Retinex-Guided Illumination Compensation

• Retinex-guided illumination compensation is a technique that decomposes an image into illumination and reflectance components, enabling targeted enhancement and noise reduction.
• It employs a range of methods from PDE-based variational models and guided filtering to Transformer and diffusion-based approaches, balancing detail recovery with efficient enhancement.
• This approach is applied to low-light, exposure correction, underwater imaging, backlit scenarios, and 360° view synthesis, demonstrating its versatility in image restoration tasks.

Retinex-guided illumination compensation denotes a class of enhancement and restoration methods that interpret a degraded image as the interaction of an illumination component and a reflectance component, then compensate illumination while preserving, denoising, or separately reconstructing reflectance-like content. In the cited literature, this principle is used not only for low-light image enhancement, but also for exposure correction, backlit enhancement, underwater restoration, illumination manipulation, diffusion-based relighting, continuous image representations, and low-light 360° novel view synthesis. The common mathematical backbone is the Retinex model $I=L\odot R$ or $I=R\odot L$, but the operational realization ranges from PDEs, guided filtering, and closed-form recursions to latent-space Transformers, state-space models, diffusion models, conditional rectified flow, and 3D Gaussian Splatting (Nnolim, 2017, Demir et al., 19 Jan 2026, Zheng et al., 16 Mar 2026, Yin et al., 27 Apr 2026).

1. Foundational formulation and problem scope

The foundational premise is that illumination degradation should be handled as a factorized problem rather than as direct global brightening. Classical Retinex-based works write the observed image as a product of illumination and reflectance, with illumination representing lighting or exposure conditions and reflectance representing intrinsic scene content. This factorization appears in low-light enhancement, exposure correction, and generalized restoration frameworks, including $I=R\circ L$, $I=R\odot L$, and related logged forms such as $S=\log(I)=\log(R)+\log(L)$ (Mu et al., 2018, Bai et al., 2024, Zheng et al., 16 Mar 2026).

The problem setting is broader than insufficient brightness alone. The cited papers repeatedly describe degraded inputs as having reduced visibility, compressed dynamic range, noise amplification, color distortion, uneven lighting, or mixed under- and over-exposed regions. This is explicit in low-light image enhancement, backlit correction, underwater imaging, and full $360^\circ$ novel view synthesis, where illumination inconsistency can propagate into geometry or rendering instability if it is not separated from actual scene radiance (Wang et al., 19 Aug 2025, Tian et al., 5 Mar 2025, Yin et al., 27 Apr 2026).

A notable implementation pattern is restriction of compensation to luminance-bearing subspaces rather than direct RGB manipulation. Representative choices include HSI intensity enhancement, HSV $V$-channel refinement, YUV luminance-domain compensation, and HVI intensity-guided decoupling. These designs are used to reduce color distortion, make illumination estimation more explicit, or suppress unreliable chromatic information in dark regions (Mu et al., 2018, Demir et al., 19 Jan 2026, Yin et al., 27 Apr 2026, Wang et al., 19 May 2026).

2. Classical, analytical, and filtering-based compensation schemes

Early and non-deep formulations treat illumination compensation as a controlled decomposition-and-correction problem. The PDE-based variational Retinex method models the logged image as $i=l+r$, combines illumination correction, color correction, and anisotropic diffusion smoothing, and selects stopping time by simultaneous entropy maximization and PQM stabilization. That report explicitly states that using HSI or HSV ensures a unique optimization solution unlike RGB, where manual iteration selection is otherwise required (Nnolim, 2017).

Guided-filter variants modify the classical Retinex surround estimator. The weighted guided image filtering method replaces the Gaussian surround with WGIF to estimate background illumination more accurately near edges, enhances only the HSI intensity channel, computes reflectance by direct division $S_{IR}(x,y)=S_I(x,y)/S_{IL}(x,y)$, applies adaptive gamma correction to illumination, denoises reflectance with WGIF, and restores RGB through a linear gain coefficient. The reported motivation is to reduce local blur, halo artifacts, noise amplification, and color distortion relative to SSR, MSR, and MSRCR (Mu et al., 2018).

