NGL: Natural Garment Language for Training-Free Sewing Pattern Estimation

Anna Badalyan; Pratheba Selvaraju; Giorgio Becherini; Omid Taheri; Victoria Fernandez Abrevaya; Michael Black

cs.CVFeb 24, 2026

NGL: Natural Garment Language for Training-Free Sewing Pattern Estimation

Authors: Anna Badalyan, Pratheba Selvaraju, Giorgio Becherini, Omid Taheri, Victoria Fernandez Abrevaya, Michael Black

Organizations: Max Planck Institute for Intelligent Systems

Abstract

Estimating sewing patterns from images is a practical approach for creating high-quality 3D garments, but it remains challenging due to the scarcity of paired real-world image and sewing-pattern data. Existing methods address this limitation by training vision-language models (VLMs) to learn low-level sewing-pattern representations from synthetic garments sampled from parametric garment models. However, they often struggle to generalize to in-the-wild images, fail to capture real-world correlations between garment parts, and are restricted to single-layer outfits. In contrast, we observe that VLMs are effective at describing garments in natural language, but mapping these descriptions into valid sewing patterns remains difficult. To bridge this gap, we propose NGL (Natural Garment Language), a novel domain-specific language that represents garments in terms aligned with VLMs' natural descriptive abilities. Leveraging NGL, we introduce a fully training-free pipeline that queries large VLMs to extract structured garment specifications and deterministically converts them into valid sewing patterns. We evaluate our method on the Dress4D, CloSe and a newly collected dataset of 253 in-the-wild fashion images. Our approach achieves state-of-the-art performance on standard geometry metrics and is preferred in both human and GPT-based perceptual evaluations compared to existing baselines. Furthermore, NGL recovers multi-layer outfits whereas competing methods focus mostly on single-layer garments, highlighting its strong generalization to real-world images even with occluded parts. These results demonstrate that an efficient garment representation is critical for sewing pattern estimation with VLMs. Our code and data will be released for research use.

Figures & tables

Figure 2: Overview of the training-free pipeline with NGL. Given an image containing a single- or multi-layer outfit, the method prompts a VLM in two stages: first, to identify garment types, then to extract attributes for each garment according to the NGL schema. A parser validates each attribute assigning a default for missing and invalid values, and then maps the results to GarmentCode design parameters. GarmentCode then generates the final sewing patterns.

Figure 3: NGL parameter schema. We show NGL’s five parameter blocks and the respective list of parameters. Bodice is activated when the type is dress, top or jumpsuit, sleeve is activated then bodice sleeve parameter yields true, pants are activated for pants and jumpsuit and skirt is activated for skirt and dress. Within the blocks, we show parameter hierarchy using bullet point levels.

	Dress4D			CloSe
Method	CD ( $\downarrow$ )	F-Score ( $\uparrow$ )	FR ( $\downarrow$ )	CD ( $\downarrow$ )	F-Score ( $\uparrow$ )	FR ( $\downarrow$ )
ChatGarment	3.99	0.78	$2.78\text{\,}\mathrm{\%}$	3.59	0.76	0
ChatGarment*	3.12	0.75	0	2.94	0.79	0
ChatGarment-GPT-5	5.41	0.78	$8.3\text{\,}\mathrm{\%}$	8.45	0.73	0.68 $\mathrm{\%}$
Design2GarmentCode-GPT-5	2.69	0.79	$2.78\text{\,}\mathrm{\%}$	2.93	0.76	1.38 $\mathrm{\%}$
NGL-GPT-5	1.75	0.81	0	2.79	0.78	0.68 $\mathrm{\%}$

Table 1: Quantitative evaluation on the Dress4D and CloSe datasets . ’ChatGarment’ reports results from our re-evaluation, and ’ChatGarment*’ reports values from the original paper. FR refers to failure rate.

Method	Single-layer (146 images)		Multi-layer (98 images)
Method	Mean ( $\uparrow$ )	FR ( $\downarrow$ )	Mean ( $\uparrow$ )	FR ( $\downarrow$ )
ChatGarment	2.678 $\pm$ 0.286	0	1.633 $\pm$ 0.133	$1\text{\,}\mathrm{\%}$
ChatGarment-GPT-5	2.842 $\pm$ 0.233	0	2.061 $\pm$ 0.221	0
Design2GarmentCode-GPT-5	4.438 $\pm$ 0.258	$5.2\text{\,}\mathrm{\%}$	1.806 $\pm$ 0.181	0
NGL-GPT-5	5.692 $\pm$ 0.253	0	5.163 $\pm$ 0.398	0

Table 2: AI study on single-layer and multi-layer images from the ASOS dataset. Mean is reported with 95% confidence intervals (CI). The scores take into account only the images that were successfully processed by each method. NGL particularly outperform others on multi-layer images. FR refers to failure rate.

