Intelligent generation and propagation prediction of cultural tourism social media content based on diffusion models and multimodal large language models

Abstract

The digital transformation of cultural tourism has intensified the demand for high-quality visual content and accurate prediction of its social media propagation potential. However, existing approaches treat content generation and propagation prediction as isolated tasks, neglecting the coupling between cultural semantic fidelity and viral dissemination. This study proposes a unified framework integrating diffusion models with multimodal large language models (MLLM) to address both challenges jointly. The framework comprises three modules: an MLLM-semantic-guided conditional diffusion generation module with a cross modal conditional injection (CMCI) mechanism for hierarchical semantic control; a content-aware dynamic hypergraph cascade prediction network (CA-DHCNet) that fuses multimodal content features with dynamic hypergraph representations of propagation topology; and a collaborative optimization strategy driven by propagation potential feedback, forming a closed-loop paradigm of "generation–prediction–feedback–regeneration." Experimental results show that, compared to the best baseline method, the proposed method improves cultural semantic consistency by 16.3% and reduces the Fréchet Inception Distance (FID) by 12.1%. The prediction module reduces Mean Squared Logarithmic Error by 17.8% on CulTour, while collaborative optimization increases propagation potential scores by 21.1%. These results indicate a positive coupling between high generation quality and high propagation potential, offering a practical solution for intelligent cultural tourism content creation and marketing.