acceptodds
Under review as a conference paper at ICLR 2027

Learning Restoration Actions from Degradation States for Multi-Weather Image Restoration

12 views

Abstract

Multi-weather image restoration aims to recover clean images from diverse weather degradations within a unified model. Existing degradation-aware methods mainly infer degradation representations and use them to condition restoration features. However, knowing what degradation is present does not fully determine how it should be restored, as images with similar degradations may require different restoration operations depending on degradation severity, spatial distribution, and image content. To bridge this gap, we propose D2AIR, an efficient framework that learns input-adaptive restoration actions from degradation states. Its core Degradation-to-Action (D2A) module consists of a Degradation State Interpreter (DSI) and a Restoration Action Executor (RAE). DSI learns degradation-discriminative channel responses to infer degradation states and adapts latent features through degradation-aware modulation. RAE further translates the inferred states into input-adaptive restoration actions through a shared operator bank. Specifically, degradation-conditioned routing provides a degradation-dependent prior over restoration operators, while instance-conditioned routing adapts their selection to the current input, enabling flexible operator composition across degradation instances. Designed to operate at the latent bottleneck, D2A can be readily integrated into diverse restoration backbones. Experiments on single-weather, composite-degradation, and real-world benchmarks demonstrate the effectiveness of the proposed method. The code will be publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image restoration, Adverse weather removal, Representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_u58DOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes D2AIR, a plug-and-play framework for multi-weather image restoration that dynamically associates degradation states with restoration actions. Its core D2A module consists of a Degradation State Interpreter (DSI), which extracts degradation-aware representations using a learnable Memory Bank, and a Restoration Action Executor (RAE), which adaptively combines shared restoration operators through dual routing mechanisms. D2A is integrated into the latent bottleneck of different restoration backbones. Experiments on three benchmarks demonstrate state-of-the-art restoration performance and architectural flexibility.

Critical strengths
  1. Plug-and-play capability at the latent level. The authors insert the D2A module only at the latent bottleneck and achieve consistent overall image restoration improvements across multiple backbone architectures, unlike many existing methods that introduce additional degradation-aware guidance into skip connections. This suggests that All-in-One restoration research should pay particular attention to latent features.

  2. Unification of effectiveness and efficiency. By explicitly establishing a dynamic association between degradation states and restoration actions, the proposed lightweight D2AIR framework achieves state-of-the-art performance in both quantitative evaluation and qualitative comparison, especially in composite degradation scenarios. These results indicate that the method unifies the two objectives of restoration quality and model efficiency.

Critical weaknesses
  1. Limited novelty and insufficient justification of the RAE design. The RAE module employs an expert-routing mechanism similar to those used in Mixture-of-Experts (MoE) architectures, raising concerns about its methodological novelty. Although ablation studies demonstrate the effectiveness of the Instance Router, the relationship between the Degradation Router and the Instance Router remains unclear. Furthermore, the paper does not justify the use of Softmax instead of Sigmoid for operator routing or discuss potential load-balancing issues arising from the routing mechanism.

  2. Unclear sources of performance improvements. Unlike existing all-in-one restoration methods that typically learn degradation representations implicitly, D2AIR employs explicit degradation classification supervision. However, it remains unclear whether the performance gains primarily stem from this degradation supervision or the proposed D2A architecture.

Questions for authors
  1. Clarification of the dual-routing mechanism. Could the authors clarify the distinct roles of the Degradation Router and the Instance Router, particularly the motivation for introducing the latter? How does the proposed routing mechanism differ from conventional MoE-based approaches? Could the routing mechanism lead to imbalanced operator utilization?

  2. Clarification of implementation details. Could the authors provide a more detailed description of the Operator Bank and the representation of degradation labels? In particular, how are labels encoded for single and composite degradations, and how are they used during training and inference? These details are important for understanding and reproducing the proposed method.

  3. Ablation on degradation classification supervision. Could the authors provide an ablation study removing the degradation classification loss while retaining the D2A module? This would help disentangle the contributions of explicit degradation supervision and the proposed architecture, and clarify whether D2A relies heavily on such supervision to achieve its reported improvements.

