Multimodal recommendation systems attract significant attention due to their ability to integrate user feedback alongside item image and text information, addressing performance limitations caused by data sparsity. While previous studies have made notable progress in improving user and item representations through various fusion and alignment techniques, two critical issues remain. First, comprehensive user preference representation is crucial for enhancing recommendation diversity and accuracy. However, previous research has focused on in-depth exploration within each modality, lacking a broader, cross-modal perspective, which limits recommendation diversity and range. Second, the insufficient capture of fine-grained item similarities in multidimensional information fusion results in semantic distances that fail to reflect true associations between items. To address these challenges, we propose a novel Comprehensive Interest Modeling and Relational Mining (CMR) approach. Specifically, we construct two types of item graphs based on interaction data and multimodal information to improve item representations. We then model users’ broad preferences by integrating graph convolutional networks with personalized features and identity information from ID embeddings. Finally, we design cross-dimensional contrastive learning tasks to minimize the semantic distance between related items, enhancing multimodal information fusion accuracy. Extensive experiments on three public datasets demonstrate the effectiveness of our model.