This paper proposes a novel diffusion-based image restoration method that decouples the data consistency step from the reverse sampling process of diffusion models. The method alternates between a reconstruction phase that maintains data consistency and a refinement phase that enforces the prior via diffusion purification. This approach is versatile and efficient, allowing for the use of both latent diffusion and consistency models. The method reduces the need for numerous sampling steps by integrating consistency models and achieves state-of-the-art performance in various image restoration tasks, including image denoising, deblurring, inpainting, and super-resolution. The proposed method is validated through comprehensive experiments, demonstrating its effectiveness in reducing inference time and improving reconstruction quality. The method is also adaptable to accelerated samplers and latent diffusion models, making it suitable for large-scale image restoration tasks. The key contributions include a framework that can easily incorporate any diffusion model and improved efficiency through the decoupling of data consistency and reverse sampling processes. The method is shown to outperform existing techniques in several quantitative metrics and is robust to measurement noise.This paper proposes a novel diffusion-based image restoration method that decouples the data consistency step from the reverse sampling process of diffusion models. The method alternates between a reconstruction phase that maintains data consistency and a refinement phase that enforces the prior via diffusion purification. This approach is versatile and efficient, allowing for the use of both latent diffusion and consistency models. The method reduces the need for numerous sampling steps by integrating consistency models and achieves state-of-the-art performance in various image restoration tasks, including image denoising, deblurring, inpainting, and super-resolution. The proposed method is validated through comprehensive experiments, demonstrating its effectiveness in reducing inference time and improving reconstruction quality. The method is also adaptable to accelerated samplers and latent diffusion models, making it suitable for large-scale image restoration tasks. The key contributions include a framework that can easily incorporate any diffusion model and improved efficiency through the decoupling of data consistency and reverse sampling processes. The method is shown to outperform existing techniques in several quantitative metrics and is robust to measurement noise.