---
title: 'InterLight: Illumination-Aware Low-Light Enhancement'
url: https://www.emergentmind.com/topics/interlight
type: topic
---

# InterLight: Illumination-Aware Low-Light Enhancement

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 [2605.19982]. 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 [2605.19982].

## 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 [2605.19982]. 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 \(I = R \odot L\) [2605.19982].

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** [2605.19982]. 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 [2605.19982].

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 [2605.19982]. 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 [2303.17147], while LightIt conditions diffusion generation on direct shading and normals to make lighting an editable control variable [2403.10615]. 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)** [2605.19982].

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 \(\tau_d = 0.05\) [2605.19982]. 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 \(d_z = 32\) [2605.19982]. 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
\[
\boldsymbol{\alpha} = \text{Softmax}(\mathbf{W}_\alpha \mathbf{z}) \in \mathbb{R}^K,
\]
and the prompt is
\[
\mathbf{p} = \phi\left(\boldsymbol{\alpha}^\top \mathbf{D}\right)
= \phi\left(\sum_{k=1}^{K} \alpha_k \mathbf{d}_k\right) \in \mathbb{R}^{d_p},
\]
where \(\phi(\cdot)\) is GELU, \(K=32\), and \(d_p=512\) [2605.19982]. 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 [2605.19982]. For input \(\mathbf{I}_{in} \in \mathbb{R}^{H \times W \times 3}\), intensity is defined by
\[
\mathcal{P} = \max_c(\mathbf{I}_{in}^c),
\]
and the density-adaptive modulation is
\[
\mathbf{C}_k = \left(\sin\left(\frac{\pi \mathcal{P}}{2}\right) + \epsilon\right)^k,
\]
where \(k\) is learnable and \(\epsilon=10^{-8}\) [2605.19982]. The HVI coordinates are then
\[
\mathcal{H} = \mathbf{C}_k \cdot S \cdot \cos(2\pi H), \quad
\mathcal{V} = \mathbf{C}_k \cdot S \cdot \sin(2\pi H), \quad
\mathcal{I} = \mathcal{P}.
\]
The paper states that \(\mathcal{I}\) is the intensity/illumination channel, \((\mathcal{H},\mathcal{V})\) encode chrominance, and \(\mathbf{C}_k\) suppresses unreliable color information in dark regions [2605.19982].

