<![CDATA[Single-image deraining for Unmanned Aerial Vehicle (UAV) imagery remains challenging due to non-uniform rain patterns, motion blur, and real-time processing requirements. Existing generative paradigms, including Generative Adversarial Networks (GAN), Diffusion, and Diffusion–GAN, each face inherent trade-offs among restoration quality, stability, and efficiency. To address the lack of unified and fair benchmarking across these paradigms, this study presents a systematic and controlled comparative evaluation of three representative models, including TBGAN, WeatherDiff, and SupResDiffGAN, to assess their relative performance in UAV deraining tasks. The models are evaluated on the UAV-Rain1K and Rain100L datasets using PSNR, SSIM, and inference efficiency metrics to support informed selection of paradigms for UAV applications. Experimental results show that WeatherDiff achieves the highest fidelity with 19.99 dB PSNR, 0.8375 SSIM on UAV-Rain1K and 29.51 dB PSNR, 0.9093 SSIM on Rain100L. TBGAN yields sharper details but lower structural consistency, whereas SupResDiffGAN offers balanced performance with 19.03 dB PSNR and 0.7053 SSIM on UAV-Rain1K and 28.51 dB PSNR and 0.8681 SSIM on Rain100L, with faster inference. These findings highlight the practical trade-offs among the three paradigms and demonstrate that diffusion–GAN frameworks provide the most practical solution for UAV deraining, combining diffusion stability with adversarial sharpness for real-time restoration.]]>
