Design of the Three Rabbit Caisson Pattern Based on the Stable Diffusion Model

To address the low efficiency in the design process of three rabbit caisson patterns and the scarcity of existing pattern samples, this research introduces a pattern generation method based on the stable diffusion model. First extracting cultural features while processing the training set, then the LoRA (Low-Rank Adaptation algorithm) fine-tuning network with the features of the three rabbit caisson, is integrated and then trained into the stable diffusion model. This integration facilitates the rapid generation of novel pattern variants employing a text-to-image methodology. The experimental results demonstrate that the lowrank adaptation approach can effectively train a model that captures the characteristics of the three rabbit caisson pattern with a limited number of photos and conclude a generalized LoRA model training method suitable for complex planar patterns. Finally, the trained LoRA is used in combination with other models to achieve the intelligent design of patterns. This technological advancement opens new avenues for the popularization and diversification of Dunhuang caisson pattern culture and provides valuable insights for the digitized preservation of other Chinese intangible cultural heritages.