Efficient image dehazing via temporal-aware diffusion

Dense haze image dehazing constitutes an ill-posed inverse problem due to the severe degradation of original image information caused by complex atmospheric scattering effects. While existing approaches have demonstrated progress, they exhibit persistent limitations in preserving structural details and chromatic fidelity under dense haze conditions. Recent developments in latent diffusion models have opened new possibilities for image restoration through their strong generative capacities, yet current implementations face efficiency bottlenecks in the diffusion process. However, achieving efficient diffusion remains challenging.In this work, we propose Efficient Image Dehazing via Temporal-Aware Diffusion. Specifically, we establish a temporal information-guided residual encoding to generate more robust conditional guidance, enabling significant reduction of Markov chain length. Additionally, we design a Temporal-Aware Dynamic Convolution Block that serves two primary purposes: adapting to shorter diffusion steps through temporal information interference and parsing haze concentration distribution from information guidance. Finally, we propose an offline Backtracking Diffusion Sampling approach that refines the transfer pathway by iteratively backtracking the diffusion steps. This enables us to achieve effective diffusion-based dehazing within 15 steps.Extensive experiments demonstrate that our method achieves SOTA performance on both synthetic and real-world haze datasets. Our code links to https://github.com/fatsotiger/E_Diff_dehaze.