Q1: Computational Efficiency

Thanks for the thoughtful comment. In response, we compared the computational cost of different fusion methods with pre-enhancement in Table r1, including parameters, FLOPs, and average runtime. Although our method is not the most efficient, it achieves acceptable inference efficiency, especially when compared with joint restoration-fusion methods such as Text-DiFuse, DRMF, and Text-IF.

Furthermore, Table r2 reports the computational overhead of individual components. Since we use the pretrained CLIP model, its parameters do not incur additional training costs, and its FLOPs and runtime remain within acceptable bounds. To reduce complexity, our model also provides a lightweight SFVA as the prompt branch, significantly reducing both parameters and runtime. In addition, the PMM module consists of only a few MLP layers and simple modulation operations, incurring negligible overhead, particularly in terms of FLOPs and runtime.

In summary, our method improves fusion performance without imposing a significant computational burden. The computational overhead introduced by the core components, including CLIP, SFVA, and PMM, in terms of FLOPs and runtime, remains within a practically acceptable range.

Table r1. Computational efficiency comparison with pre-enhancement. For fusion methods without restoration, we add the average cost of Instruct-IR and OneRestore. Bold: best; Italic: second-best.

Table r2. Computational overhead of individual components. Values in parentheses indicate the percentage relative to the full model.

Q2: Spatial-Frequency Fusion Mechanisms

Thank you for your insightful comment regarding the fusion mechanism adopted in our SFVA module for integrating spatial and frequency features. To address this concern, we conducted additional experiments by introducing two more sophisticated fusion mechanisms: 1) Weighted Fusion: We first enhance the extracted spatial and frequency features using ResBlocks, and then learn adaptive fusion weights for feature aggregation. 2) Attention Fusion: Inspired by SwinFusion [r1], we incorporate a cross-attention mechanism to facilitate more effective interaction and better integration between spatial and frequency features.

We first evaluated these fusion mechanisms by comparing the cosine similarity between the visual prompts and their corresponding textual prompts across various degradation types and severity levels. The results, shown in Tables r3 and r4, indicate that the more sophisticated fusion strategies lead to better alignment between visual and textual prompts, suggesting improved degradation characterization.

It is worth noting that, as shown in Table r5, these improvements come at the cost of significantly increased computational complexity. Both the weighted and attention fusion mechanisms introduce more than 50% additional parameters, FLOPs, and runtime compared to the simple concatenation strategy. However, the actual gains in fusion performance brought by these more complex mechanisms are marginal.

Although advanced fusion mechanisms offer some benefits, our current design strikes a practical balance between effectiveness and computational efficiency. We sincerely appreciate the reviewer’s suggestion and will explore lightweight yet effective spatial–frequency fusion strategies in future work.

Table r3. Cosine similarity between visual and textual prompts across various degradation types at medium severity across different fusion mechanisms.

Table r4. Cosine similarity between visual and textual prompts under varying degradation levels in the low-light scenario across different fusion mechanisms.

Table r5. Comparison of fusion performance and computational efficiency among different fusion mechanisms in SFVA.

[r1] Ma J, Tang L, Fan F, et al. SwinFusion: Cross-domain long-range learning for general image fusion via swin transformer[J]. IEEE/CAA JAS, 2022, 9(7): 1200-1217.

Q3: Generalization to Real-world Scenarios

Thank you for the valuable suggestion. Although our model is trained solely on the synthetic DDL-12 dataset, it generalizes well to real-world scenarios, as evidenced by the qualitative results in Figure 5 and Figure 10 (Appendix).

To demonstrate the practical utility of our method and its effectiveness in bridging the domain gap, we conducted evaluations on real nighttime scenarios from the MSRS dataset and a real weather-degraded benchmark AWMM-100k [r2], as shown in Tables r6 and r7, respectively. The MSRS dataset suffers from low-light and noise degradations, while AWMM-100k contains severe haze and snow conditions. As shown in the results, our method achieves clear advantages on these real-world benchmarks. This indicates that the proposed synthetic degradation scheme, grounded in physical imaging models, effectively reduces the domain gap between simulated and real data, thereby improving the model's real-world generalization and applicability.

Table r6. Quantitative results on nighttime scenes from MSRS.

Table r7. Quantitative results on real-world weather degradations from AWMM-100k.

[r2] Li X, Liu W, Li X, et al. All-weather multi-modality image fusion: Unified framework and 100k benchmark. arXiv preprint arXiv:2402.02090 (2024).

