See other, focus self: local-to-global cross-modal guidance for infrared and visible image fusion

Infrared and visible image fusion (IVIF) plays an important role in robust scene perception by combining complementary thermal and texture information. However, many existing methods extract modality-specific features independently and rely on general-purpose global modeling, which limits fine-grained cross-modal interaction and weakens global consistency. Moreover, their fusion mechanisms remain difficult to interpret. To this end, we propose S2Fusion, a local-to-global collaborative framework for IVIF. The framework captures fine-grained cross-modal complementarity at the local level and further coordinates the enhanced features through long-range dependency modeling to achieve globally consistent fusion. Specifically, to strengthen fine-grained local interaction, we introduce the See-Other-Focus (SOF) block. Inspired by human visual perception, SOF uses context from the complementary modality to generate dynamic kernels for adaptive local aggregation, thereby facilitating the selection and preservation of complementary information. To enhance global consistency and interpretability, we further develop the Shared Heat Conduction Operator (SHCO), inspired by the physical principle of heat conduction. SHCO models long-range cross-modal correlations in the frequency domain using shared frequency value embeddings, thereby integrating local complementary information into a coherent global representation. Extensive experiments on benchmark datasets demonstrate that S2Fusion consistently outperforms state-of-the-art methods in terms of information fidelity, structural preservation, and visual consistency.