Generative Point-Cloud Priors for Deblurring and Super-Resolution of Single-Photon LiDAR

This paper presents a plug-and-play framework for joint deblurring and reflectivity-guided super-resolution of noisy and low-resolution single-photon LiDAR data. The method alternates between an analytical deblurring update and a guided point-cloud upsampling step, enabling existing point-cloud super-resolution algorithms to be used as priors. We focus on recent generative upsampling methods based on diffusion and flow matching. To improve robustness in photon-sparse and high-background regimes, the approach incorporates a multiscale point-cloud estimation strategy. It supports arbitrary upsampling factors at inference without retraining and handles known spatial blur kernels arising from motion or long-range imaging. Experiments on simulated and real SPL data demonstrate effective denoising, deblurring, and high-resolution 3D reconstruction under challenging imaging conditions.