LAMP: Lagged Temporal Corrections Improve Diffusion Posterior Sampling for Image Restoration
The innovative technique known as LAMP (Lagged temporal correction for diffusion posterior sAMpling with residual correction) enhances diffusion-based posterior sampling aimed at image restoration. This method reexamines posterior sampling through a dynamic lens, revealing that conventional updates are essentially first-order discretizations accompanied by a residual adjustment. By employing a second-order discretization, LAMP integrates a temporal correction that accounts for the differences between successive estimates. This lagged temporal correction can be seamlessly added as a modular plug-in to current posterior sampling frameworks. Further information about this method can be found in a paper available on arXiv (2605.12573).
Key facts
- LAMP stands for Lagged temporal correction for diffusion posterior sAMpling with residual correction.
- The method improves diffusion-based posterior sampling for image restoration.
- Standard posterior sampling updates are first-order discretizations with a residual correction.
- LAMP uses a second-order discretization introducing a temporal correction.
- The temporal correction is based on the variation of consecutive estimates.
- LAMP can be implemented as a modular plug-in over existing posterior sampling backbones.
- The paper is available on arXiv with ID 2605.12573.
- The approach reinterprets posterior sampling from a dynamical perspective.
Entities
Institutions
- arXiv