Research map / 3D vision / Face and shape
Learning 3D Garment Animation from Trajectories of A Piece of Cloth
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Learns an edge-wise energy network (EUNet) from trajectories of a single pinned square cloth, then animates diverse garments by energy optimization constrained by the learned energy.
Problem
Garment animation with learned models typically needs large garment-level datasets, which are expensive to collect and may generalize poorly to unseen topologies, human shapes, and actions.
Approach
Disentangle garment learning: learn constitutive laws as an edge-wise energy (potential plus dissipation) from a single piece of cloth without analytical physics priors or differentiable simulators, using energy-change supervision and a vertex-wise contrastive loss, then embed the energy as a physics loss in a neural simulator.
Key results
- MGN-H+EUNet achieves overall Euclidean error 66.39mm on Cloth3D versus 84.85mm for HOOD and 170.34mm for MGN.
- MGN-S+EUNet reduces collision rate to 0.12% versus 0.46% for HOOD and 17.72% for MGN.
- EUNet trained on a single piece of cloth reaches low energy-change errors across silk, leather, denim, and cotton (e.g., overall 85.74 with high variance).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the dissipation energy unit to EUNet (ablation) | EUNet · Cloth trajectories test, Overall | square error: 211.76$\pm$1052.02 -> 85.74$\pm$543.06 | Helped |
| Add vertex-wise contrastive loss (ablation) | EUNet · Cloth trajectories test, Overall | square error: 101.11$\pm$671.72 -> 85.74$\pm$543.06 | Helped |
| Add vertex-wise contrastive loss (ablation) | EUNet · Cloth trajectories test, Cotton | square error: \textbf{2.38$\pm$6.31} -> 2.67$\pm$8.49 | Hurt |
| Add friction loss (ablation) | MGN-S+PHYS · Cloth3D, Overall | Euclidean error (mm): 106.63$\pm$55.48 -> 108.83$\pm$45.55 | No clear effect |
| Use the mini-network to estimate material parameters (ablation) | HOOD · Cloth3D, Overall | Euclidean error (mm): 153.78$\pm$63.75 -> 84.85$\pm$29.93 | Helped |
| Constrain MGN-H with EUNet instead of the HOOD analytical-model setup | MGN-H · Cloth3D, Overall | Euclidean error (mm): 84.85$\pm$29.93 -> 66.39$\pm$39.36 | Helped |
| Constrain MGN-S with EUNet instead of the analytical physics model | MGN-S · Cloth3D, Overall | Euclidean error (mm): 106.63$\pm$55.48 -> 89.71$\pm$147.83 | Helped |
| Remove EUNet bending energy gradients (ablation) | EUNet-constrained garment animation · Cloth3D garment animation | deformation fidelity | Hurt |
Surprising findings
- MGN-S+EUNet is worse than MGN-S+PHYS and HOOD on the Dress category (156.16mm vs 103.72mm and 92.57mm), with high variance overall (89.71±147.83).
Limitations
- Edge-wise discretization may limit performance versus polygon-wise modeling.
- Does not model self-collisions; training uses simple pinned-cloth falling motions.
- Dress animation errors are higher than some baselines, and variance is large.
Experimental setup
Datasets: Cloth3D Blender-simulated square cloth trajectories (silk, leather, denim, cotton)
Benchmarks: Cloth3D
Metrics: Euclidean error (mm) collision rate (%)
Baselines: MGN LayersNet MGN-S+PHYS HOOD
Builds on: MeshGraphNet HOOD LayersNet backward Euler incremental potential
Closest papers by meaning
- GarmentLab: A Unified Simulation and Benchmark for Garment Manipulation (NeurIPS 2024 · similarity 0.89)
- DexGarmentLab: Dexterous Garment Manipulation Environment with Generalizable Policy (NeurIPS 2025 · similarity 0.87)
- Garment Recovery with Shape and Deformation Priors (CVPR 2024 · similarity 0.87)
- GaPT-DAR: Category-level Garments Pose Tracking via Integrated 2D Deformation and 3D Reconstruction (CVPR 2025 · similarity 0.87)
- DressWild: Feed-Forward Pose-Agnostic Garment Sewing Pattern Generation from In-the-Wild Images (NeurIPS 2026 · similarity 0.87)
- GarmentPile: Point-Level Visual Affordance Guided Retrieval and Adaptation for Cluttered Garments Manipulation (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.