Learning Adaptive Lighting via Channel-Aware Guidance

We sincerely appreciate the reviewer’s insightful comments and provide detailed responses below.

Q1. What do cA, cH, cV and cD mean in equation 11?

A1: These variables represent the components of the 2D wavelet transform applied to the input image:

1. cA denotes the approximation coefficients of the wavelet transform.
2. cH denotes the horizontal detail coefficients.
3. cV denotes the vertical detail coefficients.
4. cD denotes the diagonal detail coefficients.

The wavelet transform decomposes the image into different frequency bands, which enables effective separation of illumination and detail information. These components are then leveraged in our color mixing representation to enhance adaptive lighting adjustment. We will clarify this in the revised manuscript to improve readability.

Q2. The structure of Light Guided Attention seems basic SSM(Mamba) and Attention structure, do you make any improvements?

A2: We acknowledge the reviewer’s concerns regarding the relationship between our Light State Space Module (LSSM) and Mamba. While LSSM shares similarities with state-space modeling principles, we introduce several key improvements beyond MambaIR:

1. Gated MLP instead of CAB: Unlike MambaIR, which uses the CAB module to mitigate channel redundancy, we found through activation visualizations that such redundancy was not significant in our task. Instead, we propose a Gated MLP module, which jointly processes spatial and channel-wise features, enhancing adaptability.
2. Channel separation in LSSM: We inject channel-separated features into the light adaptation process, enabling the network to better perceive channel brightness variations, thereby improving color consistency and balance across RGB channels.
3. LSSM as an optional component: While LSSM enhances feature extraction, it is not irreplaceable. We experimented with substituting LSSM with residual blocks and self-attention modules, and the model still achieved performance on par with state-of-the-art methods.

Additionally, LGA utilizes channel-separated features to guide color-mixed features, ensuring effective lighting adaptation while maintaining cross-channel consistency—a novel application not explored in prior works. We will provide clearer explanations and appropriate citations in the revision.

Q3. How about the running speed of the proposed method for each image? Does it have advantages over other methods?

A3: We recognize the importance of computational efficiency and provide a running time comparison with state-of-the-art methods.

Table R3: Inference Speed Comparison (Seconds per Image on 480p).

Our method achieves a competitive inference time of 0.018s per image, which is faster than RetinexFormer, RetinexMamba, and MambaIR, while maintaining superior enhancement quality. This demonstrates a favorable balance between efficiency and effectiveness.

Q4. What are the differences between LALNet-Tiny and LALNet?

A4: LALNet-Tiny is a lightweight variant of LALNet, designed to improve efficiency while preserving strong performance. The key differences include:

• Reduced channel dimensions: Fewer feature channels to reduce computational complexity.
• Fewer LSSM modules: A more compact design with fewer LSSM modules.

We will include a brief section describing LALNet-Tiny design and relationship to LALNet.

We appreciate the reviewer’s recognition of our contributions and their thoughtful feedback. We believe these clarifications and improvements will further strengthen our work. Thank you for your time and valuable insights.

We sincerely appreciate the reviewer’s constructive feedback and valuable suggestions. Below, we address the key concerns and outline our planned improvements.

Q1. Improving the Readability of the Method Section

A1: To improve the clarity of the Method section, we have restructured it into a more logical and intuitive organization. The revised structure is as follows:

• 2.1 Motivation: We introduce the fundamental challenges of lighting adaptation and explain why a unified framework is necessary.
• 2.2 Framework Overview: We provide a high-level description of the proposed framework, detailing its architectural design and how different components interact.
• 2.3 Color Separation Representation: Our model first extracts color-separated features to decouple illumination-related information in both the frequency and spatial domains. This step enables the network to learn distinct representations for different color channels, mitigating color distortion and channel-dependent illumination inconsistencies.
• 2.4 Color Mixing Representation: We introduce mixed channel modulation for capturing inter-channel relationships and lighting patterns that achieve a harmonious light enhancement.
  • 2.4.1 Mixed Channel Modulation: This subsection introduces mechanisms to effectively integrate cross-channel lighting information.
  • 2.4.2 Light State Space Module (LSSM): We introduce the LSSM, which leverages state-space representations to refine the global illumination representation dynamically, improving stability and adaptability under varying lighting conditions.
• 2.5 Light-Guided Attention: We propose a novel light-guided attention mechanism that utilizes color-separated features to guide color-mixed light information for adaptive lighting. This mechanism enhances the network’s capability to perceive changes in channel luminance differences and ensure visual consistency and color balance across channels.

This restructuring ensures that the methodology follows a progressive flow from problem motivation to implementation details, making it easier for readers to understand.

