FlowLPS: Langevin-Proximal Sampling for Flow-based Inverse Problem Solvers
This work addresses convergence and deviation issues in inverse problem solving using flow models, which is important for researchers in computational imaging and generative modeling, though it appears incremental as it builds on existing training-free methods.
The paper tackled the problem of existing training-free methods failing to converge to the posterior mode or suffering from manifold deviation when applied to latent flow models for inverse problems, and introduced FlowLPS, a novel training-free framework that integrates Langevin dynamics and proximal optimization, achieving superior reconstruction fidelity and perceptual quality across tasks on FFHQ and DIV2K datasets.
Deep generative models have become powerful priors for solving inverse problems, and various training-free methods have been developed. However, when applied to latent flow models, existing methods often fail to converge to the posterior mode or suffer from manifold deviation within latent spaces. To mitigate this, here we introduce a novel training-free framework, FlowLPS, that solves inverse problems with pretrained flow models via a Langevin Proximal Sampling (LPS) strategy. Our method integrates Langevin dynamics for manifold-consistent exploration with proximal optimization for precise mode seeking, achieving a superior balance between reconstruction fidelity and perceptual quality across multiple inverse tasks on FFHQ and DIV2K, outperforming state of the art inverse solvers.