InverseDraping: Recovering Sewing Patterns from 3D Garment Surfaces via BoxMesh Bridging
Authors: Leyang Jin, Zirong Jin, Zisheng Ye, Haokai Pang, Xiaoguang Han, Yujian Zheng, Hao Li
Organizations: Mohamed bin Zayed University of Artificial Intelligence · The Chinese University of Hong Kong, Shenzhen · School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, FNii-Shenzhen, and Guangdong Provincial Key Laboratory of Future Networks of Intelligence · Mohamed bin Zayed University of Artificial Intelligence and Pinscreen
Recovering sewing patterns from draped 3D garments is a challenging problem in human digitization research. In contrast to the well-studied forward process of draping designed sewing patterns using mature physical simulation engines, the inverse process of recovering parametric 2D patterns from deformed garment geometry remains fundamentally ill-posed for existing methods. We propose a two-stage framework that centers on a structured intermediate representation, BoxMesh, which serves as the key to bridging the gap between 3D garment geometry and parametric sewing patterns. BoxMesh encodes both garment-level geometry and panel-level structure in 3D, while explicitly disentangling intrinsic panel geometry and stitching topology from draping-induced deformations. This representation imposes a physically grounded structure on the problem, significantly reducing ambiguity. In Stage I, a geometry-driven autoregressive model infers BoxMesh from the input 3D garment. In Stage II, a semantics-aware autoregressive model parses BoxMesh into parametric sewing patterns. We adopt autoregressive modeling to naturally handle the variable-length and structured nature of panel configurations and stitching relationships. This decomposition separates geometric inversion from structured pattern inference, leading to more accurate and robust recovery. Extensive experiments demonstrate that our method achieves state-of-the-art performance on the GarmentCodeData benchmark and generalizes effectively to real-world scans and single-view images.
Figures & tables
Fig. 1: Examples of sewing pattern reconstruction from 3D garments obtained via multi-view capture with smartphones (left) and single-view reconstruction (right). Each triplet shows the input, the “scanned” clothed human from the input, and our predicted sewing pattern.
Fig. 2: (a) Recovering sewing patterns from a 3D garment mesh is an inverse problem of garment draping. (b) An example of the parameterization of the half hood panel (red circle in (a)).
Fig. 3: Method overview. Given an input garment mesh, our method first predicts an intermediate representation ( BoxMesh ) using a geometry-oriented auto-regressive model. A second, semantic-aware model then generates the corresponding sewing pattern, which can be draped onto the body for final garment reconstruction.
Methods
CD ↓
HD ↓
Direct Tokenization
6.46
12.66
Compressive Tokenization
3.97
9.19
TABLE I: Evaluation on Garment-to-BoxMesh on GCD testing set.
Fig. 4: Evaluation on Stage I. From left to right: (a) the 3D garment with sampled points, (b)-(d) BoxMesh of ground truth, predicted by Compressive Tokenization (our results), and Direct Tokenization, respectively.
Methods
Panel IoU ↑
Panel L2↓
Rot. L2↓
Transl. L2↓
#Panels Acc. ↑
#Edges Acc. ↑
Stitch Prec. ↑
CD BoxMesh ↓
Ours-GT
73.23%
7.46
0.0081
4.27
67.32%
71.85%
69.94%
3.64
Ours-GT*
64.36%
11.14
0.022
10.66
34.05%
62.83%
51.56%
6.09
Ours-Pred*
71.42%
9.00
0.010
5.55
57.78%
69.40%
63.47%
5.35
TABLE II: Evaluation on BoxMesh-to-Pattern on GCD testing set. * denotes that the input BoxMesh is from Stage I during testing. # denotes “the number of”.
Methods
Panel IoU ↑
Panel L2↓
Rot. L2↓
Transl. L2↓
#Panels Acc. ↑
#Edges Acc. ↑
Stitch Prec. ↑
CD BoxMesh ↓
NeuralTailor*
51.88%
22.96
0.381
25.35
17.68%
23.09%
18.69%
15.46
Ours
71.42%
9.00
0.010
5.55
57.78%
69.40%
63.47%
5.35
TABLE III: Quantitative comparisons with NeuralTailor* [ 11 ] .
