Abstract: Image restoration aims to recover high-quality visual content from degraded or incomplete images and has been widely applied in fields such as remote sensing, medical imaging, and cultural heritage restoration. However, in real-world scenarios, high-quality image restoration still faces two core challenges: insufficient representation of complex degradation and the lack of user guidance. The diffusion model, with its inherent generative stability, provides a novel solution paradigm for this pathological inverse problem. Based on a unified inverse problem framework, this paper systematically reviews the adaptation methods of diffusion models for typical image restoration tasks, including denoising, deblurring, super-resolution, and image completion. Furthermore, it analyzes the existing diffusion-based methods from the perspective of conditional constraints, model structure, and objective functions, and summarizes representative methods along with their key mechanisms. From a theoretical perspective, this paper deeply explores the convergence of posterior sampling, the accuracy of conditional modeling, and its correspondence with Bayesian inference, revealing the theoretical boundaries of this paradigm in handling pathological problems. The study shows that although there are still challenges in sampling efficiency and quantitative assessment of credibility, the diffusion model, by converting pathological restoration into a stable conditional generation process, has become an important theoretical foundation and technological frontier for achieving highly reliable and high-fidelity image restoration.