Abstract:

[Purpose/Significance] The digital preservation and transmission of intangible cultural heritage (ICH) has increasingly benefited from the development of digital technologies. However, compared with text-, audio-, and video-based documentation, research on image generation and visual reproduction of ICH remains relatively insufficient. In particular, visual-based approaches for representing, reconstructing, and creatively transforming traditional patterns have not yet fully addressed the challenges of preserving cultural characteristics during the generation process. As generative artificial intelligence has rapidly advanced, image-based style transfer has gradually become an important approach for digitally reproducing and innovatively utilizing ICH. Nevertheless, when existing style transfer methods are applied to weaving-based ICH, they frequently encounter three major instability problems: cultural semantic confusion, structural distortion of patterns, and the loss of fine-grained color regularities. These problems lead to inaccurate representation of cultural meanings, deformation of traditional pattern structures, and weakening of distinctive color features, thereby limiting the reliability of ICH feature transfer and visual reconstruction. Therefore, formal modeling of transferable cultural features and the establishment of structured constraint mechanisms are urgently required to improve the controllability and cultural authenticity of the ICH generation. Based on the four-layer cultural gene framework of ICH, this study proposes an agent-driven multi-stage controllable transfer framework for ICH patterns. [Method/Process] The proposed framework introduces digital gene-based cultural representation and structured constraint mechanisms to achieve controllable migration of ICH patterns. Specifically, the framework adopts the Hermes Agent as the cognitive center, which enables semantic understanding, reasoning, and instruction generation during the transfer process. The pattern transfer task is broken down into three layers of hierarchical instruction: semantic parsing, cultural gene hierarchy priority determination, and generation parameter compilation. Through the coordination and scheduling of these instructions, the framework integrates perceptual encoding and generative decoding modules to realize the transformation from cultural knowledge representation to controllable image generation. The four-layer cultural gene system is embedded throughout the entire framework as a hierarchical constraint mechanism. Different gene layers provide corresponding constraints for the generated results: the phenotype layer guides the preservation of visual features such as pattern forms, symbols, and colors; the technique layer constrains weaving structures and craftsmanship-related characteristics; the semantic layer maintains cultural meanings and historical contexts; and the spirit layer ensures the continuity of cultural values and emotional expressions. Through the interaction of these four layers, the framework establishes structured constraints for the generation process and improves the cultural authenticity of transferred patterns. The study selects the Zhuanghua variety of Nanjing Yun brocade as an empirical case. As a representative traditional weaving technique, Nanjing Yun brocade contains complex cultural information, pattern structures, and color organization rules, making it suitable for evaluating controllable style transfer of weaving-based ICH. The effectiveness of the proposed framework is evaluated through quantitative indicators, including pattern semantic retention rate, garment silhouette retention rate, and instance color accuracy. In addition, ablation experiments are conducted to verify the contribution of the four-layer cultural gene constraints and examine their effectiveness in reducing instability during the generation process. [Results/Conclusions] Experimental results demonstrate that the proposed framework effectively maintains the flexibility of ICH pattern generation while ensuring the reliability of cultural semantics and visual characteristics. The integration of the four-layer cultural gene system significantly improves the controllability of generated outputs and effectively suppresses the three major instability problems in ICH style transfer, including cultural semantic confusion, structural distortion of patterns, and loss of fine-grained color regularities. The results indicate that cultural gene-based structured constraints can provide an effective solution for connecting cultural knowledge representation with generative artificial intelligence models. This study establishes a structured and controllable approach for the digital reproduction and creative transformation of weaving-based ICH patterns. It not only provides a practical framework for improving the authenticity and controllability of AI-generated cultural heritage images, but also offers a potential pathway for intelligent inheritance, digital preservation, and innovative development of traditional cultural heritage.

Key words: intangible cultural heritage (ICH), agent, cultural DNA, controllable transfer, Nanjing brocade, digital humanities