Non-local transform-based decomposition provides another classical route. NLHD forms similar pixel groups through block matching and row matching, applies a separable Haar transform, reconstructs illumination from the unique low-frequency coefficient, and reconstructs reflectance from the remaining high-frequency coefficients. Illumination is then enhanced by both an exponential transform and a logarithmic transform, followed by element-wise minimum fusion $I_l^e=\min(I_{l_1}^e,I_{l_2}^e)$. The paper states that this alleviates mosaic artifacts in darker regions and reduces information loss caused by excessive enhancement in brighter regions (Hou et al., 2021).

Not all Retinex-guided methods explicitly solve intrinsic decomposition. Intrinsic Image Transfer formulates illumination manipulation as a local transfer between two illumination surfaces using image-space losses derived from illumination smoothness and reflectance-preserving local linear embedding constraints. After relaxation, the problem reduces to a sparse linear system solved by PCG or Gauss-Seidel, thereby avoiding explicit estimation of latent illumination and reflectance maps while remaining Retinex-inspired (Huang et al., 2021).

RetinexGuI is an explicitly lightweight analytical alternative. It fixes reflectance, refines only illumination, operates on the HSV $I=R\odot L$0 channel, selects cascade depth from the mean $I=R\odot L$1-channel intensity $I=R\odot L$2, and reports $I=R\odot L$3 complexity. On VE-LOL-L, RELLISUR, and LoLi-Street it reports PSNR/SSIM of $I=R\odot L$4, $I=R\odot L$5, and $I=R\odot L$6, respectively, and an average runtime of $I=R\odot L$7 s on 20 RELLISUR images of size $I=R\odot L$8, faster than the compared baselines (Demir et al., 19 Jan 2026).

3. Learned decomposition, latent-space stabilization, and intrinsic priors

A central deep-learning trend is to improve the decomposition itself before enhancement. RGT argues that direct multiplicative RGB-space decomposition causes gradient shrinkage in dark regions because $I=R\odot L$9 and $I=R\circ L$0, and therefore performs decomposition in latent log space with a 1-pixel offset: $I=R\circ L$1. It extracts an illumination prior $I=R\circ L$2, decouples reflectance and illumination with Transformer blocks, and refines the two components in a U-shaped Component Refiner using $I=R\circ L$3 for reflectance and $I=R\circ L$4 for illumination (Zheng et al., 16 Mar 2026).

RGT further introduces the Guidance Fusion Transformer Block, in which self-attention and guidance-aware cross-attention are combined through $I=R\circ L$5. The paper reports that the additive latent decomposition is more stable than multiplicative decomposition: over 10 runs, the additive version achieved mean PSNR $I=R\circ L$6 with variance $I=R\circ L$7, whereas the multiplicative version had mean PSNR $I=R\circ L$8 and variance $I=R\circ L$9. On enhancement benchmarks it reports $I=R\odot L$0 dB / $I=R\odot L$1 SSIM on LOLv2_real, $I=R\odot L$2 dB / $I=R\odot L$3 on LOLv2_syn, $I=R\odot L$4 dB / $I=R\odot L$5 on HDR+(480p), $I=R\odot L$6 dB / $I=R\odot L$7 on SDSD-indoor, and $I=R\odot L$8 dB / $I=R\odot L$9 on SICE, together with the best LPIPS on all five paired benchmarks (Zheng et al., 16 Mar 2026).

InterLight shifts emphasis from decomposition alone to intrinsic illumination priors. It uses HVI color space with $S=\log(I)=\log(R)+\log(L)$0, treats $S=\log(I)=\log(R)+\log(L)$1 as the illumination branch input, and suppresses unreliable dark-region chroma through the density term $S=\log(I)=\log(R)+\log(L)$2. Its Physics-Guided Augmentation perturbs channel-wise gamma with $S=\log(I)=\log(R)+\log(L)$3 and modulates the perturbation strength by brightness through $S=\log(I)=\log(R)+\log(L)$4, $S=\log(I)=\log(R)+\log(L)$5, $S=\log(I)=\log(R)+\log(L)$6 (Wang et al., 19 May 2026).