Q1: Effectiveness of Image-Based Degradation Prompts

Thank you for the insightful question. In our SFVA and prompt-modulated module, CLIP-based representations act as a bridge between visual and textual prompts, allowing our model to alleviate its dependence on textual input during inference. To ensure that the degradation prompts extracted from source images are effective, we design our SFVA from two perspectives. On the one hand, we incorporate a frequency-aware branch to extract informative and discriminative cues in the frequency domain. On the other hand, we introduce MSE loss and cosine similarity loss to align the visual prompts with their corresponding textual counterparts during training, thereby improving the consistency and reliability of the visual degradation representations.

As presented in Figure 8 (Appendix), using either visual or textual prompts leads to similarly consistent visual results, suggesting that visual degradation prompts from the source images are indeed effective. To further validate this, we compute the cosine similarity between visual and textual prompts across different degradation types and levels, with results presented in Tables r1 and r2. As shown in the results, our similarity scores are consistently higher than those of OneRestore in all scenarios except for low-light. This improvement primarily stems from the introduction of our frequency-aware branch, which enables better perception of degradation types. It is worth emphasizing that OneRestore’s visual prompt branch can only recognize a limited set of degradation types, such as rain, haze, and low-light, as well as their combinations, but lacks the ability to model degradation severity. In contrast, our method not only identifies the degradation types but also captures the corresponding degradation levels. Notably, when degradations become more severe, our degree-level matching becomes increasingly precise, as shown in Table r2, which is consistent with human visual perception.

Table r1. Cosine similarity between visual and textual prompts across various degradation types at medium severity.

Table r2. Cosine similarity between visual and textual prompts under varying degradation levels in the low-light scenario.

Q2: Model Complexity Analysis

Thank you for the valuable suggestion. We compare the computational complexity of various methods in Table r3, including parameters, FLOPs, and runtime. Although our method is not the most efficient, it achieves acceptable inference efficiency, particularly when compared to joint restoration-fusion methods, such as DRMF, Text-DiFuse, and Text-IF. Furthermore, Table r4 presents the overhead introduced by individual components. We observe that CLIP and SFVA, as key components for visual-text prompt alignment, significantly improve performance, as shown in Table 4 (IV), where “w/o PMM” refers to the model without aligned visual-text prompts. Particularly, this performance gain is attained with a relatively modest increase in computational overhead.

Table r3. Computational efficiency comparison with pre-enhancement. For fusion methods without restoration, we add the average cost of Instruct-IR and OneRestore. Bold: best; Italic: second-best.

Table r4. Overhead of individual components.

Q3: Prompt Design Details

We appreciate the reviewer’s valuable suggestion regarding the design details of our textual prompts. Our prompts are designed to guide the fusion network by indicating the modality affected, the degradation type, and its severity.

A typical prompt template for a single degradation is structured as follows: "We are performing infrared and visible image fusion, where the modality image suffers from grade-severity degradation type problem."

Here:

• Modality includes infrared or visible.
• Severity is an integer from 1 to 10.
• Degradation types include: rain, haze, low-light, overexposure, random noise, stripe noise, low contrast, etc.

For composite degradations, we adopt extended templates such as:

"We are performing infrared and visible image fusion. Please handle grade-severity A degradation type A problem in the modality A image, and grade-severity B degradation type B degradation in the modality B image."

or:

"We are performing infrared and visible image fusion. Please address level-severity degradation type A and degradation type B degradations in the modality image."

These templates are flexibly composed to represent both intra- and inter-modal degradation combinations.

An instantiation example is:

"We are performing infrared and visible image fusion. Please handle the grade-6 low-light degradation in the visible image, and the grade-8 stripe noise interference in the infrared image."

To cultivate a robust generalization capability, we curated a diverse set of over 100 unique textual formulations for each task. This prompt rephrasing strategy augments linguistic variability and improves the model’s robustness to diverse prompt styles during inference.

We will add these illustrations of our prompt templates and examples in the camera-ready version.

Q4: Real Degradation Evaluation

Thank you for the valuable suggestions. To address your concerns, we conducted additional evaluations on real nighttime scenes from the MSRS dataset. Due to the system’s text-only submission constraint, we regret that qualitative results cannot be presented here. Quantitative results are shown in Table r5. Our method achieves the best performance on CLIP-IQA, TRes, and SD metrics, while slightly trailing Text-DiFuse on the MUSIQ metric.

Furthermore, we conducted quantitative comparisons on the real weather-degraded benchmark AWMM-100k [r1], with the results shown in Table r6. Our method outperforms others across all metrics under real-world weather degradation conditions. Additionally, the corresponding qualitative results also demonstrate the clear advantages of our approach.

