acceptodds
Under review as a conference paper at ICLR 2027

Quality Has a Direction: Learning a Shared Quality Axis from Reasoning for Image Quality Assessment

Abstract

Reinforcement-learning-based vision–language models generate quality reasoning that provides effective supervision for generalized image quality assessment, yet the resulting text embeddings entangle perceptual quality with scene content and linguistic realization. We investigate whether transferable quality can be localized within this entangled space. We find that MOS-conditioned variation in reasoning text forms a semantically coherent quality axis despite contextual and lexical ambiguity. Analyses of contrastive training dynamics and controlled domain shifts further reveal that the quality axis is shared by language and vision across datasets. Building on this finding, we propose QAxis, which introduces the shared quality axis into contrastive learning. Symmetric RoPE makes cross-modal similarity jointly depend on semantic correspondence in the complete representations and agreement along their quality coordinates, while axis-wise ordinal learning establishes a global perceptual order. Using only a 0.1B-parameter image encoder, direct axis projection achieves state-of-the-art average PLCC/SRCC of 0.811/0.796 across seven benchmarks; incorporating the complete image representation further improves the averages to 0.816/0.800.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
image quality assessment, quality axis, Contrastive Learning

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

This paper proposes QAxis for image quality assessment,a simple and interesting IQA pipeline.

Critical strengths
  1. QAxis provides a simple and interesting IQA inference pipeline without online language generation.

  2. The component ablations compare description consensus, rotation, and ranking individually and in combination. Within the reported setup, their combination performs better than either rotation or ranking alone.

  3. Under the KonIQ-to-target protocol, QAxis improves the average PLCC/SRCC over RALI and performs better on several evaluated datasets.

Critical weaknesses

My main concern is the gap between the central interpretation and the evidence supporting it. The reported improvements suggest a useful training approach, but the current experiments leave important alternative explanations unresolved. If more detailed experiments and relevant descriptions are provided in the rebuttal stage, I will raise the score to a positive score.

  1. I am not yet convinced that the chosen text direction is special. The axis is found using MOS labels, and the ranking loss then trains image features to follow those labels along the axis. Since the visual encoder and projection layer are trainable, even a fixed random direction might become a useful scoring coordinate. Figure 7 shows that the trained coordinate predicts MOS well, but does not rule out this simpler explanation. Without random-axis, shuffled-label, and ordinary scoring-head controls, I find it hard to attribute the gains to a pre-existing quality structure in the reasoning embeddings.

  2. The claim that the axis is shared across datasets seems stronger than the evidence. Every reported axis is learned from KonIQ. The target-dataset scores show that a KonIQ-derived scorer can transfer, but do not show that comparable axes can be recovered from other domains or with less labeled data. Figure 4(b) provides an interesting gradient analysis on KADID, though its fixed distortion setting does not necessarily fix perceived quality across different image contents. I would describe the current result as transfer of one source-trained direction, with the broader shared-axis claim still open.

  3. The experiments do not cleanly separate QAxis from the supervision it inherits. The method learns from Q-Insight descriptions, while RALI already uses reasoning descriptions to train a visual scorer. I would find the improvement over RALI more convincing with the same teacher checkpoint, descriptions, backbone, and training budget. The comparison with Q-Insight is useful for scoring accuracy, but QAxis does not provide Q-Insight’s explanations, and the cost of generating its training descriptions is not reported alongside the inference cost.

  4. Lack of some more intuitive qualitative analysis, such as the CLIP's image-text similarity heatmap brought by the quality axis, whether it focuses more on the areas that better reflect "quality" or on the image foreground.

Questions for authors
  1. A comparison with several fixed random axes, a shuffled-label SIR axis, and a learnable scoring head under matched supervision and training conditions would help establish the specific contribution of the proposed direction.
  2. It is suggested to incorporate the axis results estimated from the synthetic distortion dataset and a smaller training subset. This will help clarify its dependence on KonIQ and the amount of labeled data, as well as whether the axis is domain-specific to the training domain or is common to any quality evaluation domain.
  3. Attention or attribution visualizations for ordinary alignment, QAxis, and a random-axis variant on the same images would make the proposed interpretation easier to assess.
4 comments

Thank you for your detailed review and constructive suggestions. Below, we address your concerns about the axis controls, cross-dataset transferability, inherited supervision, computational cost, and spatial responses.


Comment 1: Random-axis, shuffled-MOS, and learned-head controls (Sec. 4.3, “Effect of the axis choice” ).

Response 1: Thank you for this important observation. We have added multiple random and shuffled-MOS axes , and a learnable linear head trained with InfoNCE and the same ranking loss.