InterLight also constructs a learned degradation prompt through Adaptive Degradation Prior Generation, injects it by affine modulation and spatial gating, and uses a Luminance-Gated Intrinsic Memory in which dark regions receive stronger memory retrieval than bright regions. The paper reports $S=\log(I)=\log(R)+\log(L)$7 dB PSNR / $S=\log(I)=\log(R)+\log(L)$8 SSIM on LOL-v1, $S=\log(I)=\log(R)+\log(L)$9 dB / $360^\circ$0 on LOL-v2-Real, $360^\circ$1 dB PSNR on LOL-v2-Syn, $360^\circ$2 dB on LSRW-Huawei, and $360^\circ$3 dB on SID. Its ablation table reports baseline $360^\circ$4 dB, then $360^\circ$5 dB from ADPG, $360^\circ$6 dB from LGIM, $360^\circ$7 dB from ICDE, and $360^\circ$8 dB for the full model (Wang et al., 19 May 2026).

RetinexDualV2 extends this logic to generalized UHD restoration by conditioning the Retinex branches on task-specific physical priors. For low-light enhancement it uses a Weberized invariant $360^\circ$9, injects priors through Physical-conditioned Multi-head Self-Attention, and corrects illumination in a Fourier branch. On UHD-LL, its ablation reports $V$0 dB / $V$1 SSIM without the dual branch, $V$2 dB / $V$3 without physical prior, $V$4 dB / $V$5 without PC-MSA, and $V$6 dB / $V$7 for the full model (Kishawy et al., 30 Mar 2026).

4. Attention, state-space, and expert-routed compensation backbones

Several recent systems retain explicit illumination compensation but replace Transformer-heavy or heuristic modules with state-space or expert-routed operators. RetinexMamba divides processing into an Illumination Estimator and a Damage Restorer. The Illumination Estimator fuses the input with an illumination prior $V$8, predicts a 3-channel illumination map $V$9, and forms the illuminated intermediate image $i=l+r$0. The second stage, IFVM, is a U-shaped encoder-decoder built around the Illumination Fusion State Space Model and is explicitly tasked with suppressing noise amplification, color distortion, overexposure, and detail loss introduced by illumination boosting (Bai et al., 2024).

RetinexMamba replaces Retinexformer’s IG-MSA with Illumination-Fused Attention, where illumination features serve as the query while the content feature supplies both key and value. It also replaces self-attention-heavy global modeling with Mamba’s 2D Selective Scan. On LOL-v1 it reports PSNR $i=l+r$1, SSIM $i=l+r$2, RMSE $i=l+r$3; on LOLv2-real it reports PSNR $i=l+r$4, SSIM $i=l+r$5, RMSE $i=l+r$6. The paper states gains over Retinexformer of $i=l+r$7 PSNR on LOL-v1 and $i=l+r$8 PSNR with a $i=l+r$9 RMSE reduction on LOLv2-real (Bai et al., 2024).

ECMamba applies the Retinex split directly to multiple exposure correction. A Retinex estimator predicts approximations to inverse reflectance and inverse illumination, producing two intermediate spaces $S_{IR}(x,y)=S_I(x,y)/S_{IL}(x,y)$0 and $S_{IR}(x,y)=S_I(x,y)/S_{IL}(x,y)$1, and then restores them in two Mamba-based pathways before recomposition $S_{IR}(x,y)=S_I(x,y)/S_{IL}(x,y)$2. The core Retinex-SS2D operator injects Retinex guidance features into the state-space module and uses feature-aware rather than purely directional scanning (Dong et al., 2024).

