PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing

Chengyu Fang; Chunming He; Yuelin Zhang; Chubin Chen; Chenyang Zhu; Hongqiu Wang; Longxiang Tang; Xiu Li

cs.CVApr 8, 2026

PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing

Authors: Chengyu Fang, Chunming He, Yuelin Zhang, Chubin Chen, Chenyang Zhu, Hongqiu Wang, Longxiang Tang, Xiu Li,

Organizations: Tsinghua University · Duke University · CUHK · HKUST(GZ) ·

Abstract

Real-world image dehazing (RID) aims to remove haze-induced degradation from real scenes. This task remains challenging due to non-uniform haze distribution, spatially varying color shifts, and the scarcity of paired real hazy-clean data. In PRISM, we propose Proximal Scattering Atmosphere Reconstruction (PSAR), a physically structured framework that jointly reconstructs the clear scene and scattering variables under the atmospheric scattering model, making the restoration process more interpretable in complex real-world conditions. To bridge the synthetic-to-real gap, we design an online non-uniform haze synthesis pipeline and a Selective Self-Distillation Adaptation (SSDA) scheme for unpaired real-world scenarios, which enables the model to selectively learn from high-quality perceptual targets while leveraging its intrinsic scattering understanding to audit residual haze and guide self-refinement. Experiments on real-world benchmarks demonstrate that PRISM achieves competitive performance on RID tasks.

Figures & tables

Figure 1: Our proposed PRISM framework enables effective non-uniform haze removal, producing results with more realistic colors and fewer artifacts than prior methods.

Figure 2: Architecture of PSAR at the $k$ -th stage. Given the previous estimates $(J_{k-1},T_{k-1},A_{k-1})$ , PSAR computes proximal updates for atmospheric light, transmission, and radiance under the same ASM objective, followed by lightweight refinements for $T$ and $J$ .

Figure 3: Pipeline of Selective Self-Distillation Adaptation (SSDA). QGSD selects teacher pseudo-labels only when they outperform the student under all selected perceptual metrics, while PGSR uses the teacher to audit residual haze in the student’s output through estimated scattering variables.

Method	Venue	RTTS [ 44 ]			Fattal’s [ 5 ]
Method	Venue	FADE $\downarrow$	PAQ2PIQ $\uparrow$	CLIPIQA $\uparrow$	FADE $\downarrow$	PAQ2PIQ $\uparrow$	CLIPIQA $\uparrow$
Hazy Inputs	—	2.484	66.05	0.388	1.061	71.54	0.506
PDN [ 1 ]	ECCV 2018	0.876	67.11	0.319	0.316	73.73	0.557
MSBDN [ 10 ]	CVPR 2020	1.363	66.82	0.356	0.557	72.79	0.520
Dehamer [ 16 ]	CVPR 2022	1.895	66.70	0.365	0.667	72.73	0.496
DAD [ 34 ]	CVPR 2020	1.130	66.93	0.254	0.484	71.56	0.413

Table 1: Quantitative results on RTTS [ 44 ] and Fattal’s [ 5 ] datasets. The best result of each metric is marked in red , and the second-best result is marked in blue .

Figure 4: Visual comparison on the RTTS dataset [ 44 ] and Fattal’s dataset [ 5 ] .

Table 6

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Metric	AOD-Net [ 8 ]	MSBDN [ 10 ]	GCANet [ 51 ]	FFANet [ 15 ]	DehazeFormer [ 17 ]
NH-HAZE [ 47 ]	PSNR $\uparrow$	15.31	17.34	16.64	18.48	18.15
NH-HAZE [ 47 ]	SSIM $\uparrow$	0.458	0.556	0.558	0.618	0.607
Dataset	Metric	DehazeNet [ 7 ]	GridDehazeNet [ 9 ]	UHD [ 52 ]	DehazeDDPM [ 53 ]	PSAR
NH-HAZE [ 47 ]	PSNR $\uparrow$	11.76	17.23	16.05	19.44	20.10
NH-HAZE [ 47 ]	SSIM $\uparrow$	0.399	0.504	0.461	0.627	0.688

Appendix

Table 5: Quantitative results on the paired real-world dataset NH-HAZE [ 47 ] .

