Measurement-Consistent Langevin Corrector: A Remedy for Latent Diffusion Inverse Solvers

ICLR 2026 Conference Submission14907 Authors

Keywords: Latent Diffusion Model, Inverse Problem

TL;DR: We propose Measurement-Consistent Langevin Corrector, a plug-and-play module that stabilizes and improves latent diffusion inverse solvers.

Abstract: With recent advances in generative models, diffusion models have emerged as powerful priors for solving inverse problems in each domain. Since Latent Diffusion Models (LDMs) provide generic priors, several studies have explored their potential as domain-agnostic zero-shot inverse solvers. Despite these efforts, existing latent diffusion inverse solvers often exhibit undesirable artifacts and degraded quality. In this work, we introduce Measurement-Consistent Langevin Corrector (MCLC), a plug-and-play module that corrects reverse diffusion dynamics of the solver via Langevin updates while preserving measurement consistency. By reducing the gap to true reverse diffusion dynamics, MCLC substantially mitigates these artifacts and enhances stability of existing solvers. Our approach is grounded in a theoretical analysis of instability in solver dynamics. We demonstrate the effectiveness of MCLC and its compatibility with existing solvers across diverse image restoration tasks, achieving up to 50.7\% improvement in perceptual metrics (P-FID) over the base methods. We further analyze blob artifacts in the latent space, offering insights into their underlying causes. We highlight that the MCLC is a key step toward more robust zero-shot inverse problem solvers.

Primary Area: generative models

Submission Number: 14907