acceptodds
Under review as a conference paper at ICLR 2027

Decompose to Restore: Factor-Specific Adaptation for Image Restoration under Composite Degradations

Abstract

Image restoration under composite degradations requires a unified model to handle multiple degradation factors that often co-occur in real-world images. Existing all-in-one methods typically treat each degradation type as an independent restoration condition or use degradation cues only as implicit conditioning, which limits factor-specific restoration behavior under composite degradations. We observe that restoration models already encode degradation-factor information, but such information is not well organized for factor-wise adaptation. We propose DFLoRA, a unified restoration network that decomposes composite degradations into reusable factors with dedicated low-rank adaptations. DFLoRA explicitly organizes latent degradation information into reusable factor representations and composition coefficients, decomposes a composite degradation into its fundamental factors, and equips each factor with a dedicated low-rank restoration executor. The resulting factor-specific adaptations are then composed according to the inferred degradation composition, directly translating degradation composition into adaptation composition. On the composite degradation restoration benchmark, DFLoRA achieves a PSNR gain of 1.23 dB over the strongest competing method, and it also achieves state-of-the-art performance across multiple single-degradation tasks in the standard all-in-one restoration setting. Our code will be publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
All-in-One Image Restoration, Composite Degradation, Degradation Factorization, Low-Rank Adaptation

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_acBsOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

This paper proposes DFLoRA, a factor-specific low-rank adaptation framework for image restoration under composite degradations. Specifically, it introduces Degradation Factor Organization (DFO) to decompose composite degradations into reusable factor representations and estimate their composition coefficients. Furthermore, Factorized Restoration Adaptation (FRA) employs dedicated LoRA modules for individual degradation factors and adaptively combines their restoration responses.

Critical strengths
  1. The paper is generally well-written and logically organized.

  2. The proposed framework is reasonably designed. By integrating degradation factor organization with factor-specific LoRA adaptation, DFLoRA establishes a direct connection between degradation decomposition and restoration adaptation, enabling flexible restoration across different degradation combinations.

Critical weaknesses
  1. The proposed framework mainly combines degradation factor classification, factor-specific LoRA modules, and adaptive feature fusion. Similar techniques have been explored in existing all-in-one image restoration methods. Although their integration is reasonable, the paper does not sufficiently clarify the fundamental technical novelty beyond existing degradation-aware adaptation approaches.

  2. The experimental evaluation on real-world scenarios remains insufficient. The composite degradation experiments are primarily conducted on the synthetic CDD-11 benchmark, while the remaining evaluations mainly rely on standard restoration benchmarks. The authors are encouraged to evaluate their method on more challenging real-world datasets to further demonstrate its generalization capability and practical applicability.

  3. The effectiveness of the proposed degradation factor decomposition requires further investigation. The DFO module relies on ground-truth factor labels for supervision, and the reported factor prediction accuracy is evaluated on predefined degradation categories. It remains unclear whether the learned representations truly capture independent and reusable degradation factors or mainly perform supervised degradation classification.

Questions for authors

See the above Weaknesses parts.

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

This paper proposes DFLoRA, a framework for image restoration under compound degradations. Using degradation-factor labels for supervision, the method first maps backbone features to the presence probability of each basic degradation factor. It then assigns factor-specific LoRA adaptation parameters and combines feature responses from different branches in the decoder using predicted coefficients and spatial gating. The core idea is to handle different degradation combinations by reusing adaptation modules associated with basic degradation factors. Training requires degradation-factor labels, while the model predicts the factor combination during inference. The authors report performance improvements on the CDD-11 compound degradation benchmark and five single-degradation restoration tasks. Module ablations, degradation recognition accuracy, feature visualizations, and subspace projection analyses further assess the roles of individual components and the compositional properties of the learned representations.

Critical strengths
  1. On CDD-11, DFLoRA achieves an average PSNR of 32.01 dB, outperforming CDIR-S, the strongest comparison method in the table, by 1.23 dB overall and 1.00 dB on the compound degradation subset. The gains span single-, double-, and triple-degradation conditions. Evaluation on five single-degradation tasks also yields strong average performance. These results provide reasonably convincing evidence of the method’s practical effectiveness under different degradation conditions and form the main basis for a positive assessment.

  2. The method assigns low-rank adaptation parameters to individual basic degradation factors and combines restoration responses using the predicted factor coefficients, allowing the same set of adaptation modules to handle different degradation combinations. This design provides a clear and reusable implementation for modeling compound degradations. In the ablation study, the factor-specific LoRA component adds approximately 0.52M parameters and increases average PSNR from 31.87 to 32.01 dB, providing empirical support for its incremental value.

