Stabilizing Illumination Conditioning in SSM-Based Low-Light Image Enhancement with Bottleneck Prompt Sidecars

Wei Chen Ng; Qiyi Xie; Jian Ding Tan; Yan Kang Lee; Abdulrab Hakim Qaid Abdullah; Wei Hown Tee

eess.IVOct 16, 2026

Stabilizing Illumination Conditioning in SSM-Based Low-Light Image Enhancement with Bottleneck Prompt Sidecars

Authors: Wei Chen Ng, Qiyi Xie, Jian Ding Tan, Yan Kang Lee, Abdulrab Hakim Qaid Abdullah, Wei Hown Tee

Organizations: School of Computing and Data Science, Xiamen University Malaysia, Sepang, Selangor 43900, Malaysia · School of Artificial Intelligence and Robotics, Xiamen University Malaysia, and with the School of Engineering and Applied Science, University of Pennsylvania, Philadelphia, PA 19104, USA · School of Artificial Intelligence and Robotics and the S.M.A.R.T. Nexus Center of Excellence, Xiamen University Malaysia, Sepang, Selangor 43900, Malaysia · School of Energy and Chemical Engineering, Xiamen University Malaysia, Sepang, Selangor 43900, Malaysia ·

Abstract

Low-light image enhancement (LLIE) requires more than brightness amplification because real low-light images often exhibit spatially varying illumination, amplified noise, color distortion, and local detail loss. Recent SSM- and Mamba-based restorers provide efficient long-range modeling, yet the route for introducing illumination priors into such backbones remains insufficiently studied. This study addresses safe illumination conditioning for SSM-based LLIE through a bottleneck-side PromptAdapter. A lightweight illumination estimator first predicts a single-channel illumination map, which is then converted into prompt tokens and injected through a residual sidecar adapter at the bottleneck, thereby avoiding direct modification of the state-space core. Under a locked three-seed evaluation protocol on LOL-v1, LOLv2-Real, and LOLv2-Synthetic, the PromptAdapter improves over the Base model by 0.94 dB, 1.24 dB, and 0.93 dB in PSNR, respectively, while also improving SSIM, LPIPS, and masked dark-region metrics. The clearest real-data evidence appears on LOLv2-Real. A Learned-Fixed control on LOL-v1 indicates that the present results support the utility of the sidecar conditioning pathway more strongly than a strict claim of illumination-conditioned superiority, and the LOLv2-Synthetic result remains stability-conditioned. These findings suggest that, in SSM-based LLIE, conditioning design can be as important as the conditioning signal itself.

Figures & tables

Fig. 1: Overall architecture of the proposed SSM-based LLIE restorer with bottleneck-side Prompt-Adapter.

Fig. 2: Detail of the sidecar conditioning block with cross-attention prompt injection mechanism.

Variant	Params (M)	FLOPs (G)
Base	2.46	5.16
Prompt-Adapter	2.52	5.18
Learned-Fixed	2.47	5.17

TABLE I: Model complexity comparison across the three evaluated variants. Parameters are reported in millions (M) and FLOPs are estimated on a

128\times 128

input patch using fvcore.

Dataset	Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	mPSNR $\uparrow$	mSSIM $\uparrow$
LOL-v1	Base	25.01 $\pm$ 0.34	0.9077 $\pm$ 0.0011	0.2292 $\pm$ 0.0017	24.95 $\pm$ 0.21	0.9033 $\pm$ 0.0009
LOL-v1	Prompt-Adapter	25.95 $\pm$ 0.18	0.9094 $\pm$ 0.0010	0.2273 $\pm$ 0.0066	25.96 $\pm$ 0.28	0.9053 $\pm$ 0.0006
LOL-v1	Learned-Fixed	25.99 $\pm$ 0.59	0.9101 $\pm$ 0.0017	0.2289 $\pm$ 0.0063	26.09 $\pm$ 0.67	0.9063 $\pm$ 0.0021
LOLv2-Real	Base	20.73 $\pm$ 0.47	0.8855 $\pm$ 0.0016	0.2415 $\pm$ 0.0056	20.99 $\pm$ 0.43	0.8818 $\pm$ 0.0018
LOLv2-Real	Prompt-Adapter	21.97 $\pm$ 0.34	0.8903 $\pm$ 0.0075	0.2309 $\pm$ 0.0153	22.42 $\pm$ 0.38	0.8876 $\pm$ 0.0077
LOLv2-Synthetic	Base	25.44 $\pm$ 0.17	0.9634 $\pm$ 0.0007	0.0764 $\pm$ 0.0011	26.89 $\pm$ 0.40	0.9498 $\pm$ 0.0009

TABLE II: Controlled internal comparison under the locked three-seed Y-channel protocol (mean

\pm

std over seeds 3407, 3408, 3409).

Panel of Figure 5 — Fig. 3: Attribution-oriented qualitative comparison on LOL-v1 showing Base, Learned-Fixed, and Prompt-Adapter outputs against ground truth.

