acceptodds
Withdrawn from ICLR 2027

Learning What to Degrade: A Recoverability-Aware Curriculum for Reliable Smart-Glasses Super-Resolution

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Images captured by smart glasses often suffer from complex degradations, obscuring small text and fine-grained structures essential for visual assistance. Although super-resolution (SR) can improve perceptual quality, these gains do not necessarily translate into reliable downstream understanding by vision-language models (VLMs). We introduce a high-resolution (HR)-referenced semantic evaluation protocol that compares VLM responses to restored images with those obtained from corresponding HR references. The evaluation reveals two failure modes: generative SR can introduce misleading details, while restoration-oriented SR can leave critical details blurred or ambiguous. To improve detail recovery in restoration-oriented SR, we investigate adaptive degradation scheduling. High reconstruction error may reflect either insufficient model capability or excessive information loss, requiring different adjustments to degradation severity. We therefore propose ReDiC, a Recoverability–Difficulty Curriculum that jointly estimates input recoverability and model-relative restoration difficulty to adaptively control degradation severity during training. The scheduler increases severity under easy and recoverable conditions, maintains challenging but recoverable conditions, and reduces severity when estimated recoverability is low. ReDiC modifies only degradation generation during training and is compatible with existing SR architectures. Experiments on synthetic degradation benchmarks show consistent improvements across NAFNet, SwinIR, and MambaIR, yielding average PSNR gains of 2.36–2.65 dB over their corresponding baselines while improving downstream recognition and HR-referenced semantic agreement. Both automatic and human evaluations on real smart-glasses captures show improved downstream semantic agreement across all three backbones.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Super-Resolution, Image Restoration, Curriculum Learning, Adaptive Degradation, Recoverability-Aware Learning, Semantic Reliability

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_NNtHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ReDiC, a recoverability-aware degradation curriculum for smart-glasses super-resolution. ReDiC estimates degradation recoverability using bicubic reconstruction error and model-relative difficulty using normalized SR reconstruction error. These signals dynamically adjust training degradation severity without modifying the SR architecture. Experiments across three SR backbones demonstrate improvements in reconstruction quality, OCR performance, and VLM-based semantic agreement, including evaluations on real smart-glasses images. Presentation: 3/3

Critical strengths
  1. Practical and well-motivated framework. ReDiC provides a simple way to adapt training degradation severity and can be integrated into existing SR architectures without additional inference overhead.
  2. Consistent improvements across multiple backbones. ReDiC achieves substantial PSNR improvements of 2.36–2.65 dB across NAFNet, SwinIR, and MambaIR, alongside improvements in OCR and semantic agreement. The real-world evaluations and human study further support its practical relevance.
Critical weaknesses
  1. Insufficient analysis of hyperparameter sensitivity. ReDiC involves several important hyperparameters, including the recoverability calibration bounds, scheduling thresholds, severity update step, quantile settings, and EMA coefficient. These parameters directly affect the estimated learning state and degradation schedule. However, the paper lacks systematic sensitivity analyses, and some parameter settings are insufficiently specified. It remains unclear how robust the reported improvements are to these choices.

  2. Insufficient comparisons with alternative degradation curricula. The experiments mainly compare ReDiC against standard training and single-signal variants. Important alternatives, such as fixed-severity training, random-severity sampling, and conventional easy-to-hard curricula, are missing. Therefore, it remains unclear whether the improvements arise specifically from the proposed recoverability–difficulty feedback or simply from a more favorable training degradation schedule.

  3. Limited evaluation under unseen real-world degradations. The main quantitative experiments use synthetic degradations generated from the same degradation family as training. Although real smart-glasses captures are included, the evaluation on these images relies primarily on semantic agreement scores without quantitative restoration metrics. Additional experiments on independent real-world SR benchmarks or unseen degradation distributions would provide stronger evidence of generalization.

