acceptodds
Under review as a conference paper at ICLR 2027

Learning Instance-Adaptive Coupled-Axis Relation Prompts for All-in-One Image Restoration

Abstract

All-in-one image restoration aims to recover images corrupted by unknown and heterogeneous degradations using a single model. Existing methods primarily adapt shared features or parameters through prompts and degradation representations, while leaving the underlying attention interactions generic and often dense. This separation between degradation conditioning and relation modeling provides no structured carrier for expressing instance-specific relation changes. We propose InCaRP, an all-in-one restoration Transformer that learns instance-adaptive coupled-axis relation prompts. Specifically, low-rank relation-space attention performs structured sparsification before correlation by projecting redundant key and value responses directly into a compact set of input-dependent relation carriers, while preserving the full query and output representations. Coupled-axis prompt calibration then uses shared angular retrieval to compose complementary query-axis and relation-axis prompts from paired banks. Normalized interactions between the two prompts synthesize bounded scale and shift fields that calibrate the compact relations before value aggregation. The unified design therefore combines pre-correlation information compression with instance-adaptive relation calibration. Extensive experiments on all-in-one benchmarks demonstrate that our method achieves state-of-the-art performance while maintaining parameter efficiency and shows strong, balanced performance across heterogeneous restoration tasks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
All-in-one image restoration, Low-rank, Coupled-axis relation prompts, hyperspherical retrieval

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_zSTVOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes InCaRP, an all-in-one image restoration Transformer for handling different degradations with a single model. It introduces Low-Rank Relation-Space Attention (LRSA), which compresses key and value responses before attention computation, and Coupled-Axis Prompt Calibration (CAPC), which uses paired prompts to adapt the compact attention relations to each input. The two components aim to reduce redundant interactions while providing degradation-aware relation modeling. Experiments on AIO-3, AIO-5, and CDD-11 show strong performance across single and composite degradations, and the ablation studies support the effectiveness of the proposed components.

Critical strengths
  1. The method has a clear motivation. LRSA and CAPC provide a clear way to combine compact attention computation with input-adaptive relation modeling.
  2. The paper revisits prompt-based restoration from a practical perspective. Instead of applying prompts only to image features, it studies how prompts can directly adjust attention relations, providing a useful view of where degradation conditioning should be introduced.
Critical weaknesses
  1. Practical efficiency has not been fully demonstrated. Although the paper claims improvements in computational efficiency, it fails to report runtime, FLOPs, or peak memory usage for the complete model under a unified experimental setup.

  2. Evaluation in real-world scenarios is limited. Most experiments were conducted on paired datasets, lacking sufficient evaluation on images with unknown ground-truth degradation characteristics—such as visual comparisons involving images without ground truth.

Questions for authors
  1. The practical efficiency of the full model remains unclear because runtime, FLOPs, and peak memory usage are not reported under a consistent setting.

  2. The generalization ability on unknown real-world degradations is not sufficiently demonstrated, as most experiments are based on synthetic or paired datasets.

  3. CAPC retrieves prompts from global feature statistics, so it is unclear whether it can accurately represent spatially localized degradations such as local rain streaks or motion blur.

4 comments
AuthorsOpenReview ↗

[Part 1] Official Comment by Authors

We thank the reviewer for the constructive comments and helpful suggestions. We address the questions concerning (1) computational cost and practical efficiency, (2) generalization to real-world and unseen degradations, and (3) spatially selective restoration through global prompts. The corresponding evaluations and explanations are added to the revised main text and supplementary material, together with quantitative results, visual comparisons, and supporting analyses. Please refer to these updates for further details.

R3.1: Model complexity and efficiency

We appreciate the reviewer’s meaningful suggestion. We have additionally measured the parameter count, peak GPU memory, FLOPs, and runtime of the complete models under a unified setting. As reported in Table 4 of Section 4.3 of the revised manuscript and discussed in our response to Reviewer uJmP [W2], our method achieves the lowest parameter count, memory consumption, and FLOPs while retaining competitive runtime, demonstrating its practical efficiency. At the same time, our approach also outperforms recent state-of-the-art methods across multiple all-in-one restoration settings (Tables 1 and 2 in Section 4.2 and Table 7 in Appendix B.1 of the revised manuscript) and on real-world benchmarks (Table 3 in Section 4.2 of the revised manuscript), achieving a favorable trade-off between restoration quality and computational cost.

Computational demands comparison (Table 4 in Section 4.3 of the revised manuscript).

MethodParams.MemoryFLOPsRuntime
PromptIR36 M9884 MiB132 G70.26 ± 3.08 ms
VLU-Net35 M10767 MiB143 G110.86 ± 4.78 ms
MoCE-IR25 M6332 MiB81 G45.78 ± 2.25 ms
AdaIR29 M9701 MiB124 G75.88 ± 2.65 ms
CPLIR35 M8804 MiB132 G72.63 ± 3.67 ms
BaryIR36 M11510 MiB163 G88.64 ± 0.54 ms
InCaRP(Ours)23 M6194 MiB74 G50.24 ± 0.42 ms

Evaluation setup. We report the total number of parameters for each complete model. FLOPs, peak GPU memory usage, and inference latency are evaluated under the same setup for all methods, using 224 × 224 input images on an NVIDIA GeForce RTX 3090 Ti GPU.