NGL-GPT-5 versus:	Outfit type	N images	Mean ( $\uparrow$ )
ChatGarment-GPT-5	Multi-layer	97	$1.44\pm 0.08$
Design2GarmentCode-GPT-5	Single-layer	143	$0.56\pm 0.15$

Table 3: Human study on single-layer and multi-layer garment images from the ASOS dataset . Positive values indicate preference for NGL-GPT-5. On average, 8 participants successfully completed the study. Mean preference scores are averaged per image and reported with 95% confidence intervals (CI).

Figure 5: Empirical results on VLMs knowledge about garments. The plot shows the F1 score computed on ASOS_labeled dataset across selected set of parameters from NGL and various model sizes. All models can confidently identify intricate garment details that are commonly described on fashion websites (e.g. strapless_type straight or heart-shaped strapless neckline), but struggle with details that are not commonly described (e.g. skirt_asymm skirt with the back part longer than the front or one side longer than the other)

Figure 6: Quantitative results on Garment Attribute Accuracy across NGL LODs. We report the F1 score on our ASOS_labeled dataset to evaluate the prediction quality accross NGL parameters and model sizes. NGL-0

\cap

NGL-1 denotes the subset of attributes shared by both schema versions.

Figure 7: Prompting strategy and level of detail ablation. Only GPT-5 is able to benefit the additional details present in NGL-1. Smaller models are able to achieve reasonable reconstruction with NGL-0 under sequential prompting strategy.

Figure 8: Qualitative comparisons on the CloSe and Dress4D datasets . We compare our method (NGL-GPT-5.0) with Design2GarmentCode, ChatGarment-GPT, and ChatGarment. Columns 1-6 show results from the CloSe dataset, while columns 7-8 show from the Dress4D dataset. NGL correctly captures garment shapes and lengths, which results in a lower Chamfer Distance to the original mesh.

Figure 9: Qualitative results from the ASOS dataset . We compare our method (NGL-GPT-5.0) with Design2GarmentCode, ChatGarment-GPT, and ChatGarment. NGL accurately captures garment details on both multi-layer garments (columns 1-3) and single-layer garments (columns 4-8), particularly outperforming others on details such as neckline shapes.

Explore similar work

PatternGSL: A Structured Specification Language for Template-Free and Simulation-Ready 3D Garments
Jun 23, 2026Zhenyang Li, Lutao Jiang, Yizhou Zhao +43D Reconstruction Single-View 3D Reconstruction
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
DressWild: Feed-Forward Pose-Agnostic Garment Sewing Pattern Generation from In-the-Wild Images
Feb 18, 2026Zeng Tao, Ying Jiang, Yunuo Chen +73D Reconstruction Feed-Forward 3D Reconstruction
UCLA · Fudan University · Style3D +2
GarmentWeaver: Schema-Aware Structured Synthesis for Multimodal Sewing Patterns
Aug 31, 2026Yinwen Lu, Weihao Luo, Yueqi ZhongConstrained Generative Modeling Multimodal Generation
Donghua University · College of Textiles · Shanghai, China +3

Jun 23, 2026J/K move · Enter open · S save

PatternGSL: A Structured Specification Language for Template-Free and Simulation-Ready 3D Garments

Zhenyang Li, Lutao Jiang, Yizhou Zhao, Ying-Cong Chen+3

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

3D Reconstruction Single-View 3D Reconstruction

Open paper AskPDF arXiv

Reconstructing realistic, physically plausible garments from a single image remains a fundamental challenge. Template-free methods capture surface geometry but lack explicit sewing structure for simulation; while programmatic systems are simulation-ready but constrained by predefined templates. This reveals a fundamental representation gap between geometric reconstruction and structured garment construction. We present PatternGSL, a structured garment representation in the form of a template-free and learnable specification language that encodes complete sewing patterns, including panel boundaries, parameterized seams, and explicit stitch topology, in a compact and standardized form. PatternGSL preserves the physical rigor of pattern-based models while removing template dependence, elevating sewing structure as a first-class target for generative modeling. We further propose a vision-language framework that predicts PatternGSL specifications directly from a single image and decodes them into garments using lightweight deterministic validity handling, without optimization-based refinement or manual cleanup. In addition, we introduce PatternGSLData, the first large-scale image-to-GSL paired dataset comprising 300K samples with complete sewing pattern annotations, enabling supervised VLM training for structured garment reconstruction. Experiments demonstrate improved pattern accuracy over prior baselines, explicit sewing-structure recovery, reliable cloth simulation, and pattern-level editing through the same deterministic decoding pipeline. Code and data-processing scripts will be released at https://lagrangeli.github.io/PatternGSL/.

Abstract

Explore similar work

PatternGSL: A Structured Specification Language for Template-Free and Simulation-Ready 3D Garments

DressWild: Feed-Forward Pose-Agnostic Garment Sewing Pattern Generation from In-the-Wild Images

GarmentWeaver: Schema-Aware Structured Synthesis for Multimodal Sewing Patterns