Research on Digital Preservation Pathways for Shadow Puppetry Models Based on LoRA

Abstract

To address stylistic deviations and loss of core characteristics when general-purpose AIGC models reproduce shadow puppetry, this study explores and validates a digital heritage preservation pathway based on Low-Rank Adaptive Refinement (LoRA) fine-tuning technology. By constructing a dedicated shadow puppetry dataset and employing an “automatic annotation + manual correction” strategy for prompt processing, the foundational Stable Diffusion model undergoes targeted fine-tuning using LoRA techniques. Systematic evaluation of model convergence and generative quality was conducted using loss curves and X/Y/Z charts to select optimal models and parameters. Results indicate that the sub-model generated at the 14th training iteration, with weights at 0.8, delivers the finest performance, achieving high-fidelity reproduction of shadow puppetry’s core characteristics—including its flat, two-dimensional forms and intricate cut-out carvings. This model demonstrates outstanding stability and stylistic consistency across text-to-image generation, style transfer, and pose control applications. The research validates the efficacy of LoRA fine-tuning for intangible cultural heritage digitisation, providing technical support for shadow puppetry preservation and innovative design. It also establishes a replicable paradigm for digitising other stylistically distinct cultural heritage projects.