MethodAvg.
Random-v10.713/0.710
Random-v20.740/0.718
Random-v30.713/0.705
Shuffled-v10.737/0.716
Shuffled-v20.753/0.733
Linear head + ranking0.759/0.742
QAxis-v10.811/0.796

Random and shuffled-MOS axes are consistently less effective than our quality axis. Despite receiving ranking supervision, the learnable linear head also performs substantially worse than QAxis-v1. Ranking constrains relative scores but does not itself identify a quality-related direction in the representation. Our quality axis fixes a semantics-grounded coordinate, allowing ranking supervision to reinforce quality information from reasoning embeddings rather than learn a scoring direction from MOS ordering alone.


Comment 2: Cross-dataset transferability and reduced axis-estimation data.

Response 2: Thank you for this insightful comment. We find evidence supporting cross-dataset transferability. We have revised the manuscript accordingly.

(1) Independent cross-dataset estimation (Section 2.2, Q3).

We independently estimate quality axes from seven datasets. Their directions differ, although their projections remain strongly correlated. For KonIQ–KADID, the first five principal components account for 91.7% of the projection covariance; removing them reduces the projection correlation from 0.859 to 0.479. Projection agreement therefore depends on feature covariance and does not establish identical axis vectors. Accordingly, we revise the corresponding claims to emphasize that QAxis is transferable across datasets.

(2) Reduced axis-estimation data (Section 4.3, “Effect of axis-estimation sample size” ).

We evaluate axes estimated from 20% to 100% of the KonIQ training samples. Using only 20% retains strong cross-dataset performance, with average PLCC/SRCC of 0.805/0.786, supporting robustness to reduced data for axis estimation.


Comment 3: Inherited supervision and computational cost.

Response 3: Thank you for this suggestion. We address the two comparisons separately.

(1) QAxis versus RALI: controlling inherited supervision. (Appendix E.1)

We reproduced RALI using the same settings as our method.

MethodAvg.
RALI0.789/0.771
QAxis-v10.811/0.796

QAxis extracts a transferable quality axis from these descriptions to guide quality-conditioned alignment and ordinal learning, providing an explicit quality reference that enables more effective transfer of quality information from the same supervision.

(2) QAxis versus Q-Insight: computational cost (Section 4.5, “Efficiency Study” ).

QAxis transfers quality information from reasoning descriptions to a lightweight IQA model, providing an explicit quality direction for visual learning and enabling quality prediction without online text generation. We have added a cost analysis separating offline description generation, QAxis training, and inference.

OperationThroughputTime
TrainingReason generation5.75 reason/s8.17h
TrainingQAxis training—2.94h
InferenceQ-Insight5.75 images/s350s
InferenceQAxis317.7images/s7s

Reason generation is an offline cost, and the reasons can be reused across training runs. QAxis requires only 2% of Q-Insight's inference time, while reason generation is a one-time offline cost.


Comment 4: Spatial sensitivity to quality degradation (Sec. 4.4).

Response 4: Thank you for the suggestion. We have added local-distortion visualizations. We have applied five distortions within predefined local masks.

Figure 10 shows that Grad-CAM for the negative quality score produces stronger QAxis responses in degraded regions. We further quantify the fraction of total CAM response within the distortion mask.

BaselineQAxis
Before17.5%17.3%
Desaturation21.1%23.1%
JPEG compression30.1%38.0%
Noise20.0%26.2%
Blur32.9%47.4%
Overexposure30.4%30.3%
Avg.26.9%33.0%

Before distortion, both methods show similar ROI response shares; after distortion, QAxis achieves a higher average share, indicating greater spatial sensitivity to quality degradation beyond foreground saliency alone.


We sincerely thank the reviewer again for the careful review. We would be happy to provide further clarification if needed.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end less than two days remaining, I want to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Thank you for the additional experiments and clarifications. The axis controls, matched RALI comparison, and local-distortion visualizations meaningfully address several of my concerns. I will increase my score to reflect these improvements, while lowering my confidence because I remain uncertain about the interpretation of the central finding.

In particular, “shared” has been removed from the title, but the introduction still emphasizes a quality axis shared across reasoning and visual representations. My understanding is that the revised claim concerns cross-modal sharing after alignment, together with cross-dataset transferability, rather than a common axis independently recovered across datasets. Making this distinction explicit and consistent throughout the paper would help. The new results strengthen the case for the estimated direction as a useful supervised reference, although they only partially establish the stronger interpretation of a shared quality geometry. Clearer reporting of whether the axis and linear-head controls use the same alignment components would also help isolate the contribution of the proposed axis. Overall, I appreciate the substantial effort and find the revised empirical evidence more convincing.

