Progressive Diffusion for Single Image Dehazing

Single image dehazing remains a challenging task, particularly under dense and non-uniform haze conditions where scene structures and fine textures are severely degraded. Recently, denoising diffusion probabilistic models (DDPMs) have demonstrated strong generative capability for image restoration tasks. However, existing diffusion-based dehazing methods still suffer from two major limitations: unstable coarse-to-fine restoration under severe haze conditions and stochastic color inconsistency during iterative denoising, which often manifests as color shift artifacts. To address the first issue, we propose a progressive diffusion framework for single image dehazing, which gradually increases image resolution during training and inference to enable stable coarse-to-fine restoration. We further design a progressive backbone architecture that evolves with the input scale, allowing efficient stage-wise feature learning across different resolutions. To stabilize the progressive optimization process, a fade-in strategy is introduced to smoothly transition between adjacent restoration stages. To alleviate the second issue, we introduce a consistency-preserving group-level feature statistics module that regularizes feature distribution consistency during diffusion sampling, thereby reducing stochastic color deviations and improving global color fidelity. Extensive experiments on both synthetic and real-world datasets demonstrate that the proposed PD-Dehaze effectively restores scene structures and fine textures while significantly improving perceptual color consistency, achieving highly competitive performance across the evaluated benchmarks. The code of PD-Dehaze is available at https://anonymous.4open.science/r/PD-Dehaze.