Genpattern : dual-graph enhanced sewing pattern generation via multimodal large language model

Abstract

Customized garment production is hindered by the expert-dependent nature of sewing pattern generation—a skill-intensive process requiring years of training. While recent approaches aim to translate user intent into sewing patterns, they often struggle to interpret multimodal inputs such as text and images. Multimodal large language models (MLLMs) offer a promising path forward, as they can naturally understand diverse user intents. Yet, applying MLLMs to sewing pattern generation is challenging because conventional tokenization methods often lose the structural information of sewing patterns. To address this issue, we propose GenPattern, a novel framework that integrates structured graph modeling with MLLMs to enable more accurate sewing pattern generation. We introduce a scalable vector graphics (SVG)-style pattern tokenizer, which encodes sewing patterns into structured token sequences. Furthermore, we present SewGraphFuser, a dual-graph module that explicitly models geometric and semantic dependencies to inject structural information into MLLMs. This module combines a structure graph convolution module and a sequence graph convolution module to jointly capture multi-scale spatial and sequential features via a geometric consistency graph and a semantic dependency graph. Finally, to bridge the gap between digital design and physical fabrication, our framework drives a human-robot collaborative cutting platform, enabling expert-free, on-demand garment customization. This innovation empowers human-robot collaboration in pattern production, enhancing scalability in real-world manufacturing. Experimental results show that GenPattern achieves 86.7 % stitch accuracy and reduces panel vertex L2 error to 2.9 cm, demonstrating its potential to democratize custom fashion by enabling non-experts to reliably produce physical garments directly from their ideas.