Questions for authors
  1. Hyperparameter sensitivity: How sensitive is ReDiC to the recoverability calibration bounds, scheduling thresholds, and severity update step? A sensitivity analysis would help establish the robustness of the proposed scheduler.
  2. Curriculum comparisons: How does ReDiC compare with fixed-severity, random-severity, and conventional easy-to-hard curricula under identical training budgets?
  3. Generalization: Could the authors provide additional evaluations on unseen degradation distributions or independent real-world SR benchmarks to establish the robustness of the proposed approach?
Reviewer_k7itOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ReDiC, a degradation training curriculum for smart glass super-resolution. Bicubic reconstruction error serves as a recoverability proxy, while SR reconstruction error measures model-relative ability. These estimated signals jointly determine whether degradation severity should increase, decrease, or remain unchanged. Training with ReDiC, three SR backbones, NAFNet, SwinIR, and MambaIR, improve in evaluations of reconstruction fidelity, OCR, and semantic descriptions.

Critical strengths
  1. A practical training intervention for a relevant downstream problem. Improving the usefulness of SR images for downstream visual understanding is a meaningful motivation, particularly for small text and fine structures. ReDiC offers a simple way to modify supervised SR training while preserving the backbone.

  2. Consistent and substantial improvements across multiple backbones and tasks. The fidelity evaluation is conducted on three backbones, accompanied by OCR and semantic description. Evaluation on 200 real smart-glasses captures extends the evidence beyond synthetic degradation, and the human ratings provide additional support for the reported agreement gains. The R-only and D-only ablations also support the benefit of combining two degradation signals.

Critical weaknesses
  1. The pivotal recoverability proxy and its empirical calibration are inadequately investigated. ReDiC employs bicubic upsampling as a proxy to identify the recoverability of input images. However, poor bicubic reconstruction does not provide concrete evidence for the recoverability of input images. The paper does not validate whether this proxy reliably supports the distinction between insufficient model capability and excessive information loss. Moreover, the empirical calibration bounds are provided without specifications or analysis, despite their influence on whether the curriculum increases, maintains, or reduces degradation.

  2. The connection between the semantic-reliability motivation and the training mechanism remains insufficiently established. ReDiC’s supervision is entirely based on pixel errors, without explicitly accounting for semantically important details or using semantic consistency to guide training. Although the results demonstrate improved downstream performance in the evaluation, they do not sufficiently distinguish gains arising from better general reconstruction quality from any additional benefit specifically associated with preserving or recovering semantic information.

Questions for authors
  1. Please provide the numerical values of elowe^{\mathrm{low}} and ehighe^{\mathrm{high}}, and explain how you determined them and what evidence you used. A focused analysis would help establish how these choices affect the curriculum and final SR performance. It would also strengthen the method to validate the interpretation of bicubic reconstruction error as a recoverability proxy.

  2. Another concern is whether the semantic gains substantiate the broader reliability interpretation. Please clarify whether the semantic contribution is primarily an evaluation perspective and an empirical downstream benefit, or whether ReDiC is intended to specifically preserve semantic information. For the latter claim, a comparison at similar reconstruction quality would help assess whether the curriculum provides additional semantic benefits beyond general restoration improvements.

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

This paper first proposes a VLM-based semantic quality assessment scoring method and claim that "generative SR can introduce misleading details, while restoration-oriented SR can leave critical details blurred or ambiguous". From this, they develop a way to improve the performance of restoration-oriented SR by changing the dataset - scheduling an adaptive degradation within the dataset.

Critical strengths
  1. Practical, architecture-agnostic intervention with consistent gains. ReDiC changes only the training-time degradation schedule, yet improves PSNR, LPIPS, OCR-NED and judge scores across three different backbones (NAFNet, SwinIR, MambaIR). This matters because it is a plug-in change with essentially zero inference cost. If the gains survive proper controls (see weaknesses), it is easy to adopt.

  2. Evaluation on real smart-glasses captures with human validation. The authors collect 200 real captures and run a 20-participant human study whose ranking agrees with the LLM-judge protocol (Spearman 0.81). This matters because it grounds the semantic-reliability framing in a realistic deployment setting, rather than only synthetic benchmarks.

Critical weaknesses
  1. The two parts (observation and the method) seem separate. It is also possible to remove hallucinations from generative SR models to get better semantic score boost, but this is not done in this paper.

  2. The ablative study don't give us a full picture of whether the scheduler part is the crucial part or not. Only R (checked) D (checked) Scheduler (x) will give us a sense of the importance of the Scheduler part.

  3. There are prior works on semantic quality evaluation methods, and faithful comparison to show the goodness of the proposed semantic score is not conducted.

Questions for authors

See weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.