acceptodds
Under review as a conference paper at ICLR 2027

Attribute-Calibrated and Quality-Aware Vision-Language Learning for Image Quality Assessment

Abstract

Existing image quality assessment (IQA) methods primarily rely on score regression paradigms while neglecting semantic analysis, making score-only supervision unreliable across diverse scenarios. Additionally, the relationship between quality-aware attributes and cross-scenario quality fluctuations remains underexplored. To address these issues, an image quality assessment method based on attribute-calibrated and quality-aware vision-language learning (ACQA) is proposed. Firstly, a dual-branch vision-language model (VLM) is employed to extract content and attribute features for subsequent quality analysis. Secondly, an attribute calibration learning (ACAL) module is proposed to calibrate the VLM in attribute analysis by jointly incorporating distortion ranking, stability, and disentanglement, thereby enabling adaptive attribute information across scenarios. Thirdly, a quality-aware analysis learning (QAL) module is proposed to integrate attribute contributions with content alignment to optimize the VLM. Finally, the optimized VLM provides reliable content and attribute information for quality prediction. Experimental results demonstrate that the proposed ACAL and QAL can effectively refine content and attribute information to enhance the reliability of quality prediction. Moreover, ACQA outperforms existing IQA methods in both accuracy and generalization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Quality Assessment, Semantic Analysis, Quality-Aware, Attribute-Calibrated, Vision-Language

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6qNSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ACQA, a vision-language based framework for image quality assessment. The main idea is to use both semantic content information and quality-related attributes rather than relying only on score regression. ACQA contains a dual-branch VLM, an Attribute Calibration Learning module, and a Quality-Aware Analysis Learning module. The content branch generates image descriptions, while the attribute branch predicts several quality-related attributes. ACAL uses ranking, stability, and disentanglement rewards to improve attribute prediction. QAL further uses content alignment and attribute contribution information to optimize the VLM for quality prediction. Experiments are conducted on synthetic, authentic, and AI-generated image quality datasets.

Critical strengths

The paper considers both semantic content and quality-related attributes rather than directly predicting a quality score. This is a reasonable direction for VLM-based IQA, where semantic information may help explain quality variations across different image domains. The experimental results are strong across several types of IQA datasets. ACQA performs well on authentic, synthetic, and AI-generated images, and the ablation studies show that both the dual-branch design and the proposed ACAL/QAL components contribute to the final performance. The paper also provides relatively detailed component-level analysis. In particular, the effects of the different rewards and the changes in content and attribute information are examined separately.

Critical weaknesses

My main concern is the evaluation protocol. Some baseline results are directly taken from their original papers, but the paper does not clearly establish that all methods use the same data splits and evaluation settings. This makes the performance comparison difficult to interpret. The paper repeatedly claims improved generalization, but most reported results appear to be within-dataset evaluations. A cross-dataset experiment would be more convincing evidence for generalization. The framework is also quite complex. ACAL and QAL introduce several rewards and multiple optimization steps. Although the ablation study shows that most components help, it is still unclear whether the additional complexity is justified by the improvement over simpler alternatives. Finally, the five attributes are manually fixed to contrast, brightness, realism, saturation, and blur. These attributes may not be equally meaningful for all image domains, and some important quality factors are not covered.

Questions for authors
  1. Could you clarify the exact training/test split used for each dataset? Are the reported results averaged over multiple random splits or runs?
  2. Some baseline results in Table 1 are taken directly from the original papers. Are these results based on exactly the same dataset splits and preprocessing as ACQA? If not, could you provide a controlled comparison under the same protocol?
  3. The paper claims improved generalization, but most experiments appear to evaluate each dataset separately. Could you provide a cross-dataset experiment, for example training on one IQA dataset and testing directly on another?
  4. Why are contrast, brightness, realism, saturation, and blur selected as the five attributes? How sensitive is the method to this choice, especially since “realism” may have different meanings for natural and AI-generated images?
  5. The ranking reward assumes an exponential quality decay across synthetic distortion levels. How sensitive are the results to this assumption and to the attenuation coefficient α\alpha?
  6. The gain reward uses changes in SRCC and PLCC, while the IQA model is also trained with MOS labels. Could you clarify exactly which split is used to compute these rewards and how you avoid using validation or test information during RL optimization?
  7. ACQA contains several RL optimization stages and repeated VLM inference. What are the total training cost and inference cost compared with recent VLM-based IQA baselines?
  8. All main experiments use Qwen 2.5-7B as the VLM backbone. Does the improvement remain with other VLM backbones, or is the method sensitive to the choice of the base model?
1 comment
AuthorsOpenReview ↗

Response to Reviewer 6qNS

Dear Reviewer 6qNS,

Thank you for your thoughtful review. In our response, we use 'W' to denote weaknesses and 'Q' for questions. Additionally, we have included extensive experimental results and explanations in the Supplementary Material (SM) to support our responses.

