acceptodds
Withdrawn from ICLR 2027

Towards General Real-World Image Restoration: AIGC-Supervised Learning with Knowledge Distillation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Real-world image restoration must accommodate complex variations across scenes while supporting deployment under limited computational resources. However, acquiring paired data is costly and challenging, and synthetic degradations cannot fully capture real image formation processes, limiting the generalization of restoration models to real-world scenarios. We propose a real-world image restoration framework that integrates generative supervision, a perturbation-aware structural correction module (PSR), and knowledge distillation without requiring real clean targets. An input-anchored local risk-constrained selection strategy combines local content constraints, candidate disagreement, and high-risk region aggregation to select a pseudo-GT target with lower proxy content risk from candidates of similar perceptual quality. The selected target is paired with the degraded input to train a restoration teacher. To address variations in pseudo-GT targets across generative models or generation batches, we introduce PSR after the restoration teacher. Using selected alternative targets to guide local teacher probing, PSR learns a shared correction mapping through restoration supervision, response-weighted consistency, and input content constraints. The composite teacher incorporating PSR is then distilled into a student model. We also construct the XSMax dataset with scene-disjoint training, validation, and test splits. Extensive experiments demonstrate that our method improves the perceptual quality and content fidelity of restored images while enhancing output stability under variations in generative supervision. The proposed framework provides a practical approach to cross-scene image restoration without real clean targets and deployment in resource-constrained environments.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Real-World Image Restoration, AIGC-Supervised Learning, Knowledge Distillation.

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_C1EROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates cross-scene category image restoration using images captured by a single device when clean, real-world training reference images are unavailable. The authors introduce XSMax, containing 9900 images divided into class-disjoint training, validation, and test sets. For each input, the system generates four candidate images, evaluated by assessing their local deviation from the input image, stable text constraints, differences between candidate images, and perceptual quality tolerance. Selected candidate images provide supervision for the restoration teacher model, while a further selected target model supports local parameter probing. A residual PSR module combines restoration learning with response-weighted pixel and region consistency and constraints on the input content. The resulting teacher model is refined into a student model with 289,000 parameters.

Evaluation on 2250 test images shows that the student model outperforms competing models across all six no-reference quality metrics. Repeating supervised generation and teacher training 20 times for each generator demonstrates reduced output variability after PSR, while maintaining comparable stability after refinement. These findings suggest that perceived quality and repeatability are improved in a single-device environment, although they do not directly demonstrate the accuracy of content recovery.

Critical strengths
  1. The target selection strategy addresses a significant limitation of generative supervision: even candidates with high global quality scores can exhibit local content variations. Constraints on the input, differences between candidates, and the aggregation of large local biases constitute a reasonable approach to this problem. Table 2 further reinforces this argument by fixing the candidate pool: the complete selection process reduces the NIQE from 6.7933 to 6.0323 under quality-based selection alone, and improves the other five metrics.

  2. This study aims to quantify the sensitivity to regenerated supervision. Section 4.4 and Appendix C.4 describe 20 independent generation and retraining processes for each of the two generators, followed by a paired comparison of the pre- and post-correction results using the same checkpoint. For Nano-Banana-2, the PairLPIPS and PixelSTD ratios are 0.4501 and 0.4304, respectively, indicating that the reduction in variability extends across multiple teachers. Explicitly testing the source of this instability is a crucial component of the evaluation.

Critical weaknesses
  1. The existing evidence does not directly substantiate the claimed improvement in content fidelity. The no-reference scores in Tables 1-2 assess quality, and the measurements in Tables 3-5 assess the consistency of the output; however, neither can definitively determine whether the recovered text or structure matches the actual scene. While Figure 3 shows a clearer reconstruction, it does not provide clear reference images or independently validated content labels. The PSR ablation results also lack average quality and fidelity scores, as well as variability measurements, so the reduction in variability may come at the expense of useful detail. This issue could be addressed by blinding content errors, validating text or structural annotations, or using a subset containing real-world paired reference images.

  2. The lack of experimental detail makes the reproduction and attribution of results difficult. Sections 3-4 and Appendices B-C do not adequately describe the teacher model, student model, and PSR architecture; the encoder and OCR models; important hyperparameters; the training plan; or methods for adjusting the baseline. The selection rules also leave an unresolved issue: (F) includes candidates, but none meet the quality range. Without complete configuration and fallback behavior, readers cannot reliably reconstruct the method, nor can they determine whether the gains in Table 1 stem from the proposed components, differences in supervision, or unequal training conditions.

Questions for authors

The following points will have a substantial impact on my evaluation:

  1. Can the authors demonstrate a reduction in content errors from independent validation and report the average quality and fidelity for each variant in Table 4?

  2. Will the revised version detail all models and training configurations, explain how each baseline was tuned, and define the selection outcome for each alternative?

