acceptodds
Under review as a conference paper at ICLR 2027

From Glance to Scrutiny: Progressive Distortion Reasoning for Fine-Grained Image Quality Assessment

Abstract

Multi-modal large language models (MLLMs) have demonstrated significant potential in image quality assessment (IQA) by bridging visual perception with descriptive evaluations. However, existing approaches mainly focus on holistic quality prediction, often functioning as black boxes that provide limited insight into where distortions occur and how they affect perceived quality, hindering fine-grained analysis of localized and heterogeneous degradations. We propose GS-IQA, a framework that reformulates IQA as a progressive Where–What–How diagnosis, emulating the human perceptual process from an initial glance to closer scrutiny. Since a severity judgment is meaningful only for a correctly localized and recognized region, we realize this progression through a two-stage reinforcement learning paradigm that respects such dependencies: the glance stage uses a perception-gated reward to establish where degradations lie and what they are, activating severity feedback only once both are correct, while the scrutiny stage introduces online reward-conditioned degradation generation to synthesize hard examples targeted at the model's perceptual bottlenecks, sharpening its discrimination of subtle severity variations. To enable systematic evaluation, we construct Diag-Bench, a region-level IQA benchmark of about 25K curated samples spanning 12 distortion types and five ordinal severity levels. Extensive experiments show that GS-IQA consistently surpasses state-of-the-art methods in distortion localization, recognition, and severity estimation, and that its diagnostic representations transfer effectively to conventional global quality prediction across diverse external benchmarks. Code and data will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Quality Assessment, Perception Modeling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_bxZROpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces GS-IQA, a framework for region-level image quality assessment that jointly predicts distortion locations, categories, and ordinal severity levels. Training follows two GRPO stages: a Glance stage with severity rewards conditioned on correct localization and classification, and a Scrutiny stage that adaptively samples distortion categories for online synthesis based on recent rewards. The authors also introduce Diag-Bench, containing approximately 25K images spanning 12 distortion categories and five severity levels. Experiments demonstrate improvements in structured diagnosis, cross-dataset distortion recognition, and holistic quality prediction after additional fine-tuning.

Critical strengths
  1. Jointly predicting region–category–severity triplets addresses limitations of global quality scores and provides concrete outputs that can support diagnostic analysis and potentially guide image restoration.
  2. GS-IQA achieves an F1 of 0.930 and severity accuracy of 0.732 on Diag-Bench, compared with 0.886 and 0.669 for Refine-IQA. The comparison against random online generation is particularly informative: reward-conditioned generation improves severity accuracy from 0.694 to 0.732 while maintaining similar localization performance.
  3. The paper includes task-adapted baselines, reward and generation ablations, external evaluations, and model-scaling experiments. Splitting pristine source images before synthesis is also an appropriate precaution against source-level leakage.
Critical weaknesses
  1. Limited differentiation from closely related work. The paper acknowledges SEAGULL, Refine-IQA, Grounding-IQA, and IQA-Spider, but its novelty is narrower than the broad framing of progressive region-level quality reasoning suggests. In particular, Refine-IQA already combines distortion recognition, severity estimation, grounding, and multi-stage reinforcement fine-tuning. The principal additions appear to be joint multi-region prediction, perception-gated severity rewards, and reward-conditioned online synthesis. Their incremental value should be isolated under matched data.

  2. The total reward can favor omitting difficult regions. Eq.5 averages severity rewards over matched regions, contrary to the claim that this normalization prevents selective omission. Consider two ground-truth regions with perfectly predicted boxes and categories. Suppose one severity prediction is correct and the other differs by at least three levels. With valid formatting, the total reward in Eq.6 is Rcomplete=1+1+1+1+0.5=4.5.R_{\mathrm{complete}}=1+1+1+1+0.5=4.5. If the model omits the difficult region, its F1 reward decreases to 2/32/3, but its severity reward increases to 1: Romit=1+1+23+1+1=143>4.5.R_{\mathrm{omit}}=1+1+\tfrac23+1+1=\tfrac{14}{3}>4.5. Thus, even the F1 penalty does not always offset the incentive to omit a difficult region.

