Single image dehazing is a key but really tough task in computer vision. Especially in dense and uneven haze scenes, existing methods usually can't recover complete structures, clear details, and normal colors. To solve these problems, we propose a two-stage patch-conditioned diffusion model (TS-PCDM). It first builds the global image structure for a stable basic result, then refines local textures and edges under this guidance. We also use overlapping patch inference and EMA to reduce artifacts and make the model more stable. Experiments on three real-world haze datasets show our model clearly outperforms current top methods in PSNR and SSIM. It removes haze effectively, restores rich details without color shifts, and proves to be much better than existing approaches.
Kanghui DU (Thu,) studied this question.