Hihaze: One-Step Diffusion for High-Resolution Real-World Image Dehazing

Zheng Chen; Yuchen Huang; Xiaoyang Liu; Jue Gong; Linghe Kong; Yulun Zhang

cs.CVOct 19, 2026

Hihaze: One-Step Diffusion for High-Resolution Real-World Image Dehazing

Authors: Zheng Chen, Yuchen Huang, Xiaoyang Liu, Jue Gong, Linghe Kong, Yulun Zhang

Organizations: Shanghai Jiao Tong University

Abstract

Diffusion models have recently achieved encouraging results in real-world image dehazing. However, restoring high-resolution images remains challenging due to spatially non-uniform haze and inconsistent predictions across local patches. To address these issues, we propose Hihaze, a one-step diffusion framework for high-resolution real-world dehazing. Hihaze adopts two haze-aware modulation modules: attentive skip compensation (ASC) and a spatial degradation adapter (SDA). ASC selectively transfers reliable encoder information to the variational autoencoder (VAE) decoder, preserving structural details while suppressing haze-degraded features. SDA extracts multi-scale spatial degradation representations and injects them into the U-Net to guide restoration according to local haze variations. Furthermore, we design a training-free global-guided patch fusion (GPF) method. It uses a global prediction to align patch appearance and downweight unreliable predictions, followed by structure-guided detail compensation. Notably, this training-free design also makes GPF a plug-and-play module for other dehazing models. Experiments across four real-world benchmarks show that Hihaze surpasses state-of-the-art approaches in both visual quality and inference speed. Specifically, it delivers PSNR improvements of 3.62~~dB on Dense-Haze and 3.69~~dB on NH-HAZE compared with the strongest baselines. Code: https://github.com/zhengchen1999/Hihaze.

Figures & tables

Figure 1: Visual results on NH-HAZE 51 (left), I-HAZE 34 (upper right), and O-HAZE 01 (lower right). Each scene shows the hazy input on the left of the diagonal and the Hihaze result on the right. Bottom: matched details from the marked region of NH-HAZE 51. Hihaze removes haze while recovering scene structures and textures in the restored images.

Figure 2: Overview of Hihaze. SDA guides spatially adaptive denoising, while ASC transfers reliable encoder features to the VAE decoder. GPF combines global and local predictions for high-resolution restoration with consistent appearance and preserved fine details.

Figure 3: Visualization of the first SDA feature channel on NH-HAZE 55 (left) and 14 (right). Both examples use the same color scale; warmer colors indicate stronger feature responses.

Figure 4: Overview of global-guided patch fusion (GPF). A global prediction guides patch alignment and reliability-aware fusion. Structural cues from the hazy input guide detail compensation.

		Non-diffusion			Diffusion / Flow
Dataset	Metric	PTTD	IPC-Dehaze	BiLaLoRA	DiffUIR	FCDM	Diff-Dehazer	DiffDehaze	HazeFlow	Ours
NH-HAZE	PSNR $\uparrow$	14.53	11.18	12.07	11.79	14.04	12.99	14.26	14.38	18.22
	SSIM $\uparrow$	0.6119	0.5850	0.5989	0.4788	0.5955	0.6079	0.3067	0.6343	0.6664
	LPIPS $\downarrow$	0.3934	0.4445	0.4398	0.5424	0.4264	0.3936	0.6763	0.4180	0.2863
	DISTS $\downarrow$	0.1938	0.2400	0.2106	0.3462	0.2159	0.2094	0.3374	0.2021	0.1272
	FID $\downarrow$	187.93	200.18	199.34	215.83	198.70	151.51	170.95	162.45	90.10

Table 1: Dehazing performance of Hihaze and competing methods across four real-world benchmarks. For each metric, red denotes first place and blue denotes second place.

Panel of Figure 5 — Figure 5: Visual comparisons on O-HAZE, I-HAZE, Dense-Haze, and NH-HAZE, from top to bottom. Red boxes mark the corresponding regions shown in the enlarged views.

