LiteIE: Ultra-lightweight Low-Light Enhancement

• LiteIE is an ultra-lightweight, unsupervised framework for low-light image enhancement, featuring a two-convolution shared-weight extractor and an iterative restoration module.
• The method leverages a composite unsupervised loss (exposure control, edge-aware smoothness, and multi-scale color consistency) to ensure high-quality enhancement with minimal parameters.
• LiteIE achieves real-time on-device inference for 4K images (approximately 30 FPS) with only 58 learnable parameters and a model size of about 2.69 KB.

LiteIE, short for Lightweight Iterative Enhancement and Restoration, is an ultra-lightweight, unsupervised framework for low-light image enhancement (LLIE) designed for deployment on mobile and embedded devices. It combines a two-convolution, shared-weight feature extractor with a parameter-free Iterative Restoration Module (IRM) and an unsupervised training objective composed of exposure control, edge-aware smoothness, and multi-scale color consistency losses. In the reported configuration, the learnable portion contains 58 parameters and has a model size of approximately 2.69 KB, while the system is evaluated on paired and unpaired LLIE benchmarks and on-device inference on a Snapdragon 8 Gen 3 mobile processor (Bai et al., 6 Jul 2025).

1. Design objectives and problem formulation

LiteIE addresses the standard LLIE problem of mapping a low-light RGB input image $I_{\text{init}} \in [0,1]^{H \times W \times 3}$ to an enhanced image $\hat{I}$ with improved visibility, contrast, and color fidelity. The motivating constraint is not merely image quality, but the simultaneous satisfaction of four design requirements: ultra-lightweight parameterization, real-time high-resolution inference, unsupervised training without paired supervision, and practical edge deployment (Bai et al., 6 Jul 2025).

The framework is positioned against two common limitations in prior LLIE systems. First, many state-of-the-art methods are architecturally heavy, relying on multi-branch designs, attention modules, or large backbones, which increases latency, memory footprint, and power draw on CPUs and mobile SoCs. Second, many methods are supervised and depend on paired low/normal-light datasets, which are difficult to collect and may not generalize well across camera domains and illumination conditions. LiteIE therefore adopts an unsupervised objective while minimizing the learnable component to the point where the entire feature extractor consists of two small convolutions (Bai et al., 6 Jul 2025).

A recurrent misconception is that “ultra-lightweight” in LiteIE implies a reduction only in model size. In the paper, the term is operationalized across several deployment-facing axes: 58 learnable parameters, approximately 0.11 G FLOPs for the chosen feature extractor configuration, and mobile inference reported at 30 FPS for 4K images in the abstract. The more detailed runtime table reports 37.4 ms at 3840×2160, corresponding to approximately 27–30 FPS, which situates LiteIE in the real-time regime for 4K enhancement on modern mobile hardware (Bai et al., 6 Jul 2025).

2. Architecture and iterative processing

LiteIE has a compact but explicitly staged processing pipeline. The first stage is a backbone-agnostic feature extractor $\mathcal{F}(\cdot;W)$, implemented as two shared-weight $3 \times 3$ convolutions in a 3→1→3 configuration, each followed by BatchNorm and identity activation inside $\mathcal{F}$. This block is recursively applied three times to produce three feature maps:

$$\phi(I): \left\{ \begin{aligned} \phi_1 &= \operatorname{ReLU}\big(\mathcal{F}(I; W)\big), \ \phi_2 &= \operatorname{ReLU}\big(\mathcal{F}(\phi_1; W)\big), \ \phi_3 &= \operatorname{Tanh}\big(\mathcal{F}(\phi_2; W)\big). \end{aligned} \right.$$

The same weights $W$ are reused in all three applications of $\mathcal{F}$. The map $\phi_3$ is interpreted as an enhancement matrix, while $\phi_1$ and $\hat{I}$0 provide lower-level and intermediate structural cues (Bai et al., 6 Jul 2025).

Enhancement proceeds iteratively. Starting from $\hat{I}$1, LiteIE applies a curve-like update:

$\hat{I}$2

This term is spatially adaptive because $\hat{I}$3 is image-dependent and channel-dependent. The paper characterizes the transform as similar in spirit to Zero-DCE, but the crucial difference is that LiteIE immediately follows the enhancement step with the Iterative Restoration Module (IRM):

$\hat{I}$4

Here, the $\hat{I}$5 are fixed scalar coefficients rather than network weights. The IRM reuses the already extracted feature maps and the original input to restore fine details and stabilize repeated enhancement, and it introduces zero additional learnable parameters. That point is terminologically important: the IRM is parameter-free, whereas the overall LiteIE model still has 58 learnable parameters in its feature extractor (Bai et al., 6 Jul 2025).

The chosen 3-1-3 extractor is not an arbitrary minimal design. The paper reports an ablation over 1-block, 2-block, and 3-block alternatives, including 3-3, 3-3-3, 3-8-3, 3-16-3, 3-1-1-3, 3-3-3-3, 3-8-8-3, and 3-16-16-3. The reported observation is that PSNR, SSIM, and perceptual index are surprisingly insensitive to depth and channel count, while FLOPs and runtime scale linearly with channels. The result is an intentionally minimal “sufficient” architecture rather than a compressed version of a larger CNN (Bai et al., 6 Jul 2025).

