An MLLM-guided architecture with a mixture of frequency experts and relational alignment loss achieves state-of-the-art all-in-one image restoration, outperforming prior methods by up to 1.35 dB on the CDD11 dataset.
hub Baseline reference
Deep Retinex Decomposition for Low-Light Enhancement
Baseline reference. 78% of citing Pith papers use this work as a benchmark or comparison.
51
Pith papers citing it
Baseline
78% of classified citations
abstract
Retinex model is an effective tool for low-light image enhancement. It assumes that observed images can be decomposed into the reflectance and illumination. Most existing Retinex-based methods have carefully designed hand-crafted constraints and parameters for this highly ill-posed decomposition, which may be limited by model capacity when applied in various scenes. In this paper, we collect a LOw-Light dataset (LOL) containing low/normal-light image pairs and propose a deep Retinex-Net learned on this dataset, including a Decom-Net for decomposition and an Enhance-Net for illumination adjustment. In the training process for Decom-Net, there is no ground truth of decomposed reflectance and illumination. The network is learned with only key constraints including the consistent reflectance shared by paired low/normal-light images, and the smoothness of illumination. Based on the decomposition, subsequent lightness enhancement is conducted on illumination by an enhancement network called Enhance-Net, and for joint denoising there is a denoising operation on reflectance. The Retinex-Net is end-to-end trainable, so that the learned decomposition is by nature good for lightness adjustment. Extensive experiments demonstrate that our method not only achieves visually pleasing quality for low-light enhancement but also provides a good representation of image decomposition.
hub tools
citation-role summary
dataset 6
background 2
baseline 1
citation-polarity summary
representative citing papers
LL-Bench supplies a human-annotated dataset exposing generative model weaknesses in low-level restoration and introduces LL-Score as an MLLM evaluator that outperforms existing quality metrics and can serve as a training reward.
DRDD decouples diffusion into independent noise and residual stages to preserve domain harmonization and enable unified data-efficient I2I translation.
GGT-100K is a 103k-pair LQ-HQ dataset generated via MFMs to enhance real-world generalization of image restoration models.
EIC-LIE uses an event-illumination collaborative module and illumination-aware event filter plus a new real-world dataset to improve low-light image enhancement over prior methods.
DiSI disentangles stochastic interpolants into separate generation and regression paths, allowing controllable transitions between regression and generative image restoration with a unified few-step sampler.
RL-AWB uses reinforcement learning to optimize parameters of a statistical white-balance estimator for nighttime scenes and reports better generalization on a new multi-sensor dataset.
UnfoldLDM integrates multi-granularity degradation-aware unfolding with degradation-resistant latent diffusion priors and an over-smoothing correction transformer to achieve leading performance on blind image restoration tasks.
DACG-IR adds a lightweight degradation-aware module that generates prompts to adaptively gate attention temperature, output features, and spatial-channel fusion in an encoder-decoder network for unified image restoration.
A new framework called ERR decomposes UHD image restoration into three frequency stages with specialized sub-networks and introduces the LSUHDIR benchmark dataset of over 82,000 images.
M3D-Stereo supplies 7904 aligned stereo pairs across four multi-degradation scenarios with six progressive levels and pixel-consistent ground truths to benchmark image restoration and stereo matching.
Pre-trained diffusion models inherently support image restoration that can be unlocked by optimizing prompt embeddings at the text encoder output using a diffusion bridge formulation, achieving competitive results on models like WAN and FLUX without fine-tuning.
IQPIR uses NR-IQA-derived quality scores to condition a Transformer and dual-branch codebook for perceptually superior real-world image restoration.
EvLIR processes ordered event voxels with a lightweight ConvGRU-based TERM to generate bounded illumination corrections that improve Retinex estimation and reliability-aware restoration, topping eleven of twelve benchmark metrics.
AIGS-Net builds an input-adaptive 2D Gaussian Splatting illumination field modulated by luminance statistics, rendered via alpha compositing, plus a zero-parameter multiscale encoder and regularizers to enhance low-light images on LOL and LSRW.
Introduces self-supervised LLIE using internal low-frequency pseudo ground-truth, dual-domain structural constraints, and gain-adaptive modulation to achieve superior noise suppression and detail preservation.
SP-MoMamba uses superpixels to drive content-aware state space modeling and multi-scale mixture-of-experts for efficient single-image super-resolution.
LUMEN enhances low-light images via depth estimation, soft clustering for virtual flash simulation, and attention-based transformer fusion, reporting state-of-the-art results on LOL-v1 and LOL-v2 benchmarks.
The paper proposes the Degradation Frequency Curve (DFC) as an explicit spectral representation for quantifying degradations and develops a DFC-guided multi-scale restorer that achieves state-of-the-art performance on composite and real-world benchmarks.
