Enhancing Infrared Vision: Progressive Prompt Fusion Network and Benchmark

Rebuttal to Reviewer anvK

We sincerely thank the reviewers for recognizing the novelty and significance of our work. Below, we provide clarifications and responses to address the main concerns.

1. Clarification on the Difference Degradations Between TIR and RGB (W1):

We thank the reviewer for raising this important point. While high-order degradation models have been extensively studied in the RGB domain, thermal infrared (TIR) imaging exhibits fundamentally different physical characteristics and noise properties due to its reliance on thermal radiation rather than visible light. For RGB-based degradation models such as [1] and [2], realistic blur, noise, and compression artifacts are simulated by applying multiple degradation operations in a specific or random order, reflecting the effects of processes such as camera imaging, image editing, and Internet transmission. However, these models are inherently designed for the general image and fail to capture the modality-specific degradations present in TIR images, which mainly include the camera imaging. Specifically, unlike RGB imaging, which relies on reflected light and is sensitive to illumination and weather, TIR imaging captures emitted thermal radiation and remains stable under varying conditions. However, due to its longer wavelength and sensor characteristics, TIR images suffer from unique degradations such as stripe noise, optical noise, and radiation-caused low contrast, especially in uncooled CMOS-based systems. These structured and composite degradations are uncommon in RGB images and cannot be effectively handled by RGB-oriented models, which typically focus on random noise, bad weather, or low illumination and often fail under TIR scenarios, as shown in Figure 4 and 5. We will revise the manuscript to incorporate these RGB high-order degradation works.

2. Advanced Recent All-in-one Frameworks (W2):

Recent unified frameworks such as PromptIR [3], AutoDIR [4], and InstructIR [5] adopt various strategies for general image restoration: PromptIR uses implicit prompt tokens to improve the network in multi-task setting; AutoDIR leverages latent diffusion with degradation-specific text embeddings to automate degradation handling; InstructIR introduces natural language instructions to control restoration. While effective in the RGB domain, these methods still mainly focus on multiple single degradations, which are not designed for the composite, domain-specific degradations found in TIR images. While our method employs a prompt-guided framework with a unified backbone and fusion of degradation and scenario-specific prompts. This design enables explicit and progressively handling of diverse degradations across different scenarios. These all-in-one models will be incorporated into the revised version of the manuscript.

3. Clarification of Ambiguous Representations and Figures (W3):

We acknowledge that certain notations and figures may have caused confusion. Specifically, the “multiplication” format is used to denote diverse camera viewpoints, including variations across horizontal angles, surveillance views, and driving perspectives. Additionally, we will increase the font size and enhanced the resolution of Figure 6 to improve readability.

4. Supplementing Cost Comparison with Critical Implementation Details (W4):

To ensure fair and reproducible comparisons, we have supplemented the cost analysis section with detailed implementation information, including the hardware environment (NVIDIA RTX 4090 D GPU with 24GB memory) and the input TIR image resolution (640×512) used during inference.

Rebuttal to Reviewer VhxR

We sincerely thank the reviewer for their thorough review and constructive feedback. Below we address each of the raised concerns in detail.

1. Clarification on the hybrid degradation in TIR enhancement (W1)

We thank the reviewer for this important observation. We agree that several prior works (e.g., [1-3]) have explored thermal infrared (TIR) image enhancement under certain hybrid degradation settings. For exmaple, [1] address stripe and optical noise, [2] focus on contrast and detail enhancement, and [3] target video denoising with a physics-inspired composite noise generator. These methods typically operate under specific degradation setting, lack comprehensive degradation modeling, and are not designed to generalize across diverse and dynamically composite degradation domains. While our work is the first to systematically simulate, benchmark, and address comprehensive, real-world composite degradations in TIR images. We will revise the manuscript to clarify our novelty and include a more detailed comparison with [1-3] in the Related Work section.

2. Dataset type not clarified (W2)

We thank the reviewer for this valuable suggestion. To improve clarity, we will revise Table 1 by adding an additional row labeled “Input Type,” indicating whether each dataset is based on single images or video frames. For reference, EN, Iray, and our proposed dataset are based on single images, while SBTI, UIRD, and TIVID are constructed from video sequences.

3. Readability issues (W3)

We appreciate the reviewer pointing this out. We will increase the font size of figure and table entries in the revised version to ensure that visualizations maintain clarity when printed or viewed digitally.