R3.2: Generalization to real world and unseen distribution

Thank you for highlighting the importance of real-world evaluation. We additionally evaluate InCaRP on real-world datasets, covering deraining on SPA [1], low-light enhancement on LOL-v2-Real [2], desnowing on RealSnow [3], and shadow removal on SRD [4]. Desnowing and shadow removal extend the evaluation beyond the degradation types in AIO-5. We use PSNR, LPIPS, and CLIP-IQA to assess pixel fidelity, perceptual quality, and no-reference quality, respectively. InCaRP achieves the best results across all datasets and metrics (Table 3 in Section 4.2 of the revised manuscript). Figure 6 in Appendix B.2 of the revised manuscript presents paired real-world deraining and low-light enhancement examples. Overall, these results demonstrate that InCaRP generalizes more effectively than the competing methods to real-world images, both for degradation types covered by AIO-5 and for unseen types such as snow and shadows.

Quantitative comparison on four real-world benchmarks (Table 3 in Section 4.2 of the revised manuscript).

MethodSPA / PSNR↑SPA / LPIPS↓SPA / CLIP-IQA ↑LOL-v2-Real / PSNR↑LOL-v2-Real / LPIPS↓LOL-v2-Real / CLIP-IQA ↑RealSnow / PSNR↑RealSnow / LPIPS↓RealSnow / CLIP-IQA ↑SRD / PSNR↑SRD / LPIPS↓SRD / CLIP-IQA ↑
PromptIR29.590.2110.35220.690.2010.33524.580.2480.39517.140.1670.425
MoCE-IR28.090.2190.36327.570.0850.54024.630.2490.40517.080.1680.422
AdaIR30.100.2050.36130.980.0830.48925.100.2510.40917.250.1670.425
BaryIR30.360.2040.36632.380.0720.50925.500.2460.40717.260.1650.427
InCaRP(Ours)30.650.2010.37632.850.0640.54726.240.2320.41717.830.1560.436
AuthorsOpenReview ↗

[Part 2] Official Comment by Authors

R3.3: Concerns of spatially localized degradations representing

Thank you for this insightful question. Global conditioning helps a shared restoration model determine how to adapt to an input. For example, PromptIR [5] and CPL [6] compose input-dependent prompts, while MoCE-IR [7] uses input-dependent expert routing. CAPC also derives its conditioning from global feature statistics, and it applies this conditioning directly to the attention relations that govern feature aggregation.

The retrieved query and relation prompts generate scale and shift fields for the logits between query channels and relation atoms. These calibrated logits determine how each query channel combines spatially indexed value responses. Because the value responses depend on local image content, the same image-level calibration can affect different regions differently. Figure 5(b) in Section 4.3 of the revised manuscript illustrates this effect by comparing learned and identity calibration with the decoder-block input held fixed. The resulting response changes concentrate around elongated rain streaks in deraining examples and around blur-affected boundaries and contours in deblurring examples.

We further evaluate whether CAPC-induced output changes are spatially aligned with degradation on Rain100L [8] and GoPro [9]. For each input, we compare restored outputs with CAPC enabled and disabled at the D1-1 decoder block, keeping model weights and all other blocks fixed. We correlate the output-change map with the degradation residual computed from the degraded input and clean target. The pixel-wise Pearson correlation measures whether locations with stronger degradation residuals tend to show larger output changes when CAPC is disabled. As a spatial control, we circularly shift the residual, preserving its values and spatial structure while disrupting alignment with the output-change map. Table 13 (Appendix C.7 of the revised manuscript) reports positive correlations on Rain100L (0.5625) and GoPro (0.7547), which fall to near zero after shifting, indicating that the association depends on spatial alignment. These results support our conclusion that global prompts can modulate restoration locally through CAPC, with larger effects in regions more strongly affected by degradation.

Spatial alignment between CAPC-induced output changes and degradation residuals (Table 13 in Appendix C.7 of the revised manuscript).

DatasetReal spatial correlation ↑Shifted residual correlationCorrelation gain (Δρ) ↑Permutation p-value ↓
Rain100L0.562-0.0010.563< 0.0001
GoPro0.755-0.0080.7620.00049

In a word, CAPC leverages prompts retrieved from global feature statistics to recalibrate attention relations and change how spatially indexed value features are aggregated. This enables the model to respond more selectively to localized rain streaks and blur-affected regions.

We hope the additional evaluations address your concerns about practical efficiency, generalization to real-world and unseen degradations, and spatially selective restoration, and we welcome further discussion.

References

[1] Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, and Rynson W. H. Lau. Spatial Attentive Single-Image Deraining With a High Quality Real Rain Dataset. CVPR, 2019.

[2] Wenhan Yang, Wenjing Wang, Haofeng Huang, Shiqi Wang, and Jiaying Liu. Sparse Gradient Regularized Deep Retinex Network for Robust Low-Light Image Enhancement. IEEE TIP, 2021.

[3] Yurui Zhu, Tianyu Wang, Xueyang Fu, Xuanyu Yang, Xin Guo, Jifeng Dai, Yu Qiao, and Xiaowei Hu. Learning Weather-General and Weather-Specific Features for Image Restoration Under Multiple Adverse Weather Conditions. CVPR, 2023.

