Although significant progress has been made in daytime haze image restoration, nighttime variegated haze images remain underexplored. The primary light sources at night, such as artificial lighting and faint atmospheric light, result in low brightness, uneven illumination, and halo effects, which degrade existing algorithms performance for real-world nighttime variegated haze scenarios. To address these issues, this paper proposes a nighttime variegated haze restoration network (NVHRNet). Specially, we introduce a halo suppression module to address the halo issue, and a frequency selection module to extract effective features through frequency decomposition. In addition, we utilize a domain adaptation module to enhance the model’s generalization ability in real-world variegated haze weather conditions. Experimental results show that the proposed method achieves better restoration performance on both synthetic and real-world images. Further more, our method can significantly enhances the performance of subsequent vision tasks.