W1. Thank you for your efforts. In response to your comments, we have provided explanations in Q1–Q8 below. In addition, we have supplemented the SM with additional experiments as supporting evidence. We hope these explanations will meet with your approval.

Q1 & Q2. Firstly, we split the training, validation and test sets with a 6:2:2 ratio for each dataset. All results are averaged over ten random splits. However, the difference is that some methods (Q-Align, Compare2Score, Refine-IQA-S2) were trained on large-scale mixed datasets or self-constructed datasets to obtain their pre-trained parameters, which affects the comparison results under the same protocol. Secondly, the experimental results in S_Table 4 of the SM are illustrated to fairly compare ACQA with other methods on the BID, KonIQ-10k, and SPAQ datasets. All methods were trained with the official VLM and the same data-split protocol. ACQA performers well on individual datasets, and the average performance surpasses other methods by at least 2.5% in PLCC.

Q3. The experimental results in S_Table 8 of the SM are illustrated to analyze the performance of ACQA in cross-dataset tasks. The average performance of ACQA surpasses other methods by at least 2.5% in SRCC and 1.3% in PLCC.

Q4. Firstly, during preliminary research, we observed that image distortions across different scenarios are primarily correlated with five selected attributes, which has also been corroborated by methods such as IQA-T1 and ATTIQA, and the reference information is in SM. Secondly, the natural images are mainly degraded by processing-induced unnaturalness. And, the AI-generated images are degraded by implausible textures, illumination, and structure. In our preliminary study, we used a VLM to rate the realism attribute on the SPAQ dataset (over 10,000 natural images) and the AGIQA-20k dataset (over 20,000 images). Researchers observed and validated these ratings, confirming that the VLM’s judgments on realism were largely correct. Finally, the experimental results in S_Table 1 of the SM are illustrated to demonstrate the advantage of five selected attributes over other attributes. With the selected five attributes, ACQA achieves a performance improvement of at least 6.97% on LIVEC and 3.40% on AGIQA-3k.

Q5. We sincerely thank the reviewer for raising this question. Through preliminary experiments, we determined via sensitivity analysis that the optimal value of α is 0.3. The experimental results in S_Table 9 of the SM are illustrated to analyze the sensitivity on α across multiple distorted datasets and specific distortion types. Experimental results demonstrate that setting α to 0.3 ensures the validity of the ranking discrepancy computation with the performance improved by at least 5.44%.

Q6. The gain reward and other rewards are computed on the training and validation sets. We strictly followed a 6:2:2 training/validation/test split, and the experiments across multiple rounds, different random splits were kept independent of one another to avoid data confusion.

Q7. Firstly, S_Figure 1 of the SM has been added to analyze the performance gains and parameter increments of employing different reward mechanisms. Secondly, the experimental results in S_Table 10 of the SM are illustrated to compare the inference costs of ACQA with VLM-based IQA methods. ACQA achieve performance improvement by at least 5.21% at only a small inference cost.

Q8. The experimental results in S_Table 3 of the SM are illustrated to demonstrate the adaptability of ACQA on different VLM backbones (Gemma3-4B, mPLUG-Owl-2, Qwen2-7B, and Qwen2.5-3B). ACQA maintains stable and accurate performance across different VLM models.

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

This paper proposes ACQA, a BIQA framework based on VLM. It models the image content and quality attributes through a dual-branch structure, and designs ACAL and QAL modules to optimize attribute perception and quality prediction using RL. The paper attempts to address the issue of varying contribution of quality attributes in different scenarios, and its research motivation is reasonable.