Fig. 5: Evaluation on Stage II. From left to right: (a)-(c) the draped 3D garment with sewing pattern displayed in 2D and the corresponding BoxMesh of ground truth, Ours-GT* and Ours-Pred* , respectively; (d) the input BoxMesh predicted from Stage I.
Fig. 6: Qualitative comparisons with NeuralTailor [ 11 ] on real-scan data. First two columns are results from RenderPeople, and the rest two columns are from in-the-wild data collected by ourselves. From top to bottom: (a) the input 3D scan, (b)-(c) the simulated garment with corresponding final BoxMesh and 2D sewing pattern of NeuralTailor* and our method, respectively.
Methods
Panel IoU ↑
Panel L2↓
Rot. L2↓
Transl. L2↓
#Panels Acc. ↑
#Edges Acc. ↑
Stitch Prec. ↑
CD BoxMesh ↓
Baseline
68.70%
10.26
0.015
6.37
44.44%
65.68%
57.66%
5.13
Default_Body
58.82%
14.06
0.025
10.73
37.06%
57.56%
45.13%
8.42
Inter_Points
69.89%
9.91
0.012
6.01
50.67%
66.57%
60.03%
5.46
Extra_Output
66.11%
11.14
0.014
6.55
47.63%
64.98%
53.80%
5.81
Ours
71.42%
9.00
0.010
5.55
57.78%
69.40%
63.47%
5.35
TABLE IV: Quantitative comparisons for ablation studies.
Fig. 7: Ablation study with Baseline . From left to right: (a) the input garment mesh, (b)-(d) the draped 3D garment and corresponding BoxMesh from Baseline , Full and ground truth, respectively.
Fig. 8: Ablation study with Default_Body . From left to right: (a) the input garment mesh, (b)-(d) the draped 3D garment and corresponding BoxMesh from Default_Body , Full and ground truth, respectively.
Fig. 9: Ablation study with Inter_Points . From left to right: (a) the input garment mesh, (b)-(c) the draped 3D garment, the corresponding BoxMesh , and the input points of Stage II from Inter_Points and Full , (d) ground truth.
Fig. 10: Ablation study with Extra_Output . From left to right: (a) the input garment mesh, (b)-(d) the draped 3D garment and corresponding BoxMesh from Extra_Output , Full and ground truth, respectively.
Fig. 11: Qualitative comparisons on the application of single-view reconstruction. From left to right: (a) the input image, (b) the generated 3D clothed human mesh, (c)-(d) the draping result with corresponding BoxMesh and 2D sewing pattern of AIpparel [ 15 ] , ChatGarment [ 16 ] , and ours-SVR, respectively.
Methods
CD ↓
F1 @ 1cm ↑
F1 @ 3cm ↑
Failure Rate ↓
AIpparel
6.09
0.16
0.62
61.1%
ChatGarment*
6.77
0.17
0.64
0
Ours-SVR*
6.32
0.19
0.70
0
ChatGarment
6.58
0.19
0.63
0
Ours-SVR
5.59
0.25
0.74
0
TABLE V: Qualitative comparisons with [ 16 ] . Metrics are calculated on the intersection of successfully draped samples of all methods.
Fig. 12: Failure cases. The left example shows errors caused by noise in the scanned garment mesh, while the right example highlights limitations in the design of the current synthetic data.
Fig. 13: More results of our method. From left to right: (a) input 3D scan (the first row is from RenderPeople and the second to fourth rows are from THuman) or single-view image (last two rows), (b) segmented 3D garment with fitted body or reconstructed 3D clothed human from single view, (c) 3D garment deformed into A-pose, (d) our result after simulation.
The University of Hong Kong, Hong Kong SAR, China · The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China · Carnegie Mellon University, Pittsburgh, PA, USA +2