Fast, Physically Faithful, and High-Quality Lensless Imaging via Knowledge Distillation of Denoising Diffusion Null-Space Model

Abstract

   Lensless imaging is a computational optics technology that achieves compact imaging systems by combining coded optics with image reconstruction. Since the reconstruction involves solving an ill-conditioned inverse problem of disentangling superimposed optical information, algorithm design is critical. Classical methods are limited in reconstruction quality, while deep-learning approaches lack fidelity to the physical model. Recently, diffusion null-space models—which restrict generative completion to image components unrecoverable by inverse analysis—have achieved both physical fidelity and high quality, but remain slow due to iterative sampling.
   We propose Null-Space Diffusion Distillation (NSDD), which distills the null-space generator into a single forward pass, enabling non-iterative, physically faithful, high-quality reconstruction. Numerical experiments quantitatively showed that NSDD uniquely achieves speed, physical fidelity, and perceptual quality simultaneously. Optical experiments confirmed its effectiveness in real-world settings, including strong generalization to out-of-distribution objects. We also verify that restricting distillation to the null-space generator yields substantially better performance than naïve full-image distillation. These results position NSDD as a key enabling technology toward practical lensless cameras.