DINOLight: Robust Ambient Light Normalization with Self-supervised Visual Prior Integration

Youngjin Oh; Junhyeong Kwon; Nam Ik Cho

cs.CVMar 13, 2026

DINOLight: Robust Ambient Light Normalization with Self-supervised Visual Prior Integration

Authors: Youngjin Oh, Junhyeong Kwon, Nam Ik Cho

Organizations: Department of ECE, INMC, Seoul National University, Seoul, Korea

Abstract

This paper presents a new ambient light normalization framework, DINOLight, that integrates the self-supervised model DINOv2's image understanding capability into the restoration process as a visual prior. Ambient light normalization aims to restore images degraded by non-uniform shadows and lighting caused by multiple light sources and complex scene geometries. We observe that DINOv2 can reliably extract both semantic and geometric information from a degraded image. Based on this observation, we develop a novel framework to utilize DINOv2 features for lighting normalization. First, we propose an adaptive feature fusion module that combines features from different DINOv2 layers using a point-wise softmax mask. Next, the fused features are integrated into our proposed restoration network in both spatial and frequency domains through an auxiliary cross-attention mechanism. Experiments show that DINOLight achieves superior performance on the Ambient6K dataset, and that DINOv2 features are effective for enhancing ambient light normalization. We also apply our method to shadow-removal benchmark datasets, achieving competitive results compared to methods that use mask priors. Codes will be released upon acceptance.

Figures & tables

Figure 1: Comparison of PCA-computed DINOv2 features of pairs of images from the Ambient6K dataset, which have the same content under different lighting conditions. We observe that DINOv2 features contain degradation-dependent and -independent information that varies with layer depth, motivating us to use them for restoration. More examples can be found in the Supplementary Material.

Figure 2: Overview of DINOLight, and the two core elements: Adaptive Feature Fusion Module (AFFM), and Auxiliary Cross-Attention (ACA) of the proposed SFDINO block. (a) The first stage involves extracting features from various layers of DINOv2 and combining them using AFFM. (b) In the second stage, the fused features are integrated into the ALN process at multiple resolutions using ACA in SFDINO blocks.

Method	Task	Prior	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MACs
Degraded	-	-	13.403	0.652	0.250	-
DCShadowNet [ 22 ]	Shadow Removal	RGB	17.731	0.711	0.187	13.15G
SpA-Former [ 58 ]	Shadow Removal	RGB+Freq.	19.850	0.810	0.143	16.82G
DW-NET [ 16 ]	Dehazing	RGB+Freq.	20.969	0.805	0.137	7.54G
NAFNet [ 3 ]	Restoration	RGB	20.580	0.808	0.142	15.92G
MPRNet [ 55 ]		RGB	20.947	0.820	0.129	37.21G

Table 1: Quantitative evaluations on Ambient6K test set. We highlight the best , second , and third best results. † denotes MACs without additional models.

Figure 3: Qualitative comparison on Ambient6K test set. From left to right: input, restoration results on a versatile restoration network NAFNet [ 3 ] , previous ALN methods IFBlend [ 46 ] and PromptNorm [ 43 ] , ours, and ground truth. Red boxes highlight regions where DINOLight effectively normalizes light in comparison.

Method	$F_{shallow}$	$F_{middle}$	$F_{deep}$	AFFM	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Baseline	✗	✗	✗	✗	19.685	0.784	0.147
Model A	✓	✗	✗	✗	22.207	0.822	0.119
Model B	✗	✓	✗	✗	20.605	0.802	0.138
Model C	✗	✗	✓	✗	22.060	0.823	0.117
Model D	✓	✓	✓	✗	22.414	0.829	0.115
DINOLight	✓	✓	✓	✓	22.600	0.830	0.112

Table 2: Ablation of DINOv2 features and proposed AFFM on Ambient6K dataset.

Method	DINOv2 Feature	Spat. Attn	Freq. Attn	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Baseline	✗	✓	✓	19.685	0.784	0.147
Model 1	✓	✓	✗	22.289	0.828	0.114
Model 2	✓	✗	✓	22.487	0.829	0.113
DINOLight-p	✓	✓	✓	22.337	0.827	0.115
DINOLight	✓	✓	✓	22.600	0.830	0.112

Table 3: Ablation of proposed ACA on Ambient6K dataset.