[4] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson W. H. Lau. DeshadowNet: A Multi-Context Embedding Deep Network for Shadow Removal. CVPR, 2017.

[5] Vaishnav Potlapalli, Syed Waqas Zamir, Salman Khan, and Fahad Shahbaz Khan. PromptIR: Prompting for All-in-One Image Restoration. NeurIPS, 2023.

[6] Gang Wu, Junjun Jiang, Kui Jiang, Xianming Liu, and Liqiang Nie. Beyond Degradation Redundancy: Contrastive Prompt Learning for All-in-One Image Restoration. IEEE TPAMI, 2026.

[7] Eduard Zamfir, Zongwei Wu, Nancy Mehta, Yuedong Tan, Danda Pani Paudel, Yulun Zhang, and Radu Timofte. Complexity Experts are Task-Discriminative Learners for Any Image Restoration. CVPR, 2025.

[8] Wenhan Yang, Robby T. Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint Rain Detection and Removal from a Single Image with Contextualized Deep Networks. IEEE TPAMI, 2020.

[9] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep Multi-Scale Convolutional Neural Network for Dynamic Scene Deblurring. CVPR, 2017.

The authors' response has addressed my concerns. Therefore, I have decided to maintain my original score.

We are pleased that our response has addressed your concerns. Thank you for your thoughtful review, constructive feedback, and positive assessment and rating of our work.

Reviewer_TeEQOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes InCaRP for all-in-one image restoration, integrating Low-Rank Relation-Space Attention (LRSA) to reduce attention computation with decoder-side Coupled-Axis Prompt Calibration (CAPC).

Critical strengths

LRSA compresses K/V representations before attention computation, while CAPC employs input-conditioned paired vectors to calibrate attention logits. Together, these components improve restoration performance across multiple degradation types.

Critical weaknesses

(1) The performance gains do not fully substantiate the claimed interpretation of the prompt mechanism. Although the paper provides PCA visualizations, dehazing, deraining, and denoising are evaluated on different datasets, so the observed separation may partly arise from differences in image content or dataset distributions. Such separation does not necessarily demonstrate a distinct prompting mechanism, it merely suggests that the learned vectors capture degradation-related or dataset-specific characteristics. It remains unclear what information the banks actually encode and whether they learn the distinct query-axis and relation-axis roles claimed in the paper, rather than simply functioning as input-conditioned parameters for attention modulation.

(2) The methodological novelty appears limited. LRSA primarily reduces the K/V projection width in transposed attention, while CAPC integrates feature statistics, similarity-based retrieval, learned vector banks, and factorized scale-and-shift modulation. These designs largely adapt or combine existing mechanisms. Although their integration yields performance improvements, the paper does not convincingly establish a substantial methodological advance beyond these incremental modifications.

Questions for authors

What makes the proposed prompt banks a distinct prompting mechanism rather than a set of learnable parameters for input-conditioned attention modulation? Please clarify their methodological advantage over existing prompt designs, such as PromptIR and InstructIR, beyond differences in the overall network architecture and final restoration performance.

14 comments
AuthorsOpenReview ↗

[Part 1] Official Comment by Authors

We thank the reviewer for the detailed feedback and insightful questions. We respond point by point to concerns about (1) degradation versus content information, (2) the rationale behind attention calibration, (3) the information and roles of the two prompt banks, (4) methodological novelty, and (5) advantages over existing prompt designs. The revised main text and supplementary material include the corresponding clarifications and experiments, with further analyses, quantitative results, and visualizations. Please refer to these updates for additional details.

R2.1: Do Prompts Learn Image Content or Degradation?

We sincerely thank the reviewer for the insightful questions. We agree that differences shown in Figure 1(e) of the original manuscript in image content and dataset distributions may contribute to the separation in the original PCA. We address this concern with two experiments in the following.

(1) Preliminary experiment on noise severity.

In the original manuscript, we use image denoising as a controlled setting to test whether CAPC calibration depends on noise severity (Fig. 1(d), Section 4.3 in the original manuscript). Specifically, we keep the noisy input and model weights fixed and replace the scale and shift fields at all CAPC blocks. Using fields from a different image at the same noise level changes the average PSNR by at most 0.02 dB. Using fields from the same clean image at a different noise level reduces the average PSNR by up to 3.29 dB. These results show that CAPC calibration transfers across images at the same noise level and that matching the noise severity is important for effective restoration.

(2) Content-matched PCA on CDD-11 [1].

To further validate whether the degradation sensitivity observed in the denoising experiment is also applicable to different types of degradation, we add a content-matched PCA experiment in the revised manuscript (Fig. 5(a), Section 4.3). In the single-degradation subset of CDD-11 [1], each clean image has four degraded versions that share the same scene content and differ in degradation type: haze, rain, snow, or low light. For each clean image, we extract the prompts from its four degraded versions and project them into a common PCA space. As shown in Fig. 5(a) of the revised manuscript (Section 4.3), the prompts exhibit distinct distributions across degradation types even when the underlying image content is identical. This separation is especially clear for haze and low-light degradation, whose prompt groups are well separated from the others. In other words, even when the underlying scene is identical, the prompts still distinguish between degradation types.

