<p>Infrared and visible image fusion aims to simultaneously preserve thermal target saliency and texture details. However, most existing methods are limited to spatial domain representation learning and neglect valuable frequency domain information. In addition, the inherent semantic discrepancy between spatial and frequency domain representations further restricts the effective incorporation and fusion of frequency domain information. To address these issues, this paper proposes SF2Fuse, a wavelet-based spatial-frequency method for infrared and visible image fusion. First, shallow features are extracted from the two modalities to preserve the original signal distribution of the source images as much as possible. Then, global and local features are separately extracted in the spatial domain. In the frequency domain, a wavelet feature decomposer is introduced to explicitly obtain low- and high-frequency components complementary to the spatial domain through the Haar wavelet transform. To bridge the semantic gap between spatial and frequency features, we design a multi-scale cross-domain alignment module to map spatial and wavelet features into a unified semantic space through multi-scale orthogonal-direction rectangular convolutions and cross-attention. Furthermore, a cross-modulation fusion module is developed to adaptively integrate low- and high-frequency information, achieving a balance between texture fidelity and structural consistency. Extensive experiments show that the proposed method achieves superior fusion performance on multiple public datasets over existing state-of-the-art methods. It also demonstrates strong application potential and cross-domain generalization ability in downstream tasks, including object detection, semantic segmentation, and medical image fusion.</p>