3. Unsupervised objective and optimization logic

LiteIE is trained without paired supervision through a composite objective

$\hat{I}$6

The first component, channel-adaptive Exposure Loss $\hat{I}$7, is designed to increase brightness while preserving color ratios inherited from the input. Let $\hat{I}$8 denote the mean RGB values of the enhanced image, and let $\hat{I}$9 denote the normalized RGB ratios of the original image such that $\mathcal{F}(\cdot;W)$0. The paper defines a chromatic consistency factor

$\mathcal{F}(\cdot;W)$1

and then sets channel-specific targets through a fixed scalar $\mathcal{F}(\cdot;W)$2, with the ablation indicating the best PSNR around $\mathcal{F}(\cdot;W)$3 on LOL-v1, LOL-v2, and LSRW (Bai et al., 6 Jul 2025).

The second component, Edge-Aware Total Variation Loss $\mathcal{F}(\cdot;W)$4, regularizes the enhanced image while preserving edges. It uses gradient-dependent weights

$\mathcal{F}(\cdot;W)$5

so that flat regions are smoothed more strongly and high-gradient regions are smoothed less. The reported ablation finds the best performance around $\mathcal{F}(\cdot;W)$6, with relatively stable behavior over $\mathcal{F}(\cdot;W)$7 (Bai et al., 6 Jul 2025).

The third component, Multi-Scale Color Consistency Loss $\mathcal{F}(\cdot;W)$8, constrains both local and global chromatic behavior:

$\mathcal{F}(\cdot;W)$9

The local term preserves local chromatic relationships from the input, while the global term reduces large-scale color cast. The paper’s ablations report that removing the global color term produces strong color shifts and cast, whereas removing the local term causes local tone inconsistency (Bai et al., 6 Jul 2025).

Because the enhancement loop, IRM, and feature extraction path are fully differentiable, gradients from all three losses propagate through the unrolled iterative process back to the shared weights $3 \times 3$0. The paper therefore describes LiteIE as effectively recurrent after unrolling, with parameter sharing keeping the learnable state extremely small. A plausible implication is that the method’s compactness depends not only on network width but also on this reuse of a single feature extractor across multiple enhancement stages.

4. Efficiency, deployment, and systems characteristics

LiteIE’s most distinctive systems property is that the learnable model consists of 58 parameters. The feature extractor uses a $3 \times 3$1 convolution from 3 channels to 1 channel followed by a $3 \times 3$2 convolution from 1 channel to 3 channels, and the IRM is parameter-free. The reported model size is approximately 2.69 KB (Bai et al., 6 Jul 2025).

The paper reports FLOPs $3 \times 3$3 G for the chosen 3–1–3 configuration. Runtime measurements place LiteIE at 0.97 ms on an RTX 4090 GPU, corresponding to approximately 1030 FPS at LOL-v1 resolution, and 6.69 ms on an Intel Xeon CPU, corresponding to approximately 150 FPS. Mobile inference is reported on a Snapdragon 8 Gen 3 SoC using the TensorFlow Lite GPU delegate, with the following latencies: 5.87 ms at 1280×720, 11.2 ms at 1920×1080, 20.2 ms at 2560×1440, and 37.4 ms at 3840×2160 (Bai et al., 6 Jul 2025).

These results are coupled to an intentionally deployment-friendly operator set. The architecture uses standard convolution, BatchNorm, ReLU, tanh, and elementwise arithmetic, with no attention modules or dynamic graph components. The paper explicitly notes that this makes LiteIE amenable to mobile inference frameworks such as TensorFlow Lite, and characterizes the operator mix as friendly to NNAPI and CoreML. Memory footprint is correspondingly small because only a few feature maps need to be retained and all convolutions are $3 \times 3$4 (Bai et al., 6 Jul 2025).

The “backbone-agnostic” label has a specific meaning here. It refers, first, to the structural insensitivity of the feature extractor under substantial changes in depth and channel count, and second, to the fact that the IRM can be layered onto other iterative LLIE methods. The paper demonstrates this explicitly through experiments with Zero-DCE, where the same restoration mechanism and the new loss terms improve a separate baseline without altering its fundamental enhancement formulation (Bai et al., 6 Jul 2025).

5. Empirical performance and ablation findings

On the LOL-v1 benchmark, LiteIE reports 19.04 dB PSNR, 0.607 SSIM, 0.141 MAE, and 0.29 LOE. The paper states that this PSNR surpasses the previous state of the art by approximately 1.4 dB while using only 0.07\% of its parameters. For comparison points given in the same table, ZeroIG is reported at 17.63 dB PSNR, 0.457 SSIM, and 0.161 MAE, while NoiSER is reported at 17.31 dB, 0.682 SSIM, and 0.143 MAE (Bai et al., 6 Jul 2025).

