InterLight: Illumination-Aware Low-Light Enhancement

• InterLight is a low-light image enhancement framework that organizes its pipeline around intrinsic illumination priors, using physics-guided augmentation and self-supervised consistency.
• It employs a dual-branch U-Net design, separating luminance and chrominance with prompt-conditioned fusion and luminance-gated intrinsic memory for selective restoration.
• Empirical evaluations on datasets like LOL-v1 and LOL-v2 demonstrate improved PSNR and SSIM, mitigating issues such as over-enhancement, color distortion, and noise.

InterLight is a low-light image enhancement (LLIE) framework that organizes the enhancement pipeline around intrinsic illumination priors rather than treating low-light restoration as generic image-to-image translation. It is introduced as a method for addressing insufficient illumination, low contrast, detail loss, and noise, while targeting recurrent LLIE failure modes such as over-enhancement, color distortion, and weak robustness to non-ideal lighting and non-uniform noise (Wang et al., 19 May 2026). Its central claim is that robust enhancement requires not just estimating illumination, but constructing an illumination-aware pipeline spanning physics-guided augmentation, latent degradation representation, luminance-aware feature retrieval, and self-supervised consistency regularization (Wang et al., 19 May 2026).

1. Problem formulation and conceptual basis

InterLight is situated in the standard LLIE setting: given an underexposed image with low contrast, detail loss, color shift, and noise, the goal is to produce a visually natural well-lit image (Wang et al., 19 May 2026). The method is explicitly grounded in the usual LLIE intuition from Retinex, namely that an observed image is affected by illumination and reflectance, although it does not introduce a new explicit decomposition equation of the form $I = R \odot L$ (Wang et al., 19 May 2026).

The paper identifies several concrete shortcomings in prior methods: over-enhancement, color distortion, uniform-treatment restoration, limited handling of non-ideal lighting and sensor variation, and weak use of priors (Wang et al., 19 May 2026). In that framing, the distinctive feature of InterLight is not a single module but a systematic use of illumination-related priors at multiple levels of the model. These priors are defined as information naturally tied to how low-light images are formed and degraded, without requiring external labels, extra modalities, or pretrained semantic priors (Wang et al., 19 May 2026).

The paper distinguishes four forms of such priors. First, sensor-level illumination-response priors are injected through physics-guided augmentation. Second, latent degradation/illumination-state priors are represented by a learnable prompt derived from a global degradation code. Third, feature-level intrinsic restoration priors are implemented through an internal memory bank of global vectors and local patch patterns. Fourth, illumination-invariant priors via self-supervision are enforced through consistency under perturbations of enhanced outputs (Wang et al., 19 May 2026). This suggests that the method treats low-light enhancement as a structured restoration problem in which illumination state, region brightness, and chrominance reliability should affect the restoration pathway explicitly.

Within the broader literature, this emphasis on illumination-aware decomposition aligns with a larger trend of making light transport or illumination structure explicit rather than implicit. For example, NeILF++ represents incident light and outgoing radiance with coupled neural fields under an inter-reflection consistency relation (Zhang et al., 2023), while LightIt conditions diffusion generation on direct shading and normals to make lighting an editable control variable (Kocsis et al., 2024). InterLight differs in task and scale, but shares the premise that illumination should be operationalized as a first-class representation rather than treated as a nuisance variable.

2. Intrinsic illumination priors and staged pipeline

The InterLight pipeline is described as five stages: Physics-Guided Augmentation / Intrinsic-Consistent Data Expansion (ICDE), Latent Degradation Estimation and adaptive prompt generation (ADPG), a dual-branch HVI-space enhancement backbone, Luminance-Gated Intrinsic Memory (LGIM), and Perturbation-Invariant Consistency (PIC) (Wang et al., 19 May 2026).

