Multi-modal recommendation systems have recently gained significant attention. They enhance ID representations with multimodal features, leading to improved recommendation performance. However, existing models face several challenges: (1) applying GCN solely to ID embeddings underutilizes multimodal features; (2) training with raw features may result in noise accumulation and propagation in GCN; and (3) aligning only the final ID and modality embeddings provides insufficient supervision for representation learning. To address these issues, we propose a novel Parallel Graph Structure Framework for Multi-Modal Recommendation (ParaRec). Specifically, we first denoise the input data and enhance the modality features of items. Then, we construct a parallel graph structure to effectively learn and fuse users’ preferences. Additionally, we introduce a full-stage self-supervised learning strategy to reduce redundant information in modality features. Extensive experiments on three real-world datasets demonstrate that ParaRec achieves state-of-the-art performance, outperforming the best competitor by up to 11.32% in recall@10.