In general, Figure 1(e) in the original manuscript shows that prompts may tend to group by degradation across datasets. Figure 1(d) in the original manuscript provides complementary evidence that the resulting calibration is sensitive to noise severity even when image content is fixed. The new PCA experiment in Fig. 5(a) of the revised manuscript further shows that degradation-dependent grouping persists when both image content and dataset source are the same. Together, these results support the interpretation that the learned prompts capture degradation-related information.

AuthorsOpenReview ↗

[Part 2] Official Comment by Authors

R2.2: Design and Rationale of the Proposed Prompting Mechanism

We appreciate the reviewer’s thoughtful comment. Learning representations that capture degradation-related information is a key design objective of prompt learning in image restoration. For example, PromptIR [2] composes input-adaptive prompts and concatenates them with decoder features. DATPRL-IR [3] retrieves task and domain prompts and fuses them with decoder features through gated cross-attention, and CPLIR [4] uses sparse prompt routing and contrastive supervision on reconstructions obtained with matched and mismatched prompts. These studies demonstrate the effectiveness of prompting as a way to adapt restoration to different degradations.

(1) The characteristics of our prompt design

As for our study, we reconsider the role of prompts in image restoration by exploiting the natural ability of Transformer self-attention to construct global relations across the image. In CAPC, prompts derived from global feature statistics characterize image-level degradation and recalibrate attention logits. The calibrated attention weights determine how spatially indexed value responses are combined, allowing global degradation cues to elicit stronger responses in locally degraded regions. Direct calibration of attention relations is the critical difference between CAPC and the feature-fusion approaches described above.

Additionally, LRSA complements CAPC by projecting keys and values into a compact set of relation atoms before correlation, reducing the computational cost of the attention interactions that CAPC calibrates. The resulting model has the fewest parameters and the lowest FLOPs among the representative methods compared in Table 4 (Section 4.3 of the revised manuscript), while outperforming recent state-of-the-art methods across multiple all-in-one restoration settings (Tables 1 and 2 in Section 4.2 and Table 7 in Appendix B.1 of the revised manuscript) and real-world benchmarks (Table 3 in Section 4.2 of the revised manuscript).

(2) Effectiveness of the proposed paradigm

InCaRP uses prompts to calibrate the interactions between query channels and compact relation atoms. CAPC uses shared, input-dependent retrieval weights to compose two axis-specific prompts. Their interaction generates scale and shift fields that directly calibrate the compact QKTQK^{T} logits, allowing each query channel to adapt how it combines spatially indexed value responses. Section 3.3 of the revised manuscript describes the calibration mechanism, and Section 4.3 of the revised manuscript compares its placement with feature-level conditioning in Fig. 4(b). On the four single degradations of CDD-11 [1], the content-matched PCA in Fig. 5(a) of the revised manuscript shows degradation-dependent prompt structure with the same image content across degradation types. Low-light prompts are clearly separated, and rain and snow have nearby but distinguishable group centers, consistent with the visual similarity of their corruption patterns.

The spatial selectivity comes from the value responses that CAPC reweights. In Fig. 5(b) of Section 4.3 of the revised manuscript, we hold the input to the decoder block fixed and compare learned CAPC with identity calibration. The resulting response changes concentrate around rain streaks and blur-affected contours. When we disable CAPC at D1-1, the resulting changes in the final output align with the ground-truth degradation residuals on Rain100L [5] and GoPro [6]. This alignment drops to near zero when the residuals are circularly shifted (Table 13 in Appendix C.7 of the revised manuscript). These results show how image-level conditioning can have spatially selective effects through the aggregation of spatially indexed value features.

Spatial alignment between CAPC-induced output changes and degradation residuals (Table 13 in Appendix C.7 of the revised manuscript).

DatasetReal spatial correlation ↑Shifted residual correlationCorrelation gain (Δρ) ↑Permutation p-value ↓
Rain100L0.562-0.0010.563< 0.0001
GoPro0.755-0.0080.7620.00049

Note. The p-value is from a one-sided paired sign-flip test comparing aligned correlations with circularly shifted controls.

In summary, our proposed CAPC exploits the global aggregation capability of Transformer self-attention to translate image-level degradation information into spatially selective restoration responses. Compared with existing methods that incorporate prompts through feature fusion, CAPC assigns prompts a distinct functional role by directly recalibrating the attention relations that determine feature aggregation. Jointly designing CAPC with LRSA integrates degradation adaptation into a compact relation space, allowing InCaRP to achieve stronger restoration performance with fewer parameters and lower computational complexity than the compared methods.

AuthorsOpenReview ↗

[Part 3] Official Comment by Authors

R2.3: What Information Do the Prompt Banks Encode?

(1) What do the prompt banks encode?

Thank you for raising this question. Implicit prompts are learned vectors without predefined linguistic semantics. Prior restoration studies characterize them through their behavior. PromptIR [2] ablates the prompt module, CPLIR [4] analyzes prompt selection across degradations, and DATPRL-IR [3] examines similarities among bank entries and among instance-level prompts. Following this approach, we apply PCA to the input-conditioned bank mixtures before adding the image prompt. Fig. 1(e) of the original manuscript shows that prompts from different images tend to group by degradation condition. We further apply the same analysis to the single-degradation subset of CDD-11 [1], using different degraded versions of the same clean images (Fig. 5(a) in Section 4.3 of the revised manuscript). The prompt representations retain degradation-dependent structure with image content and dataset source held fixed. Fig. 1(d) of the original manuscript further shows that denoising is sensitive to a mismatch between the noise severity of the input and that of the calibration fields. Together, these results support the interpretation that the composed prompts encode degradation-related information.

