Learning Sewing Patterns via Latent Flow Matching of Implicit Fields

SIGGRAPH 2026

Cong Cao1,*
Ren Li1,2,*,†
Corentin Dumery3
Hao Li1,4
1MBZUAI 2SUSTech 3EPFL 4Pinscreen

*Equal contribution    †Corresponding author

Teaser showing implicit garment representation, pattern estimation, completion, and refitting

We learn sewing patterns as implicit fields in a latent flow matching framework, enabling pattern generation, image-to-pattern estimation, pattern completion, and refitting applications.

Overview Video

Project Video

Abstract

Sewing patterns define the structure of garments and are central to fashion design, fabrication, and physical simulation. Existing automated approaches still struggle with the diversity of panel shapes and seam layouts. This work introduces an implicit sewing-pattern representation in which signed distance fields describe panel boundaries and unsigned distance fields locate seam endpoints. These fields are embedded in a continuous latent space for differentiable meshing. A latent flow matching model learns distributions over panel combinations, while a stitching module predicts seam relations from extracted edge segments. The representation supports complex sewing-pattern generation, improves image-to-pattern estimation, and enables pattern completion and refitting.

Method Overview

Pipeline for latent sewing-pattern learning, generation, stitching prediction, and simulation

The pipeline learns panel latent codes through implicit fields, generates pattern layouts with latent flow matching, predicts stitching relationships, and simulates the completed garment.

Implicit Panel Fields

Panels are represented with continuous boundary and seam-endpoint fields, giving the model a smooth space for topology-aware garment structure.

Latent Flow Matching

A generative flow model learns the distribution of panel combinations in latent space, enabling diverse sewing-pattern synthesis.

Stitch Recovery

Extracted panel edges are paired by a stitching prediction module to recover seam relations and complete usable garment patterns.

Image-to-Pattern Estimation

Qualitative comparison between Apparel, our predictions, and ground-truth sewing patterns

Given a dressed garment image, the method estimates the underlying sewing pattern. Compared with vector-based Apparel predictions, the implicit representation recovers panel shapes closer to the ground truth while ensuring garment-part structure to be valid.

Pattern Editing Applications

Pattern Completion

Pattern completion examples with simulated garments

Starting from partial input panels, the model completes missing garment pieces, predicts compatible panel translations and stitches, and produces complete patterns that can be simulated as dressed garments.

Pattern Refitting

Pattern refitting examples across target bodies

Given a source garment and body, the learned pattern space supports refitting to a different target body by adapting the sewing pattern while preserving the garment style and enabling simulation on the new shape.

The learned representation can represent sewing patterns with complex structures, estimate more accurate and valid patterns from garment images than prior vector-centric approaches, and support editing tasks such as completion and refitting. These applications make the model useful for practical digital fashion workflows where pattern topology, geometry, and stitch structure must be handled together.

Citation

@article{cao2026learning,
  title   = {Learning Sewing Patterns via Latent Flow Matching of Implicit Fields},
  author  = {Cao, Cong and Li, Ren and Dumery, Corentin and Li, Hao},
  journal = {arXiv preprint arXiv:2601.17740},
  year    = {2026}
}