acceptodds
Under review as a conference paper at ICLR 2027

UniLens: Degradation-Centric Learning and Evaluation for Real-World All-in-One Image Restoration

Abstract

All-in-one image restoration aims to handle multiple restoration tasks with a single model, but its generalization to real images is still difficult. A major obstacle is deciding which training examples to add, since simply scaling up dataset size does not guarantee coverage of real-world degradation patterns. In this paper, we propose UniLens, a unified pipeline that connects degradation-aware data construction, model adaptation, and task-conditioned evaluation for real-world all-in-one restoration. First, we formulate data construction as a degradation-coverage problem: how can we represent degradations in real-world images and use this representation to build training pairs? The key is to learn a degradation space that is less sensitive to scene content and reveals which real-world degradation patterns the training data miss. We then use this space to guide data synthesis that adds training pairs to cover these patterns. Second, building on this broader coverage, we augment paired supervision with instruction-dependent targets for single, joint, and selective removal, providing richer compositional supervision for all-in-one restoration under real-world composite degradations. An instruction-conditioned mixture of LoRA experts further supports these transformations, with a shared low-rank branch capturing common restoration knowledge and sparse experts providing request-dependent adaptation. Finally, we introduce a task-conditioned vision-language evaluator that assesses what degradation remains and how much the image has improved. Our method achieves state-of-the-art performance on real-world image restoration. Compared with FoundIR-v2, it improves task-conditioned restoration scores by 24.8% on public benchmarks and by 45.3% on real images with composite degradations.

Primary area:
generative models
Keywords:
low-level vision, all-in-one image restoration

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_9tSPOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents UniLens, a unified framework for real-world all-in-one image restoration that includes data construction, instruction-conditioned adaptation, and task-conditioned evaluation. The proposed CODE encoder considers both content and degradation ordering supervision to identify degradation gaps to guide dataset construction process. Three routes are proposed, including offline construction, online synthesis, and retrieval, which are used to close the degradation distribution gap from real-world degradations. Single, joint, and selective degradation removal are all included. Experiments across multiple public benchmarks demonstrate promising results.

Critical strengths
  1. The paper presents a real-world data-construction pipeline for all-in-one image restoration, aiming to reduce the gap of degradation distribution between conventional synthetic training data and real-world images.
  2. The restoration experiments include single, composite, and selective degradation removal. The evaluation covers an extensive range of public benchmarks and demonstrates promising overall performance.
Critical weaknesses
  1. The choice of LPIPS for ordinal supervision requires further justification. Did the authors consider alternative perceptual metrics or degradation-specific statistics? It would be helpful to establish whether LPIPS-derived rankings remain reliable across the different degradation families considered.
  2. The contribution of the data-construction pipeline is difficult to separate from the capabilities of the used pretrained models. Comparisons using alternative generators would help clarify this relationship.
  3. The effective scale of training data and its influence on performance need clarification.
  4. Since data construction is guided by the selected reference images. What criteria of reference selection, and how is coverage of different scenes, degradation types, and severity levels ensured?
  5. How about the diversity of the constructed dataset in terms of degradation severity (ordering)? How do the authors ensure that all degradation levels of one of the degradations provided by reference images are covered, or the constructed training data is benchmark-oriented?
Questions for authors

The main questions are included above. In addition, the authors should better report the generation costs of data construction, it is one valuable aspect of the proposed dataset construction pipeline.

Reviewer_fV3DOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes UniLens, a unified framework for real-world all-in-one image restoration. It introduces CODE to identify degradation coverage gaps and guide training data construction, uses instruction-dependent supervision and a mixture of LoRA experts for single and composite restoration, and develops a task-conditioned evaluator (DRA) for assessing restoration quality without clean references. Extensive experiments demonstrate strong performance across public benchmarks and real-world images.