RIDE applies Retinex-based homogeneous decomposition to improve foreground-background discriminability in concealed object segmentation tasks across multiple domains.
IG-Diff adds an illumination-guided module to a diffusion model and supplies new paired datasets to restore images degraded by simultaneous low light and other factors while preserving texture.
PVRF combines zero-shot VLM-based weather perception with perception-adaptive rectified flow refinement to achieve all-in-one adverse weather removal with improved fidelity and cross-dataset generalization.
Consist-Retinex achieves one-step Retinex enhancement via a Retinex Transformer decomposition network and conditional consistency models trained with noise-emphasized dual objectives that align trajectory consistency to ground-truth components.
RDBM reformulates generalized diffusion bridge SDEs to use distribution residuals for adaptive noise modulation, unifying prior bridge models as special cases and achieving SOTA on image restoration tasks.
citing papers explorer
-
Leveraging Multimodal Large Language Models for All-in-One Image Restoration via a Mixture of Frequency Experts
An MLLM-guided architecture with a mixture of frequency experts and relational alignment loss achieves state-of-the-art all-in-one image restoration, outperforming prior methods by up to 1.35 dB on the CDD11 dataset.
-
LL-Bench: Rethinking Low-Level Vision Evaluation in the Era of Large-Scale Generative Models
LL-Bench supplies a human-annotated dataset exposing generative model weaknesses in low-level restoration and introduces LL-Score as an MLLM evaluator that outperforms existing quality metrics and can serve as a training reward.
-
Decoupled Residual Denoising Diffusion Models for Unified and Data Efficient Image-to-Image Translation
DRDD decouples diffusion into independent noise and residual stages to preserve domain harmonization and enable unified data-efficient I2I translation.
-
GGT-100K: Generative Ground Truth for Generalizable Real-World Image Restoration
GGT-100K is a 103k-pair LQ-HQ dataset generated via MFMs to enhance real-world generalization of image restoration models.
-
Event-Illumination Collaborative Low-light Image Enhancement with a High-resolution Real-world Dataset
EIC-LIE uses an event-illumination collaborative module and illumination-aware event filter plus a new real-world dataset to improve low-light image enhancement over prior methods.
-
Disentangling Generation and Regression in Stochastic Interpolants for Controllable Image Restoration
DiSI disentangles stochastic interpolants into separate generation and regression paths, allowing controllable transitions between regression and generative image restoration with a unified few-step sampler.
-
RL-AWB: Deep Reinforcement Learning for Auto White Balance Correction in Low-Light Night-time Scenes
RL-AWB uses reinforcement learning to optimize parameters of a statistical white-balance estimator for nighttime scenes and reports better generalization on a new multi-sensor dataset.
-
UnfoldLDM: Degradation-Aware Unfolding with Iterative Latent Diffusion Priors for Blind Image Restoration
UnfoldLDM integrates multi-granularity degradation-aware unfolding with degradation-resistant latent diffusion priors and an over-smoothing correction transformer to achieve leading performance on blind image restoration tasks.
-
Degradation-Aware Adaptive Context Gating for Unified Image Restoration
DACG-IR adds a lightweight degradation-aware module that generates prompts to adaptively gate attention temperature, output features, and spatial-channel fusion in an encoder-decoder network for unified image restoration.
-
From Zero to Detail: A Progressive Spectral Decoupling Paradigm for UHD Image Restoration with New Benchmark
A new framework called ERR decomposes UHD image restoration into three frequency stages with specialized sub-networks and introduces the LSUHDIR benchmark dataset of over 82,000 images.
-
M3D-Stereo: A Multiple-Medium and Multiple-Degradation Dataset for Stereo Image Restoration
M3D-Stereo supplies 7904 aligned stereo pairs across four multi-degradation scenarios with six progressive levels and pixel-consistent ground truths to benchmark image restoration and stereo matching.
-
Your Pre-trained Diffusion Model Secretly Knows Restoration
Pre-trained diffusion models inherently support image restoration that can be unlocked by optimizing prompt embeddings at the text encoder output using a diffusion bridge formulation, achieving competitive results on models like WAN and FLUX without fine-tuning.
-
Beyond Ground-Truth: Leveraging Image Quality Priors for Real-World Image Restoration
IQPIR uses NR-IQA-derived quality scores to condition a Transformer and dual-branch codebook for perceptually superior real-world image restoration.
-
EvLIR: Learning Illumination Residuals from Ordered Events for Low-Light Image Enhancement
EvLIR processes ordered event voxels with a lightweight ConvGRU-based TERM to generate bounded illumination corrections that improve Retinex estimation and reliability-aware restoration, topping eleven of twelve benchmark metrics.
