Imaging through dynamic scattering media with an adapter-enhanced diffusion model

Reconstructing objects hidden behind complex scattering media remains a fundamental challenge in optical imaging. Existing deep learning methods often require extensive training data and generalize poorly in varying scattering conditions. To address these limitations, we introduce an adapter-enhanced diffusion model (ADM) that combines the powerful denoising capability of diffusion models with a test-time adapter, allowing robust cross-domain alignment even with limited training data. Trained on only 100 paired images from a single indoor scattering condition, ADM surpasses classical Retinex optimization, U-Net, SwinUNet, DescatterNet, and a diffusion model baseline, consistently achieving superior image quality across diverse scenarios-from higher scatterer concentrations to different media types, including artificial fog, milk, and real outdoor fog. ADM demonstrates remarkable data efficiency, performing comparably to U-Net models trained with 12 times more data. Attention map analyses reveal that the test-time adapter effectively steers the diffusion process toward semantically meaningful regions, supporting ADM's robust generalization from limited data. Our results highlight ADM as a promising and efficient solution for imaging under real-world scattering conditions. Code and pre-trained models will be released for reproducibility.