GarmentImage: Raster Encoding of Garment Sewing Patterns with Diverse Topologies

TL;DR

GarmentImage introduces a raster, multi-channel grid representation for garment sewing patterns that encodes geometry, topology, and placement, addressing the latent-space discontinuities and limited generalization of vector-based patterns. By embedding front/back layers, inside/outside flags, edge types, and a local deformation matrix into grids, GarmentImage enables continuous topology transitions and CNN-friendly modeling for pattern exploration, text-driven editing, and image-to-pattern prediction. Across three applications, GarmentImage delivers more robust latent spaces and superior generalization to unseen topologies compared with vector-based representations, using simple CNN-based architectures. Limitations include encoding non-uniqueness, boundary smoothness, and reliance on automatic encoding for complex garments, guiding future work toward richer features and real-world data integration.

Abstract

Garment sewing patterns are the design language behind clothing, yet their current vector-based digital representations weren't built with machine learning in mind. Vector-based representation encodes a sewing pattern as a discrete set of panels, each defined as a sequence of lines and curves, stitching information between panels and the placement of each panel around a body. However, this representation causes two major challenges for neural networks: discontinuity in latent space between patterns with different topologies and limited generalization to garments with unseen topologies in the training data. In this work, we introduce GarmentImage, a unified raster-based sewing pattern representation. GarmentImage encodes a garment sewing pattern's geometry, topology and placement into multi-channel regular grids. Machine learning models trained on GarmentImage achieve seamless transitions between patterns with different topologies and show better generalization capabilities compared to models trained on vector-based representation. We demonstrate the effectiveness of GarmentImage across three applications: pattern exploration in latent space, text-based pattern editing, and image-to-pattern prediction. The results show that GarmentImage achieves superior performance on these applications using only simple convolutional networks.

Figures (10)

• Figure 1: Representation overview.
• Figure 2: GarmentImage encoding and decoding process. A GarmentImage is automatically encoded from a sewing pattern in vector format and can be decoded back to the vector format. Given a sewing pattern in vector format (a), we stitch the panels and fill the gaps between them before rasterizing it into a bitmap (b). During this process, we establish correspondences between the original panel curves and the vertices on the bitmap grid. This allows us to assign edge types to each grid edge and compute deformation matrices that align the grid cells with the original panels. The resulting GarmentImage representation for each cell (c) contains inside/outside flags, edge types, and a deformation matrix, visualized as a parallelogram in the cell. For the decoding process, starting from the GarmentImage, we reconstruct the sewing pattern in vector format (d), which can then be used for applications such as simulation (e).
• Figure 3: GarmentImage examples. GarmentImage can handle diverse garments and features such as darts (a), waistbands (b), and holes (c).
• Figure 4: Text-based pattern editing pipeline. (a) We first train a GarmentImage VAE encoder and decoder on the GarmentImage reconstruction task. (b) Then we train an image decoder that predicts a simulated result in $64\times64$ given a latent code from GarmentImage VAE. (c) We optimize for the best GarmentImage VAE latent code that minimizes the SDS loss pooledreamfusion to conform to the input text prompt.
• Figure 5: Text-based pattern editing results. The input GarmentImage patterns adjust their geometry (a, b) and even topology (c, d) to match the input text prompts.