-
AIGS-Net: Compact Illumination Field Modeling via 2D Gaussian Splatting for Fast Low-Light Image Enhancement
AIGS-Net builds an input-adaptive 2D Gaussian Splatting illumination field modulated by luminance statistics, rendered via alpha compositing, plus a zero-parameter multiscale encoder and regularizers to enhance low-light images on LOL and LSRW.
-
Internally Referenced Low-Light Enhancement
Introduces self-supervised LLIE using internal low-frequency pseudo ground-truth, dual-domain structural constraints, and gain-adaptive modulation to achieve superior noise suppression and detail preservation.
-
SP-MoMamba: Superpixel-driven Mixture of State Space Experts for Efficient Image Super-Resolution
SP-MoMamba uses superpixels to drive content-aware state space modeling and multi-scale mixture-of-experts for efficient single-image super-resolution.
-
Degradation Frequency Curve: An Explicit Frequency-Quantified Representation for All-in-One Image Restoration
The paper proposes the Degradation Frequency Curve (DFC) as an explicit spectral representation for quantifying degradations and develops a DFC-guided multi-scale restorer that achieves state-of-the-art performance on composite and real-world benchmarks.
-
RIDE: Retinex-Informed Decoupling for Exposing Concealed Objects
RIDE applies Retinex-based homogeneous decomposition to improve foreground-background discriminability in concealed object segmentation tasks across multiple domains.
-
IG-Diff: Complex Night Scene Restoration with Illumination-Guided Diffusion Model
IG-Diff adds an illumination-guided module to a diffusion model and supplies new paired datasets to restore images degraded by simultaneous low light and other factors while preserving texture.
-
PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
PVRF combines zero-shot VLM-based weather perception with perception-adaptive rectified flow refinement to achieve all-in-one adverse weather removal with improved fidelity and cross-dataset generalization.
-
Consist-Retinex: One-Step Noise-Emphasized Consistency Training Accelerates High-Quality Retinex Enhancement
Consist-Retinex achieves one-step Retinex enhancement via a Retinex Transformer decomposition network and conditional consistency models trained with noise-emphasized dual objectives that align trajectory consistency to ground-truth components.
-
Residual Diffusion Bridge Model for Image Restoration
RDBM reformulates generalized diffusion bridge SDEs to use distribution residuals for adaptive noise modulation, unifying prior bridge models as special cases and achieving SOTA on image restoration tasks.
-
Adapting Large VLMs with Iterative and Manual Instructions for Generative Low-light Enhancement
VLM-IMI adapts VLMs with iterative and manual instructions plus a learnable fusion module to guide diffusion-based generative low-light image enhancement, outperforming prior methods in perceptual quality.
-
SIMI: Self-information Mining Network for Low-light Image Enhancement
SIMI is an unsupervised low-light image enhancement network using bit-plane decomposition to mine self-information, reported to reach state-of-the-art performance on standard benchmarks.
-
Beyond Pixel Fidelity: Minimizing Perceptual Distortion and Color Bias in Night Photography Rendering
pHVI-ISPNet achieves state-of-the-art perceptual quality in night photography rendering by combining HVI color space with wavelet feature propagation, sample-adaptive losses, and distribution-based color constancy on the NTIRE 2025 NPR dataset.
-
Reading in the Dark: Low-light Scene Text Recognition
Introduces LSTR and ESTR low-light text datasets and shows joint LLIE-OCR training outperforms standalone models.
-
Frequency-Decomposed INR for NIR-Assisted Low-Light RGB Image Denoising
FDINR decomposes RGB-NIR pairs into frequency components via wavelets and employs dual-branch INR with cross-modal supervision and adaptive uncertainty loss to restore low-light images while enabling arbitrary-resolution output.
-
RHVI-FDD: A Hierarchical Decoupling Framework for Low-Light Image Enhancement
RHVI-FDD hierarchically decouples luminance-chrominance and then frequency components in low-light images to correct color, suppress noise, and preserve details better than prior methods.
-
Hidden-Shot: Towards One-Shot Task Generalization for Low-Level Vision Generalist Models
Hidden-Shot adds an implicit visual-task prompt and selective merging step to existing low-level vision generalist models, paired with a 3C4U/3C7U evaluation framework that reports outperformance on seven and ten datasets respectively.
-
LUMINA-26: Low-Light Understanding for Modeling and Interpreting Night-time Actions
Introduces LUMINA-26 low-light action dataset and Illumi-Net model achieving 75.95% Top-1 accuracy on it while surpassing prior SOTA on ELLAR.