Critical strengths
  1. The paper attempts to combine content semantic analysis with distortion attribute modeling, and achieve attribute calibration through RL. The overall research approach is reasonably sound.
  2. The figs in paper are beautiful and can clearly illustrate the main modules and their optimization processes. The paper also provides ablation experiments and qualitative analyses for the corresponding modules.
  3. Good to follow the story.
Critical weaknesses
  1. The reviewer finds that, the definition of attributes and the supervision mechanism are not clear. In method section, ACAL only simulates four types of distortions, but involves five quality attributes, and particularly lacks an explanation of the source of supervision for the "realism" attribute. In Fig 2, the attribute scores of different levels are all marked as "excellent", which is inconsistent with the definition in the text.
  2. The effectiveness of distortion supervision is questionable. For natural distortion datasets such as LIVEC and BID, which do not have native distortion level labels, did ACAL construct the sorting supervision through artificial distortion synthesis? This paper does not provide any explanation for this..
  3. The experiments are insufficient to support the SOTA and generalization conclusions. Table 1 only reports the complete results of four datasets, lacking the complete evaluations of KADID-10k, BID, as well as the experiments and cross-dataset generalization experiments on common datasets such as KonIQ-10k and SPAQ. Moreover, no comparisons have been made with related or representative methods such as LIQE, Compare2Score, and DeQA-Score, thus the claim of SOTA lacks credibility.
  4. The rationality of the method design is not yet fully adequate. ACAL and QAL introduce various reward mechanisms, but there are doubts regarding the optimization direction of the reward function: the KL divergence, variance, and deviation in Eq. 5, 6 and 9 usually need to be minimized, while the optimization direction and notation of PPO in formulas (7) - (8) are not clearly defined.
  5. The content alignment design and the calculation cost lack justification. The IQA model takes image patches as input, while the VLM takes the entire image as input. This results in the semantic scopes of f^I and f^T possibly being inconsistent, thereby affecting the reliability of R_car. Moreover, this method relies on multi-stage iterative training and repeated reasoning, but lacks analysis of training time, inference cost, and performance gain.
Questions for authors

See Weaknesses. The reviewer's rating of 2 is mainly based on these. When with satisfactory rebuttal, the reviewer is happy to adjust the score.

Also better could provide a comparison of other representative methods, complete results for KADID-10k, BID, KONIQ and SPAQ, cross-dataset experiments, as well as analyses of training and inference costs.

1 comment
AuthorsOpenReview ↗

Response to Reviewer inWb

Dear Reviewer inWb,

We sincerely appreciate your efforts and recognition of our work. In our response, we use 'W' to denote weaknesses and 'Q' for questions. Additionally, we have included extensive experimental results and explanations in the Supplementary Material (SM) to support our responses.

W1. Firstly, the four types of simulated distortions can be achieved by utilizing existing synthetic distortion strategy. However, the realism attribute relies on subjective judgment in current research, lacking corresponding distortion simulation. Secondly, in our preliminary study, we used the VLM to rate the realism attribute on the SPAQ dataset, which contains more than 10,000 natural images, and the AGIQA-20k dataset, which contains more than 20,000 AI-generated images. Researchers observed and validated and confirmed that judgments of the VLM on the realism attribute were largely correct. Therefore, realism attribute was included in the final selection. Thirdly, the experimental results in S_Table 1 of the SM are illustrated to analyze the effectiveness of realism attribute. Specifically, scheme #4 and scheme #9 of S_Table 1 show ACQA achieves a performance improvement of at least 6.97% on LIVEC and 3.40% on AGIQA-3k with realism attribute. Fourthly, we will explore more attribute supervision strategy in In further work. And, the attribute levels in Figure 2 have been adjusted as suggested.

W2. Firstly, t he synthetic distortions are inherently designed to simulate authentic distortions in IQA tasks. Secondly, methods such as DGQA (Li et al. (2024a)) and Q-Insight (Li et al. (2025)) have demonstrated the rationale of transferring from synthetic to authentic distortions. Therefore, we constructed ranking supervision for authentically distorted datasets like LIVEC and BID with synthetic distortions. Thirdly, the experimental results in S_Table 5 of the SM are illustrated to demonstrate the effectiveness of leveraging synthetic distortions for ranking supervision on LIVEC and BID

W3. Due to time constraints, we selected some methods for comparison on the KADID-10k, BID, and SPAQ datasets in S_Table 4 of the SM. Additionally, we have introduced extra baselines, including LIQE, Compare2Score, and DeQA-Score. We will supplement more comparison experiments in further work.