The first stage, Physics-Guided Augmentation (PGA), injects sensor-level illumination-response priors during training. A mild channel-wise gamma perturbation is sampled as

$$\gamma_c \sim \mathcal{U}(0.95, 1.05), \quad c \in \{R,G,B\},$$

and the gamma-corrected image is blended with the original using

$$\mathbf{I}_{pga} = \alpha \cdot \mathbf{I}^{\gamma} + (1 - \alpha) \cdot \mathbf{I},$$

with

$$\alpha = 3t^2 - 2t^3, \quad t = \min\left(1, \frac{\mathcal{P}}{\tau_d}\right),$$

where $\mathcal{P} = \max_c(\mathbf{I}_c)$ and $\tau_d = 0.05$ (Wang et al., 19 May 2026). The stated purpose is to avoid unrealistic augmentation of very dark, noise-dominated pixels while still simulating mild sensor-response variation.

The second stage estimates a latent degradation state. A lightweight condition network $\mathcal{C}$ extracts a feature map and global average pooling produces

$$\mathbf{z} = \text{GAP}\left(\mathcal{C}(\mathbf{I}_{pga})\right) \in \mathbb{R}^{d_z},$$

with $d_z = 32$ (Wang et al., 19 May 2026). A learnable degradation dictionary

$$\mathbf{D} = [\mathbf{d}_1,\ldots,\mathbf{d}_K]^\top \in \mathbb{R}^{K \times d_p}$$

is used to convert that latent code into a prompt. Soft assignment coefficients are computed by

$$\gamma_c \sim \mathcal{U}(0.95, 1.05), \quad c \in \{R,G,B\},$$0

and the prompt is

$$\gamma_c \sim \mathcal{U}(0.95, 1.05), \quad c \in \{R,G,B\},$$1

where $$\gamma_c \sim \mathcal{U}(0.95, 1.05), \quad c \in \{R,G,B\},$$2 is GELU, $$\gamma_c \sim \mathcal{U}(0.95, 1.05), \quad c \in \{R,G,B\},$$3, and $$\gamma_c \sim \mathcal{U}(0.95, 1.05), \quad c \in \{R,G,B\},$$4 (Wang et al., 19 May 2026). The prompt is therefore a continuous representation of the scene’s latent illumination/degradation state rather than a manually defined degradation label.

The third stage uses the HVI color space, adopted from CIDNet, to separate luminance and chrominance while suppressing unreliable color in dark regions (Wang et al., 19 May 2026). For input $$\gamma_c \sim \mathcal{U}(0.95, 1.05), \quad c \in \{R,G,B\},$$5, intensity is defined by

$$\gamma_c \sim \mathcal{U}(0.95, 1.05), \quad c \in \{R,G,B\},$$6

and the density-adaptive modulation is

$$\gamma_c \sim \mathcal{U}(0.95, 1.05), \quad c \in \{R,G,B\},$$7

where $$\gamma_c \sim \mathcal{U}(0.95, 1.05), \quad c \in \{R,G,B\},$$8 is learnable and $$\gamma_c \sim \mathcal{U}(0.95, 1.05), \quad c \in \{R,G,B\},$$9 (Wang et al., 19 May 2026). The HVI coordinates are then

$$\mathbf{I}_{pga} = \alpha \cdot \mathbf{I}^{\gamma} + (1 - \alpha) \cdot \mathbf{I},$$0

The paper states that $$\mathbf{I}_{pga} = \alpha \cdot \mathbf{I}^{\gamma} + (1 - \alpha) \cdot \mathbf{I},$$1 is the intensity/illumination channel, $$\mathbf{I}_{pga} = \alpha \cdot \mathbf{I}^{\gamma} + (1 - \alpha) \cdot \mathbf{I},$$2 encode chrominance, and $$\mathbf{I}_{pga} = \alpha \cdot \mathbf{I}^{\gamma} + (1 - \alpha) \cdot \mathbf{I},$$3 suppresses unreliable color information in dark regions (Wang et al., 19 May 2026).

3. Network architecture and restoration mechanisms

The backbone is a dual-branch four-level U-Net comprising an I-branch for illumination restoration from $$\mathbf{I}_{pga} = \alpha \cdot \mathbf{I}^{\gamma} + (1 - \alpha) \cdot \mathbf{I},$$4 and an HV-branch for chrominance refinement from $$\mathbf{I}_{pga} = \alpha \cdot \mathbf{I}^{\gamma} + (1 - \alpha) \cdot \mathbf{I},$$5, with Lightweight Cross-Attention (LCA) between branches (Wang et al., 19 May 2026). This division of labor is central to the method’s architecture: illumination restoration and chrominance restoration are coupled but not conflated.

