On the Global Photometric Alignment for Low-Level Vision

Mingjia Li; Tianle Du; Hainuo Wang; Qiming Hu; Xiaojie Guo

cs.CVApr 9, 2026

On the Global Photometric Alignment for Low-Level Vision

Authors: Mingjia Li, Tianle Du, Hainuo Wang, Qiming Hu, Xiaojie Guo

Organizations: Tianjin University

Abstract

Supervised low-level vision models rely on pixel-wise losses against paired references, yet paired training sets exhibit per-pair photometric inconsistency, say, different image pairs demand different global brightness, color, or white-balance mappings. This inconsistency enters through task-intrinsic photometric transfer (e.g., low-light enhancement) or unintended acquisition shifts (e.g., de-raining), and in either case causes an optimization pathology. Standard reconstruction losses allocate disproportionate gradient budget to conflicting per-pair photometric targets, crowding out content restoration. In this paper, we investigate this issue and prove that, under least-squares decomposition, the photometric and structural components of the prediction-target residual are orthogonal, and that the spatially dense photometric component dominates the gradient energy. Motivated by this analysis, we propose Photometric Alignment Loss (PAL). This flexible supervision objective discounts nuisance photometric discrepancy via closed-form affine color alignment while preserving restoration-relevant supervision, requiring only covariance statistics and tiny matrix inversion with negligible overhead. Across 6 tasks, 16 datasets, and 16 architectures, PAL consistently improves metrics and generalization. The implementation is in the appendix.

Figures & tables

Figure 1: An Illustration of our work. (a) We identified that there exist inconsistent global photometric shift across the paired training datasets. (b) As the photometric shift dominates the gradient, learning texture and structure content from the training data becomes difficult. Our PAL helps rebalance the gradient. (c) Equipped with our PAL, the performance gained across 6 tasks, 16 methods, and 16 datasets, with an average gain in PSNR of 0.45dB across 6 tasks.

Panel of Figure 2 — Figure 2: Comparison of alignment families applied to Low-light image pairs. The optimal transforms are computed in closed form and applied. Channel-wise mean (b) and Optimal Scalar (c) cannot correct color-temperature shifts . Optimal Diagonal (d) handles per-channel gain but not cross-channel coupling. Only PAL’s full affine model (e) closely matches the reference (f).

Figure 3: Per-pair photometric scatter plots. Each point represents one training pair, with per-channel (R/G/B) input mean on the $x$ -axis, and GT mean on the $y$ -axis. In both LOL-v2 and RESIDE-SOTS cases, the wide per-pair spread means a pixel-wise loss receives conflicting photometric supervision.

Figure 4: (Left) Decomposed photometric/content error on validation set. (Right) Gradient ratio $\rho$ . Plots are sampled from a Retinexformer trained on LOL-V1 every 1000 steps.

Panel of Figure 5 — Figure 5: Qualitative comparisons on LLIE (CIDNet on LOLv2-real). PAL produces more natural colors.

Panel of Figure 6 — Figure 6: Qualitative comparisons on nighttime dehazing (NAFNet on NHR). PAL reduces residual haze and color cast.

Panel of Figure 7 — Figure 7: Qualitative comparison on MITNet image dehazing. We use error map to highlight the difference. Darker is better. PAL reduces residual color cast relative to the baseline.

Panel of Figure 8 — Figure 8: Qualitative comparison on underwater image enhancement (EUVP). PAL produces outputs with more natural color and fewer artifacts.

Methods	LOLv1					LOLv2-syn					LOLv2-real
Methods	PSNR	SSIM	LPIPS	IQA	IAA	PSNR	SSIM	LPIPS	IQA	IAA	PSNR	SSIM	LPIPS	IQA	IAA
MIRNet ( Zamir et al., 2020 )	20.57	0.769	0.254	2.914	1.855	21.74	0.877	0.138	2.650	2.023	21.28	0.791	0.355	2.256	1.377
+PAL (Ours)	21.01	0.791	0.228	2.923	1.729	22.13	0.890	0.117	2.870	2.091	22.32	0.816	0.348	2.535	1.469
Uformer ( Wang et al., 2022 )	18.85	0.751	0.288	2.751	1.661	21.50	0.884	0.120	2.919	2.058	19.80	0.714	0.346	2.135	1.387
+PAL (Ours)	19.31	0.767	0.247	2.858	1.666	21.79	0.890	0.112	2.952	2.063	20.12	0.738	0.327	2.780	1.649
Retinexformer ( Cai et al., 2023 )	23.40	0.822	0.269	3.148	1.980	25.48	0.930	0.101	2.404	2.096	21.69	0.846	0.276	3.163	1.962

Table 1: Quantitative comparison on LOL, across PSNR (

\uparrow

), SSIM (

\uparrow

), LPIPS (

\downarrow

), IQA (

\uparrow

) and IAA (

\uparrow

Method	Loss	EUVP
Method	Loss	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Shallow-UWnet	Baseline	19.70	0.780	0.355
Shallow-UWnet	+PAL	20.35	0.790	0.327
Boths	Baseline	19.68	0.748	0.367
Boths	+PAL	19.85	0.767	0.375
LiteEnhanceNet	Baseline	20.40	0.779	0.343

Table 2: Underwater enhancement results on EUVP.

