ReWeaver: Towards Simulation-Ready and Topology-Accurate Garment Reconstruction

Ming Li1,2,3, Hui Shan1,2,3, Kai Zheng3, Chentao Shen1,3, Siyu Liu6, Yanwei Fu2,4, Zhen Chen5, Xiangru Huang3
1Zhejiang University 2Shanghai Innovation Institute 3Westlake University 4Fudan University 5Adobe 6Xidian University
Paper Code Dataset
2025 Computer Vision and Pattern Recognition (CVPR 2026)
ReWeaver teaser showing sparse-view input, reconstructed sewing patterns, and garment geometry

From as few as four input views, ReWeaver reconstructs high-precision sewing patterns with complex topology together with aligned 3D garment geometry and explicit 2D-3D correspondences.

Abstract

High-quality 3D garment reconstruction plays a crucial role in mitigating the sim-to-real gap in applications such as digital avatars, virtual try-on and robotic manipulation. However, existing garment reconstruction methods typically rely on unstructured representations, struggling to provide accurate reconstructions of garment topology and sewing structures. ReWeaver is a framework for topology-accurate 3D garment and sewing pattern reconstruction from sparse multi-view RGB images. Given as few as four input views, it predicts seams and panels as well as their connectivities in both 2D UV space and 3D space. The predicted seams and panels align precisely with the multi-view images, yielding structured 2D-3D garment representations suitable for 3D perception, high-fidelity physical simulation, and robotic manipulation. To enable effective training, we construct the GCD-TS dataset, comprising textured multi-view RGB images, 3D garment geometries, textured human body meshes, and annotated sewing patterns over more than 100,000 synthetic samples.

Structured 2D-3D Reconstruction

ReWeaver jointly reconstructs 3D patches, 3D curves, 2D panel edges, and their explicit patch-curve connectivity in a single unified representation.

Sparse Multi-View Input

The method works from as few as four RGB views by coupling a VGGT-style multi-view encoder with a bi-path transformer for geometry and topology prediction.

Simulation-Ready Assets

The reconstructed panels can be directly used for downstream simulation, while the aligned 3D predictions support perception, analysis, and robotic manipulation.

4
input views
100k+
GCD-TS samples
0.8221
panel IoU

Method Overview

ReWeaver predicts garment geometry and topology in 3D first, then recovers flattened 2D sewing patterns from the refined structured representation.

Pipeline of the ReWeaver method
Pipeline of ReWeaver. A multi-view image encoder extracts tokens from RGB inputs, a bi-path transformer predicts 3D curves, 3D patches, and connectivity, and an intra-surface attention module decodes normalized 2D panel edges that are ready for simulation.

Results

ReWeaver improves topology accuracy, geometric fidelity, and seam-panel consistency over previous pattern reconstruction baselines.

More qualitative ReWeaver results across diverse garment styles
More qualitative results. ReWeaver reconstructs structured garments across diverse styles, including strapless dresses, pencil skirts, and asymmetrical tops, while preserving usable topology and simulation-ready panel layouts.
Method Acc_p Acc_e Acc_o CD_e IoU
Sewformer 0.3761 0.4802 0.1806 0.1161 0.5844
ChatGarment 0.5557 0.8012 0.4452 0.0906 0.6533
AIpparel 0.4561 0.6774 0.3090 0.0648 0.7084
ReWeaver 0.9210 0.7175 0.6608 0.0391 0.8221
Adaptive point cloud sampling for ReWeaver patch predictions
Adaptive sampling at inference time. ReWeaver predicts smooth implicit patch geometry that can be sampled densely and uniformly according to patch scale.
Ablation study of topology and geometry refinement in ReWeaver
Topology and geometry refinement. Refinement removes redundant curves and enforces closed 2D panel boundaries, leading to cleaner structures suitable for triangulation and simulation.
Texture comparison between GCD and GCD-TS
GCD-TS dataset. Compared with the original GCD textures, GCD-TS replaces unrealistic seam-revealing textures with more realistic and diverse garment and body textures to improve generalization.

BibTeX

@inproceedings{li2026reweaver,
  title={ReWeaver: Towards Simulation-Ready and Topology-Accurate Garment Reconstruction},
  author={Li, Ming and Shan, Hui and Zheng, Kai and Shen, Chentao and Liu, Siyu and Fu, Yanwei and Chen, Zhen and Huang, Xiangru},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026},
  url={https://sii-liming.github.io/ReWeaver/}
}