NTIRE 2026 Challenge on Efficient Low Light Image Enhancement: Methods and Results
NTIRE 2026 Challenge introduces RetinexFormerRefine, achieving SSIM of 0.5654 with models under 1MB for low-light image enhancement.
Key Findings
Methodology
The study employs lightweight architectures like RetinexFormerRefine, U-Net variants, and Transformer-based models, integrating brightness-reflectance decomposition, color space separation, and structural reparameterization for optimized performance and efficiency.
Key Results
- MiVideo's RetinexFormerRefine achieved SSIM of 0.5654, LPIPS of 0.3632, with only 927,049 parameters, ranking first.
- CVPR TCD's grayscale-guided cross-attention Transformer achieved SSIM of 0.5500 and LPIPS of 0.4801 with 557,618 parameters.
- S3 team combined CLAHE preprocessing with lightweight U-Net, achieving SSIM of 0.5183 and LPIPS of 0.4157 with 741,600 parameters.
Significance
The study advances low-light image enhancement, particularly for resource-constrained environments like mobile devices, addressing computational bottlenecks through lightweight designs and efficient algorithms.
Technical Contribution
Introduced innovations such as Retinex model optimization, color space separation, lightweight Transformer modules, and structural reparameterization, significantly improving performance and efficiency.
Novelty
First to systematically combine brightness-reflectance decomposition with lightweight Transformer modules under strict sub-1MB model constraints.
Limitations
- Some methods struggle in extreme low-light scenarios, causing color distortion.
- Training requires high-performance GPUs, increasing resource demand.
- Cross-device generalization remains unexplored.
Future Work
Future work could explore cross-device generalization, robustness in extreme scenarios, and further optimization of training efficiency and resource requirements.
AI Executive Summary
The NTIRE 2026 Challenge focuses on efficient low-light image enhancement, aiming to design lightweight networks that improve image quality while meeting resource-constrained environments. The challenge attracted 207 participants, with 17 teams submitting valid solutions.
Innovative methods like RetinexFormerRefine, grayscale-guided cross-attention Transformers, and CLAHE preprocessing combined with lightweight U-Net were proposed, achieving significant improvements in enhancement quality and computational efficiency. MiVideo's RetinexFormerRefine ranked first with SSIM of 0.5654 and LPIPS of 0.3632, while maintaining model size under 1MB.
This study not only advances low-light enhancement technologies but also provides valuable insights for practical applications in mobile devices. Future directions include cross-device generalization and robustness improvements in extreme scenarios.
Deep Analysis
Background
Low-light image enhancement aims to improve visibility and contrast in dark environments, addressing issues like noise and color distortion. Deep learning methods have made significant progress but face challenges in computational cost and deployment on mobile devices.
Core Problem
Traditional methods often increase computational cost while improving quality, making them unsuitable for resource-constrained devices. Balancing performance and efficiency is a key challenge.
Innovation
The study introduces innovations like Retinex-based lightweight networks, color space separation techniques, and structural reparameterization, optimizing both performance and efficiency.
Methodology
- �� RetinexFormerRefine combines brightness-reflectance decomposition with lightweight design, using multi-branch structural reparameterization.
- �� Grayscale-guided cross-attention Transformer fuses grayscale and RGB features for structure-aware enhancement.
- �� CLAHE preprocessing combined with lightweight U-Net optimizes brightness and contrast through staged training.
Experiments
Experiments used a smartphone-captured low-light dataset with 349 training scenes, 49 validation scenes, and 102 test scenes. Metrics included SSIM, LPIPS, and DISTS, with a strict model size limit of 1MB.
Results
MiVideo's RetinexFormerRefine achieved SSIM of 0.5654 and LPIPS of 0.3632 with only 927,049 parameters, ranking first. CVPR TCD's Transformer model excelled in LPIPS with 557,618 parameters.
Applications
The findings can be directly applied to mobile device night photography optimization and provide technical support for resource-constrained environments.
Limitations & Outlook
Some methods struggle in extreme low-light scenarios, require high training resources, and lack cross-device generalization validation.
Plain Language Accessible to non-experts
Imagine taking a photo in a dark room where everything looks blurry and discolored. This research acts like a smart light bulb that not only brightens the room but also adjusts the photo's colors and details automatically. With lightweight design, this bulb fits into your phone, helping you capture clear photos anytime, anywhere.
ELI14 Explained like you're 14
Imagine playing a game where the screen is so dark you can't see your enemies. This research is like a super filter that brightens the screen while keeping the details sharp. And guess what? It's so lightweight it won't slow down your phone! Cool, right?
Glossary
Retinex model
A method to decompose image brightness and reflectance for enhancement.
Used for optimizing brightness and detail restoration.
SSIM
A metric for structural similarity between images, ranging from 0 to 1.
Evaluates enhancement quality compared to original images.
LPIPS
A perceptual loss metric based on deep features.
Measures visual quality of enhanced images.
CLAHE
Contrast Limited Adaptive Histogram Equalization, a preprocessing technique for contrast enhancement.
Used for initial brightness and contrast adjustment.
Structural reparameterization
A technique using multi-branch structures during training, converted to single-branch for inference.
Improves model efficiency.
Open Questions Unanswered questions from this research
- 1 How to improve robustness in extreme low-light scenarios?
- 2 What are the directions for optimizing cross-device generalization?
Applications
Immediate Applications
Night photography optimization
Enhances mobile phone night photography quality for general users.
Resource-constrained device applications
Provides efficient image enhancement solutions for embedded devices.
Long-term Vision
Cross-device generalization
Explores low-light enhancement technologies adaptable to diverse devices.
Abstract
This paper presents a comprehensive review of the NITRE 2026 Efficient Low Light Image Enhancement (E-LLIE) Challenge, highlighting the proposed solutions and final outcomes. This challenge focuses on mobile image enhancement under low-light conditions, aiming to design lightweight networks that improve enhancement quality while ensuring practical deployability under limited computational resources. A total of 207 participants registered, 27 teams submitted valid entries, and 17 teams ultimately provided valid factsheet. Based on these submissions, this paper provides a systematic evaluation of recent methods for E-LLIE, offering a comprehensive overview of state-of-the-art progress and demonstrating significant improvements in both performance and efficiency.