Abstract: To address the challenges of insufficient real-time performance, loss of target features, and difficulty in suppressing background artifacts during unmanned aerial vehicle (UAV) infrared-visible image fusion in complex environments, an optimized adaptive fusion algorithm based on adversarial game strategies was proposed in this study. Built on the lightweight RFN-Nest backbone, the method established a generator–discriminator adversarial framework. The generator employed channel–spatial dual attention to dynamically weight infrared thermal targets and visible textures in real time, whereas the discriminator supervised global realism with real fused images. Furthermore, an infrared-statistics-based mask weighted loss was introduced to enhance weak thermal targets, achieving high-precision real-time fusion within 20 M parameters and enabling deployment on edge devices. On the public TNO dataset, compared with the second-best method, ROIMaskCNN, the proposed approach increased entropy (EN) by 1.8%, standard deviation (SD) by 5.1%, and mutual information (MI) by 5.0%; additionally, it reduced the non-artifact blur factor (N_abf) by 9.4% and improved multi-scale structural similarity (MS-SSIM) by 0.7%, demonstrating clear superiority.