  3. The sampling description is mathematically inconsistent with the reported settings. With rewards and their EMA in [0,1][0,1], temperature τ=1\tau=1, uniform mixture weight λ=0.05\lambda=0.05, and 12 categories, Eq.8 implies $ P'_t(c)\ge 0.95\frac{e^{-1}}{11+e^{-1}}+\frac{0.05}{12} \approx0.0349. $ This assumes the EMA is initialized within the reward range. Appendix A.3.3 nevertheless states that sampling probabilities reach 0.004 and remain there.

  4. Evaluation and training details leave important attribution questions unresolved. For example, The denominators and conditioning rules for ClsAcc and SevAcc are insufficiently specified, particularly the treatment of missed regions and incorrect categories.

Questions for authors

Please see the weaknesses. I would consider raising my rating if the authors address the concerns.

1 comment

Thank you for clarifying the sampling probability bound and evaluation protocol. These address my concerns regarding W3 and W4. I would appreciate further discussion of two remaining points:

  1. For W1, do you have an ablation comparing separate predictions of location, category, and severity followed by aggregation against direct joint triplet prediction, under matched data and training budgets? Since the task-adapted baselines also use joint triplet outputs, this comparison would more directly establish the benefit of the joint formulation itself, beyond the proposed reward and sampling mechanisms.

  2. For W2, one possible explanation for the positive overall results is that a sufficiently strong Glance initialization makes this incentive less consequential. Does the gains remain for smaller VLM backbones? In particular, does subsequent training correct difficult severity predictions, or increasingly omit the corresponding regions? Tracking initially misgraded regions across training (whether they are corrected, remain incorrect, or become missed) could help clarify when this incentive materially limits diagnostic performance. Understanding this boundary would better establish the method’s applicability across model capacities and initialization strengths.

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

The paper proposes a framework called GS-IQA. It changes traditional image quality assessment (IQA) from “giving an image one overall quality score” into a step-by-step diagnosis. It first finds where the problem is, then identifies the type of distortion, and finally judges how severe it is.

Critical strengths

The paper has a clear and interesting idea. Instead of treating image quality assessment as only a score prediction task, it studies it as a “Where–What–How” diagnosis problem. It makes the results easier to understand and more useful for image repair and enhancement.

The paper explains the link between the tasks well. In this paper, the authors point out that severity cannot be judged well without first finding the bad region and identifying the distortion type. This idea is reasonable, and the reward gating design matches it well.

The second-stage training method is also well designed. Instead of only training longer, the model creates hard examples based on the distortion types it currently performs worst on. This helps the model focus on its weak points and is a smarter choice than simply using the same fixed data again.

The benchmark and experiments strengthen the paper. Diag-Bench looks like a useful dataset for future work, and the experiments include main results, ablation studies, cross-dataset testing, and transfer to traditional IQA.

Critical weaknesses
  1. Adapting baselines to the same task is reasonable, but Q-Insight† and VisualQuality-R1† are versions modified by the authors, rather than results reported in the original papers. Are all adaptation details available? Were these baselines carefully tuned, including a proper hyperparameter search, to ensure a fair comparison?
  2. Synthetic samples make up a much larger share of the dataset. The model may partly learn to judge severity from patterns introduced by the synthetic distortion operators. It is unclear how well this transfers to complex, mixed distortions in real images.
  3. The five severity levels may not be directly comparable across distortion types. Even within one type, people may disagree on the boundaries between levels. Mixed distortions in real images may also be difficult to divide into five clear levels. More details on annotation rules and agreement between annotators would help.
  4. The paper lists the backbone, RL settings, and GPU type, but does not report FLOPs, inference latency, memory use, GPU count, or training time. These details are needed to assess the method’s practical cost.
  5. The 12 distortion types may not fully cover mixed distortions, complex artifacts from camera and processing pipelines, artifacts in AI-generated images, or content-dependent quality issues. The paper should discuss these limits and how the model handles unseen distortion types.
  6. Some grammar needs correction, such as changing “MLLMs has” to “MLLMs have.” “Related Work” is the more common section title. Reference formatting should also be consistent, including capitalization, model names, conference names, and paper titles.
  7. The work is novel mainly in its integration and training design, while many core ingredients—region grounding, distortion recognition, severity estimation, and RL-based optimization—are already present in closely related prior work.
Questions for authors