ECMamba reports $S_{IR}(x,y)=S_I(x,y)/S_{IL}(x,y)$3 PSNR / $S_{IR}(x,y)=S_I(x,y)/S_{IL}(x,y)$4 SSIM on the ME dataset average and $S_{IR}(x,y)=S_I(x,y)/S_{IL}(x,y)$5 PSNR / $S_{IR}(x,y)=S_I(x,y)/S_{IL}(x,y)$6 SSIM on SICE, with stated improvements of $S_{IR}(x,y)=S_I(x,y)/S_{IL}(x,y)$7 dB / $S_{IR}(x,y)=S_I(x,y)/S_{IL}(x,y)$8 SSIM on ME and $S_{IR}(x,y)=S_I(x,y)/S_{IL}(x,y)$9 dB / $I_l^e=\min(I_{l_1}^e,I_{l_2}^e)$0 SSIM on SICE over the second-best method. On LOLv1, LOLv2-real, and LOLv2-synthetic it reports $I_l^e=\min(I_{l_1}^e,I_{l_2}^e)$1, $I_l^e=\min(I_{l_1}^e,I_{l_2}^e)$2, and $I_l^e=\min(I_{l_1}^e,I_{l_2}^e)$3. The paper also states that it outperforms LLFlow-SKF by about $I_l^e=\min(I_{l_1}^e,I_{l_2}^e)$4 dB on average PSNR while using only $I_l^e=\min(I_{l_1}^e,I_{l_2}^e)$5M parameters, about $I_l^e=\min(I_{l_1}^e,I_{l_2}^e)$6 of LLFlow-SKF’s parameters (Dong et al., 2024).

DIME-Net generalizes Retinex-guided compensation to both low-light and backlit images. It models illumination estimation as a Mixture-of-Experts over 16 S-curve expert networks, uses Top-$I_l^e=\min(I_{l_1}^e,I_{l_2}^e)$7 sparse gating, and interprets $I_l^e=\min(I_{l_1}^e,I_{l_2}^e)$8 experts as dark-region boosting and $I_l^e=\min(I_{l_1}^e,I_{l_2}^e)$9 experts as bright-region compression. Residual artifact correction is delegated to a U-Net with Illumination-Aware Cross Attention and Sequential-State Global Attention (Wang et al., 19 Aug 2025).

On MixBL, DIME-Net reports PSNR $I=R\odot L$00, SSIM $I=R\odot L$01, LPIPS $I=R\odot L$02; on LOLv1, when trained only on MixBL, it reports $I=R\odot L$03, $I=R\odot L$04, $I=R\odot L$05; on BAID it reports $I=R\odot L$06, $I=R\odot L$07, $I=R\odot L$08. Its ablation states that removing the illumination estimator lowers PSNR by about $I=R\odot L$09 dB on LOLv1 and $I=R\odot L$10 dB on BAID, while removing the damage restorer drops LOLv1 PSNR to $I=R\odot L$11 and SSIM to $I=R\odot L$12 (Wang et al., 19 Aug 2025).

5. Diffusion, flow, and training-free illumination control

Generative formulations use Retinex guidance to separate exposure correction from detail synthesis or to control lighting directly. Reti-Diff trains a Retinex-based latent diffusion model in two Siamese branches, one for reflectance priors and one for illumination priors, then injects those priors into RGformer through Retinex-guided multi-head cross attention and dynamic feature aggregation. Its central claim is that diffusion should generate compact Retinex priors rather than the final image directly (He et al., 2023).

Reti-Diff reports $I=R\odot L$13 PSNR, $I=R\odot L$14 SSIM, FID $I=R\odot L$15, BIQE $I=R\odot L$16 on LOL-v1; $I=R\odot L$17, $I=R\odot L$18, $I=R\odot L$19, $I=R\odot L$20 on LOL-v2-real; $I=R\odot L$21, $I=R\odot L$22, $I=R\odot L$23, $I=R\odot L$24 on LOL-v2-syn; and $I=R\odot L$25, $I=R\odot L$26, $I=R\odot L$27, $I=R\odot L$28 on SID. The paper states that it outperforms the second-best method by $I=R\odot L$29 on average in low-light enhancement and that RLDM converges rapidly with only 4 diffusion steps (He et al., 2023).