Method	Mask Usage	Evaluation Resolution	ISTD			ISTD+
Method	Mask Usage	Evaluation Resolution	Total $\downarrow$	Shadow $\downarrow$	Shadow-Free $\downarrow$	Total $\downarrow$	Shadow $\downarrow$	Shadow-Free $\downarrow$
Degraded	-	640 $\times$ 480	6.80	15.07	3.86	7.15	17.53	1.82
ST-CGAN [ 50 ]	Mask-Prior	256 $\times$ 256	4.05	4.83	3.44	-	-	-
SP-M Net [ 26 ]		512 $\times$ 512	-	-	-	4.37	4.79	4.27
DHAN [ 6 ]		640 $\times$ 480	3.43	4.65	3.13	3.19	4.04	2.97
PULSr [ 44 ]		512 $\times$ 512	3.33	4.48	3.03	2.82	4.12	2.39

Table 4: Quantitative evaluations on ISTD [ 50 ] and ISTD+ [ 26 ] test dataset. RMSE in the LAB color space is measured as the metric. Best results are highlighted in bold for each mask-prior and mask-free method, and unavailable results are omitted.

Image Restoration

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

See all 371 papers

CANDLE: Illumination-Invariant Semantic Priors for Color Ambient Lighting Normalization
Apr 3, 2026Rong-Lin Jian, Ting-Yao Chen, Yu-Fan Lin +4Color Constancy Image Enhancement
National Yang Ming Chiao Tung University · National Cheng Kung University · NVIDIA
UniBlendNet: Unified Global, Multi-Scale, and Region-Adaptive Modeling for Ambient Lighting Normalization
Apr 15, 2026Jiatao Dai, Wei Dong, Han Zhou +2Image Restoration Image Enhancement
McMaster University · University of Manitoba
OmniLight: One Model to Rule All Lighting Conditions
Apr 16, 2026Youngjin Oh, Junyoung Park, Junhyeong Kwon +1Shadow Removal Image Enhancement
Department of ECE, INMC, Seoul National University, South Korea · IPAI, Seoul National University, South Korea

Apr 3, 2026J/K move · Enter open · S save

CANDLE: Illumination-Invariant Semantic Priors for Color Ambient Lighting Normalization

Rong-Lin Jian, Ting-Yao Chen, Yu-Fan Lin, Chia-Ming Lee+3

National Yang Ming Chiao Tung University · National Cheng Kung University · NVIDIA

Color Constancy Image Enhancement

Open paper AskPDF arXiv

Color ambient lighting normalization under multi-colored illumination is challenging due to severe chromatic shifts, highlight saturation, and material-dependent reflectance. Existing geometric and low-level priors are insufficient for recovering object-intrinsic color when illumination-induced chromatic bias dominates. We observe that DINOv3's self-supervised features remain highly consistent between colored-light inputs and ambient-lit ground truth, motivating their use as illumination-robust semantic priors. We propose CANDLE (Color Ambient Normalization with DINO Layer Enhancement), which introduces DINO Omni-layer Guidance (D.O.G.) to adaptively inject multi-layer DINOv3 features into successive encoder stages, and a color-frequency refinement design (BFACG + SFFB) to suppress decoder-side chromatic collapse and detail contamination. Experiments on CL3AN show a +1.22 dB PSNR gain over the strongest prior method. CANDLE achieves 3rd place on the NTIRE 2026 ALN Color Lighting Challenge and 2nd place in fidelity on the White Lighting track with the lowest FID, confirming strong generalization across both chromatic and luminance-dominant illumination conditions. Code is available at https://github.com/ron941/CANDLE.

Abstract

Image Restoration

Explore similar work

CANDLE: Illumination-Invariant Semantic Priors for Color Ambient Lighting Normalization

UniBlendNet: Unified Global, Multi-Scale, and Region-Adaptive Modeling for Ambient Lighting Normalization

OmniLight: One Model to Rule All Lighting Conditions