## 3. Network architecture and restoration mechanisms

The backbone is a **dual-branch four-level U-Net** comprising an **I-branch** for illumination restoration from \(\mathcal{I}\) and an **HV-branch** for chrominance refinement from \((\mathcal{H},\mathcal{V})\), with **Lightweight Cross-Attention (LCA)** between branches [2605.19982]. 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
\[
\boldsymbol{\gamma} = \sigma(\mathbf{W}_\gamma \mathbf{p}) \in \mathbb{R}^{C}, \quad
\boldsymbol{\beta} = \mathbf{W}_\beta \mathbf{p} \in \mathbb{R}^{C},
\]
and applied to a feature map \(\mathbf{X} \in \mathbb{R}^{C \times H \times W}\) through
\[
\mathbf{X}' = \mathbf{X} \odot (1 + \boldsymbol{\gamma}) + \boldsymbol{\beta}.
\]
A spatial prompt map is then formed as
\[
\mathbf{U} = \text{Conv}\left(\text{Reshape}_{h \times w}(\mathbf{W}_s \mathbf{p})\right), \qquad
\mathbf{P}_s = \text{Upsample}_{H \times W}(\mathbf{U}) \in \mathbb{R}^{C \times H \times W}.
\]
The paper further defines
\[
\mathbf{G} = \sigma\left(\text{Conv}_{1\times1}\left(\text{DWConv}\left([\mathbf{q}_1; \mathbf{q}_2]\right)\right)\right),
\]
where
\[
\mathbf{q}_1 = \text{Conv}_{1\times1}(\mathbf{P}_s), \quad
\mathbf{q}_2 = \text{Conv}_{1\times1}(\mathbf{X}'),
\]
and uses this to construct the attention query
\[
\mathbf{Q} = \text{DWConv}\left(\mathbf{G} \odot \mathbf{q}_1 + (1 - \mathbf{G}) \odot \mathbf{q}_2\right).
\]
Key and value come from the I-branch feature \(\mathbf{Y}\):
\[
\mathbf{K}, \mathbf{V} = \text{Split}\left(\text{DWConv}\left(\text{Conv}_{1\times1}(\mathbf{Y})\right)\right).
\]
The normalized attention is written as
\[
\text{Attn}(\mathbf{Q}, \mathbf{K}, \mathbf{V}) =
\text{Softmax}\left(\frac{\bar{\mathbf{Q}}\bar{\mathbf{K}}^\top}{\tau_a}\right)\mathbf{V},
\]
and the output is
\[
\mathbf{X}_{out} = \mathbf{X} + \text{Attn}(\mathbf{Q}, \mathbf{K}, \mathbf{V}) + \text{FFN}(\mathbf{X})
\]
[2605.19982]. 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 [2605.19982]. The memory contains **global memory vectors** \(\{\mathbf{m}_l^v\}_{l=1}^{L} \in \mathbb{R}^{C}\) and **local patch memory entries** \(\{\mathbf{m}_l^p\}_{l=1}^{L} \in \mathbb{R}^{C \times r \times r}\), both learned during training. Given an input feature \(\mathbf{F}_{in}\), retrieved memory content \(\mathbf{F}_{mem}\) is fused as
\[
\mathbf{F}_{out} = \mathbf{F}_{in} + \lambda \cdot (1 + \sigma(\eta)(1 - g)) \cdot \mathbf{F}_{mem},
\]
where
\[
g = \sigma(\text{MLP}(\text{GAP}(\mathbf{F}_{in}))) \in [0,1].
\]
The paper interprets \(g\) as a learned brightness gate: if \(g \to 0\), the region is dark and memory contribution is amplified; if \(g \to 1\), the region is bright and memory acts conservatively [2605.19982]. 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 \(\lambda_{init}=1.2\), while the HV-branch uses more conservative fusion with \(\lambda_{init}=0.8\) [2605.19982].

After decoding, both branches are fused and transformed back through the inverse HVI transform to RGB with a global residual connection [2605.19982]. This architecture uses U-Net channels \([36, 36, 72, 144]\), PRFB spatial prior resolutions \(\{16, 8, 4\}\), and LGIM memory entries \(L=16\), with patch size \(k=4\) in the reported implementation [2605.19982].

## 4. Objectives, supervision, and optimization

The basic reconstruction objective is
\[
\mathcal{L}_{rec} = \mathcal{L}_{L1} + \mathcal{L}_{ssim} + \mathcal{L}_{edge} + \mu_p \mathcal{L}_{perc},
\]
combining pixel-wise \(L_1\), SSIM, Laplacian edge loss, and VGG-based perceptual loss [2605.19982]. The total loss is
\[
\mathcal{L}_{total} = \mathcal{L}_{rec}^{RGB} + \mu_{hvi}\mathcal{L}_{rec}^{HVI} + \mathcal{L}_{consistency},
\]
so supervision is applied in both RGB and HVI domains [2605.19982]. Reported coefficients are \(\mu_{hvi}=0.5\) and \(\mu_p=0.1\) [2605.19982].