Metric	Hazy Inputs	PDN [ 1 ]	MSBDN [ 10 ]	Dehamer [ 16 ]	DAD [ 34 ]	PSD [ 23 ]
NIMA $\uparrow$	4.857	4.880	4.896	4.898	4.778	4.783
MUSIQ $\uparrow$	53.77	52.97	54.04	53.78	49.92	56.52
ManIQA $\uparrow$	0.311	0.280	0.306	0.316	0.225	0.296
LIQE $\uparrow$	1.982	1.959	2.108	2.037	1.724	1.693
Metric	D4 [ 33 ]	RIDCP [ 29 ]	CORUN [ 2 ]	CoA [ 45 ]	HazeFlow [ 31 ]	PRISM
NIMA $\uparrow$	4.852	5.241	5.301	5.059	5.296	5.347

Appendix

Table 6: Supplementary quantitative results on RTTS [ 44 ] dataset using additional IQA metrics.

Datasets	Metrics	S.R. $A_{\textrm{rend}}$	S.R. $A_{\textrm{rend}}+\ell_{1}$	$A_{\textrm{rend}}+\ell_{1}$	PSAR	Fixed Teacher	w/o EMA	PRISM
RTTS	FADE $\downarrow$	0.918	0.920	0.894	0.869	0.583	0.526	0.470
	PAQ2PIQ $\uparrow$	69.63	69.11	70.18	70.43	72.67	73.12	74.05
	CLIPIQA $\uparrow$	0.388	0.376	0.396	0.436	0.438	0.447	0.456

Appendix

Table 7: Additional ablations on atmospheric-light rendering and EMA teacher update.

Figure 5: Visualization of different training phases.

Figure 6: Failure cases of our proposed PSAR.

Figure 7: Visualization of the initialized inputs and the decomposed components estimated by the proposed PRISM, including atmospheric light $A$ , transmission $T$ , and restored scene radiance $J$ .

Figure 8: More visualization results of our PRISM and other state-of-the-art methods.

Image Dehazing

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

See all 35 papers

RTE-FM-Dehazer: Radiative Transfer Equation Inspired Flow Matching for Real-World Image Dehazing
Jul 2, 2026Chenfeng Wei, Chun Wang, Boyang Zhao +3Flow Matching Radiative Transfer
Xi’an Jiaotong-Liverpool University, China · University of Liverpool, UK · Mashang Consumer Finance Co. Ltd, China +3
Hihaze: One-Step Diffusion for High-Resolution Real-World Image Dehazing
Oct 19, 2026Zheng Chen, Yuchen Huang, Xiaoyang Liu +3Image Dehazing Diffusion Models for Image Restoration
Shanghai Jiao Tong University
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

Jul 2, 2026J/K move · Enter open · S save

RTE-FM-Dehazer: Radiative Transfer Equation Inspired Flow Matching for Real-World Image Dehazing

Chenfeng Wei, Chun Wang, Boyang Zhao, Si Zuo+2

Xi’an Jiaotong-Liverpool University, China · University of Liverpool, UK · Mashang Consumer Finance Co. Ltd, China+3

Flow Matching Radiative Transfer Image Dehazing

Open paper AskPDF arXiv

Single-image dehazing aims to recover a clear scene from a hazy image and is generally formulated as an image-to-image translation task; however, it faces two limitations. Its performance depends heavily on the haze-formation priors embedded in the model. Prevailing methods adopt the Atmospheric Scattering Model (ASM), whose assumptions of single scattering and homogeneous media are often violated, leading to residual haze and color drift. Moreover, large-scale real hazy/clear pairs are impractical to collect, and existing synthesis approaches fail to reproduce the full complexity of natural haze. To address these issues, we present RTE-FM-Dehazer, a novel dehazing approach, together with a scalable data pipeline. Unlike the ASM, the Radiative Transfer Equation (RTE) jointly accounts for both scattering and absorption, naturally accommodating the non-homogeneous, multiple-scattering media that characterize real hazy scenes. Motivated by the structural similarity between the RTE diffusion-absorption term and the ODE in flow matching, we introduce a diffusion-absorption regularizer derived from a reduced RTE, to steer the flow matching trajectory at each step. Next, leveraging modern vision-language models, we build an automated pipeline and release P-HAZE, a dataset of 50000 realistic hazy/clear pairs. Extensive evaluations demonstrate that RTE-FM-Dehazer, trained solely on P-HAZE, effectively eliminates artifacts like residual haze and color drift, exhibits strong cross-domain generalization, and achieves leading results on five real-world dehazing benchmarks.

Abstract

Image Dehazing

Explore similar work

RTE-FM-Dehazer: Radiative Transfer Equation Inspired Flow Matching for Real-World Image Dehazing

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

HNDiff: Haze-Noise Diffusion for Image Dehazing