Source-linked AI summary
Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing
Ruiyi Wang, Yushuo Zheng, Zicheng Zhang, Chunyi Li, Shuaicheng Liu, Guangtao Zhai, Xiaohong Liu
TL;DR
Real-world dehazing methods often depend on pre-trained models and training data, while diffusion-based restoration of dense haze is limited in part by lengthy sampling. This paper pairs HazeGen, which generates realistic hazy training images, with DiffDehaze and its accelerated, fidelity-preserving sampling process; the authors report superior performance over existing methods.
Problem
Real-world dehazing methods depend heavily on pre-trained models and data, while limited paired real-world images and synthetic-data reliance challenge performance on real hazy scenes.
Source: §1 · pp. 1–2Method
HazeGen uses a pre-trained text-to-image diffusion prior with hybrid training and blended sampling to generate hazy training data for DiffDehaze, which uses AccSamp and AlignOp for accelerated, fidelity-preserving dehazing.
Source: §1 · p. 2Results
Takeaways & Limitations
The pipeline uses realistic generated hazy images to train DiffDehaze and uses AlignOp-based estimates to reduce sampling steps while enhancing fidelity.
Source: §1 · p. 2, §5 · p. 8The method assumes that adding Gaussian noise at timestep ω to its dehazing estimate yields a distribution close to that of the true clean image.
Source: §3.3 · p. 5
Abstract
from arXiv · showhide
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.
1. Introduction
Real-world dehazing is hindered by limited paired data and methods that depend on pre-trained models, while dense haze challenges conventional restoration and diffusion sampling can be lengthy. The proposed pipeline combines realistic hazy-image generation with accelerated, fidelity-preserving diffusion dehazing.
1. Introduction
- Limited real-world hazy/clean pairs and reliance on synthetic training data contribute to performance degradation on real-world hazy scenes.Hazy images also exhibit color distortion and detail loss, impairing downstream vision tasks.Source: §1 · pp. 1–2
- HazeGen uses a pre-trained text-to-image diffusion prior, hybrid conditional and unconditional training, and blended sampling to generate realistic hazy training images.A small fraction of unconditional predictions during sampling can enhance realism and variety.Source: §1 · p. 2
- DiffDehaze trains on HazeGen data and uses AccSamp, whose AlignOp turns an early diffusion prediction into a faithful dehazing estimate so intermediate sampling steps can be skipped.Haze-density-aware fidelity guidance further targets fidelity, especially in lightly hazy regions.Source: §1 · p. 2
- The proposed pipeline achieves superior quantitative and qualitative performance compared with state-of-the-art methods.The overview depicts HazeGen-generated images serving as DiffDehaze training data and AccSamp reducing sampling steps while enhancing fidelity.Source: Figure 1 · p. 1, §1 · p. 2
2. Related Works
Earlier dehazing approaches use handcrafted priors or learned models, but their limitations include difficulty handling real-world diversity and degradation when synthetic training data differs from real haze.
2.1. Single Image Dehazing
- Handcrafted priors and learning-based methods have limitations on diverse real-world imagery, with synthetic-trained models degrading under the real-to-synthetic domain gap.DCP can fail on prominent white objects, while CNN and Transformer methods are described as achieving superior performance on synthetic datasets.Source: §2.1 · p. 2
2.2. Real-World Image Dehazing
- Real-world methods include CycleGAN domain adaptation, prior-based fine-tuning, and inference-time adaptation, but remain dependent on pre-training data or face training and synthesis limitations.CycleGAN frameworks can have complex, unstable training and mode collapse; physical-scattering synthesis can produce unrealistic, homogeneous haze, motivating the paper's exploration of text-to-image diffusion priors.Source: §2.2 · p. 3
3. Methodology
The methodology pairs HazeGen, which generates realistic hazy training images, with DiffDehaze, which uses diffusion for dehazing. HazeGen combines hybrid training and blended sampling, while DiffDehaze uses AccSamp to improve sampling efficiency and fidelity.
