Automating garment pattern making with AI: evaluating the performance and practical utility of fine-tuned large language models in fashion production

Junyi Chen

doi:10.17918/00011002

This study investigates the effectiveness of machine learning models in automating garment pattern generation, addressing critical challenges in the fashion manufacturing industry. Traditional pattern-making methods, while refined through generations of practice, remain labor-intensive and constrained by human limitations in accommodating diverse body shapes and complex geometries. Through a comprehensive mixed-methods approach, this research develops and evaluates an AI-powered pattern generation system utilizing fine-tuned large language models (LLMs) including ChatGPT-4o, GPT-4o-mini, Gemini 1.5, and Llama 3.1. The methodology employed a two-stage AI pipeline architecture: (1) a vision model for extracting garment features from images and converting them to structured JSON format, and (2) specialized models for transforming JSON specifications into SVG pattern files. Training datasets were systematically constructed using over 10,000 garment patterns from industry partner URBN and the public GarmentCodeData repository. Three distinct fine-tuning approaches were evaluated: specialized training on 150 simple jumpsuit patterns, moderate training on 300 mixed-complexity pants patterns, and comprehensive training on 3,000 diverse garment types. Expert evaluation by professional pattern makers from URBN revealed counterintuitive findings regarding the relationship between training data characteristics and output quality. The specialized jumpsuit model achieved approximately 42% usability in manufacturing contexts, requiring only minor adjustments for production implementation. Conversely, models trained on larger, more diverse datasets (300 and 3,000 examples) produced patterns deemed unsuitable for real-world manufacturing, despite conventional machine learning expectations that larger datasets improve performance. The research identified an exponential relationship between pattern complexity and training data requirements, with simple patterns amenable to AI automation while complex designs continue to challenge current technological capabilities. Technical limitations included difficulties in capturing nuanced construction details, managing diverse human anatomical variations, and maintaining manufacturing-ready specifications. The vision component demonstrated robust performance in feature extraction, showing predictable scaling with dataset size. Key findings indicate that focused, application-specific AI models outperform general-purpose systems for pattern generation. The study provides empirical evidence that successful AI integration in pattern making requires strategic focus on specific garment categories with consistent construction principles rather than pursuing universal automation. These results suggest that AI should augment rather than replace human expertise, with optimal applications in standardized pattern production where consistency and efficiency are prioritized over creative complexity. This research contributes to understanding AI applications in creative industries by establishing performance benchmarks and identifying optimal training strategies for technical design automation. The findings offer practical guidance for fashion companies considering AI adoption, educators developing curricula that integrate AI literacy, and researchers advancing automated design technologies. Future work should explore hybrid human-AI approaches and specialized models for different garment categories while addressing material property integration and cultural design variations. Keywords: artificial intelligence, machine learning, garment pattern making, automated design, fashion technology, large language models, computer vision, human-AI collaboration.

Automating garment pattern making with AI: evaluating the performance and practical utility of fine-tuned large language models in fashion production

Files and links (1)

Abstract

Metrics

Details

Automating garment pattern making with AI: evaluating the performance and practical utility of fine-tuned large language models in fashion production

Files and links (1)

Abstract

Metrics

Details

Drexel University Social media