggle to animate loose clothing plausibly under strict runtime constraints. We present a fast and physically plausible approach for dynamic garment simulation. Our method trains a reduced-space neural dynamics simulator composed of independent coarse- and fine-level components. At the coarse level, the garment is driven by a set of virtual bones integrated with a lightweight neural network. Fine-scale wrinkle details are then recovered using a trained convolutional neural map. By decoupling identity-specific computation from real-time neural integration, our architecture maintains high performance while supporting diverse body shapes and motions. We further introduce an effective physics-supervision scheme that enables accurate results without relying on an external simulator. Experiments show that our method produces physically plausible garment dynamics, generalizes across a range of motions and body shapes, and supports a fixed set of garments. Our simulator runs at 300+ FPS on a commodity GPU, making it suitable for real-time applications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents HyperBones, a realtime neural garment simulation method using hypernetwork conditioning. It decomposes the problem into a coarse-level reduced-space dynamics model driven by a set of virtual bones integrated with a lightweight neural network, followed by a convolutional neural map to recover fine-scale wrinkle details. A physics-supervision scheme is introduced to train the model without an external simulator. The central claims are that the approach produces physically plausible garment dynamics, generalizes across motions and body shapes for a fixed set of garments, and achieves 300+ FPS on commodity GPUs.

Significance. If the physics-supervision claim holds and the quantitative results support the plausibility and generalization assertions, this work could meaningfully advance real-time garment simulation for interactive graphics applications such as games and VR. The decoupling of coarse dynamics from fine details via bones and hypernetworks, combined with the avoidance of external simulators during training, would be a practical strength for deployment and reproducibility. The method's efficiency at 300+ FPS addresses a key bottleneck in current neural garment approaches.

major comments (2)
  1. [§3.2] §3.2 (Physics Supervision Scheme): The central claim that the physics-supervision scheme produces accurate dynamics without any external simulator is load-bearing for the paper's contribution. However, the separation between the coarse bone-driven network and the fine-scale convolutional wrinkle map creates a potential point of circularity: if gradients for physical constraints (e.g., collision response or momentum) on the bone network flow through the wrinkle map, the supervision may implicitly depend on precomputed trajectories or ground-truth forces rather than purely analytic internal terms. The manuscript should provide the explicit loss equations, the computation graph, and confirmation that all physical terms are differentiable and self-contained within the network's own predictions.
  2. [Experiments] Experiments (quantitative evaluation section): The abstract asserts physically plausible results and generalization, yet no error metrics (e.g., position or velocity RMSE against ground-truth simulation), ablation studies on the supervision loss, or cross-body-shape comparisons are referenced in the provided text. If Tables 1–3 or Figures 4–6 report such numbers, they must be explicitly tied to the physics-supervision claim; otherwise the evidence for plausibility remains qualitative and insufficient to support the generalization statements.
minor comments (2)
  1. [§2.1] The definition and initialization of the 'virtual bones' (distinct from standard linear blend skinning bones) should be clarified with a diagram or pseudocode in §2.1, as this is foundational to the reduced-space model.
  2. [Method] Notation for the hypernetwork conditioning parameters is introduced but not consistently used in equations; ensure all symbols are defined before first use.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. We address each major comment below and outline the revisions we will make to strengthen the manuscript's clarity and evidentiary support.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (Physics Supervision Scheme): The central claim that the physics-supervision scheme produces accurate dynamics without any external simulator is load-bearing for the paper's contribution. However, the separation between the coarse bone-driven network and the fine-scale convolutional wrinkle map creates a potential point of circularity: if gradients for physical constraints (e.g., collision response or momentum) on the bone network flow through the wrinkle map, the supervision may implicitly depend on precomputed trajectories or ground-truth forces rather than purely analytic internal terms. The manuscript should provide the explicit loss equations, the computation graph, and confirmation that all physical terms are differentiable and self-contained within the network's own predictions.

    Authors: We appreciate the referee's careful analysis of the physics-supervision scheme. To clarify, the physical constraints (collision penalties, momentum preservation, and energy terms) are applied exclusively to the coarse-level bone-driven network outputs. The convolutional wrinkle map operates as a decoupled post-processing stage that adds high-frequency details but does not participate in the physics loss computation or receive gradients from it. All supervision terms are therefore analytic, differentiable, and computed solely from the bone network's predictions without reference to external trajectories or precomputed forces. We will revise §3.2 to include the full loss equations and a computation-graph diagram that explicitly shows the separation of the two stages. revision: yes

  2. Referee: [Experiments] Experiments (quantitative evaluation section): The abstract asserts physically plausible results and generalization, yet no error metrics (e.g., position or velocity RMSE against ground-truth simulation), ablation studies on the supervision loss, or cross-body-shape comparisons are referenced in the provided text. If Tables 1–3 or Figures 4–6 report such numbers, they must be explicitly tied to the physics-supervision claim; otherwise the evidence for plausibility remains qualitative and insufficient to support the generalization statements.

    Authors: We thank the referee for this observation. The current manuscript emphasizes visual and qualitative demonstrations of physical plausibility together with runtime measurements. We agree that explicit quantitative metrics would provide stronger support for the claims. In the revised version we will add position and velocity RMSE tables against ground-truth physics simulation, ablation results on the individual supervision loss terms, and cross-body-shape error comparisons. These new quantitative results will be directly referenced in the text and tied to the physics-supervision and generalization arguments. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper's central claims rest on a described architecture (coarse bone-driven neural network plus convolutional wrinkle map) and an explicitly stated physics-supervision scheme that operates without external simulators. No equations, fitted parameters, or self-citations are shown reducing predictions to inputs by construction. The supervision is presented as enforcing physical constraints via internal losses rather than precomputed trajectories or author-specific uniqueness theorems. The method is self-contained against the stated benchmarks of runtime performance and generalization, with no load-bearing reduction to prior self-citations or ansatz smuggling visible in the text.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the effectiveness of the physics-supervision training scheme and the ability of the hypernetwork to decouple identity-specific computation from runtime inference; these are not independently verified in the provided abstract.

free parameters (1)
  • hypernetwork conditioning parameters
    Neural weights that adapt the simulator to different body shapes and garments; trained from data.
axioms (1)
  • domain assumption Physics-based supervision during training can produce accurate dynamics without an external simulator at inference time.
    Invoked in the description of the training scheme.
invented entities (1)
  • virtual bones no independent evidence
    purpose: Drive the coarse-level garment motion through a neural network.
    Introduced as the mechanism for reduced-space simulation.

pith-pipeline@v0.9.0 · 5804 in / 1230 out tokens · 39747 ms · 2026-05-21T06:12:25.647292+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.