Image inverse problems arise when measurements are blurred, undersampled, noisy, or transformed by an imperfectly known imaging system, as in MRI, CT, microscopy, and remote sensing. Methods combine physical forward models with iterative optimization and regularization, learned or unrolled networks, and plug-and-play denoisers; newer approaches also use diffusion and flow-based generative priors for restoration and posterior sampling. Work spans tasks such as deblurring, inpainting, super-resolution, sparse-view tomography, and phase retrieval, with training-free and self-supervised methods addressing settings where paired data are scarce. A central tension is how to use strong learned priors to fill in missing information without hallucinating details, while maintaining measurement fidelity, uncertainty estimates, robustness to model mismatch, and reliable convergence at practical computational cost.