Towards Color-faithful Low-light Image Enhancement via Adaptive Color Debiasing and Saturation Rectification

Zhichen Yang, Rui Xu, Yuzhen Niu^∗, Fusheng Li, Hui Da, Ri Cheng

ACM MM 2026

Fuzhou University, Fujian, China

arXiv Paper (ACM MM 2026) Code BibTeX

TL;DR

An adaptive color transform reorganize color space and prevents embedded color bias from interfering with the enhancement pipeline, delivering a ~2.5 dB PSNR gain with only 0.07M extra Params. and less than 0.01G FLOPs .

Our motivation. Direct brightening makes color shifts more visible. Baseline enhancement improves brightness but leaves color bias and saturation abnormalities. Our CAGE delivers more faithful color restoration. — **Our motivation.** *(a→b)* Direct brightening makes color shifts more visible. *(a→c)* Baseline enhancement improves brightness but leaves color bias and saturation abnormalities. *(a→d)* Our CAGE delivers more faithful color restoration.

Gallery

Image Arena

Compare enhanced candidates, inspect paired details, and choose the result with the most faithful color restoration.

Low-light / GT-mean Enlightened Input

input.png

Low-light image

Enhanced Results

output_A.png

Enhanced image A

output_B.png

Enhanced image B

output_C.png

Enhanced image C

Left click to select Right click to compare Middle click to enlarge

Vote Statistics

Ours Others

Color Space Visualization

Inspect an image and the full color gamut in synchronized 3D mappings, with brightness-slice previews across color spaces.

Input Image

Image

Upload an image to visualize its 3D mapping and the corresponding color space voxels 3D mapping.

Receiving image...

Choose or drop an image PNG, JPEG, WebP, BMP, or GIF

Image 3D Mapping

0 points

Each sampled pixel becomes a 3D point colored by its source RGB value.

Gamut 3D Mapping

0 points

A uniformly sampled RGB gamut is mapped into the selected space.

Settings

Info Card

R ∈ [0, 1] G ∈ [0, 1] B ∈ [0, 1]

RGB coordinates retain their native channel range.

Forward conversion RGB → RGB

Inverse conversion RGB→ RGB

Sampling and geometry

Input Prefabs Image point budget RGB lattice resolution 32 samples per axis · 32,768 points HSV coordinates HVI density k

0.20

LAB coordinates

View options

Normalize 0–1 Auto rotate Show frame Show axes

1 / 1

Hold middle mouse to pan Scroll over image to zoom Click blank area to close Scroll over blank area or < > Arrow keys or 1 2 3 Number keys to switch

The Problem

01 · Input

Embedded color bias

Low signal-to-noise ratio, camera hardware, and in-camera processing introduce color shift during low-light image formation.

02 · Pipeline

Bias propagation

A color space can reorganize distorted colors, but the disturbance persists through enhancement unless the chromatic distribution is corrected.

03 · Output

Chromatic distortion

Brightness enhancement amplifies chromatic responses, causing global color bias and regional under-saturation and over-saturation.

Method Overview

CAGE integrates a low-light image enhancement backbone with the forward and inverse transforms of AdaCCT. The two transforms map images between RGB and AdaLAB while regulating the chromatic distribution before and after backbone enhancement.

Interactive Node Map

CAGE processing architecture

Select a block to inspect its operation and equation. < > to navigate.

1 / 7 Input Home End

image

Input Low-light Image

\( x \in \mathbb{R}^{3 \times H \times W} \)

The low-light input image is captured under insufficient illumination. It contains embedded color bias due to the camera hardware and in-camera processing.

CAGE architecture: a lightweight CNN predicts image-adaptive cylindrical parameters, while the input passes through the forward transform, low-light image enhancement backbone, and inverse transform to produce the enhanced image. — Click the figure to step through each block interactively.

