Abstract
This study aims to explore the generation and restoration of Kraak porcelain decorative patterns, employing a stable diffusion model that has been augmented with low-rank adaptation (LoRA) and ControlNet. Kraak porcelain, a historic blue-and-white ceramic traded along the maritime Silk Road, features intricate radial "Kaiguang" patterns that blend Chinese and Islamic artistic styles. The rules governing pattern generation have not been systematically deciphered through traditional methods, and as a result, innovation has remained constrained. The study established a multi-dimensional cultural database of 1530 Kraak porcelain images to facilitate training of a stable diffusion model. The model was fine-tuned using LoRA to adapt to small-sample data and integrated with ControlNet to impose structural constraints via edge maps. Experiments demonstrated that combining LoRA and ControlNet improved structural control over Kaiguang patterns and secondary motifs, validated by metrics like SSIM and PSNR. However, despite the improvement in pattern consistency, difficulties persisted in reproducing fine details and achieving seamless restoration of damaged patterns. Future work will focus on refining incremental learning strategies and fostering interdisciplinary collaboration, with the goal of bridging the gaps between technical implementation and cultural interpretation. Overall, this approach provides a framework for the digital preservation and innovative regeneration of cultural heritage.