acceptodds
Under review as a conference paper at ICLR 2027

Image-Internal Operator Calibration for All-in-One Image Restoration

Abstract

All-in-one image restoration requires adapting shared restoration knowledge to diverse degradations and image content. Representative prompt- and routing-based methods achieve this through feed-forward conditioning, without explicitly optimizing restoration operators on the current observation. We propose CalibIR, an image-internal operator calibration framework that uses each image to learn corrections to the operators used for its restoration. Starting from learned base operators, CalibIR performs a lightweight optimization step on an input-specific feature-prediction objective within the forward pass. To exploit image-wide information while accommodating regional differences, we develop a structured calibration rule that couples channel-specific shared updates with channel-shared local corrections. The shared component aggregates information across the image, whereas the spatially centered local component captures departures from the image-wide adjustment. Both arise from the same optimization objective, enabling coordinated adaptation without maintaining an independent kernel for every channel–location pair. Trained end-to-end under restoration supervision, CalibIR combines knowledge learned across images with operator adaptation learned within each image. Experiments on standard all-in-one image restoration benchmarks show that CalibIR achieves leading average performance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
image restoration, all-in-one image restoration

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_oFiJOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CalibIR, an image-internal operator calibration framework for all-in-one image restoration. The method introduces a lightweight optimization step within the forward pass to adapt restoration operators to each input image. Specifically, it combines channel-wise and spatial corrections derived from a shared feature-prediction objective, enabling both global and local adaptation.

Critical strengths
  1. Unlike conventional feed-forward adaptation methods, the proposed approach dynamically optimizes restoration operators based on the current input for image-specific adaptation.

  2. The experimental evaluation is comprehensive, covering multiple restoration settings, including AIO3, AIO5, CDD11, and real-world WeatherBench.

Critical weaknesses
  1. The methodological novelty appears limited. The proposed channel and spatial corrections are relatively common operations in image restoration and adaptive feature modulation. Although the authors derive these corrections from an image-specific optimization objective, the fundamental differences from existing dynamic convolution, input-adaptive filtering, and test-time training methods remain unclear.

  2. The necessity of the proposed image-internal optimization mechanism is insufficiently justified. It remains unclear why a feature-prediction objective constructed from the input itself can reliably produce better restoration operators across different degradation types.

  3. The authors evaluate the proposed method on only one real-world benchmark, WeatherBench. It is recommended to conduct additional experiments on more diverse real-world datasets to further validate the generalization capability of the proposed method. Furthermore, comparisons with several recent state-of-the-art all-in-one image restoration methods are missing.

Questions for authors

See the above Weaknesses parts.

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

This paper introduces CalibIR, an intra-image operator calibration framework for unified image restoration. The method constructs a prediction task using features from the current input and adapts learned depthwise convolution operators through a single gradient update. The update is decomposed into a spatially shared channel correction and a channel-shared, spatially centered spatial correction, combining global image information with local variations. Calibration is integrated into the network’s forward pass and requires no clean reference images at inference time. The authors evaluate CalibIR under four settings—AIO3, AIO5, CDD11, and WeatherBench—and report higher average restoration performance than the compared methods. Experiments with lightweight models, component ablations, and operator visualizations further assess the contributions of individual designs and the framework’s input-adaptive behavior.

Critical strengths

1.The paper reports better average restoration performance than the compared methods on AIO3, AIO5, CDD11, and the real-world weather dataset WeatherBench. Compared with ExDA, it improves average PSNR by 0.64 dB on AIO5 and 0.84 dB on CDD11. These results span different task combinations and compound degradations, providing reasonably convincing evidence of the method’s practical value for unified image restoration.

2.The method derives both channel and spatial corrections from the same internal prediction error and uses spatial centering to separate local variations from global adjustments. Ablation results on AIO3 show that the full method improves PSNR by 0.68 dB over static operators and outperforms variants using only channel or spatial corrections. These findings support the complementary roles of the two branches and provide component-level evidence for the performance gains.

Critical weaknesses

1.The substantive advance over existing visual TTT methods has not been sufficiently established. Constructing an internal prediction task from input features, updating a lightweight model, and then applying the updated model to query features follows the established TTT framework. ViT³, which the paper cites, has also explored single-step full-batch updates and depthwise convolutional internal models. The main novelty therefore lies in the structural design of the channel correction and centered spatial correction. Although component ablations support the effectiveness of this design, the paper does not sufficiently clarify its specific relationship to existing convolutional TTT updates or clearly map the ablation variants to these methods. This makes it difficult to assess whether the method offers a sufficiently substantial methodological advance.

