MDP: multidiffusion process plug-and-play model for image restoration

Image restoration (IR) is a key research direction in digital image processing, aiming to recover degraded images to their original state by leveraging prior knowledge. Recently, diffusion models (DMs) have demonstrated remarkable performance in image generation, and their application to IR has attracted increasing attention. However, existing methods often rely on expensive retraining or suffer from limitations in fidelity, perceptual quality, and robustness to unknown noise. To address these challenges, we propose a multidiffusion process plug-and-play model (MDP), a plug-and-play diffusion-based image restoration framework. At each sampling step of the diffusion process, MDP incorporates diffusion priors through half-quadratic splitting and integrates observed information y via multiple diffusion steps, achieving both flexibility and generality while ensuring efficient and high-quality restoration. Experimental results show that MDP achieves superior performance on deblurring and inpainting tasks across FFHQ and CelebA datasets, effectively balancing objective metrics and perceptual quality. Moreover, it exhibits strong generalization and robustness, offering a practically valuable solution for image restoration.