Prompt injection occurs in the HV-branch through the Prior-Responsive Fusion Block (PRFB). Channel-wise affine modulation is produced from the prompt by

$$\mathbf{I}_{pga} = \alpha \cdot \mathbf{I}^{\gamma} + (1 - \alpha) \cdot \mathbf{I},$$6

and applied to a feature map $$\mathbf{I}_{pga} = \alpha \cdot \mathbf{I}^{\gamma} + (1 - \alpha) \cdot \mathbf{I},$$7 through

$$\mathbf{I}_{pga} = \alpha \cdot \mathbf{I}^{\gamma} + (1 - \alpha) \cdot \mathbf{I},$$8

A spatial prompt map is then formed as

$$\mathbf{I}_{pga} = \alpha \cdot \mathbf{I}^{\gamma} + (1 - \alpha) \cdot \mathbf{I},$$9

The paper further defines

$$\alpha = 3t^2 - 2t^3, \quad t = \min\left(1, \frac{\mathcal{P}}{\tau_d}\right),$$0

where

$$\alpha = 3t^2 - 2t^3, \quad t = \min\left(1, \frac{\mathcal{P}}{\tau_d}\right),$$1

and uses this to construct the attention query

$$\alpha = 3t^2 - 2t^3, \quad t = \min\left(1, \frac{\mathcal{P}}{\tau_d}\right),$$2

Key and value come from the I-branch feature $$\alpha = 3t^2 - 2t^3, \quad t = \min\left(1, \frac{\mathcal{P}}{\tau_d}\right),$$3: $$\alpha = 3t^2 - 2t^3, \quad t = \min\left(1, \frac{\mathcal{P}}{\tau_d}\right),$$4 The normalized attention is written as

$$\alpha = 3t^2 - 2t^3, \quad t = \min\left(1, \frac{\mathcal{P}}{\tau_d}\right),$$5

and the output is

$$\alpha = 3t^2 - 2t^3, \quad t = \min\left(1, \frac{\mathcal{P}}{\tau_d}\right),$$6

(Wang et al., 19 May 2026). The stated role of PRFB is to make chrominance restoration degradation-aware and spatially adaptive.

The method’s most distinctive restoration mechanism is Luminance-Gated Intrinsic Memory (LGIM) at the bottleneck (Wang et al., 19 May 2026). The memory contains global memory vectors $$\alpha = 3t^2 - 2t^3, \quad t = \min\left(1, \frac{\mathcal{P}}{\tau_d}\right),$$7 and local patch memory entries $$\alpha = 3t^2 - 2t^3, \quad t = \min\left(1, \frac{\mathcal{P}}{\tau_d}\right),$$8, both learned during training. Given an input feature $$\alpha = 3t^2 - 2t^3, \quad t = \min\left(1, \frac{\mathcal{P}}{\tau_d}\right),$$9, retrieved memory content $\mathcal{P} = \max_c(\mathbf{I}_c)$0 is fused as

$\mathcal{P} = \max_c(\mathbf{I}_c)$1

where

$\mathcal{P} = \max_c(\mathbf{I}_c)$2

The paper interprets $\mathcal{P} = \max_c(\mathbf{I}_c)$3 as a learned brightness gate: if $\mathcal{P} = \max_c(\mathbf{I}_c)$4, the region is dark and memory contribution is amplified; if $\mathcal{P} = \max_c(\mathbf{I}_c)$5, the region is bright and memory acts conservatively (Wang et al., 19 May 2026). This is the main mechanism for reducing over-enhancement in already reliable regions while compensating more aggressively in weak-signal areas. The I-branch uses stronger fusion with $\mathcal{P} = \max_c(\mathbf{I}_c)$6, while the HV-branch uses more conservative fusion with $\mathcal{P} = \max_c(\mathbf{I}_c)$7 (Wang et al., 19 May 2026).

