A Lightweight Diffusion Framework for Cultural Pattern Generation Using LoRA

Abstract

Atlas patterns, a representative element of Uyghur intangible textile heritage from Xinjiang, are renowned for their vibrant colors and intricate symbolic structures. However, existing generative AI approaches often struggle to replicate their cultural semantics or require significant computational resources. This study addresses these limitations by proposing a low-resource, high-fidelity generation method based on Low-Rank Adaptation (LoRA) applied to the Stable Diffusion (SD) model. A categorized dataset of 103 Atlas pattern images was constructed and annotated using Peirce’s semiotic framework, including Iconographic, Directional, and Cultural Semiotic categories. LoRA modules were used to fine-tune the U-Net component of SD, substantially reducing training complexity and memory usage. Experimental results demonstrate that under optimal parameters (epoch = 100, batch size = 2, U-Net learning rate = $8 × 10^{−4}),$ the model achieves superior performance in MSE, PSNR, and SSIM compared to the base model. The generated patterns effectively reflect the geometric symmetry and cultural motifs of traditional Atlas textiles. This research highlights the practical potential of LoRA-adapted diffusion models in preserving and innovating cultural heritage, especially in the fields of digital fashion, pattern design, and virtual cultural applications.