These experimental results collectively validate the effectiveness and robustness of our method in real-world scenarios.

Table r5. Quantitative results on nighttime scenes from MSRS.

Table r6. Quantitative results on real-world weather degradations from AWMM-100k.

[r1] Li, Xilai, et al. All-weather multi-modality image fusion: Unified framework and 100k benchmark. arXiv preprint arXiv:2402.02090, 2024.

Q5: Writing and Notation Clarity

Thank you for pointing out the notation and punctuation issues. We will carefully revise the manuscript and fix these issues in the camera-ready version.

As the submission system supports only TeX format, we are unable to upload additional figures. We apologize for this limitation and will incorporate the above updates and results in the main paper or appendix.

Q1: Prompt Construction Paradigm

Thank you for the insightful comments. Some key degradation severity levels are defined as follows:

• Level 1: Barely perceptible degradation.
• Level 4: Degradation begins to interfere with human scene understanding.
• Level 7: Degradation severely hinders human perception of the scene.
• Level 10: Most useful information is completely obscured.

Other levels allow users to interpolate between these anchor levels to describe intermediate degradations.

Moreover, when users are unsure about the degradation severity, our method supports a fully automatic mode: the SFVA can be used directly to perceive the degradation type and severity from inputs. This design balances user-controllable flexibility with automated practicality, ensuring both fine-grained control and real-world usability.

Our prompts are designed to guide the fusion network by indicating the modality affected, the degradation type, and its severity.

A typical prompt template for a single degradation is: "We are performing infrared and visible image fusion, where the modality image suffers from grade-severity degradation type problem."

For composite degradations, we adopt extended templates such as: "We are performing infrared and visible image fusion. Please handle grade-severity A degradation type A in the modality A image, and grade-severity B degradation type B in the modality B image."

Further details on Prompt Construction Paradigm are provided in our response to Reviewer (iEbx) Q3.

Q2: Imaging Model Contribution

We appreciate the reviewer's suggestion. Unlike existing imaging models that focus on isolated degradations, our proposed unified imaging model integrates three physical degradation paradigms and accounts for various modality characteristics. This enables accurate representation of complex, composite degradations and provides enriched data for model training, thereby helping to bridge the gap between synthetic data and real-world degradations. We will further emphasize this key contribution and integrate the imaging model into the methodology section to better highlight its relevance to the overall framework.

Q3: Training Data Construction

Thank you for the thoughtful comment. To address potential degradations in the base datasets, we use the high-quality counterparts provided by Text-IF [r1] as clean references. Thus, the training process is not affected by inherent noise or degradation in base datasets.

[r1] Yi, Xunpeng, et al. Text-if: Leveraging semantic text guidance for degradation-aware and interactive image fusion. CVPR. 2024: 27026-27035.

Q4: Qualitative Comparison without Pre-enhancement

We regret the omission of qualitative comparisons. We have conducted qualitative assessments without pre-enhancement on the four mentioned datasets, where our method more effectively preserves salient objects from IR and retains rich texture details from VI, with these benefits also evident in qualitative comparisons with enhancement. To provide a more intuitive evaluation of fusion performance, we will include the qualitative comparisons in the Appendix.

Q5: Full-reference Metrics

Thank you for the valuable suggestion. We report full-reference metrics MI, Qabf, and VIF for various algorithms across three representative degraded scenarios in Table r1. Our method consistently outperforms others in most cases, further validating its effectiveness. We will include full-reference metrics for all degraded scenarios in the camera-ready version.

Table r1. Comparison of full-reference metrics in various degraded scenarios. Bold: best; Italic: second-best.

Q6: Real Composite Degradation

Thanks for the insightful suggestion. As shown in Table r2, we provide comparative results on nighttime scenes from the MSRS dataset. Our method achieves the best performance on CLIP-IQA, TRes, and SD metrics, while slightly trailing Text-DiFuse on the MUSIQ metric.

Furthermore, we conducted quantitative comparisons on the real weather-degraded benchmark AWMM-100k [r2], with the results shown in Table r3. Our method outperforms others across all metrics under real-world weather degradation conditions.

The above results collectively demonstrate the real-world reliability of our method in handling real composite degradations.

Table r2. Quantitative results on nighttime scenes from MSRS.

Table r3. Quantitative results on real-world weather degradations from AWMM-100k.

[r2] Li, Xilai, et al. All-weather multi-modality image fusion: Unified framework and 100k benchmark. arXiv preprint arXiv:2402.02090, 2024.

