acceptodds
Under review as a conference paper at ICLR 2027

DDPR: Dominant Degradation Planning with Latent Residual-Guided Adaptive Editing for All-in-One Image Restoration

Abstract

Generative models have shown strong potential for All-in-One image restoration. However, existing methods either process compound degradations in a single pass or lack dynamic reassessment of the dominant degradation, often leading to suboptimal restoration orders and residual degradations. They also apply uniform restoration across the image, without distinguishing regions requiring content preservation from those requiring stronger restoration, which may cause content hallucination. To address these limitations, we propose Dominant Degradation Planning for image Restoration (DDPR), a unified framework with latent residual-guided adaptive editing. Specifically, the Dominant Degradation Planner identifies the current dominant degradation and reassesses the image after each restoration step, decomposing compound degradations into a dynamically planned sequence of restoration tasks. The Degradation-Aware Latent Residual Assessor predicts a region-level degradation gate from VAE latent residuals, guiding the Degradation-Gated Dual-Branch LoRA to adaptively balance the content-preservation and degradation-restoration branches. Experiments demonstrate that DDPR achieves SOTA performance among open-source methods across both synthetic and real-world degradation benchmarks. Code will be publicly released.

Primary area:
generative models
Keywords:
all-in-one image restoration, compound degradation, image editing, degradation planning

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_hjYzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The authors propose Degradation Planning for Image Restoration (DDPR), an agentic approach for all-in-one image restoration (IR), specifically designed to tackle compound degradations. DDPR addresses the dominant degradation first and reassess the restored result after each step. Furthermore, DDPR uses degradation-aware latent residual estimator which estimates degradation information in the restored results, thereby gating the areas where restoration is required (similar to an adaptive restoration strength). DDPR achieves state-of-the-art results quantitatively.

Critical strengths
  1. The idea of using a region-level degradation gating approach is quite interesting. I was also impressed by its performance on real-world images containing compound degradations.
Critical weaknesses
  1. The qualitative results in the paper raise significant concerns about hallucination and restoration fidelity. The results of DDPR are visually appealing but examples in Fig. 5 and Fig. S2 seem to have significant content hallucinations. For instance, in Fig. S2 row 1, the shape of the red car seems to have changed in the restored result. Furthermore, the inputs showcased in the paper have extremely severe degradations, making it difficult to gauge what is the underlying structure. Thus, when the restored result is visualized, it is unclear whether the details were hallucinated or were faithful to the input.

  2. Limited novelty of the degradation planner: RestoreAgent [1] also seems to identify degradation types and severities for determining the restoration order. Furthermore, Fape-IR [2] is also a unified model which combines a planner, diffusion-based restorer and LoRA for adaptively influencing restoration. This seems relevant to the claim in Line 164.

[1] Chen, Haoyu, et al. "Restoreagent: Autonomous image restoration agent via multimodal large language models." Advances in Neural Information Processing Systems 37 (2024): 110643-110666.

[2] Liu, Jingren, et al. "FAPE-IR: frequency-aware planning and execution framework for all-in-one image restoration." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

Questions for authors
  1. Can the authors provide extensive qualitative results for the results of Table 2? Showcasing the reference would make it easier to judge restoration fidelity.

  2. Can the authors add real-world results of lower degradation severities, so that the reader can accurately judge the faithfulness of the proposed method?

  3. Could the restoration fidelity be compromised due to the usage of training data obtained from frontier AI models (Sec. S1)?

  4. Can the authors discuss the differences in novelty between RestoreAgent [1] and Fape-IR [2]?

[1] Chen, Haoyu, et al. "Restoreagent: Autonomous image restoration agent via multimodal large language models." Advances in Neural Information Processing Systems 37 (2024): 110643-110666.

[2] Liu, Jingren, et al. "FAPE-IR: frequency-aware planning and execution framework for all-in-one image restoration." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

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

Compared to existing methods, this paper addresses a critical challenge in scenarios involving compound degradation: the spatially non-uniform distribution of image degradation—a context in which current approaches struggle to distinguish between clean regions and areas with severe degradation. To address this, this paper proposes a degradation assessment mechanism based on potential residuals and designs a dual-branch gated LoRA structure.