Generalization beyond LOL-v1 is evaluated on MIT-Adobe FiveK, LSRW-Huawei, and LSRW-Nikon. The reported values are 17.73 dB PSNR and 0.81 SSIM on MIT, 18.54 dB PSNR and 0.49 SSIM on LSRW-Huawei, and 16.08 dB PSNR and 0.44 SSIM on LSRW-Nikon. The paper characterizes these results as best or second-best across the corresponding benchmarks, indicating that the compact architecture is not confined to a single dataset distribution (Bai et al., 6 Jul 2025).

On unpaired datasets including LIME, NPE, DARK FACE, and DICM, the paper reports that LiteIE achieves best or second-best PI on all datasets, the best average across all datasets with average approximately 7.36 versus approximately 7.76 for ZeroDCE++, and the best BRI scores, while remaining competitive in NIQE and CIQA. Qualitative results are described as preserving natural brightness and contrast without over-exposure, maintaining local shadows and global consistency, and avoiding strong color casts. On DarkFace, the enhanced outputs are further stated to improve downstream DSFD face detection qualitatively, with more detected boxes and fewer misses in the illustrated examples (Bai et al., 6 Jul 2025).

The ablation evidence centers on the IRM and the loss design. On a combined DICM, LIME, and NPE setting, the paper reports that removing IRM yields 14.53 PSNR, 0.58 SSIM, 3893.78 MAE, and 3.98 NIQE, whereas including IRM yields 19.04 PSNR, 0.61 SSIM, 1507.57 MAE, and 3.79 NIQE. The qualitative interpretation given in the paper is that IRM prevents cloud disappearance, color shifts, and pseudo-haze under repeated enhancement (Bai et al., 6 Jul 2025).

The same mechanism also improves another method. For Zero-DCE, the paper reports 14.97 dB PSNR as baseline, 17.46 dB for ZeroDCE + new loss, 18.29 dB for ZeroDCE + IRM, and 18.92 dB for ZeroDCE + new loss + IRM. LiteIE itself, when trained with the ZeroDCE loss, is reported at 17.86 dB with 58 parameters, while the full LiteIE system reaches 19.04 dB with the same parameter count. This indicates that the reported gains derive jointly from the IRM and the revised unsupervised objective rather than from parameter reduction alone (Bai et al., 6 Jul 2025).

6. Nomenclature, limitations, and relation to adjacent work

The name LiteIE is specific to low-light image enhancement, but it can be confused with unrelated methods that use the acronym LITE. The literature also includes LITE for compact time series classification, “Light Inception with boosTing tEchnique,” which uses depthwise separable convolutions, multiplexing, custom filters, and dilation in 1D CNNs for TSC (Ismail-Fawaz et al., 2024). A separate 2026 paper introduces LITE as a flat-direction acceleration strategy for LLM pre-training under a Riemannian ODE framework, built on top of optimizers such as Muon and SOAP (Zhu et al., 26 Feb 2026). These methods are distinct in domain, objective, and mechanism.

Within LLIE itself, LiteIE is positioned relative to several neighboring families. Compared to Retinex-based networks such as KIND++, URetinex, and RUAS, it does not perform explicit illumination/reflectance decomposition. Compared to curve-based networks such as Zero-DCE, Zero-DCE++, ChebyLighter, and Self-DACE, it retains the idea of spatially varying curve enhancement but adds an explicit parameter-free restoration stage and a multi-term unsupervised loss with multi-scale color consistency. Compared to GAN- and diffusion-based methods such as EnlightenGAN, Diff-Retinex, and LightenDiffusion, the paper frames LiteIE as better suited to real-time mobile deployment because of its lower inference cost. Among lightweight baselines such as SCLM, NoiSER, MobileIE, and COLIE, LiteIE is characterized as striking a strong balance between compactness, speed, and quality (Bai et al., 6 Jul 2025).

The limitations reported in the paper are also specific. In extremely dark regions, where the original signal is heavily degraded, LiteIE cannot reconstruct realistic colors or textures that are not present in the input, and may amplify noise while increasing structural visibility. In high-frequency regions, the model may show slight blurring, such as at the boundary between a pipe and a wall. The framework also has no explicit denoising module, so noise can remain noticeable at very low SNR. The future directions named or implied in the paper include integrating dedicated denoising or noise modeling, extending the approach to video LLIE with temporal consistency, combining the IRM principle with larger backbones, and exploring adaptive iteration counts based on scene content (Bai et al., 6 Jul 2025).

Taken together, LiteIE represents a specific line of LLIE research in which architectural minimalism is not treated as a post hoc compression step but as a first-order design principle. The paper’s central claim is not merely that low-light enhancement can be made smaller, but that a two-convolution shared-weight extractor, when embedded in an iterative enhancement-restoration loop and trained with a carefully structured unsupervised objective, can remain competitive with substantially larger models on both benchmark quality metrics and on-device deployment constraints (Bai et al., 6 Jul 2025).