Shadow puppetry style transfer via explicit structure-style decoupling

Shadow puppetry, as an intangible cultural heritage, holds significant value for the promotion of sustainable cultural development through style transfer. Although instruction-driven image editing and style transfer methods have achieved remarkable progress in general scenarios, they still struggle to simultaneously preserve geometric topology and stylistic consistency when applied to art forms with strong structural constraints such as shadow puppetry. To address this challenge, we propose an instruction-driven editing framework for shadow puppetry style transfer with explicit structure–style decoupling. Specifically, we design an instruction generation pipeline based on multimodal large language models (MLLMs), which employs structured semantic descriptions to explicitly separate the intrinsic geometric structure of shadow puppets from target costume style attributes, thereby alleviating ambiguity in semantic encoding. During the generation stage, the framework integrates a dual-branch connector module with a diffusion-based generation network, enabling global structural consistency control and fine-grained local semantic refinement for high-fidelity cross-cultural costume re-rendering. Furthermore, we construct the first dedicated image dataset for Chinese shadow puppetry and develop a multi-dimensional evaluation protocol combining the Analytic Hierarchy Process and fuzzy comprehensive evaluation. Experimental results show that the full method achieves a 4.8% improvement in structural fidelity (0.4689). This performance not only significantly surpasses existing open-source baselines but also exhibits competitive accuracy approaching that of closed-source models, establishing a robust technical foundation for the creative reuse of cultural heritage.