FuguReport

NTIRE 2026 Challenge on Efficient Low Light Image Enhancement: Methods and Results

Authors Jiebin Yan, Chenyu Tu, Weixia Zhang, Zhihua Wang, Peibei Cao, Qinghua Lin, Yuming Fang, Xiaoning Liu, Zongwei Wu, Zhuyun Zhou, Radu Timofte
Categories Evaluation / Benchmarking / Low light image enhancement challenge, Task / Image Enhancement / Enhancement for low light conditions, Method / Efficiency / Efficient enhancement techniques
License CC BY 4.0

Abstract Overview

This paper presents the NTIRE 2026 Efficient Low Light Image Enhancement (E-LLIE) Challenge, which focuses on mobile image enhancement under low-light conditions with a strict model-size constraint of less than 1 MB. The challenge used a smartphone-captured dataset (from Huawei and Apple devices) covering diverse low-light scenarios, with 349 training scenes, 49 validation scenes, and 102 test scenes at 3024×4032 resolution. A total of 207 participants registered, 27 teams submitted valid entries, and 17 teams provided valid factsheets for final evaluation. The paper summarizes the submitted methods and compares them using a multi-metric ranking based on SSIM, LPIPS, DISTS, LIQE, MUSIQ, and Q-Align, with parameter count as a tiebreaker.

Novelty

The distinctive aspect of this work is the challenge setting itself: it benchmarks low-light enhancement methods under an explicit sub-1 MB model-size constraint while evaluating them with a broad set of full-reference and no-reference quality metrics. The report also provides a cross-section of recent lightweight design strategies—including Retinex-based modeling, color-space decoupling, lightweight attention, and structural reparameterization—within a single standardized benchmark using a new smartphone-captured dataset.

Results

The final leaderboard shows MiVideo ranked first overall with the best LPIPS, DISTS, LIQE, MUSIQ, and Q-Align scores (927,049 parameters), while SYSU-FVL_ELLIE achieved the best SSIM. The results demonstrate that competitive enhancement quality can be achieved within the strict sub-1 MB size limit, with leading methods differing in their strengths across fidelity and perceptual metrics. Across the 17 ranked teams, methods range from approximately 102K to 995K parameters, confirming that strong low-light enhancement performance is attainable with compact models.

Key Points

  1. The paper is a benchmark and challenge report rather than a single-method study, evaluating 17 final teams on efficient low-light enhancement under a sub-1 MB model constraint using a new smartphone-captured dataset at 3024×4032 resolution.
  2. Ranking is based on aggregated positions across six metrics—SSIM, LPIPS, DISTS, LIQE, MUSIQ, and Q-Align—with parameter count used as a tiebreaker when overall rankings are tied.
  3. The surveyed solutions span several recurring design patterns, including lightweight encoder-decoder networks, Retinex-inspired decomposition, color-space separation (HVI, YUV, LAB), attention or transformer modules, and efficiency-oriented techniques such as structural reparameterization and partial convolution.

References

This page was created using generative AI such as GPT-5, Claude Opus 4, Gemini 3, Gemini 3.1 Flash Image, and their higher-end successor versions. No guarantee can be made regarding its contents.