Abstract: Infrared and visible image fusion integrates complementary information from different modalities to enhance salient target features and preserve textural details. The existing methods underutilize the strengths of the individual modalities and struggle to capture the complex intermodal relationships. To address this issue, an attention-guided cross-fusion method was investigated to evaluate the intermodal feature correlations and extract salient target features and texture details. L2 norm constraints were introduced during the training, and structural reparameterization optimized the model architecture to improve the efficiency. Experiments on the MSRS, TNO, and RoadScene datasets demonstrated that the model effectively combined deep features across modalities, preserving the salient target information and texture details. The visual information fidelity exceeded the highest values among the compared methods by 2.04%, 3.89%, and 4.71% for the three datasets. Structural reparameterization reduced the runtime by 23.19% relative to the model before reparameterization. The proposed method improved both the fusion performance and computational efficiency.