Method	Loss	RESIDE-SOTS-Indoor
Method	Loss	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
FocalNet	Baseline	37.97	0.989	0.012
FocalNet	+PAL	38.18	0.991	0.012
MITNet	Baseline	37.56	0.988	0.006
MITNet	+PAL	37.80	0.988	0.005
DehazeXL	Baseline	27.77	0.956	0.030

Table 3: Dehazing results on RESIDE-SOTS-Indoor.

Method	Loss	NHR
Method	Loss	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
NAFNet	Baseline	22.06	0.825	0.082
NAFNet	+PAL	22.91	0.838	0.071
Restormer	Baseline	18.30	0.794	0.118
Restormer	+PAL	18.89	0.816	0.103

Table 4: Nighttime dehazing results on NHR.

Method	Loss	ISTD
Method	Loss	PSNR $\uparrow$	SSIM $\uparrow$	RMSE $\downarrow$
RASM	Baseline	32.32	0.968	4.12
RASM	+PAL	32.65	0.968	4.16
HomoFormer	Baseline	32.02	0.968	4.24
HomoFormer	+PAL	32.49	0.968	4.17

Table 5: Shadow removal results on ISTD.

Method	Loss	Snow100K-S			Snow100K-L			Outdoor			RainDrop
Method	Loss	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Histoformer ( Sun et al., 2024b )	Baseline	37.41	0.965	0.045	32.16	0.926	0.0919	32.08	0.938	0.077	33.06	0.944	0.067
Histoformer ( Sun et al., 2024b )	+PAL	37.85	0.968	0.039	32.34	0.929	0.087	32.82	0.945	0.071	32.77	0.945	0.063
MODEM ( Wang et al., 2025 )	Baseline	38.08	0.967	0.041	32.52	0.929	0.088	33.10	0.941	0.070	33.01	0.943	0.065
MODEM ( Wang et al., 2025 )	+PAL	38.10	0.968	0.039	32.54	0.930	0.085	33.25	0.942	0.065	33.08	0.944	0.060

Table 6: All-in-one restoration results across multiple weather degradation benchmarks.

Panel of Figure 9 — Figure 9: Qualitative comparison on the all-in-one task.

Setup	The choice of $\alpha$ .						The choice of $\epsilon$ .
Setup	0.1	0.5	0.6	0.8	1.0	1.2	0.0001	0.001	0.01	0.1	1
PSNR $\uparrow$	23.07	23.37	23.95	23.41	23.19	23.45	NaN	23.41	23.08	23.17	23.12
SSIM $\uparrow$	0.822	0.840	0.870	0.861	0.857	0.831	NaN	0.861	0.842	0.850	0.834

Table 7: Ablation analysis of the weight

\alpha

for our PAL and the regularization term

\epsilon

for matrix inversion.

Method	DICM		LIME		MEF		NPE		VV
Method	IQA $\uparrow$	IAA $\uparrow$	IQA $\uparrow$	IAA $\uparrow$	IQA $\uparrow$	IAA $\uparrow$	IQA $\uparrow$	IAA $\uparrow$	IQA $\uparrow$	IAA $\uparrow$
MIRNet ( Zamir et al., 2020 )	2.680	1.892	2.409	1.793	2.174	1.557	2.193	1.710	2.580	1.669
+PAL (Ours)	2.875	2.034	2.633	1.939	2.404	1.705	2.496	1.854	2.680	1.740
Uformer ( Wang et al., 2022 )	2.548	1.852	2.499	1.986	2.308	2.017	2.435	1.832	2.567	1.710
+PAL (Ours)	2.640	1.870	2.604	2.026	2.341	2.080	2.699	1.875	2.612	1.712
Retinexformer ( Cai et al., 2023 )	2.869	2.321	3.005	2.528	3.193	2.520	2.785	2.175	2.464	1.923

