acceptodds
Withdrawn from ICLR 2027

Degradation-Adaptive Inverse Operator Learning for All-in-One Image Restoration

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

All-in-one image restoration aims to recover images corrupted by heterogeneous degradations using a single unified model. Most existing approaches directly predict clean images or residuals, often relying on degradation-specific prompts or additional conditioning modules to distinguish different restoration tasks. In this work, we explore a different perspective and formulate all-in-one restoration as input-conditioned inverse operator learning. Instead of directly regressing restored pixels, we predict spatially adaptive inverse operators that transform degraded observations into clean estimates. Based on this formulation, we propose OPIR, an all-in-one image restoration framework that infers degradation representations directly from the input and uses them to modulate a shared restoration operator. This input-conditioned modulation allows the model to adapt to heterogeneous degradations without requiring degradation labels or manually specified task prompts at inference. To progressively improve difficult regions, we further introduce an operator-uncertainty-guided refinement strategy, where the distribution and statistical responses of the predicted local operators are used to identify pixels that require stronger restoration. A second restoration stage then selectively refines these challenging regions. In addition, an efficient multi-scale operator implementation reuses compact predicted kernels across different receptive fields, substantially reducing the computational overhead of spatially adaptive filtering. Extensive experiments demonstrate that OPIR consistently achieves superior performance across multiple all-in-one restoration benchmarks, while also remaining highly competitive on task-aligned restoration settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image restorationl, Kernel-predicted

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_X46JOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OPIR, an all-in-one image restoration framework based on input-conditioned inverse operator learning. Instead of directly predicting clean images or residuals, the method predicts spatially adaptive restoration kernels and uses a task-aware module to adjust them according to the degradation information inferred from the input. It also introduces an operator-based uncertainty map to guide a second restoration stage and applies shared multi-scale kernels to improve efficiency. Experiments on multiple restoration tasks, real-world images, and unseen degradations show strong performance, while the ablation studies evaluate the contributions of the proposed components.

Critical strengths
  1. The experimental evaluation is comprehensive. The paper evaluates three-task and five-task restoration, real-world images, and unseen degradations, with consistently strong results.

  2. The method is computationally efficient. OPIR achieves competitive restoration performance with much lower runtime and FLOPs than several recent all-in-one methods.

Critical weaknesses
  1. The novelty appears limited. Some key components resemble those of EfDeRain+, and the concept of "inverse operator learning" seems similar to dynamic kernel prediction. The distinctions from existing research should be articulated more clearly.
  2. The uncertainty map has not been sufficiently validated. The paper primarily presents visualizations and ablation study results but lacks quantitative evidence demonstrating that the map reliably reflects the difficulty of restoration across different tasks.
Questions for authors
  1. What is the main technical difference between OPIR and EfDeRain+ beyond adding task-aware modulation and extending predictive filtering to multiple restoration tasks?

  2. How is the degradation representation in the task-aware module learned? Are degradation labels or task identities used during training?

  3. Can the authors provide per-task quantitative evidence showing that the uncertainty map correlates with actual restoration errors?

  4. Although the paper acknowledges the lack of evaluation on images containing multiple simultaneous degradations, such experiments are important for demonstrating the effectiveness of OPIR in complex real-world scenarios.

Reviewer_ojAQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OPIR, an all-in-one image restoration framework that formulates heterogeneous restoration as input-conditioned inverse operator learning. Instead of directly predicting clean pixels or residuals, OPIR predicts spatially adaptive restoration kernels and modulates them using degradation-aware information inferred from the input. The method is evaluated on several all-in-one and task-aligned restoration settings, including rain, snow, haze, blur, and noise, and reports competitive or superior quantitative performance over the selected baselines.

Critical strengths
  1. OPIR reports competitive performance across multiple all-in-one and task-aligned restoration settings.

  2. Modulating spatially adaptive restoration kernels using input-dependent degradation information is intuitive and computationally attractive. The proposed design achieves favorable FLOPs/runtime compared with the selected AIOIR baselines.

Critical weaknesses
  1. The methodological novelty appears limited relative to prior predictive-filtering work, especially EfDeRain+. Several major ingredients of OPIR already appear in EfDeRain+, including predictive filtering, kernel-statistics-based uncertainty estimation, uncertainty-guided cascaded refinement, and weight-shared multi-scale dilated filtering. Importantly, the authors acknowledge in Appendix A.4 that the uncertainty statistic is not new and describe the main distinction as integrating this mechanism with degradation-adaptive operator modeling in an all-in-one setting. Therefore, the main new component appears to be degradation-aware modulation of predicted kernels together with extension from single-task restoration to AIOIR. Given that degradation-aware modulation itself is already common in AIOIR, I am not fully convinced that the resulting contribution is sufficiently substantial.