Metric	IPC-Dehaze	DiffUIR	FCDM	Diff-Dehazer	DiffDehaze	HazeFlow	Ours
Runtime (s)	185.4	43.5	146.6	32.5	1,046.8	15.2	12.9
FLOPs (T)	826.391	236.731	2,022.874	453.871	9,386.778	250.991	458.955
Params (M)	43.164	36.261	155.334	1,506.159	1,666.750	87.087	975.822

Table 2: Inference time and model complexity for 3,328

\times

4,096 inputs on an RTX A6000.

Table 3: Ablation studies and GPF transferability on O-HAZE. (a) Base denotes the model without skip connections; Zero-conv uses zero-conv skip connections. (b) Base denotes the model with ASC. (c) AA: appearance alignment; RF: reliability-aware fusion; DC: detail compensation. (d) Performance comparisons of HazeFlow and Diff-Dehazer with and without GPF.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Configuration	Backbone and ASC	SDA encoder	SDA adaptation branches
Training iterations	$200{,}000$	$10{,}000$	$10{,}000$
Crop size	$512\times 512$	$512\times 512$	$1{,}024\times 1{,}024$
Total batch size	4	1	4
Learning rate	$5\times 10^{-5}$	$2\times 10^{-4}$	$10^{-4}$
Weight decay	$0.01$	$10^{-4}$	$0.01$
Learning-rate schedule	Constant	Cosine to $10\%$ of peak	Cosine to zero

Appendix

Table 4: Training configurations for Hihaze.

		Non-diffusion					Diffusion / Flow
Dataset	Metric	DehazeFormer	RIDCP	CORUN	CoA	DehazeSB	Ours
I-HAZE	PSNR $\uparrow$	16.38	15.63	17.10	13.82	16.04	19.02
	SSIM $\uparrow$	0.7835	0.7116	0.7658	0.4888	0.6135	0.7888
	LPIPS $\downarrow$	0.3360	0.4158	0.4093	0.6303	0.5019	0.3217
	DISTS $\downarrow$	0.2188	0.1645	0.2142	0.2512	0.1783	0.0946
	FID $\downarrow$	135.46	204.88	136.26	214.98	202.65	78.99

Appendix

Table 5: Additional quantitative comparisons on I-HAZE, O-HAZE, Dense-Haze, and NH-HAZE. The best and second-best values within this table are colored in red and blue , respectively.

Dataset	Setting	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	FID $\downarrow$	Q-Align $\uparrow$
O-HAZE	Raw patch	16.63	0.6484	0.4291	0.2054	152.20	3.3200
O-HAZE	+ GPF	17.95	0.6782	0.4295	0.1732	133.54	3.4142
NH-HAZE	Raw patch	14.38	0.6343	0.4180	0.2021	162.45	3.1193
NH-HAZE	+ GPF	14.72	0.6396	0.4105	0.1901	157.43	3.0556
Dense-Haze	Raw patch	12.10	0.5320	0.6374	0.3834	236.11	2.4468
Dense-Haze	+ GPF	12.57	0.5416	0.6430	0.3828	225.49	2.4267

Appendix

Table 6: Results of applying GPF to HazeFlow on all four benchmarks. The best results for each metric within each dataset are highlighted in bold.

Dataset	Setting	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	FID $\downarrow$	Q-Align $\uparrow$
O-HAZE	Raw patch	16.62	0.6332	0.4085	0.1935	144.15	3.2273
O-HAZE	+ GPF	16.78	0.6456	0.4089	0.1783	109.77	3.5442
NH-HAZE	Raw patch	12.99	0.6079	0.3936	0.2094	151.51	3.4304
NH-HAZE	+ GPF	13.11	0.6156	0.3841	0.1981	145.86	3.4129
Dense-Haze	Raw patch	10.94	0.4656	0.7149	0.5034	256.24	2.1314
Dense-Haze	+ GPF	11.25	0.4704	0.6946	0.4834	244.70	2.1626

Appendix

Table 7: Results of applying GPF to Diff-Dehazer on all four benchmarks. The best results for each metric within each dataset are highlighted in bold.

Panel of Figure 6 — Figure 6: Visual comparison of HazeFlow before and after applying GPF on O-HAZE. Rows show samples 03, 15, 17, 21, and 28, from top to bottom. Insets show matched crops.