Q7: Computational Efficiency

Thanks for the valuable suggestion. Table r4 reports the computational efficiency of various methods with enhancement, including parameters, FLOPs, and runtime. Although our method is not the most efficient, it maintains acceptable inference efficiency, especially compared with joint restoration-fusion methods such as DRMF, Text-DiFuse, and Text-IF.

Table r4. Computational efficiency comparison with pre-enhancement. For fusion methods without restoration, we add the average cost of Instruct-IR and OneRestore.

Q8: Diagram Clarity Issue

Thank you for the question. We believe the reviewer refers to Figure 1. Composite degradation does not exclusively indicate simulated data. For example, low-light degradation in VI is real. Importantly, it also refers to the cases where both IR & VI are degraded, and VI suffers from multi-type degradations simultaneously. (I), (II), and (III) respectively highlight the limitations of existing methods and our advantages under real-world scenarios, composite degradations, and varying severity levels. We will further clarify the definition and relationships in the camera-ready version.

Q9: Frequency Spectrum Comparison

Thank you for the valuable suggestion. We will add the frequency spectrum of clean images in the camera-ready version to better illustrate differences between various degradations.

Q10: Training Data Dependency

Thank you for the insightful question. To investigate the necessity and impact of the proposed DDL-12 dataset, we retrained our model on the Enhanced Multi-Spectral Various Scenarios (EMS) dataset [r1], which contains sensor- and illumination-related degradations but lacks the full range of complex degradations modeled in DDL-12.

Table r5 presents the quantitative results of the retrained model on real nighttime scenarios from the MSRS dataset and real-world weather degradations from AWMM-100k. On the MSRS dataset, the model trained on EMS shows slight performance drops in CLIP-IQA, MUSIQ, and TRes, and a more notable degradation in the SD metric. This can be attributed to the limited diversity in EMS, which weakens the model’s ability to perceive and adapt to compound degradations and different degradation levels. In addition, the fusion output tends to a uniform brightness distribution, thereby reducing contrast and structural diversity, which leads to a significant decrease in SD.

On AWMM-100k, the performance drops more significantly across all metrics. This is primarily because EMS does not provide any degradation patterns or priors related to adverse weather conditions, making it difficult for the model to generalize to scenarios involving haze or snow.

These findings highlight the importance of DDL-12, which offers diverse and physically grounded synthetic degradations, enabling better learning of degradation-aware representations and significantly improving real-world generalization.

Table r5. Quantitative comparison under real-world degradation scenarios with different training datasets.

As the submission system supports only TeX format, we are unable to upload additional figures. We apologize for this limitation and will incorporate the above updates and results in the main paper or appendix.

Q1: Effect of Severity Level Number

Thanks for the valuable comment. As shown in Table r1, we compared models trained with all 4 anchor levels (1, 4, 7, 10) against those trained with only 2 levels (1 & 10 or 4 & 7). Reducing the number of degradation levels weakens the model’s generalization ability on real-world datasets, particularly when only extreme levels (1 & 10) are used. This limits the model’s capacity to accurately perceive fine-grained degradation severity. Due to limitations in time and dataset (DDL-12 has 4 levels: 1, 4, 7, and 10), we are currently unable to explore more settings as suggested (e.g., 3 or 5-6 levels), but will investigate them in the camera-ready.

Table r1. Comparison of models trained with varying degradation levels in real-world datasets.

Q2: Pre-enhancement Method

Thanks for the thoughtful question. In Table 2, for each degradation type, we apply the same pre-enhancement method for all fusion models without built-in enhancement modules. Specifically, OneRestore is used for weather-related degradations, while Instruct-IR is applied to illumination and sensor-related degradations. For IR stripe noise, we additionally employ WDNN [r1] as preprocessing, given the scarcity of effective stripe noise removal methods. We will clarify this in camera-ready.

As reported in Table r2, we further provide comparison results using SOTA degradation-specific pre-enhancement methods under various degradation scenarios. Specifically, DRMF, Text-IF, and Text-DiFuse adopt their own built-in enhancement modules in low-light conditions. For rain, haze, and rain&haze scenarios, we employ additional pre-enhancement algorithms, as these methods lack native support for such degradations. In particular, we use HVI-CIDNet (CVPR'25) [r2] for low-light enhancement, NeRD-Rain (CVPR'24) [r3] for de-raining, and IPC-Dehaze (CVPR'25) [r4] for dehazing. While these SOTA methods can improve fusion quality to some extent under specific degradation scenarios, our method consistently achieves superior overall performance. This is mainly due to our ability to jointly exploit complementary information across modalities, enabling more accurate scene representations and better approximation of high-quality imagery. Furthermore, as shown in Tables r2(d) and r2(e), we explore the impact of sequential processing orders under compound degradations. Results indicate that the processing order does affect fusion performance to some extent, especially in the rain&haze scenario, where addressing rain before haze yields moderate performance gains.