Q2. Loss Function for Model Training

A2: Thank you for your valuable feedback. Our framework optimizes light adaptation using the following objective function: The overall objective function is formulated as follows:

$L_{\mathrm{total}} = \alpha L_{Re} + \beta L_{SSIM} + \gamma L_{HF} + \eta L_{p}$,

where:

• $L_{Re}$: reconstruction loss,
• $L_{p}$: perceptual loss,
• $L_{HF}$: high-frequency loss,
• $L_{SSIM}$: structural similarity loss.

$\alpha$, $\beta$, $\gamma$, and $\eta$ are the corresponding weight coefficients.

To justify our design, we conducted an ablation study (Table R2):

Table R2: Ablation studies on different loss functions.

These results demonstrate that incorporating all four losses leads to the best overall performance, underscoring the necessity of each component. We will clarify this in the revised manuscript.

Q3. LALNet-Tiny Design Description

A3: LALNet-Tiny is a lightweight variant of LALNet, designed to improve efficiency while preserving strong performance. The key differences include:

• Reduced channel dimensions: Fewer feature channels to reduce computational complexity.
• Fewer LSSM modules: A more compact design with fewer LSSM modules.

We will include a brief section describing LALNet-Tiny design and relationship to LALNet.

We appreciate the reviewers' valuable insights. Our key improvements include:

1. Method Section Restructuring: Introducing explicit subsections and improving explanations with figures.
2. Loss Function Clarification: Providing a detailed breakdown, justification, and coefficient selection strategy.
3. LALNet-Tiny Performance Analysis: Add a detailed description.

We believe these enhancements will further strengthen our work and address the reviewers' concerns effectively. Thank you for your constructive feedback!

We sincerely thank the reviewers for their feedback and the opportunity to clarify our work. Below, we address the concerns raised.

Q1. Motivation and Justification

A1:
As detailed in the “Introduction” section, many light-related tasks—such as exposure correction, image retouching, low-light enhancement, and tone mapping—share a common goal: adjusting scene lighting to achieve perceptual optimality. However, each task has its own subtle emphasis:

• Exposure correction balances under/overexposed regions.
• Image retouching enhances aesthetics via global illumination adjustments.
• Low-light enhancement brightens dark areas while suppressing noise.
• Tone mapping compresses HDR content while preserving details.

Existing methods typically rely on task-specific architectures and training strategies. We are not claiming that handling these tasks separately is inherently a drawback. Instead, our key motivation is that treating these tasks under a unified framework offers significant advantages:

1. Improved generalization across diverse lighting conditions (Figure 1).
2. Enhanced efficiency by reducing the need for multiple specialized models.
3. Consistent performance across tasks without significant trade-offs (Table 1, Table 2, Table 3, and Table 4).

Indeed, prior works have attempted unified treatments for some of these tasks but often sacrificed performance compared to specialized models. Along this avenue, we aim to design a unified framework that effectively generalizes across diverse light adaptation tasks. Our proposed LALNet achieves this goal by delivering robust performance across multiple light-related tasks (Figure 1).

Q2. Modular Approach Lacking Coherent Integration

A2: LALNet’s design is grounded in two key insights from light-related tasks:

(i) Different color channels have different light properties;

(ii) The channel differences reflected in the spatial and frequency domains are different.

To effectively leverage these observations, LALNet employs a dual-branch architecture:

• MCM (Mixed Channel Modulation): Capture channel-mixed features, focusing on inter-channel relationships and lighting patterns.
• DDCM (Dual Domain Channel Modulation): Extract color-separated features, focusing on light differences and color-specific luminance distributions for each channel in the spatial and frequency domains.

To ensure harmonious fusion of these complementary features, we introduce:

• LSSM (Lighting State Space Model): Integrates color-separated and color-mixed features to enhance illumination consistency across channels.
• LGA (Lighting-Guided Attention): Uses color-separated features as queries, guiding the learning of inter-channel illumination and noise differences for adaptive light restoration.

Our ablation studies (Table 5) confirm that each component is indispensable, as removing any one leads to performance degradation. We will further clarify these interdependencies in the revised manuscript.

Q4. Unified Framework

A4: We appreciate the reviewer’s concern regarding whether LALNet constitutes a truly unified framework. While it is true that each task is optimized separately within LALNet, these tasks share a common architecture and design principles. This shared structure provides a promising advantage over entirely separate task-specific models, considering practical streamlined deployment for low-level tasks. A single framework can be adapted to multiple tasks with minimal adjustments, simplifying hardware deployment. We will clarify this point further in the revised manuscript.

Q4. Writing and Presentation Improvements

A4: We sincerely appreciate the reviewer's feedback regarding the writing style and structure of our manuscript. We acknowledge that certain aspects, such as missing citation ("Mamba"), the improper use of transitions (e.g., "however"), and disjointed arguments, may have affected the clarity and coherence of our key points. To address these concerns, we will carefully revise the manuscript to enhance the logical flow, strengthen the motivation, and improve the articulation of our contributions and methodology, ensuring they are more accessible to readers.