Panel of Figure 7 — Figure 7: Visual comparison of Diff-Dehazer before and after applying GPF on O-HAZE. Rows show samples 03, 07, 21, 24, and 33, from top to bottom. Insets show matched crops.

Panel of Figure 8 — Figure 8: Full-image qualitative comparisons on O-HAZE.

Panel of Figure 9 — Figure 9: Full-image qualitative comparisons on I-HAZE.

Panel of Figure 10 — Figure 10: Full-image qualitative comparisons on Dense-Haze.

Panel of Figure 11 — Figure 11: Full-image qualitative comparisons on NH-HAZE.

Panel of Figure 12 — Figure 12: SDA feature visualizations on NH-HAZE and O-HAZE. Each pair shows the hazy input (left) and the corresponding learned feature map (right).

Image Dehazing

Image dehazing restores clear scene details and colors from images obscured by haze, fog, or smoke.

See all 35 papers

HNDiff: Haze-Noise Diffusion for Image Dehazing
Aug 11, 2026Jin-Ting He, Fu-Jen Tsai, Yan-Tsung Peng +3Physics-Informed Diffusion Models Diffusion Models for Image Restoration
Department of Computer Science, National Yang Ming Chiao Tung University, Taiwan · National Tsing Hua University, Taiwan · National Chengchi University, Taiwan +1
Efficient Real-World Dehazing via Physics-Inspired Global-Local Decoupling
Jun 24, 2026Yifei Qu, Ru Li, Junjie Chen +1Physics-Guided Image Restoration Image Restoration
Faculty of Computing, Harbin Institute of Technology, Harbin 150006, China
ZID-Net: Zero-Inference Diffusion Prior Decoupling Network for Single Image Dehazing
Apr 26, 2026Xinheng Li, Minghao Chen, Mengqing Wu +2Diffusion Models for Image Restoration Diffusion Model Distillation
College of Information Science and Engineering, Hohai University, Changzhou, 213200, Jiangsu, China

Aug 11, 2026J/K move · Enter open · S save

HNDiff: Haze-Noise Diffusion for Image Dehazing

Jin-Ting He, Fu-Jen Tsai, Yan-Tsung Peng, Min-Hung Chen+2

Department of Computer Science, National Yang Ming Chiao Tung University, Taiwan · National Tsing Hua University, Taiwan · National Chengchi University, Taiwan+1

Physics-Informed Diffusion Models Diffusion Models for Image Restoration Physics-Informed ML Image Dehazing

Open paper AskPDF arXiv

Existing diffusion-based methods have recently made significant progress in image dehazing. However, they typically neglect the physics of haze formation and reconstruct clean images from pure Gaussian noise, thereby limiting their restoration potential. To address this issue, we propose Haze-Noise Diffusion (HNDiff), a novel diffusion framework that embeds the atmospheric scattering model as an inductive bias. By grounding diffusion in physical principles, HNDiff ensures that the restoration aligns more closely with underlying mechanisms of haze formation. In its forward process, we introduce joint haze-noise diffusion with a haze-aware noise scheduler, which progressively adds both haze and noise to an image. Essentially, the scheduler adapts noise levels according to haze density, meaning that regions with heavier haze receive stronger noise injection to encourage content generation, while clearer regions receive lighter noise to better preserve details, which directly links the forward degradation process with the physics of haze. In the reverse process, we then derive a physically consistent dehazing-denoising process that simultaneously removes haze and noise to restore a clean image in a manner aligned with the forward degradation process. To further enhance practicality, we propose Latent HNDiff, which compiles clean latent priors that can be seamlessly integrated into existing dehazing networks to boost performance. Extensive experiments show that our work significantly improves leading dehazing backbones and achieves state-of-the-art results on benchmark datasets. The project page is available at https://jin-ting-he.github.io/HNDiff .

Abstract

Image Dehazing

Explore similar work

HNDiff: Haze-Noise Diffusion for Image Dehazing

Efficient Real-World Dehazing via Physics-Inspired Global-Local Decoupling

ZID-Net: Zero-Inference Diffusion Prior Decoupling Network for Single Image Dehazing