Real-World Adverse Weather Image Restoration via Dual-Level Reinforcement Learning with High-Quality Cold Start

We sincerely thank the reviewer for the valuable comments and constructive suggestions, which have played an important role in helping us further improve our work. Your review has been extremely helpful in refining our thinking and enhancing the quality of the paper. We also sincerely appreciate your recognition of the novelty and effectiveness of our work. Below, we provide detailed responses to your comments, with related issues grouped together for clarity. All corresponding discussions and additional experimental results will be included in the final camera-ready version.

[W1] Resource Overhead:

Thank the reviewer for the question. We report that cold-start training of the lightweight single-agent on eight parallel RTX 4090 GPUs required ~72 hours with a 23 GB peak memory footprint. The complete PIQO + Agent framework—trained thereafter on a single RTX 4090—took ~48 hours with a 22 GB peak. During inference, the full framework averages 0.57 s/image (17 GB peak), whereas the lightweight PIQO variant averages 0.027 s/image (4.3 GB peak). It is worth noting that our method demonstrates advantages over agent-based JarvisIR (CVPR 2025) during inference; see [Reviewer YKa4, Limitation 1]. We will include these figures and discuss deployment trade-offs in the revision.

[W2&Q2] Reward Design:

Thank you for the question. We observed that using a single IQA model as the reward could lead to overfitting—while the metric score improved, human-perceived quality often deteriorated. The intuition behind this is that the model can optimize too closely to the reward metric, without capturing the broader perceptual nuances. To mitigate this, we adopted a composite reward combining multiple IQA models with complementary strengths, which helps balance different aspects of perceptual quality and aligns well with human observation. Each score is normalized to [0, 1] to ensure fair weighting across models. Although the weights are hand-tuned, this approach stabilizes training and prevents the model from overfitting to a single metric. Additionally, if training does not converge or exhibits early signs of overfitting (e.g., performance regression despite rising metric scores), we stop early or adjust the weights to restore a better balance. We will include more details on the tuning process and stopping criteria in the revision to clarify how we ensure robust performance.

[W3] Expansibility:

Thank you for your question. To assess the impact of agent library size on performance, we expanded the weather-related agent pool from three to four by adding a Raindrop degradation agent. In real rainy scenarios, this resulted in significant improvements across multiple IQA metrics, suggesting that additional agents can enhance restoration quality. However, the larger agent pool also introduces increased scheduling complexity and inference latency. Future work will explore adaptive agent selection and dynamic scheduling to optimize the trade-off between performance gains and computational cost.

[W4] Extreme Degradation Scenarios:

Thank you for the comment. We agree that extreme degradation scenarios are important to evaluate. In fact, we have included corresponding analysis and visual results in the supplementary material (Sections B & C), where we test our model on severe fog, heavy snow, and dense rain conditions. The results show its effectiveness under extreme weather. We will clarify this in the main paper and explicitly reference the supplementary section.

[W5&Q3] Further Comparisons:

We perform the experiments with JarvisIR and DFPIR (CVPR 2025) on WReaL data and show the results below. It is worth noting that our method significantly outperforms JarvisIR in both inference speed and resource efficiency. Our method takes on average 570 ms and 17 GB peak memory per image, while JarvisIR requires 15,250 ms and 29.5 GB peak memory. The performance is evaluated by GPT-4o; see paper for details. Notably, our method demonstrates superior performance compared to these strong competitors.

Q1: Thank you for the question. The weights assigned to different IQA models in the PIQO reward were chosen based on their relevance to perceptual quality, with CLIP-IQA receiving the highest weight (w₂ = 1) due to its strong alignment with human visual quality judgments, and LIQE and Q-Align given smaller weights (w₁, w₃ = 0.2) to ensure diversity. Perturbing these weights during training showed that increasing CLIP-IQA’s weight sped up convergence but sometimes overfitted on perceptual details, while lowering weights for LIQE and Q-Align slowed convergence but maintained better balance. Large weight changes led to training instability.

Limitations: Thank you for the feedback. We acknowledge that while our method introduces additional computational overhead, the analysis of limitations and failure cases could be more comprehensive. The increased computational cost primarily stems from the multi-agent system and the iterative optimization process. Nevertheless, our method demonstrates advantages over JarvisIR. In terms of failure cases, the system may struggle when the agent pool becomes too large, leading to increased inference latency and scheduling complexity. Additionally, when faced with highly complex or unseen degradation types, the system may require more agents, which can affect runtime efficiency. To mitigate these issues, we are exploring adaptive agent selection and dynamic scheduling to optimize performance while managing computational cost.

We sincerely thank the reviewer for the valuable comments and constructive suggestions, which have played an important role in helping us further improve our work. Your review has been extremely helpful in refining our thinking and enhancing the quality of the paper. We also sincerely appreciate your recognition of the novelty and effectiveness of our work. Below, we provide detailed responses to your comments. All corresponding discussions and additional experimental results will be included in the final camera-ready version.

