Image inpainting, as a crucial task in computer vision, continues to face challenges such as insufficient inpainting accuracy and high computational cost, particularly when dealing with large and complex missing regions. To address these issues, this paper proposes a multi-stage image inpainting method based on lightweight Generative Adversarial Network (GAN) inversion. Built upon a lightweight hypernetwork, the proposed approach incorporates a multi-stage encoded feature fusion strategy to enhance contextual information integration. Additionally, a diversity regularization network is designed to ensure structural and stylistic consistency between the restored regions and the original image. Extensive experiments on datasets such as CelebA-HQ and Places2 demonstrate the superior performance of our method. It not only outperforms existing approaches in terms of objective metrics and subjective visual quality but also significantly reduces model parameters and improves inference efficiency, offering strong practical value.