
Motivation & Challenge: Synthetic hazy-clear pairs poorly represent real haze, and paired real-world data is costly to collect. Diff-Dehazer adopts a pretrained diffusion model as the image prior, and trains unpaired with cycle consistency, physical priors (DCP/BCCR), and textual guidance.
Results & Venues: Outperformed existing unpaired and several fully-supervised dehazing methods on real-world datasets.
Published at AAAI 2025.
Link: https://ojs.aaai.org/index.php/AAAI/article/download/32469/34624
1 like · 0 views · Sep 15, 2026
Motivation & Challenge: Synthetic hazy-clear pairs poorly represent real haze, and paired real-world data is costly to collect.