2.The available evidence does not clearly establish the independent contribution of optimization-based updates to the performance gains. The paper correctly derives the gradients of the internal feature prediction objective but does not directly analyze the link between optimizing this objective and final restoration quality. The operator visualizations mainly demonstrate that updates vary with the input, without explaining why these changes benefit restoration. The most relevant direct prediction baseline differs from the full method by 0.21 dB, but its prediction head architecture, whether its parameter count and computational cost are matched, and variability across repeated runs are not sufficiently reported. It therefore remains difficult to distinguish gains arising from the prediction-error-based update mechanism from those due to the specific parameterization and implementation of dynamic filtering.

Questions for authors
  1. Please clarify the specific methodological differences between CalibIR, TTT layers, and ViT³. Is the main contribution the combination of channel correction and centered spatial correction? Can the channel-only variant in Table 5 be viewed as an implementation of an existing convolutional TTT update? If not, what are the specific differences?

  2. Since both KK and VV are generated from the degraded input, please further explain why optimizing the predictive relationship between them can benefit image restoration. Is there an analysis of internal prediction error and final restoration quality before and after calibration that supports this explanation, beyond merely showing that the operators vary with the input?

  3. CalibIR-s reduces both the network width and the number of refinement blocks. Are there comparisons between static operators, direct prediction, and gradient-based calibration under the same small-model configuration? Does the available evidence support an independent benefit from calibration when the model is scaled down, or does it currently only demonstrate that the complete small model is competitive?

  4. Do CalibIR’s internal prediction objective and channel and spatial corrections rely on properties specific to image restoration? Have experiments been conducted on non-restoration tasks, such as classification, detection, or segmentation, to assess applicability across tasks?

Reviewer_Rbx4OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CalibIR, an all-in-one image restoration framework that adapts feature-space convolutional operators to the current input. Within each restoration block, learned projections produce query, key, and value features. A learned depthwise convolutional operator predicts the values from the keys, and the resulting prediction residual supplies a single gradient-based calibration update. The calibrated operator then processes the queries. The main design is a structured decomposition of the operator correction into two components: a channel-specific correction shared across spatial locations, and a spatially varying correction shared across channels. The spatial component is centered to remove its image-wide mean, separating regional deviations from the adjustment already represented by the channel component. The feature projections, base operators, and update step sizes are trained jointly using restoration supervision.

Critical strengths
  1. The full model achieves the best reported average PSNR among the compared methods in all four settings, also maintaining a pretty good efficiency.
  2. The structured correction is well motivated, and the ablations provide meaningful evidence for its components. The decomposition
\qquad \sum_p B_{p,o}=0$$ offers a simple way to combine channel-specific image-wide adaptation with regional variation. Centering has a concrete role: it assigns the spatially constant component to the channel branch and makes the spatial branch represent deviations from that component. Both branches use evidence from the same prediction residual. Table 5 supports the complementary roles of the two branches. The static operator reaches 32.54 dB, channel correction alone reaches 32.97 dB, spatial correction alone reaches 32.92 dB, and the full model reaches 33.22 dB. Thus, the combined update improves by 0.68 dB over the static operator and by 0.25–0.30 dB over either individual correction. 3. The paper writing is pretty clear and well presentations.
Critical weaknesses
  1. The technical contribution is relatively focused. The cited ViT³ work already investigates full-image, single-step inner updates and depthwise convolutional inner models. Therefore, the general idea of adapting a convolutional operator through an image-internal feature-prediction task is inherited from existing visual TTT work. There is also relevant prior work on separating dynamic filters into channel and spatial components. Decoupled Dynamic Filter Networks constructs these components through feed-forward prediction and combines them multiplicatively. CalibIR’s additive, spatially centered corrections derived from a shared prediction residual are distinct, but discussing this relationship would sharpen the account of what is new [1].

[1] Decoupled Dynamic Filter Networks. CVPR 2021.

  1. The theoretical justification is limited. I would appreciate if more analysis and theoretical motivations will be presented.
Questions for authors
  1. Can you add more theoretical analysis to the block design? Why this part can be well transfered to all-in-one image restoration task?
  2. All experiments are conducted on 128 × 128 or 256 × 256 crops. The paper does not report run-time or memory on >4K images where Restormer is usually applied.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.