Critical weaknesses
  1. The paper interprets each LoRA branch as a restoration adapter for a basic degradation factor. However, these branches receive the same global context and are jointly optimized through the final restoration objective, so the design does not guarantee a clear division of responsibilities among them. The analyses in Table 4 and Figures 5 and 6 mainly examine the separability and subspace structure of DFO representations, without directly testing the functions of the LoRA branches. Performance ablations only show that adding these branches is beneficial. Without interventions such as disabling or swapping branches, it remains difficult to determine whether the gains genuinely arise from independently reusable, factor-specific restoration capabilities. This evidence gap directly affects the credibility of the paper’s explanation of its core mechanism.

  2. According to Appendix A.1, all 11 degradation conditions in CDD-11 are included in training, so testing mainly evaluates restoration of new images under known degradation conditions. The paper does not hold out degradation combinations to test whether adapters associated with basic factors can be combined to handle conditions absent from training. The five single-degradation tasks also use separate training on their respective datasets and therefore do not provide this evidence. Furthermore, the existing real-world low-light experiments are insufficient to validate the factor composition mechanism under real-world compound degradations. The current experiments thus do not adequately distinguish generalizable compositional capabilities from adaptation to known degradation conditions, limiting the assessment of the method’s core contribution and scope of applicability.

Questions for authors

1.Do individual LoRA branches provide factor-specific restoration capabilities? The current representation analyses mainly examine DFO outputs, while all restoration branches receive the same global context. Do the authors have branch disabling or swapping experiments showing that removing the branch associated with a particular factor primarily affects restoration of the corresponding degradation, rather than restoration quality in general? This would help determine whether the branches exhibit the factor specialization emphasized in the paper.

2.All 11 degradation conditions in CDD-11 are included in training. Are there results under a setting where all basic factors have been observed during training, but a particular combination is held out? If not, how do the authors define the scope of the “composition and reuse” capability supported by the existing evidence?

3.Table 3 shows that adding factor-specific LoRA improves average PSNR by 0.14 dB. Are there comparisons against a parameter-matched shared adaptation module or branches that are not explicitly tied to degradation factors? Such comparisons would help distinguish the contribution of factor–branch correspondence from those of additional parameter capacity and conditional feature modulation.

4.Validation on real-world data currently focuses mainly on single-degradation tasks. Do the authors have tests or representative examples of real-world compound degradations, without additional fine-tuning, that show both factor predictions and restoration results? This evidence would affect the assessment of the method’s practical scope of applicability.

Reviewer_Zf4POpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DFLoRA for image restoration under composite degradations. Degradation Factor Organization predicts non-exclusive factor-presence coefficients from bottleneck features using supervised prototypes. Factorized Restoration Adaptation uses these coefficients to combine factor-specific low-rank K/V adaptations with spatially varying responses. Training transitions from ground-truth to predicted coefficients. Experiments cover CDD-11, five single-degradation tasks, progressive ablations, and representation analyses. The reported results support improved restoration over the included baselines; the stronger interpretation as reusable factor decomposition requires further validation.

Critical strengths
  1. The method achieves promising results: 32.01 dB / 0.900 SSIM on CDD-11. Table 3 shows progressive improvements from 31.65 dB for the backbone to 31.87 dB with DFO and restoration context, and 32.01 dB with factor-specific LoRA.

  2. The representation analysis provides useful evidence beyond visualization. After DFO, nearest-centroid classification accuracy increases from 67.9% to 98.8%, and factor-composition exact-match accuracy reaches 99.00%. These results support better organization of degradation information, though not necessarily functional specialization of the restoration branches.

Critical weaknesses
  1. DFO’s benefit over a simple predictor is unclear. Appendix B reports a linear-probe AUC of 0.997 on backbone features before DFO, suggesting that factor information is already readily accessible. Yet Table 3 adds DFO and restoration context together, without comparing DFO against a linear layer plus sigmoid controlling the same LoRA branches. This missing control leaves DFO’s contribution unresolved; it does not establish that DFO is unnecessary.

  2. The baseline protocols need clarification. The paper does not clearly state which results were reproduced, which were taken from prior work, or how training data and budgets compare. The five-task repetition factors—3, 120, 1, 1, and 200—also lack a clear rationale. These omissions make the performance gains harder to interpret, but do not by themselves demonstrate an unfair comparison.

Questions for authors
  1. With all other components and training settings fixed, how does DFO compare with a linear layer plus sigmoid for predicting the LoRA weights?
  2. Does disabling a factor branch selectively affect its corresponding degradation? Also, since all 11 CDD-11 combinations appear during training, the current results do not establish generalization to unseen combinations. Please clarify whether that is an intended claim and provide supporting evidence or narrow the wording.
  3. Please identify the source and training protocol of each major baseline result, and explain the choice of task-repetition factors.
  4. More related work should be included, including: Neural Degradation Representation Learning for All-in-One Image Restoration (IEEE TIP, 2024).
1 comment

I thank the authors for the additional experiments and clarifications. The linear + sigmoid comparison addresses my main concern regarding DFO’s contribution, and the clarification of the generalization claim is helpful. While some limitations in the supporting evidence remain, the response provides sufficient clarification for my assessment. Having considered both the paper and the rebuttal, my final recommendation remains Weak Accept.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.