<p>The fusion of infrared and visible light images aims to generate composite images that leverage the strengths of both modalities. To ensure extracted image features fully capture the core information of each modality and achieve comprehensive cross-modal complementarity at the feature level, we designed a fusion network named DDSEFuse. This network decomposes input images into global and local features through two core modules: transformer with spatial and channel attention and convolutional neural network with channel attention. Single-scale feature enhancement techniques are applied to progressively iterate residual processing on extracted global and local features, followed by a differentiated feature fusion strategy. The loss function incorporates a dual-driver mechanism to maximize the strengths of each module during fusion. Extensive experiments demonstrate that the proposed DDSEFuse achieves outstanding results in infrared–visible image fusion, significantly outperforming numerous state-of-the-art fusion algorithms. The training and optimization of this model underscore the pivotal role of supercomputing in advancing frontier research on complex deep learning models.</p>