Abstract

Estimating sewing patterns from images is a practical approach for creating high-quality 3D garments. Due to the lack of real-world pattern-image paired data, prior approaches fine-tune large vision-language models (VLMs) on synthetic garment datasets, limiting their generalization to real-world images.

We propose NGL (Natural Garment Language), a novel intermediate representation that restructures GarmentCode into a format more understandable to large language models, and NGL-Prompter, a training-free pipeline that queries large VLMs to extract structured garment parameters from a single image, which are then deterministically mapped to valid GarmentCode. Our approach achieves state-of-the-art performance on standard geometry metrics and is strongly preferred in both human and GPT-based perceptual evaluations. NGL-Prompter can recover multi-layer outfits whereas competing methods focus mostly on single-layer garments, highlighting its strong generalization to real-world images even with occluded parts.

Method Overview

The NGL-Prompter pipeline:

1. Garment Detection: A VLM identifies all visible garment layers and their ordering from the input image.
2. Sequential QA with NGL: Rule-based questioning with logits masking queries the VLM for structured garment attributes, ensuring schema-compliant outputs.
3. Deterministic Parsing: Custom rules map NGL attributes to valid GarmentCode parameters.
4. 3D Reconstruction: GarmentCode compiles to sewing patterns, which are assembled, textured (via FabricDiffusion), and simulated on an estimated body (via TokenHMR).

NGL comes in two variants:

• NGL-0: 27 essential reconstruction parameters (coarse schema)
• NGL-1: 46 parameters including stylistic details (extended schema)

Multi-Layer Results

NGL-Prompter is the first method to successfully reconstruct occluded, multi-layer garments without training – a significant advantage over existing approaches that are limited to single-layer garments.

Single-Layer Results

Detailed comparisons demonstrate improved garment detail capture including skirt slits, mini dresses, and pant styles.

Quantitative Results

Geometry (Dress4D Dataset):

Perceptual Evaluations (ASOS 5K):

• Single-layer AI study: 4.80/9 (NGL-0) vs. 2.54/9 (ChatGarment)
• Multi-layer AI study: 4.05/9 (NGL-0) vs. 2.33/9 (ChatGarment)
• Human preference: consistently favors NGL-Prompter on a -2 to +2 scale

BibTeX

@article{badalyan2026ngl,
  title   = {{NGL-Prompter}: Training-Free Sewing Pattern Estimation from a Single Image},
  author  = {Badalyan, Anna and Selvaraju, Pratheba and Becherini, Giorgio and Taheri, Omid and Fernandez Abrevaya, Victoria and Black, Michael},
  journal = {arXiv preprint arXiv:2602.20700},
  year    = {2026},
}