  4. Fairness of comparisons with degradation-aware methods. Since D2AIR leverages explicit degradation supervision, are the reported performance gains still maintained when compared with representative all-in-one restoration methods that use degradation supervision?

Reviewer_6Zh9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes D2AIR for multi-weather image restoration. The method introduces a Degradation-to-Action (D2A) module into the BioIR backbone, consisting of a Degradation State Interpreter (DSI) and a Restoration Action Executor (RAE). DSI learns degradation-aware representations through degradation prediction, a memory bank, and feature modulation. RAE dynamically combines multiple learnable restoration operators based on degradation states and input features. Experiments are conducted on AllWeather, CDD-11, and WeatherBench, with additional evaluations demonstrating the applicability of D2A across different backbones

Critical strengths

The paper achieves competitive performance with a relatively small number of parameters.

Critical weaknesses

1. There is a gap between the motivation for Action Formulation and its actual implementation.

The authors emphasize that identifying degradation states alone is insufficient to determine how an image should be restored, and therefore propose learning Restoration Actions from Degradation States. However, DSI (Degradation State Interpreter) mainly adopts degradation-aware representations, Memory Banks, and feature modulation, which are already common in existing studies [2, 3]. Restoration Action Executor (RAE) predicts routing weights based on degradation states and image features, and then computes a weighted combination of the outputs of multiple learnable convolutional operators.

Similar designs are already quite common in All-in-One Image Restoration. For example, MoFME [5] uses a router to select different feature modulation experts, M2Restore [6] dynamically selects restoration experts based on degradation priors, and Multi-Expert Adaptive Selection [7] further combines task prompts and image content for pixel-level expert selection. MoCE-IR [4] and MOERL [8] also investigate expert routing for image restoration.

Although the specific implementation differs, RAE is essentially still a conditional feature transformation and expert combination mechanism. The authors do not sufficiently explain what restoration decisions this Action Formulation enables that are different from existing MoE methods. Therefore, there is a gap between the emphasized Action Formulation and the actual technical contribution.

2. The lack of sufficient comparison with BioIR weakens the experimental results.

The paper uses BioIR [1] as its backbone, but does not explicitly state this in the main text. BioIR is only clearly identified in Table A1 of the appendix. The main experiments also do not directly report the performance of BioIR.

According to the results reported in the original BioIR paper [1], BioIR-B already achieves 30.54 dB on CDD-11, while D2AIR achieves 31.13 dB, a difference of only 0.59 dB. Both methods also achieve the same average SSIM of 0.893. Furthermore, the Baseline in Table 4 performs substantially worse than the original BioIR-B, and even falls below BioIR-T when averaged over the same four single-degradation cases. However, the authors do not sufficiently explain the differences in configuration and performance. Therefore, it remains unclear how much performance improvement is actually brought by D2A.

3. Insufficient validation of composite degradation handling and spatial adaptivity.

In Section 3.2 (Lines 137–138), the authors explicitly state that DSI is designed to support "both single and composite degradations." Section 4.3.2 further emphasizes the role of shared operator composition in handling composite degradations. However, the ablation study in Table 4 only reports results for four single-degradation tasks, without showing the actual contributions of DSI, RAE, and the two Routers under composite degradations. This lacks corresponding experimental validation for the composite degradation handling capability emphasized by the authors.

Furthermore, the Abstract states that the required restoration operations depend on degradation "severity, spatial distribution, and image content." Section 3.3 (Lines 206–208) also uses variations in "local degradation patterns" to motivate the Instance Router. However, RAE predicts global routing weights through GAP and applies the same operator combination weights across the entire feature map. Although DSI includes spatial modulation, the paper does not sufficiently explain how this design adapts to degradation differences across different spatial regions.

4. The description of the core method is unclear, and some implementation details are ambiguous.

Figure 2(a) labels the D2A module as "D2ABlock," which may be a labeling error. The real issue is that the Method section does not clearly introduce the complete structure of the D2A module, but only describes DSI and RAE separately. Readers need to refer back to the Introduction and subsequent equations to figure out how these two modules are connected and how the overall D2A module works. As the core contribution of the paper, this part is not clearly explained.