AdaCCT: Correct Before Enhancement, Rectify After Enhancement

Pipeline of AdaCCT. The forward transform converts RGB to LAB, applies chromatic debiasing and chroma scaling, and constructs AdaLAB. The inverse transform reverts chroma scaling, applies out-of-gamut lightness compensation, and converts LAB back to RGB. — **Pipeline of AdaCCT.** The forward transform converts RGB to LAB, applies chromatic debiasing and chroma scaling, and constructs AdaLAB. The inverse transform reverts chroma scaling, applies out-of-gamut lightness compensation, and converts LAB back to RGB.

Qualitative Results

Qualitative comparisons on paired and unpaired datasets show reduced global color bias, more stable local saturation, and more natural chromatic transitions.

Paired Dataset Results

Paired comparison on LOLv1 dataset. CAGE achieves more faithful color restoration with reduced color bias. — **Paired comparison on LOLv1 dataset.** CAGE achieves more faithful color restoration with reduced color bias.

Paired comparison on LOLv2-real dataset. Real-world low-light images with challenging color shifts. — **Paired comparison on LOLv2-real dataset.** Real-world low-light images with challenging color shifts.

Paired comparison on LOLv2-synthetic dataset. Synthetic low-light images with controlled color distortion. — **Paired comparison on LOLv2-synthetic dataset.** Synthetic low-light images with controlled color distortion.

Paired comparison on SDSD & SID datasets. Additional real-world low-light benchmarks. — **Paired comparison on SDSD & SID datasets.** Additional real-world low-light benchmarks.

Unpaired Dataset Results

Unpaired comparison on real-world images. CAGE produces more faithful colors and more balanced saturation across varied low-light scenes. — **Unpaired comparison on real-world images.** CAGE produces more faithful colors and more balanced saturation across varied low-light scenes.

User Study

Quantitative Results

CAGE consistently improves all backbones with only 0.07M additional parameters across multiple benchmarks.

Main Results

Quantitative comparisons on LOLv1, LOLv2-real, and LOLv2-synthetic datasets. **CAGE** consistently improves all backbones with only 0.07M additional parameters and less than 0.01G FLOPs. Metrics: PSNR (dB) / SSIM. **Bold** = best result. Following HVI-CIDNet , we use GT-mean evaluation on LOLv1 dataset during testing. Since LOLv1 contains only 15 testing pairs, and direct metric evaluation can be sensitive to global brightness fluctuation, which may obscure differences in color restoration and structural recovery.
Method	LOLv1	LOLv2-real	LOLv2-synthetic
RetinexNet	17.56 / 0.645	17.68 / 0.642	15.61 / 0.449
KinD	17.64 / 0.777	20.59 / 0.818	16.26 / 0.591
DRBN	20.13 / 0.830	20.29 / 0.831	23.22 / 0.927
MIRNet	26.56 / 0.853	22.68 / 0.828	25.05 / 0.923
ZeroDCE	21.88 / 0.640	16.06 / 0.580	17.71 / 0.815
EnlightenGAN	20.00 / 0.691	18.23 / 0.617	16.57 / 0.734
LLFlow	25.53 / 0.870	17.43 / 0.831	23.43 / 0.933
SNR-Net	26.81 / 0.855	21.48 / 0.849	24.14 / 0.927
BreaD	25.32 / 0.843	23.69 / 0.869	15.97 / 0.746
LLFormer	26.12 / 0.827	21.65 / 0.816	25.26 / 0.923
CWNet	26.38 / 0.864	21.65 / 0.860	25.50 / 0.937
Retinexformer	26.08 / 0.827	21.89 / 0.837	24.59 / 0.919
Retinexformer + Ours	27.09 / 0.871	24.10 / 0.847	26.33 / 0.939
Δ	+1.01 / +0.044	+2.21 / +0.010	+1.74 / +0.020
DarkIR	25.32 / 0.806	21.32 / 0.830	24.06 / 0.921
DarkIR + Ours	26.03 / 0.849	22.51 / 0.860	24.54 / 0.923
Δ	+0.71 / +0.043	+1.19 / +0.030	+0.48 / +0.002
HVI-CIDNet	26.03 / 0.832	23.94 / 0.870	25.32 / 0.929
HVI-CIDNet + Ours	27.25 / 0.857	24.24 / 0.875	25.52 / 0.930
Δ	+1.22 / +0.025	+0.30 / +0.005	+0.20 / +0.001