W4. Firstly, on the optimization direction of the reward functions. Each reward term is first converted into a common higher-is-better objective, denoted as m. The m is set to −R for the three minimization terms in Eqs. (5), (6), and (9). The QAL terms in Eqs. (10), (13), and (14) are already oriented in the higher-is-better direction and therefore used as-is. Each m is then converted into a group-normalized advantage, with statistics computed within the RL batch, and this advantage is the quantity fed into Eq. (7).

Secondly, on the PPO formulation in Eqs. (7) and (8). Eq. (7) is the standard clipped PPO surrogate objective, which is maximized; the probability ratio r_i is defined as the new-policy output probability divided by the old-policy output probability, and the clip operation constrains r_{i} within [1−ε, 1+ε]. Eq. (8) denotes gradient descent on the total loss, that is, the negated clipped objective plus a β-weighted KL penalty between the old and updated policies.

Thirdly, this convention is consistent with the reported behavior. The ACAL reward curves in Figure 3 (a), which decrease toward a stable baseline, are exactly the expected behavior of the minimization-side quantities. We have additionally summarized the parameters optimized by each reward mechanism in S_Table 6 of the Supplementary Material. In the revised version, we will incorporate the above formulation into Sec. 3.2.

W5. Firstly, images are resized to 256×256 and then randomly cropped to obtain 224×224 patches as input for the IQA model. This preprocessing strategy ensures that substantial semantic information and salient features of images are preserved. Secondly, random cropping is frequently employed in IQA methods to enhance model robustness. Thirdly, the experimental results in S_Table 5 of the SM are illustrated to demonstrate the effectiveness of patches input in ACQA. Finally, S_Figure 1 of the SM has been added to analyze the performance gains and parameter increments of employing different reward mechanisms.

Q1. We appreciate your positive response and valuable suggestions. We have responded to all of your comments and added supporting experiments in the SM. We hope these efforts will meet with your approval.

Reviewer_BWd8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ACQA, a framework for image quality assessment (IQA) that integrates a dual-branch vision-language model (VLM) with attribute-calibrated learning (ACAL) and quality-aware analysis (QAL). The approach aims to optimize attribute perception and content analysis by leveraging rewards based on ranking, stability, disentanglement, content alignment, and the contribution and gain of specific attributes.

Critical strengths
  1. The paper points out that evaluation strategies based on fixed attributes struggle to adapt to shifts in human subjective perception across different contexts; it further explores the relationship between these attributes and fluctuations in quality. This motivation offers greater value for semantic analysis than simple score regression.
  2. The paper covers datasets including KADID, TID2013, LIVEC, BID, AGIQA-3k, and AGIQA-20k, encompassing synthetic distortions, authentic distortions, and AIGC scenarios. Table 1 also provides a comparison of hand-crafted, classical deep learning-based, and VLM-based IQA (Image Quality Assessment) methods.
Critical weaknesses
  1. The paper critiques existing "fixed attribute evaluation strategies" for their inability to adapt to changing scenarios; however, ACQA itself relies on a fixed set of five attributes: contrast, brightness, realism, saturation, and blur. While ACAL employs reward calibration for attribute perception, the attribute set remains static. Are these five attributes sufficient for diverse contexts—such as AIGC, real-world distortions, or synthetic distortions? The authors should, at the very least, discuss the rationale behind the selection of these attributes or include an ablation study exploring the replacement or adaptive selection of the attribute set.
  2. In Table 2, Method #1 utilizes the image encoder from MIQA for quality prediction, whereas Methods #2, #3, and #4 employ a dual-branch VLM. Consequently, this does not constitute a rigorous ablation study under a consistent backbone or training protocol. The observed performance gains cannot be attributed solely to the "dual-branch" architecture; they may also be influenced by the semantic priors inherent to the VLMs or differences in parameter counts.
  3. ACQA is based on Qwen2.5-7B, while some VLM-based baselines may utilize different VLMs or models of varying scales. Table 1 presents a direct comparison of methods with different backbones; it is necessary to clarify whether these models share the same training data, input resolution, and pre-trained weights.
Questions for authors
  1. Include fair comparisons with VLM-based IQA baselines that utilize the same backbone/protocol, and report full results on the KADID and BID datasets.
  2. Clarify the relationship between the fixed set of five attributes and the "cross-scenario attribute calibration" process; additionally, provide ablation studies on the attribute sets and synthetic distortion types.
3 comments
AuthorsOpenReview ↗

Response to Reviewer BWd8

Dear Reviewer BWd8,

Thank you for your thoughtful review. In our response, we use 'W' to denote weaknesses and 'Q' for questions. Additionally, we have included extensive experimental results and explanations in the Supplementary Material (SM) to support our responses.

