Recent advances in garment pattern generation have shown promising progress. However, existing feed-forward methods struggle with diverse poses and viewpoints, while optimization-based approaches are computationally expensive and difficult to scale. This paper focuses on sewing pattern generation for garment modeling and fabrication applications that demand editable, separable, and simulation-ready garments. We propose DressWild, a novel feed-forward pipeline that reconstructs physics-consistent 2D sewing patterns and the corresponding 3D garments from a single in-the-wild image. Given an input image, our method leverages vision-language models (VLMs) to normalize pose variations at the image level, then extract pose-aware, 3D-informed garment features. These features are fused through a transformer-based encoder and subsequently used to predict sewing pattern parameters, which can be directly applied to physical simulation, texture synthesis, and multi-layer virtual try-on. Extensive experiments demonstrate that our approach robustly recovers diverse sewing patterns and the corresponding 3D garments from in-the-wild images without requiring multi-view inputs or iterative optimization, offering an efficient and scalable solution for realistic garment simulation and animation.
Figures & tables
Figure 1. We propose DressWild , a feed-forward sewing-pattern generation pipeline that reconstructs simulation-ready, pose-agnostic 2D sewing patterns from a single in-the-wild image, while producing the corresponding pose-fitted 3D garment.
Figure 2. Overview of our pipeline. Given a single in-the-wild image, our method reconstructs simulation-ready sewing patterns and a corresponding 3D garment.
Figure 3. Data Curation and Augmentation. We use VLM to generate novel pose and novel view images with the consistent garment.
Figure 4. Qualitative Comparison. Our method produces high-quality 2D sewing patterns and corresponding 3D garments. Here we qualitatively compare our results with baseline approaches SewFormer [ Liu et al., 2023b ] and NeuralTailor [ Korosteleva and Lee, 2022 ] .
Method
Panel Acc. ↑
Edge Acc. ↑
Shape L2↓
F-Shape L2↓
CD ↓
NeuralTailor
25.99%
29.05%
23.65
7.30
0.02837
Sewformer
28.81%
34.56%
22.94
4.53
0.02797
DressWild
94.35 %
85.41 %
6.22
5.07
0.01899
Table 1. Quantitative Comparison. Our model outperforms the other methods on most of the metrics.
Figure 5. Results from in-the-wild image with multi-layer garments. Given an in-the-wild image as input, our DressWild method can generate multi-layer pose-agnostic and simulation-ready sewing patterns.
Method
Panel Acc. ↑
Edge Acc. ↑
Shape L2↓
w/o front features
91.18%
78.58%
12.11
w/ front features
94.35%
85.41%
6.22
Table 2. Ablation of canonical-space reconstruction features. The performance of sewing pattern reconstruction is significantly improved with the inclusion of this feature.
Figure 6. Ablation of pose features. The model identifies the garment better with the assistance of the pose feature when the human pose is challenging.
Figure 7. Ablation of feature fusion. With feature fusion design, the results are more consistent with GT.
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