In addition, the Encoder preceding the Instance Router in Figure 2(c) is not sufficiently explained. Equations (10)–(11) also do not clearly specify how the Top-k weights are added to the full Softmax weights, or how the Norm operation is defined. These issues make it difficult to understand the actual computation of RAE.

[1] Cui, Y., Ren, W., and Knoll, A. Bio-Inspired Image Restoration. NeurIPS, 2025.

[2] Yao, M., Xu, R., Guan, Y., Huang, J., and Xiong, Z. Neural Degradation Representation Learning for All-in-One Image Restoration. IEEE TIP, 2024.

[3] Wu, G., Jiang, J., Jiang, K., Liu, X., and Nie, L. Learning Dynamic Prompts for All-in-One Image Restoration. IEEE TIP, 2025.

[4] Zamfir, E., Wu, Z., Mehta, N., et al. Complexity Experts are Task-Discriminative Learners for Any Image Restoration. CVPR, 2025.

[5] Zhang, R., Luo, Y., Liu, J., et al. Efficient Deweather Mixture-of-Experts with Uncertainty-aware Feature-wise Linear Modulation. AAAI, 2024.

[6] Wang, Y., Li, Y., Zheng, Z., Zhang, X.-P., and Wei, M. M2Restore: Mixture-of-Experts-based Mamba-CNN Fusion Framework for All-in-One Image Restoration. IEEE TIP, 2025.

[7] Yu, X., Zhou, S., Li, H., and Zhu, L. Multi-Expert Adaptive Selection: Task-Balancing for All-in-One Image Restoration. IEEE TCSVT, 2025.

[8] Wang, T., Xia, P., Li, B., et al. MOERL: When Mixture-of-Experts Meet Reinforcement Learning for Adverse Weather Image Restoration. ICCV, 2025.

Questions for authors
  1. Regarding BioIR: Please clarify the BioIR[1] backbone configuration used in D2AIR and provide a complete comparison between BioIR and BioIR+D2A under the same training settings. How does the Baseline in Table 4 differ from the original BioIR-T/B?

  2. Regarding Action Formulation: How do the authors define a Restoration Action? What are the substantial differences between RAE and existing MoE methods [4–8] that select experts based on degradation states or image content? If an Action mainly refers to dynamically combining learnable operators, what is the core contribution of Degradation-to-Action?

  3. Regarding composite degradations and spatial adaptivity: Why does Table 4 only report ablation results for single degradations? Could the authors provide ablation results under composite degradations? Furthermore, given that RAE uses global routing weights, how does it select appropriate restoration operations for different degraded regions within the same image?

  4. Regarding RAE implementation: Please clarify the output format of Top-k in Eq. (10), the definition of Norm in Eq. (11), and the actual number of operators executed. Please also explain the Encoder shown in Figure 2(c).

References

[1] Cui, Y., Ren, W., and Knoll, A. Bio-Inspired Image Restoration. NeurIPS, 2025.

[2] Yao, M., Xu, R., Guan, Y., Huang, J., and Xiong, Z. Neural Degradation Representation Learning for All-in-One Image Restoration. IEEE TIP, 2024.

[3] Wu, G., Jiang, J., Jiang, K., Liu, X., and Nie, L. Learning Dynamic Prompts for All-in-One Image Restoration. IEEE TIP, 2025.

[4] Zamfir, E., Wu, Z., Mehta, N., et al. Complexity Experts are Task-Discriminative Learners for Any Image Restoration. CVPR, 2025.

[5] Zhang, R., Luo, Y., Liu, J., et al. Efficient Deweather Mixture-of-Experts with Uncertainty-aware Feature-wise Linear Modulation. AAAI, 2024.

[6] Wang, Y., Li, Y., Zheng, Z., Zhang, X.-P., and Wei, M. M2Restore: Mixture-of-Experts-based Mamba-CNN Fusion Framework for All-in-One Image Restoration. IEEE TIP, 2025.