W1 & Q2. During preliminary research, we observed that image distortions across different scenarios are primarily correlated with five selected attributes, which has also been corroborated by methods such as IQA-T1 and ATTIQA. Meanwhile, the impact of five attributes on image quality fluctuations varies as the scene changes. However, existing methods mostly rely on fixed attribute weights for quality prediction. To address this, ACQA employs the contribution reward and gain reward within the QAL module to adaptively adjust attribute scores across different scenarios. To demonstrate the effectiveness of five selected attributes, we conducted ablation studies incorporating attributes such as sharpness and structural integrity in S_Table 1 of the SM. The experimental results in S_Table 1 demonstrate the effectiveness of five selected attributes. In futher work, we will explore the generalization and adaptive selection of more attributes.

W2. In Table 2, Scheme #1 is designed to demonstrate the effectiveness of the content and attribute information provided by the dual-branch VLM. And, the experimental results demonstrate that the IQA model achieves a performance improvement of at least 2.03% with the incorporation of content and attributes. We will refine this statement in the revised manuscript. Furthermore, to investigate the impact of parameter differences in the dual-branch VLM, we conducted ablation studies without updating the VLM by employing reinforcement learning. Experimental results and analysis are illustrated in S_Table 2 of the SM.

W3 & Q1. The VLM-based methods cited in Table 1 all adopt Qwen2.5-7B as the backbone, sharing the same input resolution and pre-trained weights as ACQA. Only Q-Align utilizes mPLUG-Owl-2 as its backbone. To demonstrate the effectiveness of our framework on mPLUG-Owl-2 and more VLM models, we conducted the comparison experiments presented in S_Table 3 of the SM. Moreover, we have added comparison experiments on the KADID and BID datasets in S_Table 4 of the SM. All compared methods were evaluated under a 6:2:2 training/validation/test split, with Qwen2.5-7B serving as the backbone VLM.

Thank you for the response and the additional tables. The paper criticises fixed attribute strategies, yet ACQA still uses a fixed set of five attributes and only adapts their contribution. This should be stated plainly, and Fig. 4 should be quantified across datasets. In S_Table 1, "sharpness" largely duplicates "clarity/blur", so it does not test the attribute set independently; a leave-one-out or a genuinely different attribute would be more convincing.

AuthorsOpenReview ↗

Response to Reviewer BWd8

Dear Reviewer BWd8,

Thanks for your suggestions. Firstly, we will clarify that ACQA performs contribution calibration over the fixed set of five attributes in the final manuscript. Secondly, Fig. 4 presents the quantified contributions of the different attributes on the LIVEC, AGIQA-3k, BID, and KADID datasets. In the final manuscript, we will further show the attribute contributions on more different types of images. Thirdly, we analyzed the impact of each individual attribute on performance in our work, The experimental results in table below are illustrated to analyze the impact of single attribute to ACQA. The scheme#1 to #7 show the performance of using a single attribute. Compared with using a single attribute, the five selected attributes improved performance by at least 7.8%. Finally, we will explore more attributes to improve the inference of IQA tasks. We hope these explanations can meet your recognition.

#ContrastRealismClarityBrightnessSaturationSharpnessStructuralLIVECAGIQA-3k
1✓✗✗✗✗✗✗0.7990.798
2✗✓✗✗✗✗✗0.7260.817
3✗✗✓✗✗✗✗0.7820.775
4✗✗✗✓✗✗✗0.7710.766
5✗✗✗✗✓✗✗0.7780.793
6✗✗✗✗✗✓✗0.7830.811
7✗✗✗✗✗✗✓0.7760.809
8✓✓✓✓✓✗✗0.8900.881

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.