L3IE: Lightweight Low-Light Enhancement

• L3IE is a suite of techniques that improve underexposed images using minimal parameters and efficient computation for resource-constrained scenarios.
• It employs parameter-efficient convolutions, color and frequency domain strategies, and customized loss functions to achieve real-time performance.
• L3IE supports applications in automotive safety, mobile photography, and autonomous navigation by balancing image quality with computational efficiency.

Lightweight Low-Light Image Enhancement (L3IE) refers to a family of computational methodologies, architectural paradigms, and practical implementations that deliver robust image enhancement in severely underexposed conditions under stringent memory, computation, and latency constraints. These approaches are tailored to embedded vision, mobile, and edge scenarios, where classical, high-capacity deep networks or traditional pipelines are not deployable due to resource limitations. L3IE frameworks leverage parameter minimization, aggressive architectural simplification, principled loss design, and color/frequency domain priors to recover perceptual quality, edge detail, and color faithfulness in real time for a broad spectrum of applications, including automotive safety, mobile photography, autonomous navigation, and onboard object detection.

1. Core Principles and Architectural Strategies

L3IE methods are unified by their focus on reducing learnable parameter count, memory footprint, and FLOPs, frequently trading off model expressivity for real-time or near-real-time performance on low-power devices. The principal strategies are:

• Parameter-efficient Convolutions: Dynamic or structurally re-parameterized convolutions (e.g., DSConv in UltraFast-LieNET, SCLM single-layer models) reduce the number of weights by replacing standard kernels with channel-wise or depthwise alternatives and post-training fusion of multi-branch paths (Chen et al., 2 Dec 2025, Zhang et al., 2023).
• Color Space and Domain Decomposition: Instead of direct RGB processing, recent research exploits decorrelated spaces (YUV (Yan et al., 24 Jan 2026), HVI (Yan et al., 2024), HSV) or RAW sensor domains (Wang et al., 17 Dec 2025), enabling channel-adaptive enhancement. Frequency-domain and channel-specific degradation modelling inform targeted restoration of luminance and chrominance.
• Compact Attention and Feature Extraction: Frequency/attention modules are lightweighted—employing group or depthwise convolutions, as in DSGLA and LAFA (Yan et al., 24 Jan 2026), or introducing spectral and cross-branch guidance with minimal parameter overhead.
• Modular or Plug-and-Play Design: Building blocks such as DSConv, MSRB (Chen et al., 2 Dec 2025), and parameter-free iterative modules (Bai et al., 6 Jul 2025) are instantiated in scalable backbones, allowing flexible model instantiations from tens to hundreds of parameters.
• Physically and Statistically-Informed Losses: Relative losses (Zhang et al., 2023), multi-level gradient-aware or masked degradation losses (Chen et al., 2 Dec 2025, Sun et al., 2024), and domain-specific priors (e.g., Retinex, atmospheric scattering) are used both to stabilize training and guide restoration without paired supervision.

2. Color and Frequency Domain Innovations

A central trend is the exploitation of color and frequency domain properties for more specialized and efficient restoration:

• YUV Paradigm: Analysis shows that Y channel in low light loses low-frequency content (global brightness), while U/V accumulate high-frequency color noise. Dual-branch architectures restore these separately—DSGLA (dilated self-attention and ghost-gating) for Y, LAFA (FFT-based, Y-guided masking) for UV—improving both SNR and spatial fidelity (Yan et al., 24 Jan 2026).
• HVI Trainable Color Space: HVI introduces a decoupling of brightness (intensity) and color (projected onto a trainable plane), enabling robust dual-branch enhancement without the instability of conventional color spaces. CIDNet's LCA module linearly cross-attends between intensity and color branches for artifact-free restoration at low complexity (Yan et al., 2024).
• RAW Domain and Channel Priors: For RAW inputs, ERIENet fully exploits the green channel’s dominance (twice the spatial information of R/B in Bayer mosaics). A green-guided normalization branch directly modulates latent features to achieve higher fidelity with minimal extra overhead (Wang et al., 17 Dec 2025).

3. Loss Formulations and Training Objectives

Lightweight models are particularly sensitive to loss design. Innovative loss strategies complement reduced capacity:

• Multi-level and Multi-scale Losses: Hierarchical supervision—reconstruction, MS-SSIM, and multi-level gradient (edge) losses—enforce both global and fine structure consistency across decoder stages (Chen et al., 2 Dec 2025).
• Relative and Patch-ordering Losses: FLW-Net introduces color-direction, patch brightness ordering, and structure-gradient losses, all computed relatively rather than absolutely. This circumvents the "one-to-many" mapping problem where a unique ground truth is unavailable (Zhang et al., 2023).
• Unsupervised and Task-specific Objectives: DI-Retinex deploys an unsupervised, image-adaptive masked reverse degradation loss in gamma space and a variance suppression constraint on the model's learned offset, stabilizing training and yielding SOTA performance with no supervision (Sun et al., 2024); LiteIE combines exposure control, edge-aware TV, and multi-scale color consistency losses in a self-supervised pipeline (Bai et al., 6 Jul 2025).