[7] Yu, X., Zhou, S., Li, H., and Zhu, L. Multi-Expert Adaptive Selection: Task-Balancing for All-in-One Image Restoration. IEEE TCSVT, 2025.

[8] Wang, T., Xia, P., Li, B., et al. MOERL: When Mixture-of-Experts Meet Reinforcement Learning for Adverse Weather Image Restoration. ICCV, 2025.

Reviewer_AcVrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes D2AIR for unified multi-weather image restoration. It inserts a Degradation-to-Action (D2A) module into a BioIR backbone. D2A first predicts degradation states and uses them for channel and spatial feature modulation, then combines degradation-conditioned and instance-conditioned routing to mix shared restoration operators. Experiments cover AllWeather, CDD-11, and WeatherBench, with additional component ablations and integration into other restoration backbones.

Critical strengths
  1. Promising restoration performance and efficiency. The reported results show consistent improvements across single and composite degradations and real-world imagery. D2AIR achieves an average PSNR of 31.13 dB on CDD-11 and 31.73 dB on WeatherBench, while using approximately 4M parameters. The evaluation covers relevant restoration settings and includes both quantitative and qualitative comparisons.

  2. Clear presentation and coherent design. The architecture and its two stages are generally easy to follow. The degradation-conditioned modulation and dual-routing design form a coherent pipeline, and the component ablations provide useful evidence that the proposed components contribute to restoration performance.

Critical weaknesses
  1. The contribution of D2A is insufficiently separated from that of the BioIR backbone. Figure 1 and the main comparison tables omit a directly comparable BioIR baseline. Consequently, the strong performance–efficiency trade-off of the complete model cannot be clearly attributed to the proposed module rather than the underlying backbone. Table 4 provides a baseline ablation on the four single-degradation conditions of CDD-11, which partially addresses this concern, but does not establish the source of the headline gains on WeatherBench or the full composite-degradation benchmark.

  2. The central motivation is not sufficiently distinguished from existing adaptive restoration methods. The paper emphasizes a transition from identifying degradation to deciding restoration actions. However, degradation-conditioned feature modulation and MoE routing already adapt restoration transformations to the input; for example, MoCE-IR routes inputs to experts with different computational characteristics. The paper does not clearly identify the specific limitation of these approaches that D2A resolves. This does not negate the usefulness of the proposed design, but weakens the claimed conceptual contribution.

  3. The routing ablation may conflate degradation guidance with operator participation. According to Eqs. (9)–(12), the full model combines a dense degradation router with a sparse instance router. Removing the degradation router while retaining Top-k instance routing would also reduce the number of operators contributing to the output. Therefore, the decrease from 35.08 to 33.90 dB in Table 4 may reflect both the loss of degradation guidance and a change in effective computation. The current description does not establish that this comparison isolates the benefit of degradation-conditioned routing.

Questions for authors
  1. Could you provide a controlled BioIR versus BioIR+D2A comparison on WeatherBench and the full CDD-11 benchmark? Please use the same backbone configuration, training data, training budget, and reconstruction losses, and report PSNR/SSIM together with parameter count, MACs, and inference time. This would clarify D2A’s incremental contribution to the complete model’s accuracy and efficiency.

  2. In the “w/o Degradation Router” ablation, how many operators contribute to each input, and how are the remaining routing weights normalized? Please compare the full dual-router model with a dense instance-only router using the same operator bank and executing all operators. Reporting this comparison on both single and composite degradations would help determine whether explicit degradation guidance provides benefits beyond input-conditioned routing under matched operator participation.

  3. Were the backbone baselines in Table 5 retrained under the same settings as their D2A-augmented counterparts? Their reported scores match those in the original WeatherBench benchmark at the displayed precision, whose training used 128×128 patches, whereas this submission specifies 256×256 patches and 600K iterations. Please identify which results were reused or retrained and clarify the training schedules and losses. If the settings differ, matched-training comparisons are needed to substantiate the claimed plug-and-play gains.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.