3. Methodology
- HazeGen and DiffDehaze use a fixed pre-trained Stable Diffusion model with IRControlNet for conditional information injection.HazeGen adds hybrid training and blended sampling; DiffDehaze uses AccSamp.Source: §3 · p. 3
3.1. Preliminary: Diffusion Models
- Stable Diffusion performs diffusion and denoising in a low-dimensional autoencoder latent space; HazeGen models P(x|y), while DiffDehaze models P(y|x).Source: §3.1 · p. 3
3.2. Realistic Hazy Image Generation Framework
- HazeGen’s hybrid objective combines conditional training on synthetic image pairs with unconditional training on unlabeled real-world hazy images.The conditional objective builds content relationships, while the unconditional objective helps retain realistic haze generation and prevent catastrophic forgetting.Source: §3.2 · p. 3
- Blended sampling mixes unconditional predictions into conditional noise predictions, compensating for learned defects and improving generated-image diversity.The authors distinguish its motivation from classifier-free guidance: mitigating low-quality training data and enhancing diversity.Source: §3.2 · p. 4
- HazeGen produces more realistic, visually consistent haze than OTS and RIDCP’s degradation pipeline in the visual comparison.OTS images show mild haze, while RIDCP images show abrupt nearby-object haze changes and artifacts attributed to inaccurate depth estimation.Source: §3.2 · p. 4
3.3. Diffusion-based Dehazing Framework
- AlignOp matches local patch statistics between the hazy input and an early diffusion prediction to produce a coarse dehazing estimate that retains input details.It aligns local means and variances, then assembles the patches and averages overlapping regions.Source: §3.3 · pp. 4–5
- AccSamp skips intermediate steps between its dehazing-estimate generation stage, from T to τ, and its guided-refinement stage, the final ω steps.This reduces sampling steps; refinement adds details under density-aware fidelity guidance.Source: §3.3 · p. 4, Figure 3 · p. 5
- During refinement, density-aware guidance emphasizes input fidelity in lightly hazy regions and generated content in densely hazy regions.A rough DCP transmission map serves as an inverse haze-density weighting map.Source: §3.3 · p. 5
4. Experiments
Experiments report strong quantitative and visual dehazing results, and ablations examine the HazeGen training data and AccSamp components. The evaluation also exposes a limitation of FADE as a dehazing metric.
4.1. Experiment Settings
- HazeGen was trained with about 4,800 real-world hazy images and RIDCP synthetic pairs, then generated 100,000 realistic hazy images for DiffDehaze training.RTTS, containing 4,322 images, was used for evaluation; Fattal’s dataset contains 31 classical hazy images.Source: §4.1 · p. 6
4.2. Comparison with State-of-the-Art Methods
- 23.8% improvement on Q-Align, with leading performance across the other reported metrics except FADE, is reported on RTTS.The evaluation uses no-reference metrics because real-world hazy datasets lack ground-truth clean images.Source: §4.2 · p. 6, Table 1 · p. 7
- On Fattal’s dataset, the proposed method removes haze from all regions and restores natural color, while compared methods often leave residual haze.The RTTS comparisons also report recovery of vivid details such as trees in the fifth row.Source: §4.2 · p. 7, Figure 6 · p. 7
4.3. Ablation Studies
- HazeGen ablations report that hybrid training and blended sampling are each essential for optimal performance, and all three tested models improve with HazeGen-generated training data.The three models are MSBDN, NAFNet, and DiffDehaze.Source: §4.3 · p. 7, Table 2 · p. 7, Figure 7 · p. 7
- AccSamp’s AlignOp and adaptive fidelity guidance improve sampling fidelity while preserving dehazing; the selected settings are τ = 800, ω = 600, and s = 0.1.The ablations assess AccSamp without AlignOp or haze-density weighting, and the sampling comparison is against spaced DDPM.Source: §4.3 · p. 8, Table 3 · p. 8, Figure 8 · p. 8
- FADE is insensitive to color-distorted haze residuals and can rate visually flawed blue-shifted or unclear outputs better than the proposed method.The paper illustrates these failure cases in Figure 10.Source: §4.3 · p. 8, Figure 10 · p. 8
5. Conclusion
The conclusion presents HazeGen and DiffDehaze as a controlled-Stable-Diffusion pipeline for realistic hazy-image generation and dehazing. It reports superior performance, while noting metric reliability and further fidelity improvement as limitations.
5. Conclusion
- HazeGen uses generative diffusion priors with hybrid training and blended sampling to generate realistic hazy images for DiffDehaze training.The frameworks build upon controlled Stable Diffusion models.Source: §5 · p. 8
- AccSamp uses fast dehazing estimates from AlignOp to reduce sampling steps and enhance fidelity for DiffDehaze.Source: §5 · p. 8
- The paper reports superior overall performance but identifies the need for more reliable dehazing metrics and further AccSamp fidelity improvement.Source: §5 · p. 8