RetinexDual: Dual-Branch UHD Image Restoration

• RetinexDual is a dual-branch framework that decomposes UHD images into reflectance and illumination, enabling targeted restoration of localized details and global lighting.
• The architecture employs SAMBA for reflectance recovery and FIA for illumination correction, demonstrating improved results in deraining, deblurring, dehazing, and low-light enhancement.
• RetinexDualV2 builds on the original design by integrating physical priors and attention mechanisms, allowing a unified solution for diverse restoration challenges with minimal parameter overhead.

Searching arXiv for RetinexDual and closely related Retinex restoration work. Search query: RetinexDual generalized UHD image restoration arXiv RetinexDual is a Retinex theory-based framework for generalized Ultra-High-Definition Image Restoration (UHD IR) that treats reflectance restoration and illumination or color correction as different restoration subproblems rather than as a single homogeneous mapping. In its 2025 formulation, the framework combines a Retinex decomposer with two complementary sub-networks—SAMBA for the reflectance component and FIA for the illumination component—and is evaluated on deraining, deblurring, dehazing, and Low-Light Image Enhancement (LLIE) (Kishawy et al., 6 Aug 2025). A 2026 continuation, RetinexDualV2, preserves the same key insight that reflection and illumination should be treated by different sub-networks, but adds task-specific physical priors and a physical-conditioned attention mechanism so that one architecture can handle multiple UHD restoration problems without task-specific structural modifications (Kishawy et al., 30 Mar 2026).

1. Conceptual basis and problem formulation

RetinexDual is motivated by the claim that UHD degradations are not homogeneous across tasks. The 2025 paper argues that extreme downsampling causes irreversible information loss in UHD images, while pure frequency-domain approaches are ineffective for spatially confined image artifacts because they lose degradation locality. Within this view, reflectance carries object texture, edges, and fine detail, whereas illumination or color carries exposure, lighting, haze-induced color shifts, and global tone distortions. Retinex theory therefore becomes an organizing principle for generalized restoration rather than only for low-light enhancement (Kishawy et al., 6 Aug 2025).

The clean-image model is written as

$I = R \odot L,$

where $I$ is the image, $R$ is the reflectance, $L$ is the illumination, and $\odot$ denotes element-wise multiplication. For degraded inputs, the framework uses

$$I = R_{eff} \odot L_{eff} = (R + \hat{R}) \odot (L + \hat{L}),$$

with $\hat{R}$ representing artifacts or hidden degradation in reflectance and $\hat{L}$ representing color degradation or exposure error in illumination. Restoration is then expressed as

$$R_{eff},L_{eff}=\psi_d(x),\qquad R = R_{eff} - \hat R,\qquad L = L_{eff} - \hat L,$$

with branch-specific corrections

$$\hat{R} = -\mathcal{S}(R_{eff}), \qquad \hat{L} = -\mathcal{F}(L_{eff}),$$

followed by reconstruction as $I$0 (Kishawy et al., 6 Aug 2025).

RetinexDualV2 reformulates the same idea for generalized UHD image restoration with more explicit distortion terms:

$I$1

and then

$I$2

so that

$I$3

The V2 paper states explicitly that illumination-related corruption is linked to lighting deficiency, haze, shadows, or global color bias, whereas reflection or detail-related corruption is linked to texture loss, rain streaks, drop artifacts, and local structure disruption (Kishawy et al., 30 Mar 2026).

2. Dual-branch architecture of the 2025 framework

The original RetinexDual architecture has three main parts: a Retinex decomposer, SAMBA for reflectance restoration, and FIA for illumination or color correction. The decomposer uses two convolution layers for deep feature extraction followed by a $I$4 convolution to produce $I$5 and $I$6. The paper emphasizes that the model operates directly on the decomposed components rather than first converting everything into one generic latent space (Kishawy et al., 6 Aug 2025).

SAMBA, the Scale-Attentive maMBA branch, is designed for localized artifact removal, structural refinement, and detail recovery. It is an encoder-decoder with 3 Scale Adaptive Mamba Blocks operating at scales $I$7. The block is written as

$I$8

$I$9

$R$0

$R$1

This coarse-to-fine mechanism is introduced to mitigate the causal modeling limitation of ordinary Mamba in vision. Within SAMBA, the Group State Space Block uses layer normalization, a convolutional feed-forward layer, learnable residual scaling, positional encoding, local embedding, global embedding, and an image-specific embedding $R$2, yielding the modified state-space form

$R$3

The stated purpose is to capture long-range dependencies, reduce artifacts from local corruption, preserve fine structures, and recover texture details (Kishawy et al., 6 Aug 2025).