[W1]: Thank the reviewer for this insightful comment. In addition to the IQA metrics used for training, we incorporate MUSIQ exclusively for evaluation and include GPT-4o-based perceptual scores and visual comparisons as independent assessments at test time, helping to mitigate potential reward–metric coupling and ensure fair evaluation. We will clarify this issue in revision.

[W2]: Thank you for the observation. We acknowledge that repeated agent passes may cause color shifts while optimizing IQA scores. To address this, we plan to explore adding global color consistency terms (e.g., ΔE distance) to the reward and applying a lightweight color correction module post-restoration and also plan to investigate learning a reward function that balances perceptual quality with fidelity to input color distributions.

[W3]: Thank you for the suggestion. PIQO settings like the KL threshold and reward weights have a critical impact on convergence. We set $τ = 0.01$ to balance stability and learning speed—lower values may slow or stall convergence, while higher ones can cause instability. If training fails to converge, we recommend reducing $τ$ (e.g., from 0.01 to 0.005 or 0.001) or relaxing the reward filtering criteria to retain more candidate samples. For reward weights, we prioritize CLIP-IQA (w₂ = 1) and assign smaller weights to LIQE and Q-Align (w₁, w₃ = 0.2) to ensure perceptual alignment while maintaining diversity. We will include tuning guidance in the revision.

[Q]: Thank you for the question. The gains from HFLS-Weather come from both Depth-Anything’s accurate depth and the large, diverse dataset scale. Depth improves realism in weather simulation (e.g., fog depth cues), while scale enhances generalization. As shown in Table 4, both contribute complementary benefits, and we will clarify this in the revision.

I have carefully read the authors’ response as well as the comments from other reviewers. The authors have addressed the main concerns and proposed concrete improvements. I encourage them to include the resolution of the color shift issue and relevant analysis in the revised version. Considering the current contributions and remaining issues, I decide to keep my original score of 5.

We sincerely thank the reviewer for the valuable comments and constructive suggestions, which have played an important role in helping us further improve our work. Your review has been extremely helpful in refining our thinking and enhancing the quality of the paper. We also sincerely appreciate your recognition of the novelty and effectiveness of our work. Below, we provide detailed responses to your comments, with related issues grouped together for clarity. All corresponding discussions and additional experimental results will be included in the final camera-ready version.

[W1] Related works:

Thank the reviewer for the suggestion. We will add citations to prior works that use perturbations to generate diverse outputs, including DFPIR (CVPR 2025) for degradation-aware feature perturbation [DFPIR, CVPR’25], and earlier latent-space or parameter-based methods such as Deep Mean‑Shift Priors and Autoencoding Priors for image restoration [Bigdeli et al., ICCV’17; Aittala & Durand, ICCV’17]. Unlike these, our approach applies RL-guided parameter perturbations with IQA-based reward filtering, and introduces a dual-level structure with a global meta-controller for adaptive model coordination, which distinguishes our method.

[W2&Q2] Details:

Thank you for the question. Our contribution is the novel adaptation of RL to image restoration via the proposed Perturbation-driven Image Quality Optimization (PIQO). Unlike standard perturbation or prior RL methods, PIQO injects Gaussian noise into model parameters to generate diverse outputs, filters them using no-reference IQA metrics, and applies a GRPO-inspired update based on group-wise advantages—all without paired supervision. This design enables RL-based training in a domain with deterministic outputs and sparse reward signals, making it, to our knowledge, the first framework to successfully adapt GRPO-style optimization to real-world image restoration. The novelty of our method is acknowledged by Reviewer oVun-S1 and Reviewer jNge-S2, such as “The proposed two-level reinforcement-learning strategy (PIQO and meta-agent) is novel and effective for real-world weather restoration.”

[W3] Comparison:

First, we have reported the performance of the model trained on the proposed dataset without the RL training (PIQO) and agent framework in Table 5 (basic) in paper. The results clearly show the improvement of the proposed PIQO and agent framework. Second, we re-trained WGWS (CVPR,2023) and OneRestore (ECCV,2024) on our dataset, and here are the results. Will add the new results into the paper.

[W4] Effectiveness:

Thank you for the observation. While some datasets in Table 4 focus on single degradation types, we also include more diverse datasets such as RealSnow and SPA+, which cover complex and mixed weather conditions. Additionally, our HFLS-Weather dataset supports mixed-weather synthesis (e.g., snow+haze, rain+haze), and the “Our Snow+Haze” and “Our Rain+Haze” entries reflect this capability. These results demonstrate that our dataset not only provides high fidelity but also enables robust performance across diverse and dynamic degradation scenarios. We will revise the text in Line 277 to clarify this point.

[Q1] Efficiency: Thank you for the question. While our Multi-Agent System involves running multiple specialized models, which increases latency (570ms), it offers better restoration quality for complex or mixed weather conditions compared to single-model methods like OneRestore (13ms). Though slower than single-model systems, it is more efficient than other multi-agent methods like JarvisIR (15250ms). We will include a detailed efficiency-performance comparison in the revision to highlight these trade-offs.

