Learning Sewing Patterns via Latent Flow Matching of Implicit Fields

TL;DR

Addresses sewing-pattern modeling with broad topology variation by introducing an implicit panel representation where each panel is encoded by a signed distance field $d_c$ and an unsigned distance field $d_p$, enabling differentiable meshing. A variational autoencoder learns a structured latent space for panels, and a latent flow matching model generates coherent panel combinations, complemented by a stitching-prediction module to recover seam relations. The framework supports pattern estimation from images, pattern completion, and pattern refitting across body shapes, with demonstrated improvements in panel quality and stitching accuracy over baselines. The approach provides a practical, differentiable tool for digital fashion design and physically grounded garment simulation, enabling reliable generation, reconstruction, and adaptation of sewing patterns.

Abstract

Sewing patterns define the structural foundation of garments and are essential for applications such as fashion design, fabrication, and physical simulation. Despite progress in automated pattern generation, accurately modeling sewing patterns remains difficult due to the broad variability in panel geometry and seam arrangements. In this work, we introduce a sewing pattern modeling method based on an implicit representation. We represent each panel using a signed distance field that defines its boundary and an unsigned distance field that identifies seam endpoints, and encode these fields into a continuous latent space that enables differentiable meshing. A latent flow matching model learns distributions over panel combinations in this representation, and a stitching prediction module recovers seam relations from extracted edge segments. This formulation allows accurate modeling and generation of sewing patterns with complex structures. We further show that it can be used to estimate sewing patterns from images with improved accuracy relative to existing approaches, and supports applications such as pattern completion and refitting, providing a practical tool for digital fashion design.

Figures (7)

• Figure 1: We model and generate sewing patterns using (1) an implicit garment representation, where panels are encoded as continuous distance fields and assembled into garments. Based on this representation, we enable (2.a) sewing pattern estimation from a single image, (2.b) pattern completion from partial panel inputs, and (2.c) pattern refitting for transferring garments across different body shapes.
• Figure 2: Overview of the sewing pattern modeling pipeline. Top: We learn a latent space of sewing patterns by encoding panel boundaries into continuous implicit fields, including a signed distance field (SDF) for panel shape and an unsigned distance field (UDF) for edge endpoints. Panel meshes are extracted through differentiable meshing. Bottom: Sewing pattern latent codes $\mathbf{z}$ are generated by sampling noise and mapping it through a flow matching model, optionally conditioned on an input image. The decoded 2D panels are placed in 3D using predicted translations $T$ and rotations $R$, and stitching relations are recovered by the edge-based classification model. The resulting sewing pattern is then assembled and draped in simulation to produce a 3D garment.
• Figure 3: Inferring edge endpoints. Left: We evaluate the UDF values $d_p$ on a dense grid (black dots) and retain only points near the zero roots (dark red regions inside dashed circles). Middle: These points are iteratively updated using Eq. \ref{['eq:update']} to move them toward the zero roots; black dots indicate the updated positions. Right: Cluster centers $c_i$ obtained from the converged points are used as edge endpoints, which partition the panel boundary loop (gray line) into edge segments $e_i$.
• Figure 4: Qualitative comparison on the SewFactory dataset.
• Figure 5: Qualitative comparison on the GCD dataset.