Thank you for your careful review and recognition of the additional experiments. We have revised the manuscript to distinguish cross-modal sharing after alignment from cross-dataset transferability, and clarified the control settings in Section 4.3 and Table 3.

Thank you again for helping us improve the clarity of our claims.

Reviewer_oxoUOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Building on RALI, this paper asks what is actually transferred when RL-generated quality reasoning is used to supervise a compact CLIP image encoder for NR-IQA. For each KonIQ training image, the authors collect multiple reasoning descriptions from Q-Insight and average their frozen CLIP text embeddings. They then regress these embeddings against MOS with sliced inverse regression to obtain a single direction, termed the quality axis, and build QAxis on top of it. Trained on KonIQ and evaluated on seven datasets, both QAxis variants outperform RALI on average and remain competitive with much larger MLLM-based methods, while requiring only the image encoder at inference.

Critical strengths
  1. During inference, the method reduces to a single projection of the image feature onto a fixed direction. Despite this simplicity, it is competitive on the seven-dataset average and yields the clearest gain on PIPAL.
  2. Rather than treating the reasoning embedding as a black box, the paper attempts to localize which part of the reasoning supervision is actually useful. To separate distortion from content, the authors construct three controlled groups from KADID, which is a sensible design for this question.
Critical weaknesses

The question raised by the authors is interesting. However, I am not convinced that the proposed quality axis and the symmetric rotation built on it truly answer this question. The observed gains may instead stem from empirical design choices, such as the added ranking loss and the scoring head, rather than from the mechanism the paper puts forward.

Questions for authors

The abstract, Fig. 1 and Sec. 5.1 report a 0.1B-parameter image encoder. Sec. 5.1 also states that the backbone is CLIP ViT-L/14-336, which appears to be the same backbone used by RALI. If so, the two models should have essentially the same number of parameters and the same inference cost. How was the 0.1B figure obtained, and how were the parameter counts of QAxis and RALI compared in Fig. 1? If the backbones are indeed identical, Fig. 1 overstates the efficiency advantage of QAxis and should be corrected.

4 comments

Thank you for your thorough review and thoughtful comments on our work. Your questions about the contributions of ranking supervision and the scoring head, as well as the comparison with RALI, have helped us clarify the evidence for QAxis. We address these concerns separately below.


Comment 1: The question raised by the authors is interesting. The observed gains may instead stem from empirical design choices, such as the added ranking loss and the scoring head, rather than from the mechanism the paper puts forward.

Response 1: Thank you for raising this concern. We have added controls to distinguish the contributions of ranking supervision, the scoring head, and the proposed quality axis.

(1) Does the improvement come simply from adding ranking loss? (Section 4.3, Table 3)

We evaluate a learnable linear head trained with standard InfoNCE and the same ranking loss, without the quality axis or quality-conditioned rotation.

Avg.KonIQAGIQACSIQSPAQKADIDLIVEWPIPAL
Linear head + ranking0.759/0.7420.924/0.9180.766/0.7020.733/0.7090.886/0.8840.693/0.6850.859/0.8420.449/0.453
QAxis-v10.811/0.7960.947/0.9320.790/0.7230.839/0.8150.901/0.8980.738/0.7590.882/0.8620.577/0.579

Despite receiving ranking supervision, the learnable linear head performs substantially worse than QAxis-v1. The ranking objective constrains relative scores but does not itself identify a quality-related direction in the representation. Our quality axis provides a fixed coordinate grounded in quality semantics, enabling ranking supervision to reinforce quality information extracted from reasoning embeddings rather than learn a scoring direction from MOS ordering alone. The results therefore highlight the importance of an informative quality reference for effective ranking-based learning.

(2) Does the improvement come from the scoring head?(Section 4.3, Table 2)

QAxis-v1 uses a simple scoring rule: direct projection without an additional learnable scoring head. After quality-conditioned alignment and axis-based ordinal learning, it directly projects the visual representation onto the fixed quality axis to predict quality.

The following ablations use this same head-free scoring rule.

Consensus alignmentQuality-conditioned rotationRanking lossAvg.
✓0.738/0.728
✓✓0.776/0.760
✓✓0.784/0.765
✓✓✓0.811/0.796

Rotation improves performance both without ranking supervision and when ranking is already present. Their combination achieves the best result while retaining the same direct-projection rule. These improvements support the contribution of quality-conditioned representation learning rather than an additional scoring head.

(3) Additional evidence supporting the quality axis.

