DGSolver: Diffusion Generalist Solver with Universal Posterior Sampling for Image Restoration

TL;DR

DGSolver tackles universal image restoration by unifying degradation representations through diffusion generalist models and improving inverse inference with high-order ODE solvers and universal posterior sampling. The method derives exact ODE formulations for forward and reverse processes, introduces multi-order solvers with a queue-based acceleration, and integrates UPS to provide manifold-constrained gradient guidance without training. Empirically, it delivers state-of-the-art restoration accuracy, stability, and scalability across natural and remote sensing datasets, while maintaining efficiency with a training-free pipeline. This approach offers strong practical impact for versatile, real-world restoration tasks with diverse degradation types.

Abstract

Diffusion models have achieved remarkable progress in universal image restoration. While existing methods speed up inference by reducing sampling steps, substantial step intervals often introduce cumulative errors. Moreover, they struggle to balance the commonality of degradation representations and restoration quality. To address these challenges, we introduce \textbf{DGSolver}, a diffusion generalist solver with universal posterior sampling. We first derive the exact ordinary differential equations for generalist diffusion models and tailor high-order solvers with a queue-based accelerated sampling strategy to improve both accuracy and efficiency. We then integrate universal posterior sampling to better approximate manifold-constrained gradients, yielding a more accurate noise estimation and correcting errors in inverse inference. Extensive experiments show that DGSolver outperforms state-of-the-art methods in restoration accuracy, stability, and scalability, both qualitatively and quantitatively. Code and models will be available at https://github.com/MiliLab/DGSolver.

Figures (14)

• Figure 1: Overview of the mainstream universal image restoration methods.
• Figure 2: Illustration of the DGSolver. In the forward process, we utilize a diffusion generalist model to uniformly represent diverse degradation categories into degradation-agnostic distributions. In the inverse inference, we employ a $k^{th}$-order solver to alleviate the discretization error accumulated in the multi-step sampling process. Simultaneously, universal posterior sampling is used to stabilize the solution and further minimize the discrepancies with the ideal one generated by a $\infty^{th}$-order solver.
• Figure 3: Different sampling strategies of diffusion generalist solvers ($k=2$).
• Figure 4: Visualization comparison with state-of-the-art methods on different restoration tasks.
• Figure 5: Visualization effects of varied solver and universal posterior sampling.

Theorems & Definitions (8)

• Proposition 1
• Theorem 1
• Lemma 1
• Proposition 2
• Lemma 2
• Lemma 3
• Proposition 3
• Proposition 4