[Q3]: Yes, our system restores images step by step by adaptively selecting and sequencing specialized agents based on scene analysis and prior success rates. Unlike one-shot models, it uses IQA feedback to refine execution paths, enabling more accurate restoration under complex, mixed-weather conditions.

We sincerely thank the reviewer for the valuable comments and constructive suggestions, which have played an important role in helping us further improve our work. Your review has been extremely helpful in refining our thinking and enhancing the quality of the paper. We also sincerely appreciate your recognition of the good motivation and effectiveness of our work. Below, we provide detailed responses to your comments, with related issues grouped together for clarity. All corresponding discussions and additional experimental results will be included in the final camera-ready version.

[W1] Related works:

Thank the review for pointing out the work of JarvisIR (CVPR 2025). We will include it in the revised paper. While both JarvisIR and our method adopt multi-agent RL strategies for weather-degraded image restoration, our approach differs in two key aspects: (i) we design a dual-level framework where a global meta-controller dynamically schedules restoration agents, complementing the local PIQO-based refinement, and (ii) we introduce a composite, no-reference reward function based on IQA metrics (CLIP-IQA, LIQE, Q-Align) with reward filtering to improve learning stability.

[W2] Novelty:

We propose PIQO, a novel approach that successfully extends GRPO-style reinforcement learning to image restoration by addressing the challenges of single-image input and non-deterministic rewards. The overall dual-level reinforcement learning framework for adverse weather image restoration is another contribution. The novelty of our method is acknowledged by Reviewer KNba-S2, Reviewer oVun-S1, and Reviewer jNge-S2, such as “The proposed two-level reinforcement-learning strategy (PIQO and meta-agent) is novel and effective for real-world weather restoration.”

[W3&Q2] Efficiency:

We report that our full framework (PIQO + Agent) takes on average 0.57 seconds per image on an RTX 4090 GPU with 17 GB peak memory usage. We also provide an ablation with a lightweight single-agent PIQO variant (0.027 seconds, 4.3 GB) that balances performance and efficiency. We will include these details and discuss practical deployment trade-offs. It is worth noting that our method significantly outperforms JarvisIR in both inference speed and resource efficiency. Our method takes on average 570 ms and 17 GB peak memory per image, while JarvisIR requires 15,250 ms and 29.5 GB peak memory.

[W4] Writing:

Thanks for the careful review. We have removed the duplicated sentence in the abstract, corrected grammatical issues throughout the paper, and revised Figure 2 with clearer annotations and a simplified layout for better readability in the revised paper.

Q1: We perform the experiments with JarvisIR and DFPIR (CVPR 2025) on WReaL data and show the results below. The performance is evaluated by GPT-4o; see paper for details. Notably, our method demonstrates superior performance compared to these strong competitors.

Q3: Thank you for the question. Yes, we have tested our method on real-world datasets without the HFLS-Weather synthetic pretraining (i.e., without the cold start). In these tests, we observed that the model struggles to converge when mitigating weather-induced degradations, showing inferior performance compared to the cold-start approach. Our ablation study (Table 4 in paper) highlights the superiority of HFLS-Weather as a cold start, showing substantial improvements in image quality for snow, haze, and rain. Additionally, pretraining with diverse weather types, such as combining snow and haze, enhances the model's generalization ability during fine-tuning with PIQO.

Limitation 1: While our Multi-Agent System involves running multiple specialized models, which increases latency (570ms), it offers better restoration quality for complex or mixed weather conditions compared to single-model methods like OneRestore (13ms). Though slower than single-model systems, it is more efficient than other multi-agent methods like JarvisIR (15250ms). We will include a detailed efficiency-performance comparison in the revision to highlight these trade-offs. Please also see W3 for model size.

Limitation 2: To address the potential performance degradation when a model trained on synthetic datasets is tested in real, non-synthetic environments, we propose the Perturbation-driven Image Quality Optimization (PIQO) reinforcement learning method. As demonstrated in Table 5 and Figure 4, PIQO facilitates continuous optimization by fine-tuning the model in real-world scenarios. This adaptive process significantly enhances performance, enabling the model to better handle real-world degradations and improving its robustness in actual environments.

Limitation 3: Each agent's decision is based on its specialization in a particular degradation type (e.g., snow, haze, rain), with the agents learning to optimize their performance in these specific scenarios. The meta-controller is responsible for selecting which agents to use based on their past performance in similar situations. While the decisions are driven by learned policies and experience, the selection process is determined by the needs of the scenario and the agents' historical success.

Thank you very much for your positive response and for deciding to raise the final rating. We’re pleased that our explanations have addressed most of your concerns and, as you suggested, will include a more thorough discussion of related work in the revised version.If you have any further questions or comments, we would be happy to discuss them.