Critical strengths
  1. The paper addresses an important practical problem in real-world image restoration. The idea of degradation coverage-guided data construction is well motivated, and the framework provides a comprehensive solution covering data construction, instruction-conditioned restoration, and evaluation.
  2. The experimental results are strong across multiple benchmarks, with relatively comprehensive ablation studies. The support for joint and selective degradation removal, together with the proposed real-world benchmark and task-conditioned evaluation, represents a useful contribution.
Critical weaknesses
  1. The framework mainly integrates existing techniques such as contrastive degradation representations, synthetic data expansion, and mixture-of-LoRA adaptation. Moreover, the benefits of MoLE are not fully consistent, as it improves DRA scores but performs worse than dense LoRA in PSNR and LPIPS under the same training conditions.
  2. The evaluation of the proposed framework could be more rigorous. The DRA metric is validated on only 147 human-rated outputs, and its correlation decreases when all fourteen raters are included. Additionally, selective degradation removal is not sufficiently evaluated in scenarios requiring one degradation to be removed while preserving another.
Questions for authors
  1. How does CODE compare with existing degradation-aware representations under the same data selection and restoration training settings?
  2. Could the authors provide further validation of DRA on a larger and independent human-rated dataset, particularly for composite degradations?
  3. Could the authors provide quantitative evaluations specifically for selective degradation removal, including whether non-target degradations are successfully preserved?
  4. Why does MoLE improve the DRA scores while reducing performance on some full-reference metrics compared with dense LoRA?
Reviewer_TdWSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UniLens, a unified framework for real-world all-in-one image restoration that integrates degradation-aware data construction, instruction-dependent supervision, and task-conditioned evaluation. Specifically, the authors introduce a Cross-content Ordinal Degradation Encoder (CODE) to identify underrepresented real-world degradation patterns and guide training data supplementation. An instruction-conditioned mixture of LoRA experts (MoLE) is employed to support single, joint, and selective degradation removal. Furthermore, the authors introduce a degradation-aware restoration assessment metric (DRA) and the WildRestore-Bench dataset for evaluating restoration performance on real-world images. Extensive experiments demonstrate promising results on both public benchmarks and real-world images.

Critical strengths
  1. CODE enables cross-content degradation matching and guides targeted data supplementation. Its strong retrieval performance provides evidence for the effectiveness of degradation-aware representations, offering a useful approach to addressing the mismatch between synthetic and real-world degradations.
  2. Extensive experiments on paired and real-world datasets show competitive results across multiple degradation types. The instruction-dependent compositional supervision also provides a meaningful approach to handling single and composite restoration tasks.
Critical weaknesses
  1. WildRestore-Bench evaluates single and joint restoration but lacks direct evaluation of selective removal, despite selective controllability being an important component of the proposed framework. Furthermore, the human validation of DRA covers only single-task restoration, leaving its reliability for composite degradations insufficiently established.
  2. Although CODE demonstrates strong degradation-category retrieval performance, its ability to capture fine-grained degradation similarities remains unclear. Moreover, the reported distribution-matching improvements are measured in the CODE embedding space, without a direct downstream comparison against random supplementation under identical training settings.
Questions for authors
  1. WildRestore-Bench only evaluates Single and Composite restoration, without directly assessing selective removal. Although Table 4(b) demonstrates the benefits of selective supervision, it does not establish selective controllability.
  2. Although CODE incorporates LPIPS-based ordinal learning to model within-category variations, the reported evaluations mainly focus on degradation-category matching and retrieval. It remains unclear whether CODE can reliably distinguish fine-grained degradation patterns across different scenes. Since CODE relies on these similarities for targeted data construction, the authors should provide more direct evaluations of within-category degradation matching.
  3. The advantages of MoLE are not consistently supported by the ablation results. Table 4(c) shows that MoLE reduces PSNR from 23.1269 to 22.5056 dB and worsens LPIPS from 0.2069 to 0.2133 despite improving DRA.
  4. The proposed DRA metric lacks sufficient independent validation, particularly for composite degradations. Although Table 5 shows that DRA outperforms generic image-quality metrics in correlation with human judgments, the evaluation is limited to 147 outputs from 49 images. Furthermore, the SRCC decreases from 0.537 to 0.474 when all fourteen raters are included instead of the selected seven, suggesting sensitivity to rater selection. More importantly, the human-rated development set covers only single-task restoration, leaving the reliability of DRA for composite degradation assessment insufficiently validated.
  5. Although Table 10 shows that CODE-guided supplementation achieves lower distribution discrepancy than random supplementation, the comparison is limited to metrics computed in the CODE embedding space. While Table 4(a) reports improvements in several restoration metrics after targeted supplementation, it does not establish whether CODE-guided selection outperforms random selection in restoration performance. The authors should provide a direct comparison under identical data budgets and training settings.
Reviewer_oQvNOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

