In the past decade, multi-modal image fusion has demonstrated exceptional efficacy. However, the substantial computational expense and considerable model size present obstacles for implementing these models on resource-limited devices. For this issue, we propose a based on Multi-scale Addition Feature Decomposition (MAFD) network, for efficient image fusion tasks in this paper. By using addition instead of traditional multiplication, our network reduces the parameters and enhances computational efficiency while maintaining performance. In MAFD, we propose a Multi-scale Additive Self-attention Module (MASM) and a Routing Mix Attention Module (RMAM). Specifically, the attention mechanism in MASM employs addition instead of multiplication to compute queries and keys and captures context information at different scales. RMAM leverages routing mechanisms to handle semantic differences between different modalities and establishes internal dependencies through channel attention, allowing the module to prioritize more relevant information during shallow feature extraction. The experimental results point out that our model excels on infrared-visible image datasets and improves the efficacy of downstream tasks, including detection and segmentation.