Application and Evaluation of Stable Diffusion-Based Generative AI in the Digital Reconstruction of Cultural Heritage Patterns

This study investigates the application of Stable Diffusion in the digital reconstruction of degraded or incomplete cultural heritage patterns. Focusing on traditional Chinese visual artifacts from murals, ceramics, and textiles, we propose a hybrid generative framework that integrates ControlNet for spatial constraint and LoRA for style-specific tuning. A culturally-informed prompt vocabulary and modular style embeddings are employed to ensure that the reconstructed outputs preserve both visual coherence and symbolic authenticity. To evaluate the quality and appropriateness of the generated patterns, we construct a multi-criteria expert assessment system grounded in the Analytic Hierarchy Process (AHP), comprising five key dimensions: structural integrity, stylistic fidelity, semantic accuracy, visual coherence, and cultural appropriateness. Expert evaluations reveal that pattern continuity, texture consistency, and avoidance of anachronisms are critical indicators of perceived reconstruction quality. This work contributes a scalable pipeline for culturally-aware AI restoration, proving extensible to new domains via modular fine-tuning, and provides empirical insight into the strengths and limitations of applying diffusion models in heritage preservation. The findings underscore the importance of combining algorithmic innovation with domain expertise to ensure that AI-assisted reconstructions remain both technically sound and culturally responsible.