Challenging Datasets

Quantitative comparisons on challenging datasets (SDSD-indoor, SDSD-outdoor, and SID). **CAGE** achieves consistent improvement with PSNR gains up to 1.78 dB on SDSD-indoor. Metrics: PSNR (dB) / SSIM. **Bold** = best result.
Method	SDSD-indoor	SDSD-outdoor	SID
3DLUT	24.78 / 0.718	23.29 / 0.703	16.97 / 0.591
SNR-Net	26.13 / 0.815	19.22 / 0.657	21.35 / 0.550
LEDNet	27.29 / 0.876	26.66 / 0.850	21.47 / 0.638
FourLLIE	24.74 / 0.826	24.67 / 0.787	18.42 / 0.513
RetinexMamba	28.44 / 0.894	28.52 / 0.859	22.45 / 0.656
MIRNet	28.64 / 0.888	28.99 / 0.869	21.36 / 0.632
Restormer	28.49 / 0.892	27.99 / 0.868	22.01 / 0.645
MambaIR	25.14 / 0.876	27.53 / 0.851	22.02 / 0.658
CWNet	30.28 / 0.904	28.54 / 0.857	20.97 / 0.623
Retinexformer	28.20 / 0.877	29.57 / 0.874	22.05 / 0.636
Retinexformer + Ours	29.98 / 0.906	30.12 / 0.877	22.64 / 0.660
Δ	+1.78 / +0.029	+0.55 / +0.003	+0.59 / +0.024
DarkIR	29.94 / 0.897	28.81 / 0.848	21.50 / 0.588
DarkIR + Ours	30.36 / 0.899	29.36 / 0.858	22.33 / 0.639
Δ	+0.42 / +0.002	+0.55 / +0.010	+0.83 / +0.051
HVI-CIDNet	28.74 / 0.895	28.70 / 0.866	22.21 / 0.631
HVI-CIDNet + Ours	29.91 / 0.900	29.80 / 0.877	22.46 / 0.653
Δ	+1.17 / +0.005	+1.10 / +0.011	+0.25 / +0.022

Key Takeaways

New Perspective

Low-light embedded color bias

Max Gain

a ~2.5 dB PSNR gain

Minimal Tradeoff

0.07M Params. and < 0.01G FLOPs

Plug & Play

w/o arch. modifying.

BibTeX

For the implementation and project materials, visit the project repository. If you find the work useful, please cite:

@inproceedings{yang2026cage,
  title     = {Towards color-faithful low-light image enhancement via adaptive color debiasing and saturation rectification},
  author    = {Yang, Zhichen and Xu, Rui and Niu, Yuzhen and Li, Fusheng and Da, Hui and Cheng, Ri},
  booktitle = {Proceedings of the ACM International Conference on Multimedia (ACMMM)},
  year      = {2026}
}

Acknowledgement

If you find our work helpful, please also consider citing the following related work:

@inproceedings{xu2025urwkv,
  title     = {{URWKV}: Unified {RWKV} model with multi-state perspective for low-light image restoration},
  author    = {Xu, Rui and Niu, Yuzhen and Li, Yuezhou and Xu, Huangbiao and Liu, Wenxi and Chen, Yuzhong},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {21267--21276},
  year      = {2025}
}

We thank the authors of the PointDiT, ReasonX, and Lotus project pages, whose page designs inspired this website.

CAGE is heavily inspired by HVI-CIDNet and OKLab. We sincerely thank the authors for their elegant work.