ZeroIDIR is zero-reference and explicitly decouples adaptive illumination correction from diffusion-based reconstruction. It performs Retinex decomposition $I=R\odot L$30, predicts two gamma maps and two spatial weight maps in an Adaptive Gamma Correction Module, constructs $I=R\odot L$31, and then forms $I=R\odot L$32 as an intermediate state for a Perturbed Consistency Diffusion Model. The illumination branch is regularized by a histogram-guided loss $I=R\odot L$33, where the prior is built from around $I=R\odot L$34 normal-light images (Jiang et al., 12 May 2026).

IllumFlow separates illumination and reflectance, but models illumination change as a continuous trajectory through conditional rectified flow. Using a pretrained Transformer Decomposition Network from Diff-Retinex, it decomposes $I=R\odot L$35 and $I=R\odot L$36 into $I=R\odot L$37, then learns the illumination velocity $I=R\odot L$38. At inference it supports both one-step enhancement $I=R\odot L$39 and progressive updates, while reflectance is denoised separately by CRFR (Wei et al., 4 Nov 2025).

IllumFlow reports PSNR $I=R\odot L$40, SSIM $I=R\odot L$41, LPIPS $I=R\odot L$42 on LOLv1 and PSNR $I=R\odot L$43, SSIM $I=R\odot L$44, LPIPS $I=R\odot L$45 on LOLv2 real-captured. The paper emphasizes continuous, bidirectional brightness control, including illumination suppression by moving $I=R\odot L$46 downward, and states that it provides much faster inference than CLE-Diffusion (Wei et al., 4 Nov 2025).

Retinex-Diffusion addresses illumination control rather than paired restoration. It treats a pretrained diffusion model as a black-box renderer, extracts a multi-scale Retinex illumination estimate from the current denoised sample, decomposes guidance into illumination-related and reflectance-related energies, and steers reverse diffusion toward Gaussian-mixture illumination prompts. For real-image relighting it adds cross-color ratios as an illumination-invariant reflectance cue. The reported effects include cast shadow, soft shadow, inter-reflections, light direction control, and lamp activation, all without additional training, intrinsic-decomposition supervision, or latent-direction search (Xing et al., 2024).

6. Continuous representations, 3D synthesis, and cross-domain extensions

Retinex-guided illumination compensation is increasingly used outside conventional 2D image enhancement. MERID-GS moves low-light handling out of the 3D representation and into an explicit scene-adaptive image-domain stage for sparse-view $I=R\odot L$47 novel view synthesis. It converts RGB to YUV, models only the luminance channel $I=R\odot L$48, defines spatial-domain reflectance as $I=R\odot L$49, and predicts a bounded gain

$I=R\odot L$50

with $I=R\odot L$51 and $I=R\odot L$52. It then constructs an illumination-state feature $I=R\odot L$53, uses Illumination-State-Guided Frequency Gating to modulate value features by frequency band, applies a lightweight Reflection Head $I=R\odot L$54, and feeds the corrected reflectance views into 3D Gaussian Splatting (Yin et al., 27 Apr 2026).

MERID-GS states that it can adapt to unseen scenes using only about 10 normally lit input views and synthesize $I=R\odot L$55 views in roughly 5 minutes end-to-end. Its ablation reports that removing IS-FGA reduces PSNR by about $I=R\odot L$56 dB, while the full model achieves average zero-shot performance of PSNR $I=R\odot L$57 dB, SSIM $I=R\odot L$58, and LPIPS $I=R\odot L$59 across NeRF360, LOM_full, and LLD. The paper attributes its gains to explicit illumination-reflectance decoupling, bounded structure-aware gain, illumination-aware frequency gating, and lightweight scene adaptation (Yin et al., 27 Apr 2026).

CGS-Retinex likewise adopts explicit-implicit joint modeling, but for 2D enhancement. It estimates a spatially continuous illumination field with continuous Gaussian splatting, computes a base reflectance by division $I=R\odot L$60, and restores residual reflectance with an implicit neural representation guided by Fourier positional encoding and shallow high-frequency features. It further regularizes the system with grayscale brightness consistency, anisotropic TV on illumination, perceptual loss, and color loss (Chen et al., 15 Jun 2026).

