Source-linked AI summary
Low Light Image Enhancement Challenge at NTIRE 2026
George Ciubotariu, Sharif S M A, Abdur Rehman, Fayaz Ali Dharejo, Rizwan Ali Naqvi, Marcos V. Conde, Radu Timofte, Zhi Jin, Hongjun Wu, Wenjian Zhang, Chang Ye, Xunpeng Yi, Qinglong Yan, Yibing Zhang, Zaynab Ali, Saiprasad Meesiyawar, Varda I Pattanshetty, Varsha I Pattanshetty, Nikhil Akalwadi, Padmashree Desai, Ramesh Ashok Tabib, Uma Mudenagudi, Hao Yang, Ruikun Zhang, Liyuan Pan, Furkan Kınlı, Donghun Ryou, Inju Ha, Junoh Kang, Bohyung Han, Wei Zhou, Yuval Haitman, Ariel Lapid, Reuven Peretz, Idit Diamant, Leilei Cao, Shuo Zhang, Praful Hambarde, Prateek Shaily, Jayant Kumar, Hardik Sharma, Aashish Negi, Sachin Chaudhary, Akshay Dudhane, Amit Shukla, MoHao Wu, Lin Wang, Jiachen Tu, Guoyi Xu, Yaoxin Jiang, Jiajia Liu, Yaokun Shi, Raul Balmez, Alexandru Brateanu, Ciprian Orhei, Cosmin Ancuti, Codruta O. Ancuti, Bilel Benjdira, Anas M. Ali, Wadii Boulila, Kaifan Qiao, Bofei Chen, Jingyi Xu, Duo Zhang, Xin Deng, Mai Xu, Shengxi Li, Lai Jiang, Harini A, Ananya N, Lakshanya K, Ying Xu, Xinyi Zhu, Shijun Shi, Jiangning Zhang, Yong Liu, Kai Hu, Jing Xu, Xianfang Zeng, Jinao Song, Guangsheng Tang, Cheng Li, Yuqiang Yang, Ziyi Wang, Yan Chen, Long Bao, Heng Sun, Mohab Kishawy, Jun Chen, Wan-Chi Siu, Yihao Cheng, Hon Man Hammond Lee, Chun-Chuen Hui
TL;DR
Low-light enhancement needs robust evaluation because existing datasets are constrained and real images contain severe degradation. This paper reviews a challenge built on diverse smartphone data, its tracks and evaluation, and participating solutions. The challenge advances real-world assessment, while extreme conditions still produce artifacts and expose a gap between benchmark metrics and practical performance.
Problem
Existing low-light datasets are limited and controlled, making it difficult to evaluate or train models for complex real-world images with degradation, detail loss, and poor contrast.
Method
The paper reviews the NTIRE 2026 challenge, which evaluates denoised enhancement and joint denoising-enhancement using a diverse smartphone dataset and analyzes submitted architectures and results.
Results
The review reports substantial solution development across the challenge, including LSD pretraining contributing +4.51 dB and progressive enhancement adding +0.60 dB in BAPE.
Takeaways & Limitations
The challenge shifts low-light enhancement evaluation toward uncontrolled real-world data and supports comparison of methods addressing visibility, color, noise, and structure.
Takeaways & Limitations
Current methods still produce artifacts in extreme conditions, indicating a gap between benchmark metrics and real-world performance.
Abstract
from arXiv · showhide
This paper presents a comprehensive review of the NTIRE 2026 Low Light Image Enhancement Challenge, highlighting the proposed solutions and final results. The objective of this challenge is to identify effective networks capable of producing clearer and visually compelling images in diverse and challenging conditions by learning representative visual cues with the purpose of restoring information loss due to low-contrast and noisy images. A total of 195 participants registered for the first track and 153 for the second track of the competition, and 22 teams ultimately submitted valid entries. This paper thoroughly evaluates the state-of-the-art advances in (joint denoising and) low-light image enhancement, showcasing the significant progress in the field, while leveraging samples of our novel dataset.
1. Introduction
Low-light image enhancement remains difficult because real images suffer degradation, detail loss, and poor contrast, while existing datasets are limited and controlled. The challenge addresses this gap with two real-world tracks built on diverse smartphone captures spanning extreme lighting conditions.
- Low-light enhancement supports applications from consumer photography to autonomous navigation, but archived images often have severe degradation, detail loss, and poor contrast.
- Existing SLLIE datasets contain few samples from controlled or artificially dimmed settings, limiting generalization to complex real-world environments.
- The challenge uses the Low-light Smartphone Dataset, captured with 15 smartphone cameras across uncontrolled environments and 0.1–200 lux lighting.
- Track 1 enhances visibility and color in denoised low-light images, whereas Track 2 jointly performs low-light denoising and enhancement while preserving authentic sensor noise.
2. Challenge
The challenge provides large-scale patch training data, paired and unpaired test images, and a two-phase evaluation process. Final rankings combine normalized objective metrics, with weighting determined separately for reference and non-reference tracks.
- Training data comprises 115,376 Track 1 patches and 64,318 Track 2 patches extracted from 6,025 high-resolution scenes.