-
Enlight: Fast Low-Light Image Enhancement via Multi-Objective Optimization and Shadow-Aware Refinement
ENLIGHT is a zero-shot optimization framework for low-light image enhancement using global illumination adjustment followed by shadow-aware local refinement to achieve competitive quality with lower inference time.
-
Spiking Pyramid Wavelet Transformation for High-efficient and Low-energy Image Restoration
SPWM introduces spiking dual pyramid wavelet blocks to lower computational costs and energy use in image restoration while keeping quality comparable to prior methods.
-
Gaussian Light Field Splatting: A Physical Prior-Driven Vision Transformer for Unsupervised Low-Light Image Enhancement
GLFS represents illumination via anisotropic Gaussian basis functions, adds physics-guided biases to self-attention in a Vision Transformer, and introduces color-vector angular and luminance-edge losses to achieve SOTA unsupervised low-light enhancement.
-
Learning Reference-Guided Exposure Correction with Hybrid Illumination Characteristics
HICNet is a reference-guided exposure correction network that distills images into illumination embeddings, uses their differences to drive FiLM-based modulation and photometric channel rebalancing, and employs cross-batch contrastive loss, all trained without ground truth.
-
M2Retinexformer: Multi-Modal Retinexformer for Low-Light Image Enhancement
M2Retinexformer improves low-light images by progressively refining RGB data with depth, luminance, and semantic modalities through cross-attention and adaptive gating, showing gains on LOL, SID, SMID, and SDSD benchmarks.
-
TPGDiff: Hierarchical Triple-Prior Guided Diffusion for Image Restoration
TPGDiff introduces hierarchical triple-prior guidance in a diffusion network, placing degradation priors throughout, structural priors in shallow layers, and semantic priors in deep layers for improved all-in-one image restoration.
-
Towards Lightest Low-Light Image Enhancement Architecture for Mobile Devices
LiteIE proposes a two-layer backbone-agnostic feature extractor and parameter-free Iterative Restoration Module for unsupervised low-light enhancement, claiming 19.04 dB PSNR on LOL with 0.07% of SOTA parameters and 30 FPS 4K on Snapdragon 8 Gen 3.
-
Unifying Deep Stochastic Processes for Image Enhancement
Stochastic image enhancement methods are shown to be variants of a shared SDE differing in drift, diffusion, terminal distributions and boundary conditions, with controlled experiments revealing no single dominant family and a new modular library released.
-
SmartPhotoCrafter: Unified Reasoning, Generation and Optimization for Automatic Photographic Image Editing
SmartPhotoCrafter performs automatic photographic image editing by coupling an Image Critic module that identifies deficiencies with a Photographic Artist module that generates edits, trained via multi-stage pretraining, reasoning supervision, and reinforcement learning.
-
Naka-GS: A Bionics-inspired Dual-Branch Naka Correction and Progressive Point Pruning for Low-Light 3DGS
NAKA-GS combines bionics-inspired Naka chroma correction with progressive point pruning to boost restoration quality and efficiency in low-light 3D Gaussian Splatting.
-
Deep Light Pollution Removal in Night Cityscape Photographs
A deep learning method with an enhanced physical degradation model incorporating anisotropic light spread and hidden skyglow, trained via generative models and synthetic-real coupling, removes light pollution from night cityscape images more effectively than prior restoration techniques.
-
Lighting-Aware Representation Learning under Controllable Lighting Variation
Lighting-aware extension of contrastive learning adds an auxiliary objective for illumination variation and reports improved classification and detection performance on ImageNet, ExDark, and PASCAL VOC.
-
FLARE-BO: Fused Luminance and Adaptive Retinex Enhancement via Bayesian Optimisation for Low-Light Robotic Vision
FLARE-BO uses Bayesian optimization over an eight-parameter space to fuse luminance and adaptive Retinex techniques, reporting marked improvements on the LOL low-light dataset compared to untrained baselines.
-
ELoG-GS: Dual-Branch Gaussian Splatting with Luminance-Guided Enhancement for Extreme Low-light 3D Reconstruction
ELoG-GS integrates geometry-aware initialization and luminance-guided photometric adaptation into Gaussian Splatting, achieving PSNR 18.66 and SSIM 0.69 on the NTIRE 2026 Track 1 low-light 3D reconstruction benchmark.
-
Attention Is not Everything: Efficient Alternatives for Vision
A survey that taxonomizes non-Transformer vision models and evaluates their practical trade-offs across efficiency, scalability, and robustness.
-
Low Light Image Enhancement Challenge at NTIRE 2026
Report on the NTIRE 2026 Low Light Image Enhancement Challenge that evaluates 22 team submissions for joint denoising and enhancement on a new dataset.
- E-VLA: Event-Augmented Vision-Language-Action Model for Dark and Blurred Scenes