Table 8: Quantitative results on unpaired datasets across IQA (

\uparrow

) and IAA (

\uparrow

), evaluated using Q-Align.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Methods	LOLv1					LOLv2-syn					LOLv2-real
Methods	PSNR	SSIM	LPIPS	IQA	IAA	PSNR	SSIM	LPIPS	IQA	IAA	PSNR	SSIM	LPIPS	IQA	IAA
MIRNet ( Zamir et al., 2020 )	20.57	0.769	0.254	2.914	1.855	21.74	0.877	0.138	2.650	2.023	21.28	0.791	0.355	2.256	1.377
+GT-Mean Loss	20.79	0.787	0.233	2.892	1.728	21.92	0.888	0.118	2.840	2.084	21.77	0.800	0.365	2.285	1.331
+PAL (Ours)	21.01	0.791	0.228	2.923	1.729	22.13	0.890	0.117	2.870	2.091	22.32	0.816	0.348	2.535	1.469
Uformer ( Wang et al., 2022 )	18.85	0.751	0.288	2.751	1.661	21.50	0.884	0.120	2.919	2.058	19.80	0.714	0.346	2.135	1.387
+GT-Mean Loss	19.01	0.759	0.279	2.728	1.558	21.62	0.894	0.107	2.996	2.104	19.92	0.729	0.331	2.154	1.416

Appendix

Table 9: Direct comparison of Baseline, GT-Mean Loss ( Liao et al., 2025 ) , and PAL (Ours) on LOL datasets. Best and second-best results are in bold and underlined . PAL achieves the best PSNR and SSIM across all backbones and datasets, demonstrating that the full affine alignment consistently outperforms mean-based alignment even on GT-Mean’s home domain (LLIE).

Figure 10: Per-pair photometric analysis across nine datasets spanning four task families. Each plot shows the per-channel (R/G/B) mean intensity of the input vs. ground truth for every training pair, with per-channel linear fits and KDE density contours. The gray dashed line denotes $y{=}x$ (identity). All datasets exhibit per-pair scatter away from any single trajectory, confirming the ubiquity of per-pair photometric inconsistency that injects conflicting supervision into pixel-wise losses (cf. Section 3.1 of the main paper). The separation between per-channel regression lines further demonstrates that the inconsistency is channel-dependent, requiring a full affine color model to discount.

Image Restoration

Image restoration uses computational methods to recover useful visual content from noisy, blurred, low-resolution, or otherwise degraded images.

See all 372 papers

PASDiff: Physics-Aware Semantic Guidance for Joint Real-World Low-Light Face Enhancement and Restoration
Mar 26, 2026Yilin Ni, Wenjie Li, Zhengxue Wang +3Diffusion Model Guidance Low-Light Image Enhancement
College of Automation, Nanjing University of Posts and Telecommunications, Nanjing, China · School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing, China · PCA Lab, School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China +1
Retrieval-Augmented Generation-Based Color Restoration for Low-Light Image Enhancement
Aug 8, 2026Li-Wei Lu, Shaou-Gang MiaouLow-Light Image Enhancement Image Restoration
Department of Electronic Engineering, Chung Yuan Christian University, Taoyuan 320, Taiwan
Geometric Similarity in VLM Low-Level Vision Representations
Oct 1, 2026Shao-Jun Xia, Huixin Zhang, Zhen Lei +3Vision-Language Models Representational Similarity Analysis
Florida State University · UNC Chapel Hill

Mar 26, 2026J/K move · Enter open · S save

PASDiff: Physics-Aware Semantic Guidance for Joint Real-World Low-Light Face Enhancement and Restoration

Yilin Ni, Wenjie Li, Zhengxue Wang, Juncheng Li+2

College of Automation, Nanjing University of Posts and Telecommunications, Nanjing, China · School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing, China · PCA Lab, School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China+1

Diffusion Model Guidance Low-Light Image Enhancement All-in-One Image Restoration Image Restoration

Open paper AskPDF arXiv

Face images captured in real-world low light suffer multiple degradations-low illumination, blur, noise, and low visibility, etc. Existing cascaded solutions often suffer from severe error accumulation, while generic joint models lack explicit facial priors and struggle to resolve clear face structures. In this paper, we propose PASDiff, a Physics-Aware Semantic Diffusion with a training-free manner. To achieve a plausible illumination and color distribution, we leverage inverse intensity weighting and Retinex theory to introduce photometric constraints, thereby reliably recovering visibility and natural chromaticity. To faithfully reconstruct facial details, our Style-Agnostic Structural Injection (SASI) extracts structures from an off-the-shelf facial prior while filtering out its intrinsic photometric biases, seamlessly harmonizing identity features with physical constraints. Furthermore, we construct WildDark-Face, a real-world benchmark of 700 low-light facial images with complex degradations. Extensive experiments demonstrate that PASDiff significantly outperforms existing methods, achieving a superior balance among natural illumination, color recovery, and identity consistency. Code and dataset will be available at https://github.com/IVIPLab/PASDiff.

Abstract

Image Restoration

Explore similar work

PASDiff: Physics-Aware Semantic Guidance for Joint Real-World Low-Light Face Enhancement and Restoration

Retrieval-Augmented Generation-Based Color Restoration for Low-Light Image Enhancement

Geometric Similarity in VLM Low-Level Vision Representations