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.
Figures & tables
Figure 1: DINOv3 [ 24 ] features remain highly consistent under colored illumination. Despite severe appearance shifts between the colored-light input and the ambient-lit GT, DINOv3 PCA visualizations remain highly consistent across domains. Patch-wise cosine similarity maps (brighter = stronger consistency) show that DINOv3 preserves substantially higher similarity both on average ( P ) and in the hardest regions ( W ), outperforming CLIP [ 27 ] and supervised ResNet-50 [ 12 ] .
Figure 2: Quantitative representation consistency over 10 input/GT pairs. DINOv3 [ 24 ] achieves the highest average patch-wise cosine similarity and the strongest robustness on the hardest 10% of regions ( W ), consistently outperforming CLIP [ 27 ] and ResNet-50 [ 12 ] .
Figure 3: Overview of CANDLE. Multi-layer features from a frozen DINOv3 [ 24 ] ViT-L/16 are adaptively selected and fused via PSF and injected into successive encoder stages through DRFB (D.O.G., Sec. 3.1 ), replacing surface-normal guidance with illumination-robust semantic priors. At the decoder, BFACG decouples structural and chromatic restoration under edge-aware modulation, while SFFB suppresses illumination-corrupted skip features via Haar wavelet gating (Sec. 3.2 ).
Figure 4: Qualitative comparison with state-of-the-art methods on CL3AN [ 35 ] . General restoration methods [ 3 , 42 ] exhibit washed-out colors and incomplete correction under strong chromatic shifts. ALN-specific methods [ 34 , 31 ] partially recover global color balance but remain inconsistent across materials, as frequency and geometric priors are insufficient when chromaticity shifts dominate. CANDLE recovers object-intrinsic color more faithfully with sharper boundaries, enabled by D.O.G. and color-frequency refinement.
Method
Venue
MACs(G.)
PSNR ↑
SSIM ↑
LPIPS ↓
NAFNet [ 3 ]
ECCV 2022
4.04
18.547
0.6121
0.3804
SFNet [ 7 ]
ICLR 2023
30.59
16.778
0.5489
0.6183
Uformer [ 37 ]
CVPR 2022
20.89
17.785
0.6815
0.3437
Restormer [ 42 ]
CVPR 2022
35.31
18.244
0.6570
0.3541
HINet [ 15 ]
ICCV 2021
39.46
17.728
0.6035
0.3887
IFBlend [ 34 ]
ECCV 2024
24.65
19.418
0.7581
0.2505
Table 1: Quantitative comparison on the CL3AN benchmark. 1st 2nd 3rd indicate top-three methods by PSNR.
Configuration
PSNR ↑
SSIM ↑
LPIPS ↓
w/o guidance
17.449
0.7131
0.3019
Normal Guidance [ 31 ]
19.216
0.7485
0.2590
CANDLE (D.O.G., ours)
21.066
0.7788
0.2325
Table 2: Ablation on guidance representation. Semantic guidance outperforms geometric normal guidance for color ALN.
Integration
PSNR ↑
SSIM ↑
LPIPS ↓
Input concat
20.682
0.7649
0.2380
Feature/value fusion
20.475
0.7613
0.2408
Query-based
20.403
0.7628
0.2418
CANDLE (D.O.G., ours)
21.066
0.7788
0.2325
Table 3: Ablation on guidance integration strategies. All variants include BFACG and SFFB. D.O.G. with adaptive multi-layer fusion achieves the best performance.
Configuration
PSNR ↑
SSIM ↑
LPIPS ↓
Baseline (w/o guidance)
17.449
0.7131
0.3019
+ D.O.G.
20.840
0.7705
0.2334
+ D.O.G. + BFACG + SFFB (CANDLE)
21.066
0.7788
0.2325
Table 4: Architecture build-up. Each component contributes consistently to final performance.
Figure 5: Training dynamics of ablation variants. (a) D.O.G. sustains a peak gain of +3.68 dB over the unguided baseline, with the gap maintained across all training epochs, confirming a structural rather than incidental improvement. (b) Query-based fusion converges more stably than input-concat or value-fusion variants in later epochs.
Configuration
PSNR ↑
SSIM ↑
LPIPS ↓
CANDLE
20.779
0.7250
0.2159
+ DINO-CLIP Retrieval [ 24 , 27 ]
21.079
0.7291
0.2079
+ Extra synthetic data
21.376
0.7365
0.2021
+ NAFNet [ 3 ] Joint Training
21.432
0.7431
0.1964
Table 5: Effect of auxiliary strategies on the NTIRE 2026 ALN Color Lighting track.
Configuration
PSNR ↑
SSIM ↑
LPIPS ↓
DenseSR [ 20 ]
23.785
0.8466
0.1007
CANDLE
24.370
0.8516
0.1078
+ DINO-CLIP Retrieval [ 24 , 27 ]
24.542
0.8533
0.1072
+ DenseSR [ 20 ] Ensemble
24.875
0.8639
0.0892
Table 6: Effect of ensemble strategies on the NTIRE 2026 ALN White Lighting track.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
FID ↓
Fidelity
Perceptual
MiPorAlgo
21.305
0.6294
0.2478
92.900
4
1
SNU-ISPL-B
22.038
0.7528
0.1861
82.013
1
2
ACVLAB (Ours)
21.432
0.7431
0.1964
88.200
3
3
SNU-ISPL-A
21.900
0.7436
0.2038
86.654
2
4
OUT_OF_MEMORY
20.175
0.6637
0.3338
124.439
5
5
Amber
17.863
0.6812
0.2835
149.271
6
6
Table 7: Ranking on NTIRE 2026 ALN Color Lighting Challenge. Results follow the official NTIRE 2026 challenge report [ 33 ] .
Method
PSNR ↑
SSIM ↑
LPIPS ↓
FID ↓
Fidelity
Perceptual
SNU-ISPL-A
24.275
0.8536
0.0986
54.169
6
1
MILab_ALN
26.680
0.8728
0.0832
50.218
1
2
ACVLAB (Ours)
24.875
0.8639
0.0892
49.838
2
3
MiPorAlgo
23.141
0.7631
0.1687
59.441
11
4
OUT_OF_MEMORY
25.416
0.8604
0.0938
54.496
3
5
SNU-ISPL-B
24.300
0.8561
0.1006
51.883
5
6
Table 8: Ranking on NTIRE 2026 ALN White Lighting Challenge. Results follow the official NTIRE 2026 challenge report [ 33 ] .