Abstract: To mitigate the inherent ill-posedness of inverse scattering problems and improve reconstruction accuracy, this paper introduces generative priors and systematically investigates the influence of various generative model driven inversion strategies on reconstruction performance. Initially, a variational autoencoder (VAE) tailored for inverse scattering tasks is developed by fine-tuning the VAE parameters of Stable Diffusion. Subsequently, within a framework integrating contrast source prediction operators with VAE, three inversion strategies are proposed: frozen cascade strategy, fine-tuned cascade strategy, and pixel projection strategy. The performance of these strategies is systematically evaluated using synthetic and experimental data, focusing on reconstruction accuracy, noise robustness, and computational efficiency. Numerical results show that all three strategies effectively alleviate the ill-posedness of the inverse problem and enhance reconstruction quality. The frozen cascade strategy offers higher computational efficiency, the fine-tuned cascade strategy improves reconstruction accuracy by adjusting generative model parameters, and the pixel projection strategy provides superior reconstruction quality and stability in complex scenarios. This study demonstrates that effective integration of generative models enables the incorporation of prior information, significantly improving inversion accuracy and offering a reliable and robust solution for electromagnetic imaging.