Critical strengths
  1. The paper addresses two practical problems: choosing the restoration order when multiple degradations interfere with one another, and accounting for differences in the restoration strength required across spatial regions. Combining dynamic replanning with spatially adaptive editing provides a reasonable approach to addressing these problems.
  2. The proposed method is relatively simple, clear, and easy to understand.
Critical weaknesses
  1. This paper argues that the global weights in traditional LoRA cannot distinguish between different types of degradation, and the visualization of the "degradation gating" mechanism appears effective. It is recommended to include an ablation study visualizing the use of the "content-preservation branch" versus the "composite branch" in isolation to verify whether the content-preservation branch alone is insufficient to address the degradation.
  2. The model is constrained to remove only the dominant degradation while retaining others. Since strongly coupled composite degradations are not merely simple superpositions, does this constraint introduce optimization challenges when applied to real-world images? Could it lead to incomplete restoration or interference between different types of degradation?
  3. Additional information should be provided regarding the number of iterations required and the inference efficiency of the model across different datasets.
  4. There is a fairness issue in the AgenticIR benchmark comparisons: the baseline methods were trained on "Group A" data, whereas this paper utilizes a self-constructed dataset, meaning the training sets are not consistent.
  5. The paper lacks experimental comparisons with the latest literature [1][2].
  6. The evaluation of the planner is inadequate: it relies solely on final image quality metrics without assessing decision-making performance. Specifically, it lacks metrics for the accuracy of dominant degradation identification and the reliability of the "stop" signal; furthermore, there is no comparison against simpler planning strategies.

[1] OPERA: An Agent for Image Restoration with End-to-End Joint Planning–Execution Optimization [2] TIR-AGENT: TRAINING AN EXPLORATIVE AND EFFICIENT AGENT FOR IMAGE RESTORATION

Questions for authors

Please refer to Weaknesses.

1 comment

The authors does not submit the rebuttal. I keep my original score.

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

This paper proposes DDPR, focusing all-in-one image restoration, and taking more consideration on restoration order and spatially variant corruptions. Specifically, the authors propose combining a Dominant Degradation Planner with a unified image editing restorer. The planner is used to repeatedly identify the degradation and determine when to stop. In the restorer backbone, DLRA is proposed to preserve content fidelity and help improve degradation removal. In experiments, RealComp-Bench is introduced as another benchmark for composite degradations. Extensive experiments on various datasets demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The proposed method alleviates the limitations caused by the restoration order. Reusing a unified restorer across steps provides a way to handle different degradation combinations without maintaining a separate model for each task.
  2. The proposed DLRA, which learns local restoration requirements from the latent residual, enables task-dependent spatial guidance.
  3. The experiments cover comprehensive restoration settings, a new benchmark dataset `RealComp-Bench' is introduced. The results on comparative experiments and ablation studies demonstrate the effectiveness of the proposed method and each component.
Critical weaknesses
  1. The claimed spatial preservation benefits require more direct validation. In experiments, the authors do not directly quantify whether the resulting gates reduce unnecessary changes in relatively clean regions while improving restoration in degraded regions. The implementation and trainable parameter count of the w/o DGDB variant also need clarification to help distinguish the observed gains can be attributed to spatial gating beyond differences in adaptation capacity.

  2. The evaluation on real degradations has limitations in assessing content fidelity. Equation (16) clips negative changes to zero, so CDRS does not directly penalize worsening an existing degradation. Its restriction to the input degradation set also excludes newly introduced degradation types, while CDFS contains no explicit content-fidelity term. Moreover, four IQA metrics used to construct the planning labels also appear in the evaluation, motivating additional assessment using criteria not involved in label generation.

  3. The validation of synthetic planning supervision on actual restoration trajectories remains not fully established. In Section 3.1, candidate states are constructed by removing a degradation from the synthesis sequence and resynthesizing the image from the clean source. These states do not model the incomplete degradation removal, unintended content changes, or newly introduced artifacts. Thus, the action preferences derived from idealized quality improvements may not convert into the most effective actions during inference. The ablation study in Table 4 provides useful evidence for order selection, but does not directly isolate the benefit of dynamic reassessment over an order planned once from the initial image.

Questions for authors

The detailed questions are listed in Weakness. The authors could provide small-scale experiments to improve their validation further.

  1. How well do the synthetic planning labels predict effective actions by the trained restorer?
  2. Could the authors directly evaluate spatial preservation and the contribution of gating?
  3. Could the authors provide stronger multi-round comparisons and report their computational costs?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.