NeuralGarSim: Geometry-Agnostic Garment Simulation with Neural Fields

Arihant Gaur^1,2 Navami Kairanda¹ Christian Theobalt¹ Vladislav Golyanik¹

¹ Max Planck Institute for Informatics, SIC ² Saarland University, SIC

European Conference on Computer Vision (ECCV), 2026

Paper Video Code

NeuralGarSim supports UDF, mesh, point-cloud, and Gaussian inputs and produces consistent deformations across resolutions and materials

NeuralGarSim is a quasi-static neural garment simulator that operates directly on the 3D garment surface without a 2D parameterisation. It accepts UDFs, triangle meshes, point clouds, or 3D Gaussians and produces continuous, physically consistent deformations across garments with complex topology.

Abstract

Most existing mesh-based methods for garment simulation suffer from sensitivities tied to mesh discretisation and resolution. Recent approaches such as NeuralClothSim employ continuous neural fields, but remain constrained by a 2D curvilinear parameterisation that models a single cloth panel. We introduce NeuralGarSim, a quasi-static garment simulator formulated directly in 3D Euclidean space. It accepts distance fields, meshes, point clouds, and 3D Gaussians as undeformed garment states. We represent garment deformation as a neural field and define a nonlinear Kirchhoff–Love shell model directly in R³ using tangential differential calculus. By minimising a potential energy functional, NeuralGarSim predicts physically consistent deformations across garments with multiple panels, seams, and holes. It supports multiple material models while retaining continuous, consistent, and memory-adaptive behaviour.

Supplementary Video

Method

Overview of the NeuralGarSim garment simulation pipeline

NeuralGarSim obtains the surface normals and curvature from UDFs, meshes, point clouds, or 3D Gaussians. It combines these geometric quantities with material properties, boundary constraints, and external forces, then minimises the total potential energy to recover the equilibrium garment through a continuous neural deformation field. Please refer to the paper for the complete formulation.

Experimental Results

Discretisation Robustness

NeuralGarSim vs. DiffCloth and DiffARCSim

Under identical forces and pinned regions, NeuralGarSim preserves fold and wrinkle structure across input discretisations, while the mesh-based baselines remain sensitive to connectivity.

NeuralGarSim and DiffCloth multi-resolution consistency — NeuralGarSim maintains consistent wrinkle structure while DiffCloth changes with mesh resolution.

Wrinkle and full-surface Chamfer distances across discretisations — Wrinkle-region and full-surface Chamfer distances across three discretisations; lower values indicate better consistency.

NeuralGarSim and DiffARCSim across three input discretisations — Comparison with adaptive DiffARCSim across three discretisations.

Representation agnostic

One Simulator, Four Input Modalities

The same garment can be supplied as a UDF, triangle mesh, point cloud, or 3D Gaussians. Large-scale folds, drape direction, and fine wrinkles remain consistent.

Material models

Material-Dependent Drape

Linear isotropic, cotton StVK, and silk StVK parameters produce distinct fold orientation, stiffness, and wrinkle density without changing the simulation architecture.

Qualitative gallery

Complex Garments and Topologies

Results on T-shirts, dresses, and jackets demonstrate continuous simulation across multiple panels, seams, and holes.

Application

Gaussian Garments Integration

Surface-registered 3D Gaussians reconstructed from multi-view video are passed directly to NeuralGarSim. The simulator uses their orientations for surface normals and predicts a quasi-static drape under gravity and pinned shoulders—without constructing a 2D parameter domain.

Quantitative Results

NeuralGarSim closely matches analytical thin-shell benchmarks and trains consistently faster than NeuralClothSim.

Runtime comparison with NeuralClothSim — Training-time comparison with NeuralClothSim across benchmark and garment experiments.

Belytschko benchmark validation — Maximum displacement on the square plate, Scordelis–Lo roof, and pinched cylinder benchmarks.

Citation

@inproceedings{gaur2026neuralgarsim,
  title     = {NeuralGarSim: Geometry-Agnostic Garment Simulation with Neural Fields},
  author    = {Gaur, Arihant and Kairanda, Navami and Theobalt, Christian and Golyanik, Vladislav},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

Contact

For questions or clarifications, please get in touch with:

Arihant Gaur agaur@mpi-inf.mpg.de

Vladislav Golyanik golyanik@mpi-inf.mpg.de