To enhance the accuracy of image-based recommendation systems, this study employs Graph Neural Networks (GNNs) to effectively model the complex relationships among users, items, and their multi-modal features. The proposed approach aims to leverage latent representations extracted from multiple modalities, including textual descriptions, product images, and user–item interaction behaviors. In this work, we introduce PCAGCN (Parallel Cross-Attention Graph Convolutional Network), a hybrid model that integrates a graph-based structure with a parallel cross-attention fusion mechanism. The model enables comprehensive feature learning and cross-modal correlation enhancement on the user–item graph, leading to improved recommendation performance in image-based scenarios.