acceptodds
Under review as a conference paper at ICLR 2027

Quality Signal Is Already There: Training-Free Cross-Domain IQA by Amplifying Latent Quality Direction in Frozen CLIP

Abstract

Cross-domain image quality assessment (CD-IQA) typically requires collecting target-domain images, aligning source and target distributions, and retraining a model whenever the deployment domain changes. Frozen contrastive vision–language models such as CLIP provide a training-free alternative, but their semantic-dominant representations are relatively insensitive to perceptual quality variations. By probing frozen CLIP across heterogeneous IQA datasets, we find that independently estimated quality directions exhibit non-random alignment and concentrate in a few common components, while the associated quality signal occupies only a small fraction of the overall feature energy. Perceptual quality is therefore latent rather than absent: transferable quality structure is already encoded in CLIP, but is underweighted in its default similarity space. This partial cross-domain commonality further suggests that a direction instantiated from an accessible synthetic source can serve as a transferable prior for strengthening quality-related components in unseen domains, without reproducing each target distribution. Based on this observation, we introduce a simple quality-directional feature amplification framework. A source-instantiated quality direction is constructed by contrasting high and low quality feature centroids, either using subjective ratings when available or, notably, without MOS by ranking synthetic distortions according to their affinity to pristine references. At inference, we amplify both the image-text feature using this source-derived quality prior before the standard cosine-similarity readout. The resulting method requires no target-domain data, parameter optimization, or additional network modules. Despite being training-free, it remains competitive with trained conventional cross-domain and VLM-based IQA methods, while consistently improving the frozen CLIP baseline across synthetic, held-out algorithmic, authentic, and AIGC quality benchmarks. Further analyses show that amplification strengthens distortion sensitivity and quality discrimination while largely retaining CLIP's semantic organization. The core code can be found in the Supplementary Materials.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Cross-Domain IQA, CLIP, Training-Free Feature Amplification

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_MqJCOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors addressed cross-domain blind image quality assessment, which is a valuable topic in the community in recent years. The underlying idea is to leverage the latent quality-aware knowledge (visual anchors) in frozen CLIP backbone and perform quality prediction by projecting image feature to the anchors based on cosine similarity. The idea of training cross-domain BIQA model without using target data and no training is interesting. The authors have done extensive experiments and comparisons to demonstrate the advantage of the proposed model.

Critical strengths
  1. The proposed method is simple and effective. The quality prior is a one-time closed-form feature operation. An obvious advantage is that it does not need extra training and target-domain data. The directional update accounts for only about 2.5% of total inference latency. The MOS-free KADIS variant consistently exceeds the frozen CLIP baseline on authentic and AIGC datasets.
  2. The paper provides diagnostics beyond final correlation coefficients. Independently constructed quality directions have a mean cross-dataset cosine similarity of 0.60, compared with about 0.03 for random directions. The image/text branch ablations, random-direction controls, and semantic-task evaluations suggest that the amplification primarily increases quality sensitivity without severely destroying CLIP's semantic structure.
  3. Extensive experiments and analysis were conducetd.
Critical weaknesses
  1. The evidence for shared cross-domain structure is not sufficient to support the generality claims. The main geometry analysis uses MOS-derived directions from seven datasets and performs an uncentered SVD on these directions. Although the authors stated that these directions are used only for diagnosis, the analysis still depends on target-domain annotations and does not provide cross-dataset independence tests, confidence intervals, or statistical significance tests.
  2. A single global direction and fixed linear amplification are the main bottlenecks of the methods. No source is uniformly optimal across targets, gains are weaker for color/contrast distortions and PIPAL, and global pooling can miss localized defects. The KADIS direction also reduces performance on aesthetic quality datasets. The authors could quantify this risk with breakdowns by distortion type, and local versus global defects.
  3. Some of the state-of-the-art BIQA models, particularly MLLM-based models, are missing in the comparison, such as Q-Align, DepictQA, DeQA-Score, etc. More comparisons should be included, especially models published in the last two years.
Questions for authors
  1. Please clarify whether the KADIS source bank and each target dataset share content, reference images, or distortion generation. In particular, please explain whether near-duplicate images exist between KADIS and datasets such as KADID or PIPAL, so that potential data leakage during source construction can be ruled out.
  2. It is better to clarify which baselines in Tables 1 and 2 were rerun with the authors' code and preprocessing, and which numbers were transcribed from prior work? Could the authors report, input preprocessing, and target-data-free protocol, the per-dataset results for frozen CLIP, CLIP-IQA/MA-CLIP, and the main cross-domain methods, together with the mean/standard deviation? This would determine whether the reported gains are statistically meaningful and can be fairly attributed to directional amplification.
  3. Could the authors provide a genuinely held-out direction evaluation: estimate the shared subspace or direction using only a subset of sources/datasets, then evaluate it on datasets not used in the direction analysis?
  4. More latest BIQA models, especically MLLM-based one, could be included in the comparisons.
5 comments

We sincerely thank the reviewer for recognizing the simplicity, effectiveness, and target-data-free nature of our approach. We appreciate the constructive feedback and have conducted additional experiments to strengthen the geometric evidence, examine distortion-specific robustness, verify source–target independence, and assess statistical reliability. We also clarify several experimental details already documented in the submission.