  • Cross-modal quality reference and cross-dataset transferability — Section 2.2, Q2–Q3.

    Although derived from reasoning embeddings, the axis exposes quality information in visual representations: alignment concentrates gradient energy along it, raising the visual–MOS SRCC from 0.413 to 0.872. Across seven datasets, independently estimated axes differ in direction but yield strongly correlated projections, indicating a common quality reference and cross-dataset transferability.

  • Spatial sensitivity to degradation — Section 4.4, Figure 10 and Table 5.

    We introduce controlled local distortions to separate responses to quality degradation from responses to foreground content. QAxis concentrates Grad-CAM(-q) responses more strongly within degraded regions, achieving an average ROI response share of 33.0% versus 26.9% for consensus alignment. This provides complementary spatial evidence that QAxis's learned quality coordinate responds specifically to degradation-related evidence.

  • Importance of an informative quality direction — Section 4.3, Table 3.

    We evaluate multiple random axes and shuffled-MOS axes. Their best average PLCC/SRCC values are 0.740/0.718 and 0.753/0.733, respectively, compared with 0.811/0.796 for QAxis-v1. Despite ranking supervision, neither matches QAxis, indicating that QAxis's transferable quality information cannot be replaced by an arbitrary scoring coordinate.


Comment 2: Parameter counts and the efficiency comparison with RALI

Response 2: Thank you for identifying this inconsistency. The 0.1B figure was a reporting error; the correct inference-time parameter count is 0.3B. We have corrected the parameter count in the abstract, Figure 1, and Section 4.1, and revised Figure 1 to place QAxis and RALI at approximately **0.3B parameters. We apologize for the incorrect parameter count and the misleading comparison in Figure 1.


We sincerely thank the reviewer again for the careful review. We would be happy to provide further clarification if needed.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end less than two days remaining, I want to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Thanks for the responses. I also went through other reviewers' comments and the updated manuscript. I think the added experiments and the rebuttals are convincing, therefore I will raise my score to weak accept.

Thank you for carefully considering our responses, the updated manuscript, and the other reviewers’ discussions. We appreciate your positive assessment of the additional experiments and your constructive feedback, which has helped us strengthen the evidence and improve the clarity of the paper.

Reviewer_TX4hOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether transferable perceptual quality can be represented by a shared direction in language and vision embeddings. The authors aggregate quality reasoning generated offline by Q-Insight and estimate a MOS-supervised quality axis using sliced inverse regression on frozen CLIP text features. They then propose QAxis, which combines symmetric quality-conditioned rotations with axis-wise ordinal learning. The symmetric construction factorizes image–text similarity into the original feature cosine and a quality-coordinate agreement term. At inference, QAxis-v1 scores images through direct axis projection, while QAxis-v2 uses a PCA–ridge readout. Training on KonIQ and evaluating on its held-out test set and six external datasets, the methods achieve seven-dataset average PLCC/SRCC of 0.811/0.796 and 0.816/0.800, respectively.

Critical strengths

S1. The motivation is interesting. Previous work has shown that reasoning-based supervision can improve clip-style IQA, but it is less clear what information is actually transferred. This paper takes a further step by looking for a shared quality direction in the representation space. I think this is a reasonable and interesting perspective.

S2. The method is well motivated but also simple. Although most of the individual components are based on existing techniques, they fit together reasonably well. The symmetric RoPE design is also interesting, introduces quality information into cross-modal similarity while preserving the original semantic similarity.

S3. The experimental results seem good and solid. The paper evaluated on multiple datasets, the improvements are fairly consistent. The ablation studies also show that both quality-sensitive alignment and ordinal learning work.

Critical weaknesses

W1. The claim about a shared quality axis could be better supported. Since the axis is extracted using MOS supervision, finding a direction correlated with quality is somewhat expected. The cross-domain results are encouraging, but the author may provide analysis of how stable this direction is across different datasets or training subsets.

W2. The paper relies on Q-Insight, which is based on the relatively old Qwen2.5-VL model. Have the authors considered using more recent and stronger reasoning models to generate quality descriptions?

Questions for authors

Q1. Why does QAxis show larger improvements on PIPAL and KADID, but only marginal gains on AGIQA and even worse performance on LIVEW compared to RALI? Could the authors provide some explanation for these differences?

Q2. The model size is somewhat confusing. The paper states that it uses CLIP ViT-L/14-336 but reports only 0.1B parameters at inference time, which seems inconsistent. Could the authors clarify this?

3 comments

Thank you for your careful review and constructive suggestions. We have considered these concerns and provide our responses and the corresponding revisions below.


