Image restoration via optimizing fidelity with diffusion and gradient regularization

Denoising diffusion models have achieved remarkable success in various image restoration tasks. However, most existing approaches rely solely on diffusion priors, which inherently limit their representational capacity and restoration performance. To address this limitation, we propose a novel image restoration method, which optimizes the fidelity with the diffusion and gradient regularization (FDGR). FDGR establishes a zero-shot, parameterized restoration framework by integrating the complementary strengths of diffusion priors and gradient priors. Specifically, by introducing a non-convex gradient regularization term, it effectively suppresses artifacts and preserves structural edges in noise-interfered environments, thereby enhancing overall image fidelity. Meanwhile, we extend this regularization term to the frequency domain to strengthen global texture capture and improve perceptual naturalness. The core of FDGR addresses the trade-off between structural fidelity and perceptual naturalness through a multi-mode restoration scheme that optimizes these objectives separately. Based on the proposed framework, we propose a novel plug-and-play sampling method that adapts to multi-scenario restoration tasks via parameterized adjustments without model retraining. Extensive qualitative and quantitative experiments on popular datasets show that the G ( + ) and G ( − ) modes of our method generally achieve superior results on the distortion-based and perception-based quality metrics, respectively. The ablation studies validate the efficacy of the proposed method. Our code is available at https://github.com/Anonymous-CodeSub/FDGR .