Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing

TL;DR

The paper tackles real-world hazy image dehazing by bridging data realism and dehazing fidelity through diffusion priors. It introduces HazeGen to generate realistic hazy images using a fixed latent diffusion model and DiffDehaze to recover clean content with an Accelerated Fidelity-Preserving Sampling process (AccSamp) that employs AlignOp to create a faithful dehazing estimate with fewer steps. The key contributions are the hybrid training and blended sampling in HazeGen, the AccSamp framework with two-stage sampling and patch-wise alignment, and extensive experiments showing superior performance on real-world datasets like RTTS and Fattal compared to state-of-the-art methods. This approach reduces reliance on pre-trained models, narrows the synthetic-real domain gap, and enables efficient, high-quality dehazing in real-world scenarios.

Abstract

Existing real-world image dehazing methods primarily attempt to fine-tune pre-trained models or adapt their inference procedures, thus heavily relying on the pre-trained models and associated training data. Moreover, restoring heavily distorted information under dense haze requires generative diffusion models, whose potential in dehazing remains underutilized partly due to their lengthy sampling processes. To address these limitations, we introduce a novel hazing-dehazing pipeline consisting of a Realistic Hazy Image Generation framework (HazeGen) and a Diffusion-based Dehazing framework (DiffDehaze). Specifically, HazeGen harnesses robust generative diffusion priors of real-world hazy images embedded in a pre-trained text-to-image diffusion model. By employing specialized hybrid training and blended sampling strategies, HazeGen produces realistic and diverse hazy images as high-quality training data for DiffDehaze. To alleviate the inefficiency and fidelity concerns associated with diffusion-based methods, DiffDehaze adopts an Accelerated Fidelity-Preserving Sampling process (AccSamp). The core of AccSamp is the Tiled Statistical Alignment Operation (AlignOp), which can provide a clean and faithful dehazing estimate within a small fraction of sampling steps to reduce complexity and enable effective fidelity guidance. Extensive experiments demonstrate the superior dehazing performance and visual quality of our approach over existing methods. The code is available at https://github.com/ruiyi-w/Learning-Hazing-to-Dehazing.

Figures (10)

• Figure 1: Overview of the proposed pipeline. HazeGen utilizes a pre-trained text-to-image diffusion model to generate realistic hazy images, which serve as the training data for DiffDehaze. DiffDehaze adopts an Accelerated Fidelity-Preserving Sampling process (AccSamp) that effectively reduces sampling steps while producing superior dehazing results with enhanced fidelity.
• Figure 2: Visual comparisons between hazy images generated by HazeGen and synthetic images from OTS reside and the phenomenological degradation pipeline of RIDCP ridcp.
• Figure 3: Overview of the AccSamp sampling process. The accelerated sampling process consists of two stages: the dehazing estimate generation stage and the guided refinement stage. In the initial stage (timesteps $T$ to $\tau$), AlignOp transforms a blurry early diffusion prediction into a detailed and faithful dehazing estimate. In the subsequent refinement stage (the final $\omega$ steps), additional vivid details are generated under density-aware fidelity guidance. Intermediate sampling steps between $\tau$ and $\omega$ are skipped to enhance efficiency.
• Figure 4: Visualization of AlignOp’s effect. By aligning the local patch statistics of the hazy image with those of an early diffusion prediction, AlignOp produces a clean and faithful dehazing estimate.
• Figure 5: Visual comparisons on the RTTS dataset reside. Zoomed-in for details.