(2) What distinct roles do the two banks serve?

The roles are defined by the two axes of the compact relation logits. The query-bank mixture produces one factor per query channel, which determines how that channel’s row is calibrated. The relation-bank mixture produces scale and shift factors for each relation atom, which determine how that atom’s column is calibrated. Each logit is thus calibrated by the interaction of its query-channel and relation-atom factors. Fig. 1(f) of the original manuscript shows that the two sets of factors respond differently across degradation conditions, and the bank ablation shows that paired banks are more effective than a shared prompt or either bank alone (Fig. 4(a) in Section 4.3 and Table 11 in Appendix C.5 of the revised manuscript). The local degradation experiments demonstrate the effect of their joint calibration. With the decoder input fixed, CAPC-induced response changes concentrate around rain streaks and blur-affected contours (Fig. 5(b) in Section 4.3 of the revised manuscript). Disabling CAPC at D1-1 produces output changes that align with degradation residuals, and this alignment falls to near zero under circular shifts (Table 13 in Appendix C.7 of the revised manuscript). By reweighting spatially indexed value responses, the two axis-specific prompts allow image-level conditioning to produce spatially selective restoration effects.

The above analyses and results indicate that the prompt banks encode degradation-related information with complementary roles in attention calibration. The query bank conditions how individual feature channels aggregate information, while the relation bank conditions the contributions of the compact responses supplying that information. Shared retrieval aligns both prompts with the same degradation condition, and their distinct representations jointly determine the calibration of each query–relation pair. The advantage of paired banks over shared and single-axis alternatives supports the functional benefit of learning separate representations for these roles, while the spatial experiments demonstrate that their coordinated calibration produces restoration responses aligned with local degradation.

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

This paper proposes InCaRP, an all-in-one image restoration framework that adapts attention relations to individual degradations without requiring degradation labels at inference. It introduces two modules: Low-Rank Relation-Space Attention (LRSA), which compresses key and value representations into input-dependent relation carriers while preserving query and output dimensions; and Coupled-Axis Prompt Calibration (CAPC), which uses shared hyperspherical retrieval with paired prompt banks to generate bounded, low-rank scale and shift fields for attention-logit modulation. The modules are integrated into a multi-scale encoder-decoder Transformer, with CAPC applied in the decoder. Experiments on AIO-3, AIO-5, and CDD-11 evaluate restoration performance across individual and compositional degradations, supported by ablations of the main design choices.

Critical strengths
  1. The paper presents a clear motivation and coherent architecture. LRSA constructs a compact relation space, while CAPC adapts attention interactions to individual inputs. Their complementary roles provide a well-motivated, integrated design.
  2. The method achieves competitive results across AIO-3, AIO-5, and CDD-11, covering multiple degradation types with a relatively modest parameter count. Extensive ablations further support the effectiveness of the main components.
Critical weaknesses
  1. The incremental contribution of the specific coupled-axis design remains uncertain. Paired prompts and shared retrieval improve average PSNR by only 0.09 dB and 0.08 dB over simpler alternatives, respectively. Without multi-seed results or uncertainty estimates, the robustness and significance of these gains remain unclear.
  2. LRSA's practical efficiency benefits are not fully verified. The paper lacks end-to-end FLOPs, inference latency, and peak memory measurements. Moreover, Appendix Table 7 appears not to control consistently for CAPC across variants, making it difficult to isolate LRSA's advantages over alternative sparsification designs.
Questions for authors
  1. Appendix Table 7 reports 33.06 dB for LRSA with (r=32), matching the full model rather than the LRSA-only result of 32.38 dB in Table 3. Were CAPC and hyperspherical retrieval consistently controlled across the Table 7 variants? This clarification is important for determining LRSA's independent contribution.
  2. How consistent are the 0.09 dB and 0.08 dB improvements from paired prompts and shared retrieval across different training seeds? This would help establish the significance of the specific coupling mechanism.
3 comments
AuthorsOpenReview ↗

[Part 1] Official Comment by Authors

We thank the reviewer for the careful assessment and constructive suggestions. We address each concern below: (1) the robustness of the prompt-design gains across training seeds; (2) the settings and costs of the attention variants; and (3) model efficiency. The corresponding experiments and clarifications are incorporated into the revised main text and supplementary material, with additional analyses and result tables. Please refer to these updates for further details.

R1.1: Incremental contribution of the coupled-axis design

We thank the reviewer for careful reading and constructive suggestion. To address this concern, we repeated both comparisons with two training seeds on AIO-3 and AIO-5. We also redesigned the retrieval ablation to separate the effects of sharing retrieval weights and hyperspherical matching. The revised manuscript summarizes these results in Section 4.3 and reports the results for each seed in Table 11 (Appendix C.5) and Table 12 (Appendix C.6).

(1) Paired-prompt design.