- Each track includes 24 high-resolution validation pairs, 26 paired test images for quantitative evaluation, and 60 unpaired images for perceptual assessment.
- The challenge first provides training data and validation inputs with server feedback, then evaluates submissions on hidden test images without ground truth.
- Final scores use Min-Max-normalized metrics, equally weighting PSNR, SSIM, and LPIPS in the Reference Track.
- The Non-Reference Track assigns 50.00% to Mean Opinion Score and 16.67% each to NIQUE, MUSIQUE, and CLIPIA.
3. Challenge Methods
The challenge methods span ensemble fusion, hierarchical refinement, multi-scale representation, dynamic expert routing, wavelet-based propagation, domain adaptation, and physically grounded illumination modeling. These solutions target robustness, detail preservation, brightness consistency, and generalization across real low-light conditions.
- Architectural strategies: The challenge report compares independently trained enhancement models, weighted fusion, hierarchical refinement, multi-scale encoding, dynamic experts, and globally conditioned patch processing.SYSU-FVL fuses ESDNet and DarkIR; WHU-MVP uses coarse enhancement as guidance; other teams target adaptive local restoration and patch-consistent brightness.
- BAPE: 0.75 dB is the remaining PSNR gap after progressive patch escalation reduces the patch-to-full-image discrepancy from 3.7 dB.The schedule increases patch size from 384 to 512 to 768, while multi-source bridging uses 200 full-resolution pairs from four LSD subsets.
- BAPE: +4.51 dB comes from LSD pretraining, while progressive patch escalation adds +0.60 dB and multi-source bridging narrows independent-validation versus competition-validation performance to 0.75 dB.Inference uses 768×768 tiled processing with 128-pixel overlap, Hanning blending, and 8× geometric test-time augmentation.
- Physically grounded restoration: RetinexDualV2 adds a task-specific physical grounding module that extracts degradation-aware, structure-aware illumination priors from noisy low-light inputs.The method retains dual reflectance and illumination branches while introducing physical vectors for restoration guidance.
- Attention-based restoration: Gated LCA adaptively weights attention-branch features to emphasize information reconstruction while reducing noise disturbance.The module is designed to adjust feature importance rather than treating all extracted features equally.
4. Discussion and Conclusion
The challenge review highlights diverse low-light enhancement solutions while noting that current methods still produce artifacts under extreme conditions. Future editions aim to better align benchmark evaluation with real-world performance.
- Discussion and Conclusion: Current methods still produce artifacts in extreme conditions, exposing a gap between benchmark metrics and real-world performance.The challenge shifted evaluation from synthesized to real-world data, but this gap remains.
- Challenge Solutions: The reviewed contributions span hierarchical enhancement, expert routing, wavelet processing, brightness awareness, latent flow matching, diffusion, and physics-guided dual-domain modeling.These approaches are represented by the listed challenge submissions and their corresponding method titles.
- Related Work: The paper also covers methods for joint denoising and low-light enhancement, including diffusion-based and physically grounded restoration approaches.Examples include Exploring Large Diffusion Models for Joint Denoising and Low-Light Enhancement and Physically-Grounded Dual Retinex for Generalized UHD Image Restoration.
B. Implementation Details
Implementation details describe progressive training and full-resolution or patch-based processing across submitted systems. The reviewed materials also include architectural summaries for the UHDM model.
- Training Details: Both ESDNet and DarkIR use progressive training with 512 × 512 crops initially and larger patches during later fine-tuning.ESDNet and DarkIR are trained from scratch with Adam and cyclic cosine annealing.
- Training Details: ESDNet is trained for 291,500 iterations, while DarkIR is trained for 301,500 iterations.Their fine-tuning stages use progressively larger patch sizes and adjusted batch sizes.
- Training Details: One implementation trains on full-resolution LSD images using patch sizes from 960 to 1600 and batch sizes from 4 to 1 across five progressive stages.The training uses four NVIDIA RTX 4090 GPUs.
- UHDM: UHDM is summarized as a hierarchical encoder–decoder that extracts multi-scale features representing global illumination and fine spatial structures.Its decoder aggregates hierarchical features to reconstruct the enhanced image.
BITssvgg
BITssvgg’s DERNet combines global contextual modeling with dynamic expert routing to address spatially varying low-light degradation. Its implementation includes GPU inference with adaptive resizing and padding.
- BITssvgg: During inference, input images are normalized and adaptively resized according to their short-side resolution.Images with short sides at least 3k pixels are resized to 0.5×, while those above 2k pixels are resized to 0.8×.
- BITssvgg: The resized images are padded so height and width become multiples of 8 before network processing.This padding follows the adaptive resizing step during GPU inference.
BAU-Vision
The supplied BAU-Vision materials describe brightness-adaptive training and several restoration pipelines, but provide limited detail about Wave-P’s architecture or quantitative outcomes.
- Training: Sample-adaptive coefficients balance batch gradients across absolute brightness levels and reduce optimization bias toward high-intensity samples.
- Architectures and evaluation: The supplied figures identify Wave-P and MB-LPFR as separate restoration architectures, while Table 3 is an ablation study.
