EnlightenGANs with DRU Framework

• EnlightenGANs are generative models for low-light image enhancement that use an autoencoder-based generator with skip connections and dual discriminators to preserve image content and aesthetics.
• They incorporate the DRU framework, which quantifies illumination uncertainty through learned relativistic probabilities, dynamically re-weighting adversarial objectives akin to quantum wave–particle duality.
• Empirical studies on anime scenery datasets show that DRU-enhanced EnlightenGANs outperform baseline methods, achieving better BRISQUE, PIQE, and NIMA scores while reducing color bias and artifacts.

EnlightenGANs are a class of Generative Adversarial Networks originally developed for unsupervised low-light image enhancement with a focus on preserving content, structure, and perceptual aesthetics. They employ an autoencoder-based generator with skip-connections and a dual discriminator setup, and emphasize self-feature preservation to avoid content loss. The introduction of the Data Relativistic Uncertainty (DRU) framework, as realized in recent work on anime scenery enhancement, fundamentally extends EnlightenGANs by incorporating explicit representations of illumination uncertainty analogous to quantum wave–particle duality, dynamically re-weighting objectives using learned sample-wise confidence scores—termed "relativistic probabilities"—to improve performance and robustness under diverse conditions (Gao et al., 26 Dec 2025).

1. Theoretical Foundations: From Wave–Particle Analogy to Relativistic Uncertainty

The DRU framework underlying extended EnlightenGANs draws inspiration from both quantum formalism and Relativistic GANs (RaGAN). In quantum theory, the wave–particle duality requires every measurement to resolve a probabilistic superposition into a definite outcome. DRU leverages this analogy: each input image’s illumination is treated as a superposition between perfect “dark” and “bright” states, with a learned classifier producing a two-component probability vector $(RP_d, RP_b)$ encoding confidence in each state, $\sum RP=1$.

The “relativistic” aspect, in contrast to RaGAN’s real–fake comparison, quantifies how far each training example is from an (unobtainable) ideal dark or bright prototype. This confidence is then used to weight the adversarial and perceptual objective functions during training, so that higher-confidence samples exert greater influence—mirroring the collapse of a quantum amplitude to a classical measurement (Gao et al., 26 Dec 2025, Skilling et al., 2020).

2. Mathematical Formulation of EnlightenGANs under DRU

Illumination Uncertainty Quantification

Images are partitioned via quartile mean brightness into "confident-dark," "confident-bright," and "uncertain" sets: $$\text{mean}(q) = \frac{1}{wh} \sum_{x=0}^{w-1} \sum_{y=0}^{h-1} q(x, y)$$

$$\begin{cases} \text{mean}(q) < B_\text{low} \implies \text{confident-dark}\ \text{mean}(q) > B_\text{high} \implies \text{confident-bright}\ \text{otherwise} \implies \text{uncertain} \end{cases}$$

with typical thresholds $B_\text{low}=50$, $B_\text{high}=150$. A classifier $F_q$ (e.g., ResNet, ViT) is trained on confident subsets to output

$$F_q(x) = \text{softmax}(\omega_q x + \beta_q) = [p_d, p_b]$$

where $p_d$, $p_b=1-p_d$ are the probabilities for "dark" and "bright" respectively.

DRU-weighted Losses

Loss terms for global relativistic (RaGAN) and local patch (LSGAN) discriminators are modulated per-sample by $RP_d$, $RP_b$ as follows:

$$\begin{aligned} L_D^{\mathrm{global}} &= RP_b\,RP_d\left[\left(D_{\mathrm{RGAN}}(I_b, G(I_d)) - 1\right)^2 + D_{\mathrm{RGAN}}(G(I_d), I_b)^2\right] \ L_D^{\mathrm{local}} &= RP_b\left(D_{\mathrm{LSGAN}}(\overline{I_b})-1\right)^2 + RP_d D_{\mathrm{LSGAN}}(\overline{G(I_d)})^2 \ L_G^{\mathrm{global}} &= RP_d\,RP_b\left[\left(D_{\mathrm{RGAN}}(G(I_d), I_b) - 1\right)^2 + D_{\mathrm{RGAN}}(I_b, G(I_d))^2\right] \ L_G^{\mathrm{local}} &= RP_d\left(D_{\mathrm{LSGAN}}(\overline{G(I_d)})-1\right)^2 \ \end{aligned}$$