FIA, the Frequency Illumination Adaptor, is the illumination or color branch. It contains a compact Fourier Correction Block. Given $R$4, the block applies layer normalization, FFT, amplitude and phase processing with $R$5 convolutions and ReLU, inverse FFT, adaptive scaling, and a final $R$6 convolution:

$R$7

$R$8

The paper argues that frequency processing is effective for illumination correction because low frequencies capture global structure and lighting, while amplitude and phase allow the block to adjust brightness, color intensity, global illumination, and structural alignment. FIA is described as efficient, with about $R$9M parameters in its FCB design (Kishawy et al., 6 Aug 2025).

3. Optimization, benchmarks, and empirical profile of RetinexDual

The original framework is trained with a multi-term loss

$L$0

with $L$1, $L$2, $L$3, and $L$4. It also uses multi-level supervision:

$L$5

Training is performed on 4 NVIDIA H100 GPUs with AdamW, an initial learning rate of $L$6, a final learning rate of $L$7, cosine annealing, a patch size of $L$8, and batch size $L$9. Evaluation is reported on UHD-LL, UHD-Haze, UHD-Blur, and 4K-Rain13K using PSNR, SSIM, parameter count, and inference time (Kishawy et al., 6 Aug 2025).

Benchmark	RetinexDual	Reported comparator
UHD-LL	28.79 / 0.934	ERR 27.57 / 0.932
4K-Rain13K	34.50 / 0.967	ERR 34.48 / 0.952
UHD-Haze	26.63 / 0.956	ERR 25.12 / 0.950
UHD-Blur	30.71 / 0.886	D2Net 30.46 / 0.872

The ablation studies are central to the framework’s interpretation. Removing multi-level supervision lowers UHD-LL performance from $\odot$0 to $\odot$1. Using FIA for both reflectance and illumination yields $\odot$2; using SAMBA for both gives $\odot$3; removing FIA gives $\odot$4; and removing SAMBA gives $\odot$5. Within SAMBA, the reported variants are $\odot$6 without multi-scale, $\odot$7 without GSSB, and $\odot$8 without both. Removing Fourier processing from FIA lowers performance to $\odot$9. These results support the paper’s central claim that reflectance and illumination require different architectures rather than a single shared backbone (Kishawy et al., 6 Aug 2025).

4. RetinexDualV2 and physically grounded generalization

RetinexDualV2 is presented as a physically grounded continuation of RetinexDual. It keeps the dual-branch design but adds a Task-Specific Physical Grounding Module (TS-PGM) and Physical-conditioned Multi-head Self-Attention (PC-MSA) so that one architecture can handle low-light enhancement, dehazing, deraining, shadow removal, and NTIRE 2026 challenge settings without task-specific structural modifications (Kishawy et al., 30 Mar 2026).

TS-PGM extracts a physical prior map $$I = R_{eff} \odot L_{eff} = (R + \hat{R}) \odot (L + \hat{L}),$$0 through a fusion of gradient and intensity or structure cues:

$$I = R_{eff} \odot L_{eff} = (R + \hat{R}) \odot (L + \hat{L}),$$1

The specific prior depends on the task. For deraining, the paper uses a residual rain intensity mask predicted by a shallow UNet $$I = R_{eff} \odot L_{eff} = (R + \hat{R}) \odot (L + \hat{L}),$$2, with an offline target

$$I = R_{eff} \odot L_{eff} = (R + \hat{R}) \odot (L + \hat{L}),$$3

For low-light enhancement, it uses a structure-aware illumination prior derived from a Gaussian color model:

$$I = R_{eff} \odot L_{eff} = (R + \hat{R}) \odot (L + \hat{L}),$$4

For nighttime dehazing, it uses the dark channel prior

$$I = R_{eff} \odot L_{eff} = (R + \hat{R}) \odot (L + \hat{L}),$$5

For shadow removal, it uses a log-chromaticity shadow-resistant prior based on $$I = R_{eff} \odot L_{eff} = (R + \hat{R}) \odot (L + \hat{L}),$$6, $$I = R_{eff} \odot L_{eff} = (R + \hat{R}) \odot (L + \hat{L}),$$7, $$I = R_{eff} \odot L_{eff} = (R + \hat{R}) \odot (L + \hat{L}),$$8, and a learnable projection angle $$I = R_{eff} \odot L_{eff} = (R + \hat{R}) \odot (L + \hat{L}),$$9, followed by percentile-based normalization. The paper’s interpretation is that these priors are explicitly tied to the image formation process of each corruption type rather than to generic semantic features (Kishawy et al., 30 Mar 2026).