We repeat the comparison between paired prompts and a single prompt shared across both axes using two training seeds. On AIO-3, paired prompts improved average PSNR by 0.09 and 0.13 dB for the two seeds, respectively. On AIO-5, the corresponding gains are 0.29 and 0.33 dB. These results show that the original gain is reproduced under a second seed, with consistent improvements across both settings. These consistent gains support the specific design choice of learning distinct prompts for the query and relation axes, with the larger gains on AIO-5 indicating that this distinction is particularly beneficial when the model handles a broader range of degradations.

Prompt design ablation (Table 11 in Appendix C.5 of the revised manuscript).

VariantAIO-3 / seed 1AIO-3 / seed 2AIO-5 / seed 1AIO-5 / seed 2
w/o Prompts32.38/0.91332.35/0.91429.82/0.91229.80/0.911
Query only32.86/0.91932.85/0.92030.56/0.91730.58/0.916
Relation only32.92/0.92032.91/0.91930.66/0.92230.64/0.922
Shared prompt32.97/0.91832.95/0.92131.02/0.92431.00/0.923
Paired prompts33.06/0.92233.08/0.92231.31/0.92631.33/0.926

Variant settings. All variants use the same backbone. w/o Prompts disables prompt-based calibration. Query only and Relation only use an input-dependent prompt on the query-channel axis and the relation-atom axis, respectively. Shared prompt feeds the same prompt to both axes. Paired prompts (Ours) use the same retrieval weights to compose two distinct prompts from separate query and relation banks, one for each axis. The single-axis variants test the contribution of conditioning each axis individually, and the comparison between Shared prompt and Paired prompts tests the benefit of giving the two axes distinct prompt representations.

(2) Prompt retrieval ways.

We redesigned the retrieval ablation to separate the effect of sharing retrieval weights from that of hyperspherical matching, and repeated the experiments with two training seeds. With hyperspherical matching fixed, shared retrieval improved average PSNR by 0.25 and 0.23 dB on AIO-3. On AIO-5, shared retrieval yields larger gains of 0.66 and 0.68 dB across the two seeds, and the gains are similar across seeds in both settings, supporting the benefit of coordinating the query and relation prompts through shared retrieval when both variants use hyperspherical matching.

Prompt retrieval ablation (Table 12 in Appendix C.6 of the revised manuscript).

Retrieval variantAIO-3 / seed 1AIO-3 / seed 2AIO-5 / seed 1AIO-5 / seed 2
Independent retrieval32.73/0.91832.70/0.91730.41/0.91730.43/0.918
Independent hyperspherical retrieval32.81/0.91932.82/0.91930.65/0.91930.62/0.919
Shared retrieval32.96/0.92032.94/0.92031.17/0.92331.18/0.924
Shared hyperspherical retrieval33.06/0.92233.05/0.92231.31/0.92631.30/0.926
AuthorsOpenReview ↗

[Part 2] Official Comment by Authors

R1.2: Model configurations and computational costs in the original Appendix Table 7

We thank the reviewer for careful reading and for pointing out the ambiguity in original Appendix Table 7. We apologize for not clearly specifying the model configuration for each variant in the original Table 7. The original row labels identify the attention design and relation rank, which leaves the status of CAPC and hyperspherical retrieval unclear. The first three rows provide attention-only controls for dense attention, post-correlation Top-32 selection, and the symmetric bottleneck. The four remaining rows form a rank sweep of the full LRSA+CAPC model with hyperspherical retrieval. Within this sweep, we keep the query width and calibration design fixed and vary the number of K/V relation carriers r. The updated row labels and caption in Table 8 (Appendix C.1 of the revised manuscript) make these settings explicit.

Therefore, the average performance of 33.06 dB at r=32 corresponds to the full model. To assess the independent contribution of LRSA, Table 5 in Section 4.3 of the revised manuscript compares dense attention with LRSA while keeping CAPC and hyperspherical retrieval off in both configurations. The 32.38 dB result belongs to that LRSA-only configuration. We clarify this distinction in Appendix C.1 of the revised manuscript so that readers can identify which comparison isolates LRSA and which evaluates the choice of relation rank within the complete model.

In addition, following the reviewer’s suggestion, we report each variant’s parameter count, peak memory usage, FLOPs, and inference latency in Table 8 (Appendix C.1 of the revised manuscript). The results support our choice of r=32 as a favorable balance between restoration accuracy and computational cost among the rank variants.

Model configurations and computational costs across different attention variants (Table 8 in Appendix C.1 of the revised manuscript).

VariantQ widthK/V widthScore sizeSelectionAvg. P/SParams.MemoryFLOPsRuntime
Full Attention484848×48None32.09/0.91321.25 M6480 MiB79 G48.70 ± 0.46 ms
Sparse Attention484848×48Top-3232.23/0.91621.31 M6998 MiB85 G51.48 ± 0.61 ms
Symmetric bottleneck323232×32None31.91/0.91320.09 M4294 MiB36 G39.92 ± 0.47 ms
LRSA+CAPC, r=16481648×16None32.73/0.91821.09 M5899 MiB68 G49.29 ± 0.54 ms
LRSA+CAPC, r=24482448×24None32.98/0.92121.93 M6057 MiB71 G49.90 ± 0.55 ms
LRSA+CAPC, r=32483248×32None33.06/0.92222.77 M6194 MiB74 G50.24 ± 0.42 ms
LRSA+CAPC, r=48484848×48None33.01/0.92124.44 M6494 MiB80 G51.15 ± 0.44 ms

Evaluation setup. For each variant, we report the total parameter count and measure FLOPs, peak GPU memory usage, and inference latency using a 224 × 224 input on an NVIDIA GeForce RTX 3090 Ti GPU.