  2. The experiments do not sufficiently validate the paper's central claim or isolate the source of the gains. The paper motivates OPIR by arguing that adaptation should occur in the restoration-operator space rather than primarily in the feature space, but there is no controlled comparison between matched feature modulation and operator modulation. Thus, the current ablations show that TAM is useful, but do not establish that operator-space adaptation is preferable. In addition, the proposed TDAS/task-balanced training strategy alone improves average PSNR by 0.47 dB, which is substantial relative to the reported gains over strong baselines. It is unclear whether competing methods were retrained using exactly the same balancing/augmentation strategy, making it difficult to attribute the improvements specifically to the proposed operator modeling.

Questions for authors
  1. Could the authors clarify the exact methodological novelty of OPIR over EfDeRain+ beyond introducing task/degradation-aware modulation and extending the framework to the all-in-one setting? In particular, predictive filtering, kernel-statistics-based uncertainty, two-stage uncertainty-guided refinement, and weight-shared multi-scale filtering all appear closely related to EfDeRain+.

  2. The main motivation argues that adapting the restoration operator is preferable to adapting intermediate features. Could the authors provide a controlled experiment using the same backbone and conditioning signal, comparing feature modulation, operator modulation, and potentially their combination? This would directly test the central motivation of the paper.

  3. The paper states that deep learning baselines are fine-tuned or retrained, and several reported results appear different from the numbers commonly reported in the corresponding papers. Could the authors clarify which results are reproduced versus taken from the original papers, and provide the exact training protocol for each baseline? In particular, were all baselines trained using the same task-balanced sampling and augmentation strategy as OPIR?

  4. Could the authors explain the omission of highly relevant degradation-adaptive AIOIR methods such as AdaIR from the main quantitative comparison? Since the paper's primary claim concerns degradation-aware adaptation, comparison against matched recent feature-modulation/adaptation approaches seems important for assessing

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

OPIR predicts per-pixel restoration kernels without degradation labels at test time. Building on EfDeRain+, it uses a task-aware module to modulate kernels and kernel statistics to guide a second refinement stage. Weight-shared multi-scale dilated filtering enlarges the receptive field at low cost, while balanced sampling and augmentation address haze-dominated training data. OPIR reports higher average PSNR on two self-composed benchmarks covering three and five degradations. Further experiments test real-world and unseen degradations, include component ablations, and compare runtime.

Critical strengths
  1. OPIR extends predictive filtering to all-in-one restoration by adapting per-pixel kernels to the input degradation within a shared model. I find this extension interesting, although it builds substantially on EfDeRain+.
  2. OPIR is efficient: Tab.15 reports 47 GFLOPs and 0.174 s per image, about a quarter of Perceive-IR's runtime, alongside higher average PSNR.
Critical weaknesses
  1. UM's interpretation as task-agnostic uncertainty is insufficiently supported. Eq. 6 depends only on the kernel-weight sum, so identity and normalized averaging kernels get the same score. The visualization and ablation show that UM helps refinement, but do not validate it as a measure of reconstruction difficulty across tasks.
  2. No matched-backbone comparison isolates kernel prediction from pixel or residual regression. TDAS improves average PSNR by 0.47 dB but reduces dehazing by 1.62 dB (Tab.13), which shows a clear task trade-off. Because it is unclear whether the baselines use the same sampling and augmentation, I cannot tell how much of the gain comes from the restoration formulation.
Questions for authors
  1. What constraints, if any, are imposed on the kernels used in Eq. 6? Beyond the refinement gain, can you provide quantitative evidence that higher UM values correspond to larger stage-1 reconstruction errors for each degradation type? 2.Were the baselines trained with the same task sampling and augmentation as OPIR? Is there a controlled comparison with direct pixel or residual regression using the same backbone, stage configuration, and training protocol? 3.Sec. 3.3 refers to “task embeddings” and task-specific parameters θ_t, while Sec. 3.2 states that task information is inferred from the input. How are these embeddings computed and used in each stage at inference? 4.Do Tab.1 and 5 use the same training and evaluation settings for R+S+H? If so, why do their results differ? For Tab. 14, please specify what was replaced in EfDeRain+, since A.4 states that it already uses multi-scale filtering.

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

Related papers

Discussion (0)

Sign in to comment.