DINOLight: Ambient Light Normalization

• DINOLight is an ambient light normalization framework that leverages self-supervised DINOv2 features to restore images degraded by non-uniform shadows and lighting.
• It employs an Adaptive Feature Fusion Module (AFFM) to fuse multi-depth DINO features and an Auxiliary Cross-Attention mechanism to inject these cues into a Transformer restoration backbone.
• DINOLight achieves competitive performance on benchmarks such as Ambient6K and ISTD with mask-free restoration, improving PSNR, SSIM, and LPIPS over baseline methods.

DINOLight is an ambient light normalization framework that integrates self-supervised DINOv2 features into image restoration as a visual prior in order to restore images degraded by non-uniform shadows and lighting caused by multiple light sources, complex scene geometries, and diverse surface materials (Oh et al., 13 Mar 2026). It is explicitly mask-free, end-to-end, and organized around two coupled mechanisms: an Adaptive Feature Fusion Module (AFFM) that fuses multi-depth DINOv2 features with point-wise softmax weights, and an Auxiliary Cross-Attention (ACA) mechanism that injects the fused prior into a Transformer restoration backbone in both spatial and frequency domains. In subsequent work on lighting-related image restoration, DINOLight is also used as a task-specialized baseline for shadow removal and for ambient lighting normalization under both white and colored illumination regimes (Oh et al., 16 Apr 2026).

1. Task formulation and scope

Ambient light normalization (ALN) aims to restore images degraded by non-uniform illumination and shadows arising from multiple light sources, complex scene geometries, and diverse surface materials. In DINOLight, this problem is framed more broadly than conventional shadow removal. The target degradations include non-uniform shadows and soft shadows, multi-source lighting that yields overlapping or interacting shadow fields, and complex geometries involving self-shadows, occlusions, inter-reflections, and complex materials (Oh et al., 13 Mar 2026).

This broader framing is important because many shadow-removal pipelines are built around explicit shadow mask priors. Such priors are laborious to collect, difficult to estimate accurately in the wild, and often unstable around soft or ambiguous boundaries. DINOLight departs from that design in two stated ways: it does not require ground-truth shadow masks, and it integrates robust self-supervised priors from DINOv2 to exploit geometric and semantic information learned from large-scale, uncurated data. A plausible implication is that DINOLight treats localization as an implicit inference problem mediated by foundation-model features rather than as an explicit binary-mask estimation problem.

Within the broader lighting-restoration setting studied later, DINOLight is applied not only to shadow removal but also to two ALN regimes: ALN under white lighting, which seeks white-balanced normalization while preserving true object colors, and ALN under color lighting, which additionally requires per-pixel chromatic adaptation to separate colored illumination from surface reflectance (Oh et al., 16 Apr 2026). This situates DINOLight at the intersection of shadow removal, illumination normalization, and reflectance-preserving restoration.

2. DINOv2 visual priors and adaptive feature fusion

DINOLight uses a distilled DINOv2 ViT-B/14 with registers as a frozen feature extractor. Given a degraded image $I_D \in \mathbb{R}^{H \times W \times 3}$, it extracts features from layers 1, 6, and 12, denoted as shallow, middle, and deep features, each with shape $\mathbb{R}^{(H/14) \times (W/14) \times 768}$ (Oh et al., 13 Mar 2026). The reported empirical interpretation is that shallow features retain finer positional and geometric details together with lighting-dependent cues, while deep features provide semantic-aware and more lighting-invariant representations.

The AFFM is designed to combine these heterogeneous cues adaptively at each spatial position. Each feature tensor $F_l$, with $l \in \{s,m,d\}$, is passed through a SiLU activation followed by a $1 \times 1$ convolution:

$$G_l = \mathrm{Conv}_{1\times 1}(\mathrm{SiLU}(F_l)).$$

The resulting tensors are concatenated along the layer dimension, and a point-wise softmax produces scalar fusion weights for shallow, middle, and deep features. For each spatial position $x$,

$$M_l(x) = \frac{\exp(z_l(x))}{\exp(z_s(x)) + \exp(z_m(x)) + \exp(z_d(x))}, \quad l \in \{s,m,d\}.$$

