Self-DACE++: Real-Time LLIE Framework

• Self-DACE++ is an unsupervised, lightweight low-light image enhancement framework that employs Adaptive Adjustment Curves for dynamic-range modification.
• It integrates a fusion of Disordered Modules and a dedicated denoising block to achieve superior restoration quality, rapid inference, and compact model size.
• The framework uses a physics-grounded objective function and a randomized training regimen to enhance robustness and maintain color and structural fidelity.

Self-DACE++ is an unsupervised and lightweight framework for Low-Light Image Enhancement (LLIE), specifically designed to balance computational efficiency with high-quality restoration. Building upon the original Self-Reference Deep Adaptive Curve Estimation (Self-DACE), Self-DACE++ introduces novel components in curve modeling, architecture, training methodology, and objective functions, including a physics-grounded loss and a dedicated denoising module. It outperforms state-of-the-art methods in terms of enhancement quality, speed, and compactness, making it suitable for real-time and low-resource deployment (Wen et al., 28 Apr 2026).

1. Adaptive Adjustment Curves (AACs)

Self-DACE++ utilizes Adaptive Adjustment Curves (AACs) to perform efficient, interpretable dynamic-range modification at the pixel and channel level. For each channel $c$, enhancement is performed as

$$\mathrm{AAC}^c(\alpha^c,\beta^c;I^c) = I^c + \alpha^c \otimes \frac{1}{\beta^c} \otimes C(\beta^c;I^c)$$

where $I^c$ is the normalized input, and $\alpha^c$, $\beta^c$ are trainable per-pixel, per-channel maps. The function $C(\cdot)$ is defined in two forms for low-light area enhancement (LAEC) and high-light area suppression (HASC):

$$\text{LAEC:}\quad C(\beta^c;I^c) = S\bigl(-k\,(I^c-\beta^c+\delta)\bigr) \,\otimes\,I^c\,\otimes\,(\beta^c-I^c)$$

$$\text{HASC:}\quad C(\beta^c;I^c) = S\bigl(+k\,(I^c-\beta^c-\delta)\bigr) \,\otimes\,(1-I^c)\,\otimes\,(I^c-\beta^c)$$

with $S(x)=1/(1+e^{-x})$, $k=15$, $$\mathrm{AAC}^c(\alpha^c,\beta^c;I^c) = I^c + \alpha^c \otimes \frac{1}{\beta^c} \otimes C(\beta^c;I^c)$$0. AACs enable the network to flexibly stretch or compress the dynamic range, maintain monotonicity, and preserve color and structural fidelity using only two maps ($$\mathrm{AAC}^c(\alpha^c,\beta^c;I^c) = I^c + \alpha^c \otimes \frac{1}{\beta^c} \otimes C(\beta^c;I^c)$$1, $$\mathrm{AAC}^c(\alpha^c,\beta^c;I^c) = I^c + \alpha^c \otimes \frac{1}{\beta^c} \otimes C(\beta^c;I^c)$$2) per channel.

2. Network Architecture and Model Compression

The central component is the Illuminance Adjustment (IA) block, which interleaves

1. Low-Light Area Enhancement (LLAE) — 9 iterations,
2. High-Light Area Suppression (HLAS) — 3 iterations,

Each iteration uses a Disordered Module (DM) to regress AAC parameters from the current feature map. During training, LLAE and HLAS employ $$\mathrm{AAC}^c(\alpha^c,\beta^c;I^c) = I^c + \alpha^c \otimes \frac{1}{\beta^c} \otimes C(\beta^c;I^c)$$3 and $$\mathrm{AAC}^c(\alpha^c,\beta^c;I^c) = I^c + \alpha^c \otimes \frac{1}{\beta^c} \otimes C(\beta^c;I^c)$$4 independent DMs, respectively, applied in randomized order within each mini-batch, thereby reducing specialization and improving convergence.

After training, these independent DMs are fused by averaging their weights:

$$\mathrm{AAC}^c(\alpha^c,\beta^c;I^c) = I^c + \alpha^c \otimes \frac{1}{\beta^c} \otimes C(\beta^c;I^c)$$5

During inference, the single fused DM ($$\mathrm{AAC}^c(\alpha^c,\beta^c;I^c) = I^c + \alpha^c \otimes \frac{1}{\beta^c} \otimes C(\beta^c;I^c)$$6) is applied recurrently for all iterations, minimizing memory and model size without degrading performance.

