The task of hazy image generation in the autonomous driving scenario has draws more and more attention in recent years. Existing research usually relies on software rendering, monocular depth estimation and GAN-based generative networks as their main approaches. However, their generated results could be limited by the following factors: weak sense of hazy image hierarchy and the attenuation of physical properties. To address these challenges, this paper proposes a two-stage framework, named Glow-Diffusion, for generating realistic hazy images based on the atmospheric multiple scattering model. In the first stage, the Hazy-Diffusion, with HD En-Decoder Resblock and Global Constrained Interaction Attention as the auxiliary wings, model hazy images with rich global information and strong geometric constraints using a conditional diffusion model. In the second stage, the prior module introduces the glow flare prior and the atmospheric point spread function prior to achieve the superposition of the glowing gradation phenomenon caused by artificial light sources. Extensive experiments have demonstrated that the proposed framework achieves state-of-the-art performance on unsupervised image quality metrics and subjective evaluation systems.