The fused feature is then

$$F_{\mathrm{AFFM}}(x) = M_s(x)\,F_{\mathrm{shallow}}(x) + M_m(x)\,F_{\mathrm{middle}}(x) + M_d(x)\,F_{\mathrm{deep}}(x).$$

In the paper’s notation,

$$F_{\mathrm{AFFM}} = W_s \odot F_{\mathrm{shallow}} + W_m \odot F_{\mathrm{middle}} + W_d \odot F_{\mathrm{deep}},$$

where $\mathbb{R}^{(H/14) \times (W/14) \times 768}$0 denotes element-wise multiplication and the point-wise masks are broadcast along the channel dimension.

The softmax is computed independently at each spatial location, the masks sum to one per location, and gradients flow through SiLU, the $\mathbb{R}^{(H/14) \times (W/14) \times 768}$1 convolutions, the softmax, and the weighted sum. For multi-scale injection, $\mathbb{R}^{(H/14) \times (W/14) \times 768}$2 is projected by a $\mathbb{R}^{(H/14) \times (W/14) \times 768}$3 convolution to match a stage’s channel dimension and then bilinearly upsampled to the stage’s spatial resolution. This makes the DINO prior compatible with hierarchical feature maps at multiple scales. The later OmniLight study describes the same prior source more generally as frozen DINOv2 features $\mathbb{R}^{(H/14) \times (W/14) \times 768}$4 that guide restoration without a distillation loss or DINO-specific supervision term (Oh et al., 16 Apr 2026).

3. SFDINO blocks, auxiliary cross-attention, and dual-domain processing

DINOLight introduces SFDINO blocks, which augment conventional self-attention with auxiliary cross-attention to merge internal restoration features with external DINO priors (Oh et al., 13 Mar 2026). In the original report, the restoration network is hierarchical and multi-scale; in the later OmniLight study, the same architecture is described as a U-Net-style encoder-decoder with skip connections whose main computational units are SFDINO blocks (Oh et al., 16 Apr 2026). Both descriptions are consistent with a Transformer-based, multi-resolution restoration backbone.

Let $\mathbb{R}^{(H/14) \times (W/14) \times 768}$5 and $\mathbb{R}^{(H/14) \times (W/14) \times 768}$6 denote the internal feature map from the previous Transformer block and the fused DINO feature aligned to the same stage. After LayerNorm,

$\mathbb{R}^{(H/14) \times (W/14) \times 768}$7

Attention projections use $\mathbb{R}^{(H/14) \times (W/14) \times 768}$8 point-wise and $\mathbb{R}^{(H/14) \times (W/14) \times 768}$9 depth-wise convolutions:

$F_l$0

$F_l$1

Attention follows the standard scaled dot-product form, with transposed attention as in Restormer:

$F_l$2

DINOLight computes internal self-attention

$F_l$3

and auxiliary cross-attention from the DINO prior

$F_l$4

The contribution of the auxiliary path is regulated by a learnable channel-wise parameter $F_l$5, normalized to $F_l$6 by a sigmoid:

$F_l$7

The output then passes through the block’s feed-forward network. ACA is applied at multiple resolutions, and the DINO prior is aligned at each stage by $F_l$8 projection and bilinear upsampling before injection.

DINOLight also includes a frequency-domain path inspired by IFBlend and frequency-domain Transformers. The exact transform operator is not specified in the main text; it is denoted generically by $F_l$9 and $l \in \{s,m,d\}$0. Conceptually,

$l \in \{s,m,d\}$1

and

$l \in \{s,m,d\}$2

Spatial and frequency attentions are alternated sequentially across SFDINO blocks. A parallel variant, DINOLight-p, was tested and found inferior to the sequential design. This suggests that the method benefits from staged rather than simultaneous aggregation of spatial and spectral dependencies.