AuthorsOpenReview ↗

[Part 3] Official Comment by Authors

R1.3: Model efficiency

We thank the reviewer for the concern about the model efficiency. We compare InCaRP with six representative all-in-one image restoration methods (PromptIR [1], VLU-Net [2], MoCE-IR [3], AdaIR [4], CPLIR [5], and BaryIR [6]) and report their model size, FLOPs, peak memory usage, and inference latency. As shown in Table 4 of Section 4.3 in the revised manuscript, our method achieves the lowest parameter count, FLOPs, and peak memory usage among the compared methods, while maintaining inference latency comparable to MoCE-IR.

Model complexity and inference efficiency of representative all-in-one image restoration methods. (Table 4 in Section 4.3 of the revised manuscript)

MethodParams.MemoryFLOPsRuntime
PromptIR36 M9884 MiB132 G70.26 ± 3.08 ms
VLU-Net35 M10767 MiB143 G110.86 ± 4.78 ms
MoCE-IR25 M6332 MiB81 G45.78 ± 2.25 ms
AdaIR29 M9701 MiB124 G75.88 ± 2.65 ms
CPLIR35 M8804 MiB132 G72.63 ± 3.67 ms
BaryIR36 M11510 MiB163 G88.64 ± 0.54 ms
InCaRP(Ours)23 M6194 MiB74 G50.24 ± 0.42 ms

Evaluation setup. We report the total number of parameters for each complete model. FLOPs, peak GPU memory usage, and inference latency are evaluated under the same setup for all methods, using 224 × 224 input images on an NVIDIA GeForce RTX 3090 Ti GPU.

We hope the additional experiments and clarifications address your concerns about the robustness of the gains, model configurations, and computational efficiency, and we welcome further discussion.

References

[1] Vaishnav Potlapalli, Syed Waqas Zamir, Salman Khan, and Fahad Shahbaz Khan. PromptIR: Prompting for All-in-One Image Restoration. NeurIPS, 2023.

[2] Haijin Zeng, Xiangming Wang, Yongyong Chen, Jingyong Su, and Jie Liu. Vision-Language Gradient Descent-Driven All-in-One Deep Unfolding Networks. CVPR, 2025.

[3] Eduard Zamfir, Zongwei Wu, Nancy Mehta, Yuedong Tan, Danda Pani Paudel, Yulun Zhang, and Radu Timofte. Complexity Experts are Task-Discriminative Learners for Any Image Restoration. CVPR, 2025.

[4] Yuning Cui, Syed Waqas Zamir, Salman Khan, Alois Knoll, Mubarak Shah, and Fahad Shahbaz Khan. AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation. ICLR, 2025.

[5] Gang Wu, Junjun Jiang, Kui Jiang, Xianming Liu, and Liqiang Nie. Beyond Degradation Redundancy: Contrastive Prompt Learning for All-in-One Image Restoration. IEEE TPAMI, 2026.

[6] Xiaole Tang, Xiaoyi He, Jiayi Xu, Xiang Gu, and Jian Sun. Learning Continuous Wasserstein Barycenter Space for Generalized All-in-One Image Restoration. IEEE TPAMI, 2026.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Overview of the Discussion

We thank the reviewers, ACs, and SACs for their careful evaluation and constructive suggestions. We appreciate the reviewers' positive assessment of our clear motivation, integrated LRSA–CAPC design, insight into directly conditioning attention relations with degradation-aware prompts, and promising restoration performance with a relatively modest parameter count, supported by extensive experiments. We outline the principal concerns and the evidence and clarifications provided in our responses below.


Reviewer uJmP

Key concerns: (1) robustness of paired prompts and shared retrieval, (2) attention-ablation settings and costs, (3) overall model efficiency.

Response summary:

(1) We provide ablations across seeds and task settings, showing consistent benefits from paired prompts and shared retrieval (Appendix Tables 11–12).

(2) We clarify the attention controls and rank settings, separating component gains and documenting accuracy–cost trade-offs (Table 5 and Appendix Table 8).

(3) We compare model size, computation, memory, and runtime under a unified setup, demonstrating strong overall efficiency (Table 4).


Reviewer TeEQ

Key concerns across three rounds: (1) degradation versus content information, (2) rationale for attention calibration, (3) information and roles of the two banks, (4) methodological contributions of LRSA and CAPC, (5) advantages over existing prompting methods, (6) LRSA’s contribution and connections to prior attention designs, (7) prompt-bank terminology versus degradation representations, (8) mixed-degradation behavior and channel specialization, (9) contribution positioning.

Response summary:

(1) We provide content-matched PCA and calibration-swap analyses, showing that prompts capture degradation type and severity (Figures 1(d) and 4(a)).

(2) We explain degradation-aware aggregation and compare conditioning locations, showing the advantage of direct attention-logit calibration (Appendix Figure 9(b)).

(3) We clarify the banks’ complementary query and relation roles and how shared retrieval coordinates their conditioning (Section 3.3, Figure 2).

