<p>Data sparsity remains a critical challenge that hinders the performance of recommendation systems. Multimodal recommendation, which aims to alleviate this issue by enriching item and user representations through heterogeneous modality information, has garnered widespread attention. While existing studies have achieved notable success, their early or late fusion strategies inadequately balance modality-specific characteristics with cross-modal correlations. Moreover, static fusion paradigms are unable to effectively capture the dynamic changes in modality complementarity. To address these limitations, we propose a fused modality-enhanced graph convolutional network for multimodal recommendation (FM-GCN). Our method innovatively treats fused multimodal features as a distinct modality, achieving refined feature integration and noise suppression via a cross-attention mechanism and a self-prompted denoising diffusion model. Moreover, a decoupled feature distillation framework is employed to extract task-specific modality representations. Additionally, we construct a fused modality-enhanced user–item graph and a behavior-augmented item–item graph to synergistically model user interaction patterns and multimodal semantic relationships. Comprehensive experiments on four real-world datasets demonstrate that FM-GCN outperforms baselines with average improvements of 6.18%–8.10% in Recall@20 and NDCG@20 metrics, validating its effectiveness.</p>