Inference Pipeline Overview

$I^c$9

3. Physics-Grounded Objective Function

The optimization objective is based on a Retinex decomposition of illumination ($$\mathrm{AAC}^c(\alpha^c,\beta^c;I^c) = I^c + \alpha^c \otimes \frac{1}{\beta^c} \otimes C(\beta^c;I^c)$$7) and reflectance ($$\mathrm{AAC}^c(\alpha^c,\beta^c;I^c) = I^c + \alpha^c \otimes \frac{1}{\beta^c} \otimes C(\beta^c;I^c)$$8), with several specialized regularization terms:

• Reflectance Consistency: Matches enhanced and original reflectance,

$$\mathrm{AAC}^c(\alpha^c,\beta^c;I^c) = I^c + \alpha^c \otimes \frac{1}{\beta^c} \otimes C(\beta^c;I^c)$$9

• White-Balance: Prevents color channel saturation,

$I^c$0

• Illuminance Consistency: Enforces plausible brightness via target illumination,

$I^c$1

• Curve Smoothness: Promotes spatial smoothness,

$I^c$2

• Denoising Loss: Jointly maximizes SSIM and penalizes residual gradients after pseudo-noise injection during training.

The total loss is

$I^c$3

with specified weights.

4. Denoising Module

A lightweight convolutional neural network (typically 4–6 convolutional layers) is positioned after IA to eliminate noise accentuated by strong enhancement. During training, pseudo-Gaussian noise is added to simulate real-world degradations, and the denoiser is optimized using a mixture of SSIM and gradient-based losses. During inference, this denoising block is applied once to the enhanced output.

5. Training Regimen and Randomized-Order Strategy

Training utilizes real-world low-light samples from the SCIE Part 1 dataset, rescaled to $I^c$4, with standard data augmentation (random flips, crops). The LLAE and HLAS modules are trained jointly for 100 epochs with random ordering of DMs per mini-batch, enforcing module flexibility. After fusing the DMs for inference, the denoising module is trained for an additional 200 epochs. The network is optimized using Adam with a learning rate of $I^c$5 and batch size 16.

A core feature is the randomized application order of DMs during training, which regularizes learning and permits all DMs to operate effectively at different iterative depths. A plausible implication is enhanced model robustness and reduced parameter redundancy.

6. Experimental Evaluation and Comparative Analysis

Self-DACE++ achieves a strong balance of compactness, efficiency, and restoration quality across multiple scales:

Model	Params	LOL-test (PSNR/SSIM)	SCIE-part2 (PSNR/SSIM)
ZeroDCE	0.079 M	14.86 / 0.56	14.81 / 0.69
RUAS	0.003 M	16.40 / 0.50	14.98 / 0.67
SCI	0.00035 M	14.78 / 0.52	14.07 / 0.65
Ours-Tiny	0.00034 M	17.65 / 0.61	19.85 / 0.74
Ours-Small	0.023 M	18.91 / 0.59	21.03 / 0.75
Ours	0.654 M	19.69 / 0.78	21.02 / 0.75

On downstream face detection tasks (DarkFace + RetinaFace), Ours-Small attains [email protected] = 0.666, outperforming all prior unsupervised and many supervised systems.

Ablation studies confirm the necessity of each objective component: omitting illuminance consistency ($I^c$6) collapses PSNR to 7.9 dB, removal of reflectance-consistency ($I^c$7) yields 10.2 dB, and absence of curve smoothness ($I^c$8) results in significant artifacts (PSNR ~ 14.2 dB).

Qualitative evaluations indicate that Self-DACE++ yields balanced luminance, accurate color reproduction, and minimal amplified noise relative to generative-based or vanilla curve-based approaches. Failure cases include persistently noisy outputs under severe non-Gaussian corruption and imperfect correction in regions with strong color casts outside white-balance priors.

7. Significance and Deployment Considerations

Self-DACE++ achieves real-time inference speeds on GPUs and edge devices. The architecture can be scaled down to ultra-compact variants (as low as 340 parameters at 51 FPS) without catastrophic quality loss. The demonstrated generalization to cross-domain data and efficacy as a pre-processing step for downstream vision tasks position Self-DACE++ as a practical solution for low-light enhancement in both academic and deployment contexts (Wen et al., 28 Apr 2026).