4. Fixed Degradation Order Assumption (Q1)

We argue that in TIR imaging, degradations typically follow a consistent order driven by imaging processes. Low contrast comes first from limited thermal radiation, followed by blur from defocus or turbulence, and finally hybrid noise such as fixed pattern noise and random noise. Our framework follows this order, and we evaluate on real-world TIR images, as shown in Figure 5 and Table 2, this design leads to strong performance in real-world scenes. To test robustness, we also evaluated the model under incorrect degradation removal orders. As shown in Figure 10, performance drops significantly in some case, confirming the model works with a fixed processing order. In future work, we plan to model and predict degradation order to handle more complex scenarios.

5. Ambiguity in “multiplication” notation (Q2)

Thank you for pointing out this ambiguity. The “multiplication” format is used to denote diverse camera viewpoints, including variations across horizontal angles, surveillance views, and driving perspectives. We will revise the table and its caption to clarify this interpretation explicitly.

6. Coverage of extreme weather conditions (Limitation)

In Limitations section, we admit that our current degradation model does not fully cover all possible TIR degradations. Although we did not consider the extreme weather conditions, we believe our framework retains some degree of generalization to such scenarios. Specifically, in rainy or foggy environments, the increased scattering and absorption of thermal radiation reduce the energy reaching the sensor, especially from low-emissivity or distant objects. This results in compressed temperature distributions and weaker radiation intensity, manifesting as low contrast. Since our framework is trained to handle such contrast-degraded scenarios, it can partially generalize to these extreme weather-induced degradations, even without direct exposure during training. Consequently, although our current model does not explicitly simulate all atmospheric effects like rain or fog. We will clarify this point in the revised manuscript to highlight the potential of our approach to generalize beyond the explicitly simulated degradations.

The authors have already responded to my concerns. I will maintain my score.

Rebuttal to Reviewer 3U2T

We thank the reviewer for their positive feedback and for recognizing the effectiveness of our proposed framework in addressing complex degradations in TIR images. Below, we provide detailed responses to the concerns regarding degradation order, noise handling, prompt design, and TIR-RGB differences.

1. Fixed Degradation Order Assumption (W1):

We argue that in TIR imaging, degradations typically follow a consistent order driven by imaging processes. Low contrast comes first from limited thermal radiation, followed by blur from defocus or turbulence, and finally hybrid noise such as fixed pattern noise and random noise. Our framework follows this order, and we evaluate on real-world TIR images, as shown in Figure 5 and Table 2, this design leads to strong performance in real-world scenes. To test robustness, we also evaluated the model under incorrect degradation removal orders. As shown in Figure 10, performance drops significantly in some case, confirming the model works with a fixed processing order. In future work, we plan to model and predict degradation order to handle more complex scenarios.

2. Handling Simultaneous Multiple Noise Types in Infrared Images (W2) :

Regarding diverse noise patterns, we observe that stripe noise and optical distortion often coexist in real-world thermal infrared images due to the radiation inside the imaging body and environmental temperature.[1] Consequently, rather than handling each type of noise independently, we consider their coexistence as a unified degradation pattern and model it as hybrid noise (see Figure 2(a), Noise Part). A degradation prompt is employed to guide the model in conditional learning to suppress this pattern during both training and inference. The specific degradation simulation formulations are provided in Supplementary Material A1.1.1.

3. Ambiguity in Prompt Design (W3) :

We acknowledge that the current manuscript lacks a clear explanation of how the prompts are defined, which may have caused confusion. In our framework, two types of prompts are used as conditional inputs to guide the network. These prompts are randomly initialized and fixed for each prompt type, and during training, they gradually encode task-relevant conditions through model optimization under specific prompts. We manually define the prompt corresponding specific degradation conditions. Specifically, the prompts are designed to represent the degradation type and processing scenario. As part of future work, we plan to explore more flexible and scalable prompt designs. In particular, we will investigate the use of image-based self-prompting mechanisms, where the model dynamically generates degradation-aware prompts from the input itself. Additionally, we are interested in integrating language-based prompts to express complex degradation descriptions in a more interpretable and user-controllable manner.

4. Differences Between TIR and RGB Imaging and Artifact Causes (W4):

Unlike RGB imaging, which relies on reflected light and is sensitive to illumination and weather, TIR imaging captures emitted thermal radiation and remains stable under varying conditions. However, due to its longer wavelength and sensor characteristics, TIR images suffer from unique degradations such as stripe noise, optical noise, and radiation-caused low contrast, especially in uncooled CMOS-based systems.[1-2] These structured and composite degradations are uncommon in RGB images and cannot be effectively handled by RGB-oriented models, which typically focus on random noise, bad weather, or low illumination and often fail under TIR scenarios, as shown in Figure 4 and 5. Moreover, modeling all degradations within a single network still yields unsatisfactory results when dealing with complex, dynamic, and compound degradations, as shown in Figure 7 and Table 3. Consequently, we propose a TIR-specific framework that explicitly and decomposed models thermal degradations using prompt-based learning. Our method integrates a prompt fusion network and sequential tuning strategy to effectively handle both single and composite degradations, demonstrating robust performance where RGB-based models fall short, as demonstrated in Figure 4, Figure 5, and Table 2.

