LAMP: New diffusion method sharpens image restoration without extra cost

Standard diffusion posterior samplers rely on instantaneous data-consistent estimates, which introduce temporal variability in the reverse dynamics. Evangelista et al. reinterpret this process from a dynamical perspective, showing that the standard update is essentially a first-order discretization with a residual correction. Their proposed method, LAMP, employs a second-order discretization that naturally incorporates a lagged temporal correction based on the variation of consecutive estimates. This correction is modular and can be plugged into existing posterior sampling backbones like DiffPIR or DDRM without architectural changes. The authors perform a one-step risk analysis to characterize the bias-variance trade-off that explains when LAMP improves the reverse transition.

Experiments across multiple imaging tasks—including super-resolution, inpainting, and deblurring—show that LAMP delivers consistent improvements over strong baselines such as DiffPIR and DDRM. Critically, these gains come without increasing the number of denoising evaluations, meaning LAMP enhances performance at the same computational cost. The method preserves the structure of a posterior sampler, making it easy to adopt. With 9 figures and 9 tables supporting the results, LAMP represents a practical advancement for anyone using diffusion models for image restoration, offering sharper outputs without trade-offs in speed or complexity.