Method	LOL-v1	LOLv2-Real	LOLv2-Syn	Relevance	Protocol note
SNR-Net	24.61 / 0.842	21.48 / 0.849	24.14 / 0.928	degradation-aware CNN	PSNR/SSIM only
Retinexformer	25.16 / 0.845	22.80 / 0.840	25.67 / 0.930	illumination-guided Transformer	official ICCV 2023 table
RetinexMamba †	24.03 / 0.827	22.45 / 0.844	25.89 / 0.935	SSM/Mamba-family reference	protocol footnote a
Base (this study)	25.01 / 0.9077	20.73 / 0.8855	25.44 / 0.9634	reproduced base model	Y-channel (BT.601)
Prompt-Adapter (this study)	25.95 / 0.9094	21.97 / 0.8903	26.37 / 0.9661	reproduced safe-conditioning route	Y-channel (BT.601); synthetic caveat

TABLE III: Cross-protocol context for representative LLIE methods. External rows are cited from their primary sources under the original evaluation protocols; rows from this study use the Y-channel protocol of Section III . The two groups are not directly comparable, and no ranking is implied.

Panel of Figure 7 — Fig. 4: Qualitative comparison between Base and Prompt-Adapter across LOL-v1, LOLv2-Real, and LOLv2-Synthetic test samples.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Dataset	Variant	Seed	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	mPSNR $\uparrow$	mSSIM $\uparrow$
LOL-v1	Base	3407	25.2991	0.9081	0.2288	25.1429	0.9035
LOL-v1	Base	3408	24.6306	0.9065	0.2310	24.7266	0.9023
LOL-v1	Base	3409	25.1019	0.9086	0.2277	24.9871	0.9041
LOL-v1	Prompt-Adapter	3407	25.9464	0.9098	0.2218	25.9856	0.9059
LOL-v1	Prompt-Adapter	3408	26.1289	0.9100	0.2346	26.2300	0.9052
LOL-v1	Prompt-Adapter	3409	25.7692	0.9083	0.2255	25.6722	0.9047

Appendix

TABLE IV: Per-seed results underlying the locked internal comparison. All rows are evaluated under the locked Y-channel protocol with VGG-based LPIPS. This table provides the raw seed-level evidence for the mean

\pm

standard deviation results reported in Table II .

Low-Light Image Enhancement

Low-light image enhancement restores visibility, color, and detail in underexposed images while limiting noise, blur, and color distortion.

See all 95 papers

Towards Controllable Low-Light Image Enhancement: A Continuous Multi-illumination Dataset and Efficient State Space Framework
Mar 26, 2026Hongru Han, Tingrui Guo, Liming Zhang +3Image Enhancement Low-Light Image Enhancement
Department of Computer and Information Science, University of Macau, Taipa, Macau 999078, China · School of Automation, China University of Geosciences, Wuhan 430074, China · Department of Electromechanical Engineering, University of Macau, Taipa, Macau 999078, China
Diffusion-Based Low-Light Image Enhancement with Color and Luminance Priors
Feb 27, 2026Xuanshuo Fu, Lei Kang, Javier Vazquez-CorralLow-Light Image Enhancement Conditional Diffusion Models
Computer Vision Center, Cerdanyola del Vallès, Spain and Universitat Autònoma de Barcelona, Cerdanyola del Vallès, Spain
InterLight: Leveraging Intrinsic Illumination Priors for Low-Light Image Enhancement
May 19, 2026Ziqi Wang, Xu Zhang, Laibin Chang +3Image Enhancement Low-Light Image Enhancement
National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University · Department of Computer Science, University of Macau · Department of Computer Vision, Mohamed bin Zayed University of Artificial Intelligence +1

Mar 26, 2026J/K move · Enter open · S save

Towards Controllable Low-Light Image Enhancement: A Continuous Multi-illumination Dataset and Efficient State Space Framework

Hongru Han, Tingrui Guo, Liming Zhang, Yan Su+2

Department of Computer and Information Science, University of Macau, Taipa, Macau 999078, China · School of Automation, China University of Geosciences, Wuhan 430074, China · Department of Electromechanical Engineering, University of Macau, Taipa, Macau 999078, China

Image Enhancement Low-Light Image Enhancement Visual SSMs

Open paper AskPDF arXiv

Low-light image enhancement (LLIE) has traditionally been formulated as a deterministic mapping. However, this paradigm often struggles to account for the ill-posed nature of the task, where unknown ambient conditions and sensor parameters create a multimodal solution space. Consequently, state-of-the-art methods frequently encounter luminance discrepancies between predictions and labels, often necessitating "gt-mean" post-processing to align output luminance for evaluation. To address this fundamental limitation, we propose a transition toward Controllable Low-light Enhancement (CLE), explicitly reformulating the task as a well-posed conditional problem. To this end, we introduce CLE-RWKV, a holistic framework supported by Light100, a new benchmark featuring continuous real-world illumination transitions. To resolve the conflict between luminance control and chromatic fidelity, a noise-decoupled supervision strategy in the HVI color space is employed, effectively separating illumination modulation from texture restoration. Architecturally, to adapt efficient State Space Models (SSMs) for dense prediction, we leverage a Space-to-Depth (S2D) strategy. By folding spatial neighborhoods into channel dimensions, this design allows the model to recover local inductive biases and effectively bridge the "scanning gap" inherent in flattened visual sequences without sacrificing linear complexity. Experiments across seven benchmarks demonstrate that our approach achieves competitive performance and robust controllability, providing a real-world multi-illumination alternative that significantly reduces the reliance on gt-mean post-processing.

Abstract

Low-Light Image Enhancement

Explore similar work

Towards Controllable Low-Light Image Enhancement: A Continuous Multi-illumination Dataset and Efficient State Space Framework

Diffusion-Based Low-Light Image Enhancement with Color and Luminance Priors

InterLight: Leveraging Intrinsic Illumination Priors for Low-Light Image Enhancement