The self-supervised regularizer is **Perturbation-Invariant Consistency (PIC)**. Starting from an enhanced output \(\mathbf{I}_e\), the method forms
\[
\mathbf{I}_w = \text{CenterCrop}_{s}(\mathbf{I}_e), \quad
\mathbf{I}_s = \mathcal{G}_{\sigma}(\mathbf{I}_w),
\]
where \(s=16\) and \(\mathcal{G}_{\sigma}\) is Gaussian blur with kernel size 9–21 and \(\sigma \sim \mathcal{U}(0.1,5)\) [2605.19982]. The consistency loss is
\[
\mathcal{L}_{consistency} = \beta(t)\cdot \|\mathbf{I}_w - \mathbf{I}_s\|_2^2,
\]
with cosine-decayed weight
\[
\beta(t) = \frac{\beta_0}{2}\left(\cos\left(\frac{\pi t}{T}\right)+1\right),
\]
where \(\beta_0=0.1\) [2605.19982]. 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
\[
\mathcal{L}_{dual} = \mathcal{L}_{total}(\hat{\mathbf{I}}_{base}) + \lambda_{lgim}\mathcal{L}_{total}(\hat{\mathbf{I}}_{mem}),
\]
with \(\lambda_{lgim}=1.0\) [2605.19982]. The stated intent is to stabilize enhancement while ensuring that LGIM contributes complementary rather than destabilizing gains.

Training uses Adam with \(\beta_1 = 0.9\), \(\beta_2 = 0.999\), initial learning rate \(2 \times 10^{-4}\), cosine annealing, 1500 epochs, batch size 8, and crop size \(256 \times 256\) [2605.19982]. Random horizontal and vertical flips are used. The framework is implemented in PyTorch and trained on NVIDIA RTX 4090 GPUs [2605.19982]. 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 [2605.19982]. The reported model size is **10.91M parameters** and **8.41G FLOPs** [2605.19982].

## 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** [2605.19982]. Baselines listed in the paper include RetinexNet, KinD, RUAS, Zero-DCE, EnlightenGAN, LLFormer, SNR-Aware, Bread, PairLIE, Retinexformer, LightenDiff, CWNet, and CIDNet [2605.19982].

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 [2605.19982]. Compared with **CIDNet**, it improves from 23.81 to 24.78 PSNR and from 0.857 to 0.862 SSIM [2605.19982]. On **LOL-v2-Real**, it reports **PSNR 24.06** and **SSIM 0.866**, slightly surpassing CIDNet’s 23.90 / 0.865 [2605.19982]. On **LOL-v2-Syn**, it reports **PSNR 25.73** and **SSIM 0.935**, edging out CIDNet in PSNR though not in SSIM [2605.19982].

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** [2605.19982]. On **SICE**, it reaches **13.56 dB PSNR**, the best among listed methods [2605.19982]. On **LSRW-Huawei**, it obtains **PSNR 21.39** and **SSIM 0.625**, the best PSNR in the table though not the best SSIM [2605.19982]. The paper interprets the LSRW-Huawei result as evidence that physics-guided augmentation helps transfer across different sensor characteristics [2605.19982].

The principal ablation is performed on **LOL-v1**. The baseline model gives **23.46 PSNR** and **0.842 SSIM** [2605.19982]. 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** [2605.19982]. The full model reaches **24.78 PSNR** and **0.866 SSIM**, for a total gain of **+1.32 dB** over baseline [2605.19982]. The paper explicitly concludes that the architectural illumination-prior modules, especially **ADPG** and **LGIM**, are the major contributors, while **ICDE** adds additional robustness [2605.19982].

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 [2605.19982]. 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 [2605.19982].

## 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 [2511.22857]. NeILF++ uses an incident light field and an outgoing radiance field coupled by an inter-reflection consistency constraint [2303.17147]. In generative modeling, LightIt controls diffusion outputs with direct shading and normal maps so that lighting becomes an editable condition [2403.10615]. 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** [2605.19982]. 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 [2605.19982]. Second, the **physics-guided augmentation still assumes relatively linear degradation behavior** and may not capture all real sensor/image formation complexities [2605.19982]. 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 [2605.19982].

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 [2605.19982]. 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 [2605.19982].

Source: https://www.emergentmind.com/topics/interlight