Table r2. Quantitative results in various degraded scenarios with degradation-specific SOTA pre-enhancement methods. Bold: best; Italic: second-best.

[r1] Guan J, et al. Wavelet deep neural network for stripe noise removal[J]. IEEE Access, 2019, 7: 44544-44554.

[r2] Yan Q, et al. Hvi: A new color space for low-light image enhancement[C]. CVPR. 2025:5678-5687.

[r3] Chen X, et al. Bidirectional multi-scale implicit neural representations for image deraining[C]. CVPR. 2024: 25627-25636.

[r4] Fu J, et al. Iterative Predictor-Critic Code Decoding for Real-World Image Dehazing[C]. CVPR. 2025: 12700-12709.

Q3: Real-world Datasets

Thanks for the insightful comment. In dataset construction, we carefully modeled degradations across sensor-related artifacts, illumination changes, and weather conditions to approximate real-world scenarios. Still, some degradations (e.g., extreme weather), remain challenging to simulate accurately with our current imaging model, which is an open issue shared across existing methods.

In Eq.(5), $D_{vi}^{w}=P_{w}(I_{vi})=I_{vi}\cdot t+A(1-t)+R$, we provide a general formulation for weather-related degradations, where $R$ denotes the rain component. It typically manifests as rain streaks or raindrops. While we implement rain streaks, the model is readily extensible to raindrops or other patterns, ensuring flexibility to simulate more complex weather effects.

We adopt a standard homogeneous haze model, a common practice in many methods [r5, r6], due to its physical soundness, controllability, and reproducibility. To enhance diversity, we simulate a range of haze effects by using various depth maps, varying transmission functions, and multiple parameter settings (e.g., $\beta\in[0.5,2.0],A\in[0.3,0.9]$). Despite its simplicity, this model exhibits promising generalization in real-world hazy scenes, as shown in Figure 10 (Appendix) and Table r4(AWMM-100k [r7]). We acknowledge its limitations, and will explore more complex variants, such as non-uniform haze (e.g., patch-wise variable $\beta$) or learning haze distributions from real foggy images.

Although DDL-12 covers only specific degradation types, it spans diverse degradation patterns from 3 complementary dimensions, i.e., sensor, illumination, and weather. This structured modeling provides our model with diverse and robust priors for generating high-quality fusion results. Besides, rather than tailoring to specific degradations in the training data, our model is designed to generalize beyond specific synthetic priors. For instance, although our dataset includes only rain and haze, the learned weather priors enable the model to handle unseen weather-related degradations such as snow. As shown in Table r3, our method achieves promising performance on snow scenarios from the AWMM-100k benchmark, demonstrating its robustness and generalization capability.

Table r3. Quantitative results on snow scenarios from AWMM-100k.

[r5] Fattal R. Single image dehazing[J]. ACM TOG, 2008, 27(3):1-9.

[r6] He K, et al. Single image haze removal using dark channel prior[J]. IEEE TPAMI, 2010, 33(12):2341-2353.

[r7] Li X, et al. All-weather multi-modality image fusion: Unified framework and 100k benchmark. arXiv preprint arXiv:2402.02090 (2024).

Q4: Weather Degradations

Thank you for the question. As shown in Figure 10 (Appendix), we provide results under haze conditions, a common weather-related degradation. Existing multi-modal datasets do not include rain scenarios, and due to exposure time limitations, our multimodal sensors cannot effectively capture real rain streaks. Therefore, we are currently unable to present results under real rain degradation. To further demonstrate our method's effectiveness under real-world weather degradations, we conducted experiments on AWMM-100k [r7], a multi-modal dataset with real-world adverse weather. As shown in Table r4, our method achieves clear advantages under real-world adverse weather conditions.

Table r4. Quantitative results on real-world weather degradations from AWMM-100k.

Q5: Sensor-Specific Priors

Thank you for your insightful comment. Our method does not solely rely on enlarging the training dataset but strategically leverages sensor-specific priors. First, we incorporate infrared-specific characteristics, such as stripe noise, when constructing degraded infrared data, enabling the model to effectively handle modality-specific degradations. Second, our fusion model implicitly exploits high-quality infrared properties, such as high contrast, to guide visible image restoration tasks like low-light enhancement and dehazing, where contrast enhancement is a key objective. These designs ensure that sensor-specific information is effectively utilized, beyond merely scaling the dataset.