Abstract: To address the issue of visual degradation in multi-modal image fusion under nighttime or low-light conditions, this paper proposes a dual-path adversarial generation framework that integrates an adaptive illumination compensation network and a cross-modal feature interaction module. First, a luminance-aware visible image correction unit is designed, which generates region enhancement weights by quantifying illumination intensity, thereby enhancing visual features in dark areas. Second, during the feature extraction stage, a cross-attention mechanism and residual dense blocks are introduced to strengthen the collaborative representation of infrared target structural features and visible high-frequency textures through channel-level feature recalibration. Finally, a dual-path discriminative network is constructed to separately model infrared contrast features and visible detail distributions, mitigating the modal feature bias caused by a single discriminator. Experiments conducted on the LLVIP dataset demonstrate that the proposed method achieves competitive performance in three objective metrics: information entropy, correlation coefficient, and visual information fidelity.