Lensless imaging is promising for miniature applications owing to its compact, lightweight, and low-cost design. However, reconstruction quality severely degrades under low-light conditions, where the pervasive interference between structural information and measurement residuals (e.g., noise and brightness variations) poses a critical challenge for existing methods to recover clean, high-fidelity details. To address this, our work fundamentally revisits the lensless measurement paradigm, introducing a novel structure-guided model for low-light lensless imaging that embeds an explicit structure–residual decomposition within the forward process. This formulation breaks with the convention of holistic scene treatment by redefining the scene as a combination of intrinsic structural and residual components, enabling precise and targeted component-wise processing to achieve enhanced fidelity and noise reduction. Building on this model, we develop a two-stage physics-aware reconstruction approach: (1) a multilevel, multiscale extraction module embedded with the forward model initially extracts the components from the measurements, reducing noise impact on the structure through feature extraction across multiple scales and levels; (2) conditional diffusion modules, trained bidirectionally for stability, refine structures and optimize residuals generatively to boost detail recovery before fusion. Experiments on low-light datasets from our custom lensless camera (22,000 images with phase/amplitude masks) show that our approach outperforms state-of-the-art methods in both objective assessments and visual inspection, validating its advantages in denoising, structural fidelity, and overall image quality.
Liu et al. (Fri,) studied this question.