Learning Latent Transmission and Glare Maps for Lens Veiling Glare Removal

TL;DR

The paper tackles the hard problem of joint aberration and veiling glare in compact optical systems by introducing VeilGen, a physics-informed diffusion generator that uses Latent Optical Transmission and Glare Map Predictor (LOTGMP) to estimate transmission and glare maps, guiding a Veiling Glare Imposition Module (VGIM) to synthesize realistic degraded data under the forward model $I_{de}^p = (I_c^p \otimes K^p) \cdot T^p + I_g^p$. It couples VeilGen with DeVeiler, a reversible restoration network trained via a distilled forward model (DDN) and a reversibility constraint to invert the learned degradation through a Veiling Glare Compensation Module (VGCM). The approach demonstrates state-of-the-art restoration performance on two challenging lens types (SL and MRL), with strong generalization to real-world glare conditions and data-efficient domain adaptation. By grounding data synthesis and restoration in physically interpretable latent maps, the method provides robust, practical solutions for high-quality imaging in space-, mobile-, and AR/VR-enabled compact optics.

Abstract

Beyond the commonly recognized optical aberrations, the imaging performance of compact optical systems-including single-lens and metalens designs-is often further degraded by veiling glare caused by stray-light scattering from non-ideal optical surfaces and coatings, particularly in complex real-world environments. This compound degradation undermines traditional lens aberration correction yet remains underexplored. A major challenge is that conventional scattering models (e.g., for dehazing) fail to fit veiling glare due to its spatial-varying and depth-independent nature. Consequently, paired high-quality data are difficult to prepare via simulation, hindering application of data-driven veiling glare removal models. To this end, we propose VeilGen, a generative model that learns to simulate veiling glare by estimating its underlying optical transmission and glare maps in an unsupervised manner from target images, regularized by Stable Diffusion (SD)-based priors. VeilGen enables paired dataset generation with realistic compound degradation of optical aberrations and veiling glare, while also providing the estimated latent optical transmission and glare maps to guide the veiling glare removal process. We further introduce DeVeiler, a restoration network trained with a reversibility constraint, which utilizes the predicted latent maps to guide an inverse process of the learned scattering model. Extensive experiments on challenging compact optical systems demonstrate that our approach delivers superior restoration quality and physical fidelity compared with existing methods. These suggest that VeilGen reliably synthesizes realistic veiling glare, and its learned latent maps effectively guide the restoration process in DeVeiler. All code and datasets will be publicly released at https://github.com/XiaolongQian/DeVeiler.

Figures (15)

• Figure 1: Compact optical systems suffer from residual aberrations and veiling glare (a), caused by design-induced blur and stray-light scattering from non-ideal surfaces and coatings. (b) Existing methods fail under this compound degradation: a Computational Aberration Correction (CAC) model liang2021swinir retrained on aberration-only data fails on unseen veiling glare, while a cascaded state-of-the-art dehazing model wang2025learning introduces inconsistent artifacts. Our DeVeiler restores a clean image by jointly correcting the compound degradations.
• Figure 2: Overall architecture of the proposed VeilGen. In Stage i, VeilGen is trained to synthesize compound degradations by using a Latent Optical Transmission and Glare Map Predictor (LOTGMP) to estimate latent maps. These maps then guide the diffusion process via the Veiling Glare Imposition Module (VGIM). $Z_{de}^\mathcal{T}$ denotes the target degraded latent representation. $Z_{t}$ denotes the noisy latent representation at timestep $t$ of the forward diffusion process. $Z_{null}$ represents an all-zero latent representation. $txt^{\mathcal{S}/\mathcal{T}}$ denotes the text prompts for the source and target domains, respectively.
• Figure 3: The distillation and restoration networks. (a) The Distilled Degradation Net (DDN), trained in Stage ii, models the forward degradation, using VGIM to apply the veiling glare prior. (b) The DeVeiler, trained in Stage iii, is designed to reverse this process: it first removes the veiling glare using the VGCM, and then feeds the intermediate result into its main bottleneck to correct the aberrations.
• Figure 4: Visual results on Screen-Compound domain. Zoom in for the best view.
• Figure 5: Visual results on Realworld-Compound domain. Zoom in for the best view.