After decoding, both branches are fused and transformed back through the inverse HVI transform to RGB with a global residual connection (Wang et al., 19 May 2026). This architecture uses U-Net channels $\mathcal{P} = \max_c(\mathbf{I}_c)$8, PRFB spatial prior resolutions $\mathcal{P} = \max_c(\mathbf{I}_c)$9, and LGIM memory entries $\tau_d = 0.05$0, with patch size $\tau_d = 0.05$1 in the reported implementation (Wang et al., 19 May 2026).

4. Objectives, supervision, and optimization

The basic reconstruction objective is

$\tau_d = 0.05$2

combining pixel-wise $\tau_d = 0.05$3, SSIM, Laplacian edge loss, and VGG-based perceptual loss (Wang et al., 19 May 2026). The total loss is

$\tau_d = 0.05$4

so supervision is applied in both RGB and HVI domains (Wang et al., 19 May 2026). Reported coefficients are $\tau_d = 0.05$5 and $\tau_d = 0.05$6 (Wang et al., 19 May 2026).

The self-supervised regularizer is Perturbation-Invariant Consistency (PIC). Starting from an enhanced output $\tau_d = 0.05$7, the method forms

$\tau_d = 0.05$8

where $\tau_d = 0.05$9 and $\mathcal{C}$0 is Gaussian blur with kernel size 9–21 and $\mathcal{C}$1 (Wang et al., 19 May 2026). The consistency loss is

$\mathcal{C}$2

with cosine-decayed weight

$\mathcal{C}$3

where $\mathcal{C}$4 (Wang et al., 19 May 2026). The paper interprets this as distilling illumination-invariant or perturbation-invariant features intrinsic to the data.

InterLight also uses dual-path supervision: a baseline path without memory and a memory-enhanced path. This is written as

$\mathcal{C}$5

with $\mathcal{C}$6 (Wang et al., 19 May 2026). The stated intent is to stabilize enhancement while ensuring that LGIM contributes complementary rather than destabilizing gains.

Training uses Adam with $\mathcal{C}$7, $\mathcal{C}$8, initial learning rate $\mathcal{C}$9, cosine annealing, 1500 epochs, batch size 8, and crop size $$\mathbf{z} = \text{GAP}\left(\mathcal{C}(\mathbf{I}_{pga})\right) \in \mathbb{R}^{d_z},$$0 (Wang et al., 19 May 2026). Random horizontal and vertical flips are used. The framework is implemented in PyTorch and trained on NVIDIA RTX 4090 GPUs (Wang et al., 19 May 2026). During inference, PGA and PIC are bypassed; the image is transformed to HVI, the prompt is estimated, the dual-branch network runs, LGIM contributes retrieval-based compensation, and the inverse HVI transform reconstructs the RGB output (Wang et al., 19 May 2026). The reported model size is 10.91M parameters and 8.41G FLOPs (Wang et al., 19 May 2026).

5. Empirical evaluation and ablation evidence

InterLight is evaluated on LOL-v1, LOL-v2 (Real and Synthetic subsets), SICE, SID (Sony-Total-Dark), and LSRW-Huawei, using PSNR and SSIM (Wang et al., 19 May 2026). Baselines listed in the paper include RetinexNet, KinD, RUAS, Zero-DCE, EnlightenGAN, LLFormer, SNR-Aware, Bread, PairLIE, Retinexformer, LightenDiff, CWNet, and CIDNet (Wang et al., 19 May 2026).

On LOL-v1, InterLight achieves PSNR 24.78 and SSIM 0.862, which the paper describes as second-best PSNR behind Retinexformer’s 25.16 and best SSIM in the table (Wang et al., 19 May 2026). Compared with CIDNet, it improves from 23.81 to 24.78 PSNR and from 0.857 to 0.862 SSIM (Wang et al., 19 May 2026). On LOL-v2-Real, it reports PSNR 24.06 and SSIM 0.866, slightly surpassing CIDNet’s 23.90 / 0.865 (Wang et al., 19 May 2026). On LOL-v2-Syn, it reports PSNR 25.73 and SSIM 0.935, edging out CIDNet in PSNR though not in SSIM (Wang et al., 19 May 2026).