[1] Cai et al. Exploring video denoising in thermal infrared imaging: physics-inspired noise generator, dataset and model. IEEE TIP (2024).

[2] Danaci and Akagunduz. A survey on infrared image & video sets. Multimedia Tools and Applications (2024).

Rebuttal to Reviewer RuCv

We sincerely thank the reviewer for the positive feedback and for recognizing the novelty of our method, the benchmark construction, and the clarity of our presentation. We address the concerns in detail below.

1. Lack of Infrared-Specific Design (W1, Q1):

We argue that our framework is specifically designed to address the complex degradations found in thermal infrared (TIR) images by adopting general prompt-based techniques. These degradations, including varieties of fix pattern noises and low contrast, are fundamentally different from those in RGB images and have been insufficiently explored in existing literature. Prior methods often assume simple or isolated degradations, while our method introduces a progressive prompt fusion network that enables progressive removal in composite degradation patterns or directly handle single degradation scenario. In addition, we establish a dedicated TIR benchmark (HM-TIR) and design a selective progressive training strategy to improve robustness across varying degradation scenarios. While our approach can apply more broad domain in principle, it provides a targeted solution to a domain-specific problem that is not well addressed by existing models.

2. Insufficient Comparison with Existing Methods (W2):

The key different is that, instead of modeling all degradations jointly within a single network, focusing on specific degradation, or using existing all-in-one approaches that address multiple single degradations, our method employs a prompt-guided framework with a unified backbone and fusion of degradation and scenario-specific prompts. This design enables explicit and progressively handling of diverse degradations across different scenarios. For comparison, our manuscript includes extensive evaluations demonstrating the effectiveness of our method. For example, the Figure 4 shows our clear advantage over TIR-specific and RGB all-in-one models on synthetic composite data. Figure 5 and Table 2 extend this to real-world TIR images, where our method consistently achieves better visual and metric results. Figure 7 and Table 3 compare baselines with and without our prompt fusion module. These comparisons demonstrate the novelty of our model over TIR-specific methods, RGB all-in-one models, and baseline variants.

3. Architectural Complexity and Redundancy (W3, Q2):

Although our method performs iterative processing, it does not cascade multiple independent networks. Instead, it employs a unified backbone across iterations, modulated by degradation and scenario prompts, enabling progressive removal of each degradation type in different scenarios. This design avoids structural redundancy, and all iterations reuse the same network parameters, with only lightweight prompt modules introduced. While iterative inference may require more steps than single-pass baselines, the added cost is slight and results in significantly improved performance, as demonstrated in Figure 7, Table 3, and Table 5.

4. Limited Generalization to Unknown Degradations (W4, Q4):

We acknowledge that our current framework is based on a predefined set of degradation types commonly observed in TIR images, which may limit its ability to handle unknown degradations. However, it is important to note that we have already evaluated our method on real-world TIR data (Figure 5, Table 2), demonstrating its strong practical effectiveness. Additional experiments on other real-world degradations are provided in the supplementary material, primarily to demonstrate that our strategy can effectively enhance TIR images in broader real-world scenarios. The results further support the effectiveness and generalization capability of our method, with consistent visual improvements in noise reduction, contrast enhancement, and structure preservation. Given the already significant visual improvements and performance gains, we believe the current results sufficiently validate the robustness and practicality of our approach. While our current setup enables effective TIR enhancement, we agree that improving generalization to unknown degradations or adapting to newly emerging degradation types remains a critical and promising direction for future work.

5. Ambiguity in Prompt Design (Q3):

We acknowledge that the current manuscript lacks a clear explanation of how the prompts are defined, which may have caused confusion. In our framework, two types of prompts are used as conditional inputs to guide the network. These prompts are randomly initialized and fixed for each prompt type, and during training, they gradually encode task-relevant conditions through model optimization under specific prompts. We manually define the prompt corresponding specific degradation conditions. Specifically, the prompts are designed to represent the degradation type and processing scenario. As part of future work, we plan to explore more flexible and scalable prompt designs. In particular, we will investigate the use of image-based self-prompting mechanisms, where the model dynamically generates degradation-aware prompts from the input itself. Additionally, we are interested in integrating language-based prompts to express complex degradation descriptions in a more interpretable and user-controllable manner.