UniLens addresses all-in-one image restoration in real-world scenes, focusing on insufficient training coverage of complex degradation patterns. It trains an encoder with cross-content contrastive learning and within-scene degradation ranking to compare degradations in real images and training samples, then constructs training pairs through offline synthesis, online degradation, and targeted supplementation. The restoration model augments a pretrained image-editing model with a shared low-rank branch and instruction-conditioned sparse experts, using single, joint, and selective degradation-removal targets to support different restoration requests. The paper also introduces a real-image benchmark and the task-conditioned evaluator DRA, which assesses residual degradation and improvement relative to the input. Across multiple public benchmarks, the method reports overall gains in pixel fidelity and perceptual metrics, with component experiments examining data construction, supervision targets, and adapter design.

Critical strengths
  1. The paper formulates insufficient degradation coverage as a testable representation and data-construction problem. In Table 1, CODE achieves a 98.45% degradation-matching rate when content and degradation cues conflict. Table 3 shows reductions in distribution discrepancy and nearest-neighbor distances under a fixed budget of 24K samples per task.
  2. Table 2 reports both pixel fidelity and perceptual metrics across 36 public benchmarks, with a 3.53 dB increase in average PSNR and a 31.3% reduction in DISTS over FoundIR-v2. Figures 4–5 further illustrate rain, haze, and flare removal. The complementary metrics and visual results indicate that the overall restoration gains extend beyond higher scores from the custom DRA evaluator.
Critical weaknesses
  1. The 45.3% gain on real composite degradations in Table 2 is measured by DRA, whereas its correlation with human ratings in Table 5 is assessed only on a development set of 49 inputs and 147 outputs. Although DRA correlates better than generic quality metrics, method rankings on composite tasks have not been independently validated, and the examples in Figure 4 cannot replace this validation. Independent human evaluation is therefore needed to establish whether the metric gain consistently corresponds to better restoration of real composite degradations.
  2. Contrastive degradation representations have precedents such as CONTRIQUE, while the perceptual distance learned in Section 3.1 does not necessarily correspond to the process that generates a degradation. Section 3.2 adds training pairs based on real queries, which does differ from FoundIR-v2's resampling of an existing data pool. However, the principle for translating coverage gaps into a choice of generation operations remains unclear, making it difficult to identify the mechanistic advance beyond degradation embeddings and nearest-neighbor selection.
Questions for authors
  1. Can the method rankings on real composite tasks in Table 2 be reproduced through blinded human evaluation independent of the development set? Please clarify how the validation data in Table 5 are separated from DRA development and calibration, and report ranking agreement on composite tasks.
  2. How are the real reference images used for targeted supplementation in Section 3.2 separated from the evaluation inputs in Section 4.1? Page 7 reports the proposed method's results on the remaining 9,491 images; how do the baselines compare on that same set?
  3. Does CODE distance characterize visual symptoms, the degradation-generating process, or the inverse mapping required for restoration? When these differ, how does the coverage diagnosis determine the generation operations, and what principle distinguishes this procedure from generic nearest-neighbor data selection?
  4. Section 3.3 introduces selective degradation removal, while Table 4(b) mainly shows its benefit for joint restoration. Can the instruction be varied for the same composite input to separately assess removal of the target degradation, retention of the other degradation, and content fidelity?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.