For harder or more sensor-specific settings, the narrative emphasizes SID and LSRW-Huawei. On SID, InterLight achieves 22.98 dB PSNR, beating CIDNet by 0.08 dB and LLFlow by over 6.7 dB (Wang et al., 19 May 2026). On SICE, it reaches 13.56 dB PSNR, the best among listed methods (Wang et al., 19 May 2026). On LSRW-Huawei, it obtains PSNR 21.39 and SSIM 0.625, the best PSNR in the table though not the best SSIM (Wang et al., 19 May 2026). The paper interprets the LSRW-Huawei result as evidence that physics-guided augmentation helps transfer across different sensor characteristics (Wang et al., 19 May 2026).

The principal ablation is performed on LOL-v1. The baseline model gives 23.46 PSNR and 0.842 SSIM (Wang et al., 19 May 2026). Adding ADPG yields 24.21 PSNR and 0.859 SSIM, a gain of +0.75 dB. Adding LGIM yields 24.27 PSNR and 0.859 SSIM, a gain of +0.81 dB. Adding ICDE yields 23.87 PSNR and 0.848 SSIM, a gain of +0.41 dB (Wang et al., 19 May 2026). The full model reaches 24.78 PSNR and 0.866 SSIM, for a total gain of +1.32 dB over baseline (Wang et al., 19 May 2026). The paper explicitly concludes that the architectural illumination-prior modules, especially ADPG and LGIM, are the major contributors, while ICDE adds additional robustness (Wang et al., 19 May 2026).

Qualitatively, InterLight is reported to produce clearer textures, more stable brightness, more coherent colors, improved dark-detail recovery, and more realistic enhancement than methods that may oversmooth or distort (Wang et al., 19 May 2026). The paper specifically attributes improvements in color stability to the HVI representation and prompt-guided chrominance restoration, and attributes improvements in over-enhancement control to luminance-gated memory fusion (Wang et al., 19 May 2026).

6. Position in illumination-aware vision research, limitations, and significance

InterLight belongs to a broader family of methods that make illumination explicit in the representation rather than leaving it entangled with appearance. In inverse rendering, GLOW addresses dynamic co-located light and camera capture by explicitly modeling global illumination and inter-reflections with a dynamic radiance cache (Wu et al., 28 Nov 2025). NeILF++ uses an incident light field and an outgoing radiance field coupled by an inter-reflection consistency constraint (Zhang et al., 2023). In generative modeling, LightIt controls diffusion outputs with direct shading and normal maps so that lighting becomes an editable condition (Kocsis et al., 2024). InterLight is not an inverse-rendering or relighting method, but it shares the same methodological tendency: light should be represented, conditioned on, and regularized structurally.

The method’s significance within LLIE is described less as a single isolated block and more as a coherent architectural thesis: illumination-aware priors should influence data augmentation, degradation representation, feature conditioning, selective restoration, and self-supervision simultaneously (Wang et al., 19 May 2026). This suggests a shift away from scaling model size alone and toward more structured low-light restoration.

The paper states two explicit limitations. First, the dual-branch architecture plus LGIM adds complexity compared with simpler baselines (Wang et al., 19 May 2026). Second, the physics-guided augmentation still assumes relatively linear degradation behavior and may not capture all real sensor/image formation complexities (Wang et al., 19 May 2026). A plausible implication is that the method’s sensor-transfer advantages may remain bounded by the simplifications in the augmentation model. The paper proposes model compression / lightweight deployment and extension to video low-light enhancement as future work (Wang et al., 19 May 2026).

In technical summary, InterLight is a low-light image enhancement framework in which illumination priors are embedded at four levels: sensor-response augmentation, latent degradation prompting, luminance-gated intrinsic memory, and perturbation-invariant consistency (Wang et al., 19 May 2026). Its architecture combines HVI-space decomposition, prompt-conditioned cross-branch interaction, and brightness-aware memory fusion to treat dark and bright regions differently. Empirically, the reported gains on LOL-v1, LOL-v2, SICE, SID, and LSRW-Huawei, together with the ablations, support the claim that illumination-aware conditioning and selective intrinsic restoration are the main sources of improvement (Wang et al., 19 May 2026).