PC-MSA injects those priors into attention by modulating the value stream. Given $\hat{R}$0, standard projections produce

$\hat{R}$1

and the modification is

$\hat{R}$2

Attention is then computed as

$\hat{R}$3

The reflectance branch becomes PG-SAMBA, which injects TS-PGM priors into Mamba-style aggregation via PC-MSA and uses a lightweight group state-space component for long-range semantic consistency. The illumination branch becomes PG-FCB, which applies PC-MSA conditioning before compact Fourier-domain processing and is described as having a small parameter footprint of about $\hat{R}$4M (Kishawy et al., 30 Mar 2026).

The V2 training objective uses deep supervision at three output resolutions $\hat{R}$5, $\hat{R}$6, and $\hat{R}$7:

$\hat{R}$8

with $\hat{R}$9, $\hat{L}$0, $\hat{L}$1, and $\hat{L}$2. The implementation uses AdamW, an initial learning rate of $\hat{L}$3, cosine annealing down to $\hat{L}$4, patch size $\hat{L}$5, batch size $\hat{L}$6, and training on an NVIDIA H100 GPU. On 4K-Rain13K, RetinexDualV2 reports $\hat{L}$7 PSNR, $\hat{L}$8 SSIM, and $\hat{L}$9M parameters; on UHD-LL it reports $$R_{eff},L_{eff}=\psi_d(x),\qquad R = R_{eff} - \hat R,\qquad L = L_{eff} - \hat L,$$0 PSNR, $$R_{eff},L_{eff}=\psi_d(x),\qquad R = R_{eff} - \hat R,\qquad L = L_{eff} - \hat L,$$1 SSIM, and $$R_{eff},L_{eff}=\psi_d(x),\qquad R = R_{eff} - \hat R,\qquad L = L_{eff} - \hat L,$$2M parameters. In NTIRE 2026, it reports 4th place in the Day and Night Raindrop Removal challenge with PSNR $$R_{eff},L_{eff}=\psi_d(x),\qquad R = R_{eff} - \hat R,\qquad L = L_{eff} - \hat L,$$3, SSIM $$R_{eff},L_{eff}=\psi_d(x),\qquad R = R_{eff} - \hat R,\qquad L = L_{eff} - \hat L,$$4, LPIPS $$R_{eff},L_{eff}=\psi_d(x),\qquad R = R_{eff} - \hat R,\qquad L = L_{eff} - \hat L,$$5, and $$R_{eff},L_{eff}=\psi_d(x),\qquad R = R_{eff} - \hat R,\qquad L = L_{eff} - \hat L,$$6M parameters, and 5th place in the Joint Noise Low-light Enhancement challenge with PSNR $$R_{eff},L_{eff}=\psi_d(x),\qquad R = R_{eff} - \hat R,\qquad L = L_{eff} - \hat L,$$7, SSIM $$R_{eff},L_{eff}=\psi_d(x),\qquad R = R_{eff} - \hat R,\qquad L = L_{eff} - \hat L,$$8, and LPIPS $$R_{eff},L_{eff}=\psi_d(x),\qquad R = R_{eff} - \hat R,\qquad L = L_{eff} - \hat L,$$9. It also reports Nighttime Dehazing at PSNR $$\hat{R} = -\mathcal{S}(R_{eff}), \qquad \hat{L} = -\mathcal{F}(L_{eff}),$$0, SSIM $$\hat{R} = -\mathcal{S}(R_{eff}), \qquad \hat{L} = -\mathcal{F}(L_{eff}),$$1, LPIPS $$\hat{R} = -\mathcal{S}(R_{eff}), \qquad \hat{L} = -\mathcal{F}(L_{eff}),$$2, and Shadow Removal at PSNR $$\hat{R} = -\mathcal{S}(R_{eff}), \qquad \hat{L} = -\mathcal{F}(L_{eff}),$$3, SSIM $$\hat{R} = -\mathcal{S}(R_{eff}), \qquad \hat{L} = -\mathcal{F}(L_{eff}),$$4, LPIPS $$\hat{R} = -\mathcal{S}(R_{eff}), \qquad \hat{L} = -\mathcal{F}(L_{eff}),$$5. The paper notes a discrepancy in that the conclusion mentions “6th place” for JNLLIE, whereas the main results table shows 5th place; the table is the more direct evidence in the paper body. Its UHD-LL ablation further reports $$\hat{R} = -\mathcal{S}(R_{eff}), \qquad \hat{L} = -\mathcal{F}(L_{eff}),$$6 without dual-branch design, $$\hat{R} = -\mathcal{S}(R_{eff}), \qquad \hat{L} = -\mathcal{F}(L_{eff}),$$7 without physical prior, $$\hat{R} = -\mathcal{S}(R_{eff}), \qquad \hat{L} = -\mathcal{F}(L_{eff}),$$8 without PC-MSA, and $$\hat{R} = -\mathcal{S}(R_{eff}), \qquad \hat{L} = -\mathcal{F}(L_{eff}),$$9 for the full model (Kishawy et al., 30 Mar 2026).