4. Optimization, data, inference, and computational profile

DINOLight is trained end-to-end with a fidelity-driven objective that combines $l \in \{s,m,d\}$3 and MS-SSIM:

$l \in \{s,m,d\}$4

The original DINOLight report states that no adversarial loss and no perceptual loss such as VGG or LPIPS are used in training (Oh et al., 13 Mar 2026). It also reports Adam optimization with $l \in \{s,m,d\}$5, $l \in \{s,m,d\}$6, an initial learning rate of $l \in \{s,m,d\}$7 cosine annealed to $l \in \{s,m,d\}$8, batch size 4, 400K training iterations, random flips and rotations, and implementation in PyTorch on NVIDIA GeForce RTX 3090 GPUs.

Ambient6K is the central ALN benchmark in the original study. It contains 5,000 high-resolution training pairs at $l \in \{s,m,d\}$9 and 500 test pairs at $1 \times 1$0. Training uses $1 \times 1$1 crops, and test-time inference uses a sliding window with the same patch size. Evaluation is reported with PSNR, SSIM, and LPIPS. DINOLight is also evaluated in a mask-free configuration on the ISTD and ISTD+ shadow-removal benchmarks using LAB-space RMSE, typically at $1 \times 1$2 (Oh et al., 13 Mar 2026).

A later study reuses DINOLight as a specialized baseline across three NTIRE 2026 tracks—Shadow Removal, ALN (White Lighting), and ALN (Color Lighting)—and reports track-specific training on WSRD+ plus the corresponding NTIRE and public datasets, including Ambient6K and CL3AN (Oh et al., 16 Apr 2026). In that setting, training uses random $1 \times 1$3 crops, random flips and rotations, AdamW with $1 \times 1$4 and $1 \times 1$5, and cosine scheduling with warm-up similar to OmniLight. The reported inference scheme remains sliding-window evaluation at $1 \times 1$6 with configurable overlap.

The computational profile reported for DINOLight is relatively compact for the task. The restoration network requires 7.89 GMACs, or 16.86 GMACs when the external ViT-B/14 DINOv2 is included; it has 9.5M parameters, or 95.0M including DINOv2 (Oh et al., 13 Mar 2026). DINOLight-L is reported at 17.48 GMACs for the restoration network. In the later benchmark report, end-to-end runtimes on an RTX 3090 range approximately from 1.6 s to 3.1 s for inputs of $1 \times 1$7 and from 1.6 s to 2.6 s for $1 \times 1$8, depending on overlap (Oh et al., 16 Apr 2026).

5. Empirical results and ablation analyses

On the Ambient6K test set, the original study reports the following results: IFBlend reaches PSNR 21.443, SSIM 0.819, and LPIPS 0.128; PromptNorm reaches PSNR 22.116, SSIM 0.822, and LPIPS 0.124; DINOLight reaches PSNR 22.788, SSIM 0.838, and LPIPS 0.107; and DINOLight-L reaches PSNR 23.059, SSIM 0.839, and LPIPS 0.107 (Oh et al., 13 Mar 2026). The reported difference between DINOLight-L and PromptNorm is +0.943 dB PSNR, with LPIPS reduced from 0.124 to 0.107. Qualitatively, DINOLight is reported to distinguish more reliably between shadows and objects in ambiguous regions, reduce artifacts and color inconsistencies, and handle self-shadowed regions without over-normalization.

The shadow-removal experiments emphasize the mask-free setting. On ISTD, DINOLight reports LAB RMSE values of Total 3.00, Shadow 3.63, and Shadow-Free 2.84, compared with IFBlend at Total 3.52, Shadow 4.02, and Shadow-Free 3.38; ShadowFormer, a strong mask-prior reference, reports Total 2.43, Shadow 3.25, and Shadow-Free 2.38. On ISTD+, DINOLight reports Total 2.34, Shadow 3.41, and Shadow-Free 1.99, compared with IFBlend at Total 2.40, Shadow 3.18, and Shadow-Free 2.14; ShadowFormer reports Total 1.93, Shadow 2.93, and Shadow-Free 1.66 (Oh et al., 13 Mar 2026). The reported interpretation is that DINOLight achieves competitive performance without masks and clearly outperforms prior mask-free methods on ISTD while performing favorably on ISTD+.