Comment 1: Stability and cross-dataset transferability of the quality axis.

Response 1: Thank you for this helpful suggestion. We agree that correlation with MOS is expected when the axis is estimated using MOS supervision, and does not alone establish a shared quality direction. We have added analyses of cross-dataset consistency and sensitivity to the axis-estimation subset, and revised the claim to emphasize cross-dataset transferability.

(1) Cross-dataset consistency — Section 2.2, Q3, Figure 5(a–c).

We independently estimate quality axes from seven datasets using the same procedure. Their directions differ, although their projections on the consensus features remain strongly correlated. We further examine this agreement through covariance decomposition: for KonIQ–KADID, the first five principal components contribute 91.7% of the projection covariance, and removing them reduces projection correlation from 0.859 to 0.479. These results distinguish agreement on observed features from geometric identity and motivate our more precise interpretation of a transferable quality axis.

(2) Sensitivity to training subsets — Section 4.3, “Effect of axis-estimation sample size,” Table 4.

We estimate axes using 20%, 40%, 60%, 80%, and 100% of the KonIQ training samples. Their cosine similarities to the full-data axis are 0.768, 0.854, 0.930, 0.976, and 1.000, respectively. The axis estimated from only 20% retains strong cross-dataset performance, achieving average PLCC/SRCC of 0.805/0.786. This supports robustness to reduced axis-estimation data. Visual alignment and ranking use the full training set throughout, isolating the effect of the subset used to estimate the axis.


Comment 2: Quality descriptions from newer models.

Response 2: Thank you for your suggestion. We evaluated Qwen3-VL-8B-Instruct and Qwen3.5-9B without IQA-specific fine-tuning, and VITAL, a quality-specialized model based on InternVL3-8B-Instruct.

Description generatorAvg. PLCC/SRCC
Qwen3-VL-8B-Instruct0.793/0.777
Qwen3.5-9B0.801/0.784
VITAL0.803/0.785
Q-Insight0.811/0.796

QAxis remains effective with these alternatives, while Q-Insight achieves the best average performance. A same-image comparison suggests why: models without IQA-specific fine-tuning may mix quality judgments with stylistic interpretations or provide conflicting degradation assessments, while VITAL’s descriptions can be relatively brief. These observations highlight the importance of task-relevant quality evidence. We added this analysis to Appendix D.


Comment 3: Differences in performance across target datasets.

Response 3: Thank you for this question. We interpret the different gains in terms of how QAxis and RALI transfer quality information to each target domain.

(1) PIPAL and KADID. Images with similar semantic content can differ substantially in perceptual quality. RALI’s contrastive alignment does not explicitly enforce quality ordering, whereas QAxis introduces a quality-axis reference and ordinal supervision. This helps distinguish distortion and restoration artifacts, consistent with SRCC improvements of 0.051 and 0.034, respectively, over RALI.

(2) AGIQA. Generated images introduce structural and semantic quality variations beyond photographic degradation. RALI’s reasoning representations already provide relevant semantic information, while the KonIQ-derived axis may contribute less additional guidance for generation-specific defects. This may explain the smaller gain over RALI.

(3) LIVEW. Although both training and target datasets contain authentic distortions, their acquisition and filtering differ. LIVEW includes 149 nighttime images with severe low-light distortions (12.8%), whereas KonIQ filters some heavily underexposed images with little visible content. Compared with RALI’s full-representation alignment, QAxis more explicitly reinforces the source-domain quality structure. Differences in distortion mixtures may therefore limit its additional benefit on LIVEW.


Comment 4: Clarification of the inference-time parameter count.

Response 4: Thank you for pointing this out. The reported 0.1B was a typographical error; the correct inference-time parameter count is approximately 0.3B for the CLIP ViT-L/14-336 image encoder. The description generator and CLIP text encoder are not used at inference. We have corrected this number throughout the manuscript. We apologize for the confusion caused by this error.


We sincerely thank the reviewer again for the careful review. We would be happy to provide further clarification if needed.

Thank you for your response and the additional experiments. I will maintain my original rating.

Thank you for reviewing our responses and the additional experiments. We sincerely appreciate your time and constructive suggestions, which have helped us strengthen the experimental analysis and clarify the scope of our claims.

Comments on the paper

1 comment
Area_Chair_fQTuOpenReview ↗

Reviewer-author discussion period is ending. Please feedback.

Dear Reviewers,

Thanks for your time and effort in reviewing ICLR2027 submissions.

As the reviewer-author discussion period will end very soon, please take time to read the author’s rebuttal and give your feedback. Your continuous engagement is highly appreciated.

Best regards,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.