AI-Assisted Yunjin Pattern Generation Using Fine-Tuned Diffusion Models for the Digital Exploration of Intangible Cultural Heritage

Nanjing Yunjin is an outstanding representative of human silk weaving techniques, named for its colorful and dignified decorations, which resemble colored clouds. However, the preservation and modern adaptation of Nanjing Yunjin have been hampered by traditional hand-weaving methods, material costs, and a lack of design resources. To support the digital exploration and contemporary design application of this Intangible Cultural Heritage (ICH), this study proposes an AI-assisted method that uses fine-tuned diffusion models to generate Yunjin pattern candidates. First, addressing the lack of specialized training data, we construct a dedicated dataset comprising 3495 annotated Yunjin images, categorized by motif themes and color styles. Second, we employ LoRA-based fine-tuning with DreamBooth-style prior preservation, incorporating motif types and color styles through textual conditioning to support domain-specific Yunjin pattern generation. Third, we establish a systematic GCSFBPFLA workflow and evaluate the generated patterns through both subjective scoring and objective image quality assessment (IQA) metrics, including SSIM, FID, and CLIP Score. Experimental results from the ablation study demonstrate progressive improvements from Configuration A to Configuration D, generating patterns with higher stylistic alignment, structural similarity, and text–image consistency. Finally, we present conceptual application scenarios of the generated patterns on daily products to illustrate their practical design potential. By supporting design exploration and diversifying visual references, this method provides a technical pathway for the digital exploration and contemporary design application of Yunjin patterns and can be extended to digital exploration and design-support tasks for other intangible cultural heritage domains.