  3. When the training seed is controlled, does PSR still maintain its stability advantage over random parameter probing with a matching norm? Also, please confirm whether the instructors used for evaluation have been excluded from the PSR fit.

Reviewer_DsK5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method that learns real-world image restoration without real clean targets. Specifically, the authors propose using input-anchored local risk constraints to select generative pseudo-targets and a PSR module to correct a frozen restoration teacher network. The composite teacher is then distilled into a student with 0.289M parameters. Additionally, this paper introduces the XSMax dataset, and experimental results demonstrate the effectiveness of the proposed methods.

Critical strengths
  1. The paper alleviates a limitation of generative supervision. It explicitly considers both content changes in pseudo-targets and instability across generation rounds, providing more robust supervision.
  2. The introduced dataset, XSMax, provides a comprehensive evaluation setting.
  3. Extensive experiments demonstrate the effectiveness and stability of the proposed method.
  4. By distillation, only a 0.289M-parameter student network at inference is attractive for practical deployment.
Critical weaknesses
  1. The claim of improved content fidelity is not directly validated. The results of Tables 1 and 2 support improved perceptual quality but do not establish whether generated text, object identities, and local structures remain faithful to the original scene. Similarly, the selection criteria combine reasonable proxies, but their relationship to actual content errors is not measured.

  2. The stability evaluation does not establish the associated quality–stability trade-off, e.g., Table 1 does not include the original teacher without PSR, making its effect on average restoration quality difficult to assess. The stability gains should be accompanied by quality and content-fidelity measurements for the same configurations and rounds.

  3. The experiments do not fully isolate the contributions of the proposed pipeline. The gains in Table 1 are difficult to attribute separately to the method and device-specific training data. Additional ablation studies or comparisons would clarify the value of each proposed component.

  4. The parameter count alone does not establish practical efficiency at the evaluated image resolution. On deployment, FLOPs and latency on practical hardware are also important. Furthermore, it is better to know the costs of constructing supervision, it would help assess the overall practicality of the proposed method.

Questions for authors

The detailed questions are listed in Weakness. In addition, several issues should be addressed to further improve the quality of the paper. 1. Complete the network architectures and training settings. Some implementation details are missing. 2. How does the same student architecture perform when trained directly on pseudo-targets or distilled from the teacher without PSR?

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

To address the challenge of obtaining paired training data in real-world scenarios, this paper utilizes AIGC to generate multiple candidate pseudo-reference images for model training. Furthermore, a Perturbation-Aware Correction (PAC) module is designed to mitigate output instability stemming from AIGC-based supervisory signals, and knowledge distillation is employed to derive a lightweight student model.

Critical strengths
  1. Rather than directly accepting the output from the generative model, the authors devised a candidate selection strategy to mitigate the risk of artifacts and content drift in the pseudo-reference targets.
  2. The student model, obtained through distillation, contains only 0.289M parameters yet achieves impressive restoration performance.
Critical weaknesses
  1. The method relies on local inconsistencies among multiple candidates to identify unreliable regions; however, since multiple candidates may exhibit the same errors in the same area, this could lead to a systematic bias in the restored results.
  2. The PSR module introduces perturbations at the decoder’s non-linear blocks and the output projection layer; the design rationale for selecting these specific locations for perturbation requires further explanation.
  3. The frozen feature encoder used to measure content discrepancies should be clearly specified. Given that the original images contain degradation-related perturbations, the validity of calculating content discrepancies directly between the degraded original images and the candidate images needs to be justified.
  4. The model is trained and tested exclusively on a self-constructed dataset, lacking validation on publicly available datasets.
  5. Only the parameter count for the student model is provided; the architecture, parameter count, and key hyperparameters of the teacher model are not disclosed.
  6. The training protocol for the baseline models must be clarified to ensure fair comparisons. Additionally, the source of the ground truth (GT) data used by methods such as AdaIR should be specified. If GT data is available, PSNR/SSIM metrics should be included, and experiments on public datasets—such as SIDD and RealSR—should be added.
  7. The XSMax dataset consists of images captured at a single resolution and from a single device; therefore, the model’s performance across different resolutions and devices remains unknown. The current results are insufficient to support the claim of "general" capability as suggested in the title.
Questions for authors

Please refer to Weaknesses.

Comments on the paper

1 comment
Area_Chair_ri2jOpenReview ↗

AC Reminder: Please Follow Up on the Authors’ Responses

Dear Reviewers,

Thank you for your time and thoughtful reviews. We are now about halfway through the discussion period, and the authors have posted responses to your questions and concerns.

If you have not yet done so, please read their responses and any relevant revisions, and leave a follow-up comment visible to the authors indicating which concerns have been addressed and which remain unresolved. Please update your review and rating if warranted. If your assessment remains unchanged, a brief explanation would still be helpful.

Thank you for your continued engagement in the discussion.

Best regards,

Area Chair

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

Related papers

Discussion (0)

Sign in to comment.