Cross-domain applications preserve the same decomposition logic. The underwater enhancement framework uses a hybrid illumination model that combines Gamma correction, CLAHE, and Retinex, then applies Wavelet-Guided Adaptive Filtering and adaptive color compensation through RCP, DCP, and MUDCP. The paper states that on U45 the method achieves the best results across UCIQE, UIQM, and UISM, but does not provide a standalone ablation isolating the Retinex block (Tian et al., 5 Mar 2025).

A related complex-illumination framework estimates illumination with GDWGIF, processes both the original image and its inverted version, decomposes into illumination and reflection, corrects illumination by adaptive gamma, denoises reflection with GDWGIF, and then fuses under-exposed, over-exposed, and original reconstructions before linear stretching. It reports the best average NIQE of $I=R\odot L$61, average NIQMC of $I=R\odot L$62, and denoising scores of SSIM $I=R\odot L$63 and PSNR $I=R\odot L$64 on synthetic noisy images (Tao et al., 9 Dec 2025).

7. Recurring design patterns, misconceptions, and open issues

A recurrent misconception is that Retinex-guided illumination compensation is equivalent to global brightening. The cited works do not support that reading. Even the simplest methods separate illumination and reflectance conceptually, and more recent systems add bounded gain, edge-aware filtering, physical priors, dual-branch restoration, frequency gating, or explicit damage restoration because brightness amplification alone tends to amplify noise, distort color, or wash out details (Demir et al., 19 Jan 2026, Bai et al., 2024, Kishawy et al., 30 Mar 2026).

A second misconception is that explicit intrinsic decomposition is mandatory. Several methods remain Retinex-guided while avoiding full intrinsic-image recovery: IIT derives image-space losses and solves a closed-form linear system without explicit decomposition, and Retinex-Diffusion uses a multi-scale Retinex approximation as a guidance signal inside diffusion sampling rather than training an intrinsic network (Huang et al., 2021, Xing et al., 2024). This suggests that the Retinex contribution can lie either in explicit factorization or in the structure of the guidance prior.

The literature also presents divergent choices about the illumination domain. MERID-GS avoids a logarithmic Retinex formulation because it is described as numerically unstable in very dark pixels, whereas RGT uses a log transform plus a 1-pixel offset precisely to stabilize decomposition and training in latent space (Yin et al., 27 Apr 2026, Zheng et al., 16 Mar 2026). These are not mutually exclusive claims; they reflect different numerical strategies for different modules and tasks.

Efficiency remains a major axis of differentiation. RetinexGuI emphasizes $I=R\odot L$65 per-pixel complexity and the best runtime in its experiments, ECMamba stresses linear-time selective state-space modeling with a 1.75M-parameter model, and RLDM-based Reti-Diff reports convergence with only 4 diffusion steps. By contrast, IIT explicitly notes that it is time-consuming because it requires large-scale LLE weights and a sparse linear solve (Demir et al., 19 Jan 2026, Dong et al., 2024, He et al., 2023, Huang et al., 2021).

Open issues are stated directly in several papers. RetinexGuI notes that robustness under extreme illumination conditions remains open; the WGIF-based HSI method states that enhancement in very dark local regions is still limited when illumination is very uneven; IIT depends on an exemplar and is time-consuming; ZeroIDIR depends on an empirical histogram prior and on the usefulness of the Retinex decomposition; and supervised low-light enhancement models are described in MERID-GS as often requiring retraining and struggling to generalize to new scenes (Demir et al., 19 Jan 2026, Mu et al., 2018, Huang et al., 2021, Jiang et al., 12 May 2026, Yin et al., 27 Apr 2026).

Taken together, the literature indicates that Retinex-guided illumination compensation has evolved from surround estimation and variational optimization into a broad design principle: explicitly or implicitly isolate illumination, constrain its correction by physical or structural priors, and decouple low-frequency exposure normalization from high-frequency reflectance recovery. A plausible implication is that future progress will continue to depend less on brightness amplification itself than on how illumination priors are represented, regularized, and coupled to detail restoration across 2D, generative, and 3D settings.