acceptodds
Under review as a conference paper at ICLR 2027

CONTROLIR: LEARNING TO CONTROL STATE SPACE MODELS VIA SPATIAL EVIDENCE FOR IMAGE RESTORATION

Abstract

Image restoration aims to remove degradations and reconstruct clear scene content, but complex scene structures and spatially varying degradations make it difficult to suppress interference while preserving details. Existing methods use scene structure or degradation cues to guide restoration, but how to jointly describe the structural and degradation properties of different regions and use them to regulate the restoration process remains a key challenge. We propose ControlIR, a compact restoration network that extracts multidimensional spatial evidence from the input image to describe regional scene structure and degradation characteristics while preserving the spatial correspondence among evidence fields. It then uses these fields to condition the state-space restoration process according to their roles, allowing each location to adapt how it uses cross-region information while removing local interference and preserving both fine details and overall structure. Experiments show that ControlIR achieves higher image restoration quality than the compared methods. Controlled studies verify the contributions of multidimensional spatial evidence and operation-level control while requiring only small increases in parameters and computation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Restoration, State Space Models, Spatial Evidence, Spatially Adaptive Control, Image Deraining

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_Cqo6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes ControlIR, a compact image restoration framework that utilizes multidimensional spatial evidence to standardize state-space modeling. The proposed ControlSSM module adjusts state propagation based on spatial structure and degradation characteristics. Experiments demonstrate the effectiveness of the proposed method.

Critical strengths

This method has strong recovery performance and relatively low parameters and computational costs.

Critical weaknesses
  1. The novelty of the method is incremental, as spatial guidance and conditional state space modulation have already been studied in existing restoration methods.
  2. The evaluation mainly focuses on rain removal, which limits the universality of the proposed restoration framework.
  3. The main benchmark comparison relies on previously published results rather than a uniformly retrained baseline.
  4. There is insufficient research on the robustness of predicted spatial evidence under hybrid degradation.
Questions for authors

Can ControlSSM be extended to other recovery tasks such as deblurring, denoising, and low light enhancement? How sensitive is this method to errors in estimating geometric and reliability maps? Can the author provide additional comparisons under the same training and evaluation scheme?

Reviewer_1R5MOpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ControlIR, a lightweight network for image restoration. The method extracts spatial cues from degraded images, such as depth, horizontal/vertical geometric variations, local geometric changes, degradation intensity, and geometric reliability, and uses three control interfaces to respectively adjust Mamba2's discretization steps, state writing, and post-scan feature correction. The authors conducted experiments on RainCityscapes++ and LHP-Rain, and verified the effectiveness of different control mechanisms and spatial cues through ablation studies. ...

Critical strengths
  1. The design of ControlSSM is quite targeted. It directly uses spatial information to control different operations of SSM, rather than just simple feature concatenation, which makes the design reasonable.
  2. The ablation experiments are quite systematic. They analyzed different control routes and evidence assignments and reported results from multiple runs, which can somewhat support the method design.
Critical weaknesses
  1. The credibility of the main SOTA comparison is insufficient. Some of the baseline results and evaluation protocols in Table 1 are not fully explained, especially the results of DEMore-Net need to be verified. If different methods used different training/evaluation protocols, the performance improvements claimed in the paper are not fully convincing.
  2. The role of “multi-dimensional spatial evidence” itself has not been sufficiently isolated. Experiments like individual evidence, spatial shuffle/misalignment, and oracle depth are missing, so it’s currently hard to tell whether the performance improvements come from the proposed spatial evidence or from the additional network capacity and adaptive modulation mechanisms.
Questions for authors
  1. Were all baselines in Table 1 trained and evaluated using exactly the same protocols? Please clarify the specific source and evaluation setup for DEMore-Net results.
  2. Could you report the separate ablation results for D, Gx, Gy, L, Rdeg, and Crel, as well as the experiments with spatially shuffled evidence?
  3. Crel is used in multiple control modules. How does performance change if Crel is removed? Is there experimental evidence showing that Crel reliably reflects geometric quality?
  4. Please explain the testing environment, batch size, precision, and number of repetitions for the 0.019s latency to validate the efficiency advantage.
Reviewer_ifm9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ontrolIR is an image restoration framework that constructs spatial evidence describing geometry, local variation, degradation-related responses, and prediction reliability. These fields condition state-space discretization, input writing, and local–global feature fusion. However, the submission does not sufficiently establish its methodological novelty relative to existing degradation-conditioned state-space models, and the evaluation remains narrow.

Critical strengths
  1. Coherent design. The method provides a clear mechanism for incorporating spatially varying information into state propagation, input writing, and feature fusion.
  2. Positive restoration results. ControlIR improves restoration quality on both RainCityscapes++ and LHP-Rain. The additional control heads introduce relatively small overhead over the depth-equipped state-space baseline.
Critical weaknesses
  1. The incremental contribution over closely related work is insufficiently established. The paper explicitly acknowledges that DPMambaIR already uses degradation information to modulate state-space discretization, input, and readout mappings. ControlIR introduces richer spatial evidence, predefined field assignments, and post-scan correction, but does not directly compare against this closely related approach. Comparisons with concatenation and affine modulation are informative, yet do not establish the advantage over existing operation-level SSM conditioning.
  2. The evaluation does not establish the breadth of the proposed restoration framework. Both evaluated datasets concern mixture-of-rain removal. There are no experiments on other restoration tasks or direct cross-dataset generalization. Consequently, the results establish effectiveness in the tested deraining setting but provide limited evidence that the evidence construction and assignment rules generalize beyond it.
  3. External comparisons are not fully controlled. Table 1 combines published restoration scores and complexity figures from their original protocols with internally measured results for ControlIR. Differences in training, preprocessing, evaluation, and FLOPs accounting can affect the claimed quality–efficiency advantage. The unified training comparison on LHP-Rain is helpful, but covers only a limited selection of methods and omits the closest degradation-conditioned SSM baseline.
  4. The necessity of the proposed evidence semantics remains unclear. The degradation-response field measures local intensity variation and therefore also responds to scene texture and edges. Similarly, reliability supervision measures agreement with teacher-generated geometry rather than independently validated geometric confidence. The paper acknowledges some of these limitations, but does not adequately distinguish the benefit of these semantically motivated fields from that of a generic learned spatial conditioning representation under comparable capacity and supervision.
  5. The evidence-assignment study supports a configuration choice more than a general principle. Table 5 shows that the selected assignment outperforms several expanded-input alternatives on the main benchmark. However, this does not establish that the prescribed field-to-operation mapping is intrinsically preferable to learned assignment or other comparably constrained mappings. This limits the strength of the assignment scheme as a central methodological contribution.
Questions for authors
  1. How does ControlIR compare with DPMambaIR under matched training and evaluation settings? Can the proposed spatial evidence also improve that baseline, separating the contribution of evidence construction from the control architecture?
  2. Which Tab. 1 results were reproduced, and which were taken from published sources? Are image resolution, data splits, border cropping, and metric implementations identical for the main quality comparisons?
  3. Is there a capacity-matched learned spatial-conditioning baseline using comparable supervision? Such a comparison would help establish whether the explicit evidence decomposition is necessary.
  4. Were the field-assignment rules fixed before evaluation, and are they retained unchanged on LHP-Rain? Are ablation results available there to show that their benefits extend beyond the primary benchmark?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.