4. Notable Model Families and Benchmark Performance

A spectrum of architectures demonstrates the current L3IE performance–complexity frontier:

Representative Model	Parameter Count	FLOPs (input size)	Peak PSNR / Dataset	FPS / Hardware
UltraFast-LieNETMax (Chen et al., 2 Dec 2025)	180	14M / 600x400	26.51 dB (LOLI-Street)	372 (Jetson AGX Orin)
FLW-Net (Zhang et al., 2023)	0.02M	0.05G / 256x256	26.61 dB (LOL-v2)	1000+ on GPU
YUV-based (Yan et al., 24 Jan 2026)	0.03M	1.45G / 256x256	27.16 dB (LOL-v1)	6.5 ms (3090 GPU)
LiteIE (Bai et al., 6 Jul 2025)	58	<0.01G	19.04 dB (LOL), unsup.	30 (4K, Snapdragon 8)
SCLM (Zhang et al., 2023)	87	0.17G / 1080x1920	21.15 dB (LOL, est.)	0.02 s (1080 GPU)
ERIENet (RAW) (Wang et al., 17 Dec 2025)	1.42M	39.3G / 512x512	29.12 dB (SID, RAW)	146 (4K, RTX 3090)
DI-Retinex (Sun et al., 2024)	0.04M	20G / 512x512	21.54 dB (LOL-v1, unsup.)	950 (1080Ti GPU)

Most L3IE models (<0.05 M parameters) outperform much larger SOTA baselines in reference-based (PSNR, SSIM, LPIPS) and no-reference (NIQE, EME, LOE) evaluations, while running at real-time rates on embedded CPUs and GPUs. Several models (LiteIE, SCLM, ERIENet) scale efficiently to 4K and demonstrate explicit practical deployment scenarios.

5. Applications, Downstream Integration, and Future Directions

• Automotive Vision: UltraFast-LieNET’s real-time operation with ~180 parameters meets the latency, robustness, and size constraints of embedded automotive SoCs, with demonstrated superior edge and color recovery under challenging driving scenes (Chen et al., 2 Dec 2025).
• Mobile Photography: Models such as LiteIE, DI-Retinex, and YUV-based approaches maintain performance on ARM CPUs, Snapdragon, and other mobile platforms, supporting edge-side image enhancement for live viewfinder and post-capture correction (Bai et al., 6 Jul 2025, Yan et al., 24 Jan 2026, Sun et al., 2024).
• RAW Pipeline Integration: ERIENet fuses with ISP backends for direct RAW image enhancement, leveraging parallel multi-scale and channel-aware dense features for full-resolution throughput on powerful but memory-limited hardware (Wang et al., 17 Dec 2025).
• Downstream Task Synergy: Latent disentanglement (e.g., LDE-Net) improves not only image quality but also the performance of subsequent tasks such as object detection (YOLOv3 on ExDark), UAV tracking (DarkTrack2021), and face detection under low-light, with minimal impact on overall system latency and throughput (Zheng et al., 2024, Sun et al., 2024).
• Ultra-Low Parameterization: Methods such as LiteIE and SCLM illustrate that effective enhancement, measured both quantitatively and qualitatively, is achievable with parameter budgets well below 100. This demonstrates the viability of enhancement as a “pre-processing primitive” on devices with deeply constrained memory and compute (Bai et al., 6 Jul 2025, Zhang et al., 2023).

6. Open Problems and Design Guidelines

Although recent L3IE advances establish new SOTA trade-offs, the following challenges remain:

• Generalization Across Domains: Performance is not yet uniform across unpaired, non-photographic, or highly non-stationary domains. Data-driven and self-regularizing loss schemes (e.g., relative, edge, or variance suppression losses) are critical for robustness.
• No-Reference/Zero-Shot Learning: Unsupervised objectives decouple training from large-scale paired data dependencies, but are still evolving to match supervised reference-based quality in all scenarios (Bai et al., 6 Jul 2025, Sun et al., 2024).
• Quantization and Post-training Compression: While INT8/binary weight quantization is a straightforward extension for further deployment gains, robustness under quantized weights is still an open empirical concern.
• Component Reuse and Modularity: Plug-and-play DSConv, MSRB, GFE, LAFA/DSGLA modules are now routinely shared as backbone augmentations, but cross-benchmark, cross-sensor modularity is an area of ongoing work.
• Optimal Loss Scheduling and Hyperparam Tuning: Lightweight models are especially unstable with naïve loss weighting or brightness scaling; adaptive objective scheduling and automated brightness/contrast control remain significant for in-field deployment (Zhang et al., 2023).

Recommended best practices for L3IE include: choosing decorrelated or physically-interpretable color spaces (YUV, HVI, RAW), aggressively minimizing parameter and FLOP footprints using group/depthwise/fused convolutional structures, employing cross-domain/self-regularizing losses, leveraging channel and frequency-aware modules for targeted denoising/brightening, and maintaining modularity for edge-to-edge integration with other vision stack components. Integrations with downstream perception modules, quantization pipelines, and formal robustness evaluation in adverse environments constitute promising venues for ongoing research (Chen et al., 2 Dec 2025, Wang et al., 17 Dec 2025, Zheng et al., 2024, Yan et al., 24 Jan 2026).