Spatially Adaptive Frequency-Decoupled Diffusion for Image Dehazing

Abstract

Single-image dehazing remains challenging when haze degradation and scene structures vary spatially, while diffusion-based restoration is computationally expensive when iterative denoising is performed at full resolution. This work proposes a frequency-decoupled conditional diffusion framework built around an adaptive low-pass residual decomposition. A predefined bank of Gaussian low-pass filters is combined through lightweight spatial routing to construct an input-dependent reduced representation and a complementary full-resolution residual. The reduced representation is restored by a conditional diffusion model using implicit sampling, whereas a deterministic high-frequency refinement branch preserves local structures and textures. The two restored pathways are subsequently integrated by a spatial feature fusion module through directional cross-frequency attention and adaptive spatial refinement. This design confines iterative denoising to a compact representation while retaining a full-resolution route for residual information. Experiments on synthetic and real-world benchmarks demonstrate consistent improvements over representative prior-based, deterministic, and diffusion-based methods. Ablation and sampling analyses further confirm the effectiveness of spatially adaptive decomposition, high-frequency refinement, cross-frequency fusion, and the resulting quality–efficiency trade-off, supporting the applicability of the proposed framework to practical image dehazing scenarios.