Multimodal recommendation systems enhance recommendation performance by integrating heterogeneous information such as text and images. However, they still face challenges in modeling complex dependencies between modalities and addressing data sparsity. To address these issues, this paper proposes a Dynamic Spectral Fusion Causal Graph Propagation (DSFCGP) architecture to achieve high-quality fusion of heterogeneous modalities and deep modeling of user-item interactions. Specifically, DSFCGP introduces a spectral convolution module that uses the Fourier transform to map modality features into the frequency domain. It highlights effective information by combining dominant frequency filtering and complex weights. A noise estimator is employed to dynamically assess the quality of modalities, enabling noise filtering and adaptive fusion. To mitigate modality interference caused by selection bias and exposure bias, the model incorporates a residual-enhanced causal preference modeling mechanism. This mechanism combines causal inference with contrastive learning to uncover users’ true interests. Finally, an attention-driven dynamic graph propagation mechanism is used to dynamically propagate personalized preferences among neighbors, effectively improving long-range modeling and robustness to cold start problems. Extensive experiments on multiple benchmark datasets from Amazon demonstrate that, compared with existing state-of-the-art methods, DSFCGP achieves an average significant performance improvement of 3% in key evaluation metrics, verifying the effectiveness and superiority of the proposed approach.