Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired Training

TL;DR

This work tackles real-world image dehazing under unpaired training by embedding diffusion priors into a CycleGAN framework, yielding Diff-Dehazer. It introduces Text-Aware Guidance and Physics-Aware Guidance to leverage high-level semantics and physical priors (DCP/BCCR), plus a Perceptual Fusion Model to fuse priors and a physical loss based on the Atmospheric Scattering Model. Areal real-world hazy/clear dataset is leveraged for training and evaluation, with extensive experiments showing superior performance on RTTS, Haze2020, and OHAZE compared to state-of-the-art methods. The approach demonstrates improved generalization to diverse real-world hazy scenes and offers a practical, multi-modal dehazing solution, albeit with potential diffusion-noise-induced variability in extreme haze.

Abstract

Unpaired training has been verified as one of the most effective paradigms for real scene dehazing by learning from unpaired real-world hazy and clear images. Although numerous studies have been proposed, current methods demonstrate limited generalization for various real scenes due to limited feature representation and insufficient use of real-world prior. Inspired by the strong generative capabilities of diffusion models in producing both hazy and clear images, we exploit diffusion prior for real-world image dehazing, and propose an unpaired framework named Diff-Dehazer. Specifically, we leverage diffusion prior as bijective mapping learners within the CycleGAN, a classic unpaired learning framework. Considering that physical priors contain pivotal statistics information of real-world data, we further excavate real-world knowledge by integrating physical priors into our framework. Furthermore, we introduce a new perspective for adequately leveraging the representation ability of diffusion models by removing degradation in image and text modalities, so as to improve the dehazing effect. Extensive experiments on multiple real-world datasets demonstrate the superior performance of our method. Our code https://github.com/ywxjm/Diff-Dehazer.

Figures (8)

• Figure 1: (a) Previous dehazing methods with paired training. (b) Previous CycleGAN-based dehazing methods with unpaired training. (c) Existing stable diffusion-based dehazing methods with paired training. (d) Our stable diffusion-based dehazing method with unpaired training.
• Figure 2: Overview of our method. Black arrows represent the training process and blue arrows denote the inference process.
• Figure 3: Orange Area: Backbone network of our framework. Note that hazing and dehazing backbones share the same U-Net and employ two individual VAEs. Green Area: Text-Aware Guidance (TAG). Blue Area: Physics-Aware Guidance (PAG). Dark Channel Prior (DCP) and Boundary Constraint and Contextual Regularization (BCCR) are two physical prior-based dehazing methods. ASM is the Atmospheric Scattering Model. Gray Area: The structure of Perceptual Fusion Model (PFM).
• Figure 4: Visual comparison of samples from Haze2020 and RTTS. Our method can effectively remove haze and generate high-quality images with natural color and realistic contrast. More visual results are presented in our supplementary materials.
• Figure 5: Visual comparison of samples from Haze2020. Areas where our method works better are boxed out and zoomed in. Our method can generate clear images with high fidelity and discriminative textures.