With additional self-feature preserving losses from the original EnlightenGAN, the total generator objective is:

$$L_{\mathrm{overall}} = L_D^{\mathrm{global}} + L_D^{\mathrm{local}} + L_G^{\mathrm{global}} + L_G^{\mathrm{local}} + L_{\mathrm{SFP}}^{\mathrm{global}} + L_{\mathrm{SFP}}^{\mathrm{local}}$$

3. Network Architecture and Dynamic Training Pipeline

EnlightenGAN under DRU employs an autoencoder with skip connections for the generator, and both a global relativistic GAN discriminator ($D_{RGAN}$) and a local patch-based ($D_{LSGAN}$) discriminator. The DRU extension introduces a separate “probability head” network $F_q$ (ResNet, ViT, or similar), pre-trained on luminance-labeled subsets to output per-sample $(RP_d, RP_b)$. During adversarial training, each sample in a minibatch is passed through $F_q$, and its output is immediately used to proportionally weight the adversarial objective terms.

This pipeline allows emphasis to dynamically shift toward higher-confidence samples at every iteration, with gradient flow adaptively directed according to the current state of the uncertainty quantifier. As $F_q$ adapts through training, sample scores and therefore their impact on optimization continually recalibrate.

4. Dataset Construction, Data Partition, and Evaluation Metrics

The DRU-augmented EnlightenGAN is evaluated on an 18,804-image unpaired anime scenery dataset compiled from Scenimefy frames, AnimeGAN synthetic images, and CycleGAN-translated natural scenes. The quartile-based mean-brightness test partitions the data, and a ResNet18 classifier (trained on confident dark/bright) is used to reclassify ambiguous cases, yielding larger, relabeled training splits:

Subset	Count	Description
trainDark	8,240	High-confidence dark frames
trainBright	8,501	High-confidence bright frames
testDark	2,063	Hold-out confident dark for evaluation

No ground-truth normal-light anime frames are available; therefore, evaluation relies exclusively on no-reference quality and aesthetics metrics:

• BRISQUE: Lower is better (blind reference-less perceptual quality)
• PIQE: Lower is better (perceptual image quality)
• PI: Lower is better (Perceptual Index, combines NRQM and Ma metrics)
• NIMA: Higher is better (image aesthetics)

5. Empirical Performance and Analysis

Quantitative evaluation of DRU-enhanced EnlightenGAN using ViT-B16 as $F_q$ produces improvements over the vanilla baseline and prior SOTA unsupervised/zero-reference methods (SCI, ZeroDCE++, RUAS, CLIP-LIT). On the held-out test set:

Model	BRISQUE	PIQE	PI	NIMA
Vanilla EG	27.28	45.11	4.39	4.75
DRU (ViT)	26.45	42.72	4.30	4.79

Color bias is reduced, with better restoration of shadows and textures. Qualitative analysis demonstrates superior handling of ambiguous illumination and reduced artifacts compared to baselines such as ZeroDCE++ and RUAS (Gao et al., 26 Dec 2025).

Ablation studies reveal that combining confident and uncertain subsets yields optimal BRISQUE/NIMA trade-offs. Robustness is further substantiated: label noise degrades DRU performance less than vanilla EnlightenGAN due to the attenuation effect of uncertainty-based weighting.

6. Data-Centric Learning and Broader Impact

The DRU-enhanced EnlightenGAN exemplifies a data-centric paradigm in LLIE research: sample-level ambiguity is not ignored or thresholded but encoded, quantified, and dynamically injected into training. All data, including those in "grey zone" mid-tones, contribute according to confidence, sidestepping brittleness associated with hard labeling.

This approach generalizes to other visual domains with inherent annotation uncertainty (e.g., segmentation of occluded regions, style–content disentanglement, low-contrast medical imaging) and even to language domains—where probabilistic quantizers can adapt objectives based on gradations of sentiment or formality, thus mitigating the negative effect of noisy supervision.

The DRU framework thus extends unsupervised generative learning for ill-posed scenarios, clarifies the relationship between quantum-inspired probabilistic coding and adversarial objectives, and sets a foundation for systematic data-centric strategies across modalities (Gao et al., 26 Dec 2025, Skilling et al., 2020).