- Training: Random gamma adjustments are incorporated during training to improve robustness across lighting conditions.
YuFans
YuFans’s solution combines staged training with brightness-aware augmentation and additional full-resolution data to address the challenge’s low-light domain gap. Its reported components produce measurable validation gains and a small gap between independent and competition validation.
- Training: Training proceeds in two stages, beginning with NAFNet w64 pretraining on 115,376 LSD DLL patches and followed by progressive fine-tuning on 200 full-resolution pairs with BAA.
- Training: The progressive training schedule is documented in Table 4.
- Results: LSD pretraining contributes +4.51 dB, while PPE adds +0.60 dB.
- Results: MSDB yields 24.54 dB on independent validation versus 23.79 dB on competition validation, a gap of 0.75 dB.
AAIR-ARM
The supplied AAIR-ARM and related materials cover latent-space flow matching, hybrid multi-objective losses, multi-scale inference, and complementary expert-fusion strategies. They also document augmentation and training configurations across several submitted solutions.
- AAIR-ARM: AAIR-ARM trains an HDiT-based flow-matching network in latent space and iteratively transforms a low-light latent during inference before decoding the enhanced image.
- Training objectives: The hybrid loss combines mean-intensity reconstruction, luminance-chrominance, perceptual, gradient, and frequency-domain constraints.
- Inference: Inference uses sliding windows of 2048 and 512 pixels together with eight geometric augmentations whose predictions are inversely transformed and averaged.
- Results: The supplied materials report quantitative and qualitative advantages for TranssionAI’s solution but provide no numerical result in the cited passage.
- Expert fusion: A weighted late-fusion architecture routes shared features to fidelity-, perceptual-, and naturalness-oriented experts before aggregating their outputs.
NTR
The NTR materials describe a composite pixel-and-structural training objective, tiled full-resolution inference, and detailed optimization settings. The supplied passages specify processing cost but do not report a quality metric.
- Training: The model is fine-tuned on 115,376 paired 512×512 LSD patches using a composite pixel loss with a Sobel edge term for structural detail.
- Evaluation: Table 6 presents DH-XHDL-Team’s ablative experiments, while the supplied NTR passages do not state their outcomes.
- Inference: Full-resolution 4080×3060 images are processed with 512×512 tiles, 256-pixel overlap, and Hann-window blending without test-time augmentation or ensembling.
- Inference: Inference takes approximately 9 seconds per full-resolution image on a single NVIDIA A10.
- Optimization: The training schedule uses 300k iterations with Adam, gradient clipping, and a constant-then-cosine learning-rate schedule.
PSU
The reviewed PSU solutions span diverse low-light enhancement strategies, including U-Net architectures, diffusion-based methods, training-free denoising, tailored training, and ensemble inference.
- Architecture: DUSKAN uses a symmetric four-level U-Net with DUSKANBlock stages, strided downsampling, PixelShuffle upsampling, and global residual learning.Its blocks combine FFT-based global-feature modulation, local multi-scale depthwise convolutions, polynomial-basis activations, and selective gating.
- Training strategy: DNDiff is training-free and avoids extensive retraining for specific noise distributions.
- Training strategy: Another solution fine-tunes a pre-trained model using challenge-provided training data, with training taking approximately 30 minutes.
- Inference and optimization: A customized augmentation and loss-weighting strategy prioritizes PSNR restoration, while cross-model weighted averaging improves inference robustness and predictive accuracy.
APRIL-AIGC
The APRIL-AIGC section covers diffusion-based conditioning, tiled high-resolution inference, denoising and enhancement pipelines, multi-loss training, and architectures designed to address noise and detail recovery.
- APRIL-AIGC: APRIL-AIGC fine-tunes FLUX.2 on released training pairs using 512 × 512 crops, eight H20 GPUs, and a fixed prompt emphasizing denoising, shadow recovery, brightness, and color fidelity.
- APRIL-AIGC: Its diffusion backbone uses a rectified-flow objective.
- Related pipelines: The reviewed pipelines include DNATT, MiDLLIE, and RetinexDualV2, each presented through an overall pipeline or architecture figure.
- APRIL-AIGC: APRIL-AIGC’s fixed 2 × 3 partition with moderate overlap yields fewer illumination discontinuities than denser window layouts.
- APRIL-AIGC: High-resolution inference decomposes images into six overlapping tiles in a fixed 2 × 3 grid, with five sampling steps and guidance scale 2.0.For a 3060 × 4080 image, tiles are approximately 1600 × 1456 pixels with overlaps close to 128 × 144 pixels.
- Training objectives: One denoising-enhancement pipeline resizes training data to 256 × 256 and combines reconstruction, frequency-domain, perceptual, and IQA losses for low-light enhancement.Its denoising stage supervises NAFNet with Restormer predictions and ground truth.
- WIRNet: WIRNet is trained with weighted RGB- and HVI-domain losses combining L1, SSIM, edge, and LPIPS terms.
- WIRNet: WIRNet builds on CIDNet, whose noise sensitivity and limited fine-detail recovery motivate three proposed contributions.