5. Broader Retinex dualities and terminological scope

The term “RetinexDual” is not uniform across the literature. In the UHD restoration papers it denotes a formal framework family, whereas other Retinex papers use “dual” in different senses.

Usage	Meaning	arXiv id
RetinexDual	Dual-branch generalized UHD image restoration with SAMBA and FIA	(Kishawy et al., 6 Aug 2025)
RetinexDualV2	Physically grounded continuation with TS-PGM and PC-MSA	(Kishawy et al., 30 Mar 2026)
Retinex-based dual formulation	Nonlocal Retinex variational model and deep unfolding twin	(Torres et al., 10 Apr 2025)
Retinex/dehazing duality	$I$00	(Galdran et al., 2017)

A foundational antecedent is the 2017 dehazing paper “On the Duality Between Retinex and Image Dehazing,” which proves that

$I$01

Under the white-balanced assumption $I$02, the haze model

$I$03

becomes, after intensity inversion,

$I$04

which the authors interpret as a Retinex-style multiplicative factorization. The paper also shows that because threshold-free Retinex increases brightness, the transformed operator must decrease intensity, consistent with dehazing. This duality is operator-theoretic and should not be conflated with the later dual-branch UHD architectures (Galdran et al., 2017).

Another related usage appears in the 2025 paper on a nonlocal Retinex-based variational model and its deep unfolding twin for low-light image enhancement. There, the relevant “dual” idea is a paired framework rather than a named UHD architecture. The variational model uses

$I$05

adds color correction preprocessing, a nonlocal total variation prior on reflectance, a local TV prior on illumination, an explicit noise variable, a nonlocal gradient-type fidelity term, and automatic gamma correction. Its deep unfolding counterpart replaces proximal steps with ProxNet, CARNet, and cross-attention modules over 5 unfolded stages, trained for 1000 epochs with Adam and learning rate $I$06 (Torres et al., 10 Apr 2025).

A nearby classical line of work on simultaneous enhancement and noise suppression under complex illumination conditions also uses Retinex decomposition, parallel illumination and reflection processing, a dual-channel strategy on the original and inverted image, and inversion-based haze removal. That method is centered on the gradient-domain weighted guided filter rather than on the RetinexDual name, but it shows that dual-branch and dual-domain reasoning extends beyond the specific UHD framework family (Tao et al., 9 Dec 2025).

6. Limitations, misconceptions, and significance

A common simplification is to treat RetinexDual as merely a low-light enhancement model or as a single generic restoration backbone. That description is inconsistent with the 2025 paper, which evaluates on deraining, deblurring, dehazing, and LLIE, and with its ablations showing that using the same sub-network for both reflectance and illumination is inferior to branch-specific processing. The original framework is therefore better understood as a generalized UHD restoration architecture whose central thesis is that spatially localized reflectance corruption and global illumination or color distortion should not be restored in the same way (Kishawy et al., 6 Aug 2025).

A second misconception is that Retinex decomposition alone is sufficient for complex scenes. RetinexDualV2 argues directly against that position: the paper states that Retinex decomposition alone is not enough and adds physically meaningful guidance to locate and suppress task-specific corruption. At the same time, its own limitations are explicit. TS-PGM depends on knowing the degradation type and designing an appropriate prior for it, so the method is not fully task-agnostic in the strictest sense. Its generalization is described as strong but not universal, and the evaluation is mostly benchmark- and challenge-driven, with limited discussion of extreme out-of-distribution failures, mixed degradations, or incorrect prior estimation. The parameter count also rises slightly relative to RetinexDual, from $I$07M to $I$08M (Kishawy et al., 30 Mar 2026).

Across the broader Retinex literature, “dual” should also be read with care. In the dehazing duality paper, the clean equivalence requires white-balanced atmospheric light, and per-channel Retinex can produce unnatural colors when atmospheric light is not neutral. In the low-light variational and unfolding paper, the variational model is computationally heavy, while the unfolding approach, although more efficient and accurate, does not explicitly guarantee a reliable decomposition. These caveats suggest that RetinexDual is best situated as one branch of a wider research program: using Retinex factorization to separate restoration into physically or structurally distinct components, then choosing architectures or operators that match those components rather than collapsing them into a single undifferentiated model (Galdran et al., 2017, Torres et al., 10 Apr 2025).