(4) We present LRSA’s aggregation constraint and CAPC’s degradation-aware relation calibration as distinct contributions, supported by component and prompt-design ablations (Table 5, Appendix Tables 11–12).

(5) We compare prompt sharing, retrieval, and injection strategies, supporting paired prompts, shared retrieval, and direct relation calibration (Appendix Tables 11–12 and Figure 9).

(6) We clarify LRSA’s sparse-attention motivation, discuss its connections to Linformer/PVT, and explain its response compression and distinctions from GRL/ATD (Sections 2 and 3.2).

(7) We clarify prompt-bank terminology and identify shared retrieval and paired logit calibration as the distinctive use of degradation representations (Sections 2 and 3.3).

(8) We analyze mixed-degradation prompts and retrieval, showing adaptive prompt composition, and clarify the interpretation of channel roles (Appendix Figures 7–8).

(9) We refine the contribution statements, identifying CAPC as the main innovation and LRSA as the complementary aggregation constraint (Abstract, Sections 1–3).

Discussion update: Following our initial response, Reviewer TeEQ raises the score. The subsequent follow-up maintains this rating and recognizes the effectiveness of LRSA, while retaining concerns about technical novelty. Our final response addresses these concerns through comparisons with prior work and a clearer account of each contribution.


Reviewer zSTV

Key concerns: (1) practical efficiency, (2) real-world and unseen-degradation generalization, (3) spatially localized restoration through global prompts.

Response summary:

(1) We report practical resource and runtime measurements, demonstrating low computational costs and competitive inference speed (Table 4).

(2) We evaluate real-world and unseen degradations without fine-tuning, demonstrating strong generalization across fidelity and perceptual metrics (Table 3).

(3) We explain spatially varying value aggregation and provide controlled interventions, showing that global prompts induce locally aligned restoration responses (Figure 4(b), Appendix Table 13).

Discussion update: The reviewer confirms that our response addresses the concerns raised and maintains the original positive score.


The corresponding updates are marked in blue in the revised manuscript. We hope this overview, together with the detailed responses, assists the area chairs in assessing the revised submission.

AuthorsOpenReview ↗

Clarification of Methodological Novelty

We sincerely thank all reviewers, ACs, and SACs for their time, constructive feedback, and thoughtful evaluation. To address the remaining questions about methodological novelty, we further explain the distinct contributions of LRSA and CAPC and their relationship to representative restoration methods. The proposed method introduces two complementary mechanisms, LRSA and CAPC, with CAPC forming the core methodological contribution. Extensive experiments and ablations demonstrate strong restoration performance, while comparisons with representative methods show fewer parameters and lower computational costs (Tables 1–4). To contextualize these contributions, representative restoration methods advance prompt representation, conditioning, or attention through the following designs:

  • PromptIR (NeurIPS 2023) [1] → Adaptive prompt composition for feature conditioning.
  • InstructIR (ECCV 2024) [2] → Human-written instructions for restoration across multiple degradations.
  • NDR (TIP 2024) [3] → Degradation dictionary learning with bidirectional training.
  • DATPRL-IR (ICLR 2026) [4] → Task and domain prompts with distilled priors.
  • CPLIR (TPAMI 2026) [5] → Sparse prompt selection with contrastive regularization.
  • BaryIR (TPAMI 2026) [6] → Wasserstein barycenter alignment with orthogonal degradation-specific subspaces.
  • DRSformer (CVPR 2023) [7] → Top-k attention selection with mixed-scale feature processing.
  • AST-v2 (TPAMI 2025) [8] → Dense attention modulation using sparse scores from squared ReLU.

These works illustrate the value of targeted advances in how restoration networks represent degradations and aggregate information. InCaRP advances degradation-conditioned aggregation through two complementary mechanisms. LRSA constrains the available K/V responses while preserving full-width queries and outputs. CAPC composes distinct query-axis and relation-axis prompts through shared hyperspherical retrieval, and their factorized interactions generate pair-specific scale and shift fields before softmax. This coordinated prompt construction and interaction forms our principal methodological contribution. Component and prompt-design ablations support these choices and show that CAPC remains effective without LRSA (Table 5, Appendix Tables 11–12).

We hope this clarification of the respective roles of LRSA and CAPC supports a fuller assessment of the novelty and technical value of our work.


References

[1] Vaishnav Potlapalli et al. PromptIR: Prompting for All-in-One Image Restoration. NeurIPS, 2023.

[2] Marcos V. Conde et al. InstructIR: High-Quality Image Restoration Following Human Instructions. ECCV, 2024.

[3] Mingde Yao et al. Neural Degradation Representation Learning for All-in-One Image Restoration. IEEE Transactions on Image Processing, 2024.

[4] Guanglu Dong et al. Learning Domain-Aware Task Prompt Representations for Multi-Domain All-in-One Image Restoration. ICLR, 2026.

[5] Gang Wu et al. Beyond Degradation Redundancy: Contrastive Prompt Learning for All-in-One Image Restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026.

[6] Xiaole Tang et al. Learning Continuous Wasserstein Barycenter Space for Generalized All-in-One Image Restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026.

[7] Xiang Chen et al. Learning a Sparse Transformer Network for Effective Image Deraining. CVPR, 2023.

[8] Shihao Zhou et al. Learning an Adaptive Sparse Transformer for Efficient Image Restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.