The ablation studies isolate the contribution of DINO priors, AFFM, and ACA. On Ambient6K, a baseline without DINO features reaches PSNR 19.685, SSIM 0.784, and LPIPS 0.147. Using only shallow features raises performance to 22.207/0.822/0.119, only middle features to 20.605/0.802/0.138, and only deep features to 22.060/0.823/0.117. Simple addition of all three DINO feature levels yields 22.414/0.829/0.115, while AFFM-based fusion yields 22.600/0.830/0.112 (Oh et al., 13 Mar 2026). The paper’s stated observation is that all DINO features help, shallow features provide the largest single-layer gain, and adaptive point-wise fusion improves over simple addition.

A second ablation examines the dual-domain ACA design. The baseline without DINO remains at 19.685/0.784/0.147; spatial-only attention with DINO reaches 22.289/0.828/0.114; frequency-only attention reaches 22.487/0.829/0.113; DINOLight-p, which uses spatial and frequency ACA in parallel, reaches 22.337/0.827/0.115; and sequential DINOLight reaches 22.600/0.830/0.112. The study states that injecting DINO features through ACA yields at least +2.80 dB PSNR over the baseline and that sequential dual-domain integration outperforms purely spatial, purely frequency, and parallel variants (Oh et al., 13 Mar 2026).

The later OmniLight study extends the empirical picture to additional datasets. There, DINOLight reports on CL3AN 21.107 dB PSNR, 0.764 SSIM, and 0.188 LPIPS, which is described as the top PSNR among compared methods on that benchmark. On WSRD+, DINOLight reports 27.17 dB PSNR and 0.849 SSIM, again the best among listed methods. These results are used to characterize DINOLight as a strong in-domain, track-specialized baseline (Oh et al., 16 Apr 2026).

6. Relation to mask-prior methods, specialization, and limitations

DINOLight is frequently contrasted with methods that use explicit shadow masks. The reported advantage of mask-prior methods is accurate localization on well-defined hard shadows, but the corresponding disadvantages are difficulty with soft shadows and ambiguous boundaries and the burden of obtaining masks. DINOLight’s alternative is to replace explicit masks with DINOv2-derived geometric and semantic priors injected through ACA. The reported advantages are no dependency on masks, better handling of complex and non-uniform lighting, and the use of priors derived from large-scale self-supervised training (Oh et al., 13 Mar 2026). A common misconception, therefore, is that explicit shadow masks are necessary for strong performance; DINOLight’s benchmark results indicate that competitive restoration can be achieved without them.

The trade-off is that mask-free restoration still relies on implicit localization and on the quality of prior integration. The original DINOLight report does not enumerate explicit failure cases, but it notes that extremely challenging lighting conditions, including severe saturation, highly reflective or transparent materials, or strong texture ambiguities, may remain difficult (Oh et al., 13 Mar 2026). This is framed not as a specific failure mode list but as a limitation typical of mask-free approaches.

Within the OmniLight framework, DINOLight is positioned as the specialized end of a specialization–generalization spectrum. It is optimized per track and per dataset, which yields high fidelity on target distributions but reduced robustness under domain shift, particularly for out-of-distribution lighting such as complex multi-colored illumination. The later report explicitly states that a DINOLight model trained for one track cannot be expected to perform well on another without retraining (Oh et al., 16 Apr 2026). In NTIRE 2026, the DINOLight submission ranked 6th Perceptual and 5th Fidelity on ALN (White Lighting), 1st in Fidelity and 2nd in Perceptual on ALN (Color Lighting), and 3rd overall on Shadow Removal. These outcomes reinforce its status as a strong specialized model rather than a unified cross-domain solution.

Later work identifies several directions that bear directly on DINOLight’s design limits: stronger reflectance–shading disentanglement, more principled routing or uncertainty-aware gating, and broader priors such as depth, normals, or temporal cues (Oh et al., 16 Apr 2026). This suggests that DINOLight’s core idea—foundation-model prior injection for lighting restoration—has room to expand beyond the original fixed DINOv2 prior and beyond single-domain specialization.