see above comments

Reviewer_w9BAOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper propose a method to decompose the Image Quality Assessment (IQA) task into one that predicts where (e.g., which regions), what (type of artifact), how-much (severity) artifacts are present -- using a training approach that is meant to mimic human assessment by providing a quick overview (“glance”) and then refining that with more “scrutiny”.

The approach uses MLLM to predict the set of regions, type of artifact, and level of distortion (as a set of triplets) in an image, and clearly establishes the reward functions used to penalize differences between elements of these triplets, where first a bipartite matching is performed to match generated vs training lists of triplets that may have different lengths for each image.

A new 25K data set (combining authentic and synthetic distortions) is used for training the static and online phases of the model (using GRPO). The online phase of training is meant to be focus on hard / subtle case that are hard to distinguish (“reward conditioned”).

Experiments demonstrate that the model outperforms alternate models on their dataset, but also that the trained models generalize well in several cross-dataset experiments. The ablation studies sufficiently cover various decision points of the approach (from reward formulation, static to online training)

Critical strengths

The theme of the paper is easy to follow, written well, with a clear narrative, and important aspects of the main contribution being clearly described at sufficient detail (the "where-what-how" of GS-IQA). The reward-driven adaptive sample distribution in the later stage of training is novel for the IQA task.

Experiments are sufficient, model performance is good, and ablation studies cover the most meaningful dimensions.

And many open questions I had, e.g., on data set construction, were easy to track down in the appendix.

Critical weaknesses

The synthetic degradations applied to "regions" seems very unrealistic (Practically, regions and distortions will much more non-linear, content-dependent, and the perceptual "severity" may not align with a predefined synthetic parameter). That said, the experiments do seem to indicate that the method works well in practice (e.g., Table 4 is on holistic IQA). The authors could provide more insight on what specific part of their base-model training (the initial dataset, or the "glance" or "scrutiny" phases) are best at aiding the authentic task.

Questions for authors

349: “Human experts discard perceptually ambiguous candidates?” What does this mean? Is this based on content being ambiguous (e.g., black screen) or the artifacts degradations being many and ambiguous or that raters disagree?

Some other feedback (not going to affect my assessment):

  • Table 5. What is meaning of bold as in many columns the best row is not the cell that is bolded (esp the top table, except for the SevAcc column).
  • Surely there is a better reference for OpenCV than “G. Bradski. The OpenCV Library. Dr. Dobb’s Journal of Software Tools, 2000.”
  • Perhaps consider moving the "Overall Multi-Task Reward" at the beginning of section 3.1? As it stands, the reader encounters " We set R_fmt = 1" on line 204, without knowing exactly what R_{fmt} is or how it is going to be used (Although, in the context of the section it is in, it is not that hard to infer what it is going to be). Having the overall reward listed up front gives the reader an expectation of what is to come.

Comments on the paper

1 comment

Dear Reviewers,

As the discussion period is coming to a close, I would like to encourage everyone to review the authors' responses and engage in any remaining discussions. Please indicate whether your concerns have been adequately addressed and update your scores if your assessment has changed.

Your active participation is greatly appreciated and will be valuable in reaching a fair and well-informed decision.

Thank you for your time and effort!

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.