W1 & Q3: Independent and Statistical Validation of the Shared Quality Structure

We appreciate the reviewer's suggestion to further validate the independence and statistical significance of the observed shared quality structure.

First, we clarify that the seven MOS-derived quality directions in Fig. 4 are constructed exclusively for diagnostic analysis and are never used by our proposed IQA method. Our deployed quality prior is derived solely from MOS-free KADIS, without accessing target images or annotations. Moreover, Appendix C.5 (Fig. 11) already evaluates independently constructed KADIS and KADID directions against the seven-dataset diagnostic subspace, without including either source direction in its estimation.

To determine whether the observed low-dimensional organization generalizes beyond in-sample SVD, we conduct a seven-dataset Leave-One-Dataset-Out (LODO) experiment. For each dataset, we construct a normalized quality direction using the highest- and lowest-MOS 20% of images. In each LODO run, we estimate a shared subspace using uncentered SVD on the other six directions and evaluate the excluded direction without including it in subspace estimation.

Statistical validation. We perform 200 image-level bootstrap repetitions, reconstructing the quality directions and shared subspace in each repetition to obtain 95% percentile confidence intervals (CIs). We additionally perform 200 within-dataset MOS-label permutations to establish a null distribution that preserves image features while disrupting their association with subjective quality.

Table R1. Leave-one-dataset-out validation of the shared quality structure.

Held-out DatasetPC1 CosinePC1 EnergyTop-3 Energy
LIVE0.621038.56%82.28%
CSIQ0.671145.03%73.75%
TID20130.692047.88%73.56%
LIVEC0.730553.36%81.02%
KonIQ0.726652.79%70.71%
BID0.747055.80%80.60%
FLIVE0.689747.57%64.59%
Mean0.696848.71%75.22%

PC1 Cosine measures alignment with the leading direction estimated from the other six datasets. PC1/Top-3 Energy denotes the squared projection energy of the held-out normalized direction onto the corresponding subspace.

The results provide three complementary findings:

  1. Consistent out-of-sample alignment. All seven held-out directions show positive PC1 cosine similarities (0.6210-0.7470), with a mean of 0.6968 (95% CI: [0.6737, 0.7011]).

  2. Strong low-dimensional generalization. The first three components explain 75.22% of held-out directional energy on average, substantially exceeding the MOS-permutation null mean of 5.07% (p≈0.005p\approx0.005).

  3. Cross-dataset geometric consistency. Independently constructed directions exhibit a mean pairwise cosine similarity of 0.5726 (95% CI: [0.5448, 0.5762]), with 90.86% of their in-sample energy explained by the first three components.

Importantly, the shared structure remains observable even when each target direction is excluded from subspace estimation, demonstrating that the low-dimensional organization is not merely an in-sample SVD effect. The consistent alignment across heterogeneous datasets, together with bootstrap and permutation evidence, further supports the statistical reliability of this observation.

W2: Distortion-Specific Performance and Robustness of the Global Quality Direction

We agree that a single global quality direction cannot fully characterize every perceptual degradation. Our method deliberately employs a simple source-derived direction as an approximation to a richer, partially domain-dependent quality structure. This limitation is explicitly discussed in Appendix H, while Appendix A.6 and Appendix F.3 already provide distortion-specific sensitivity and spatial-response analyses. We additionally conducted two complementary experiments to quantify distortion-specific performance and examine whether the global direction depends excessively on particular source degradations.

(1) Distortion-specific performance. We evaluated our original, unchanged global quality direction separately on all 25 KADID-10K distortion types, each containing 405 images. The same KADIS-derived prior, fixed amplification factor, and text calibration are used across all distortion categories, without category-specific adaptation.

Our method achieves positive SRCC improvements over frozen CLIP-IQA on all 25 distortion types, increasing overall KADID SRCC from 0.4739 to 0.6737. The improvements vary across degradation categories. For example, color quantization improves from −0.2132 to 0.4343, while contrast change improves from 0.0877 to 0.1996.

These results demonstrate that directional amplification improves quality discrimination even for challenging color/contrast distortions. However, certain categories remain difficult in terms of absolute performance, consistent with the limitations acknowledged in Appendix H.

(2) Source distortion composition ablation. To investigate whether the global quality direction is dominated by particular degradation types, we constructed additional source quality priors under two settings:

  • Single-type: Construct a quality prior using only one of the 25 source distortion types.
  • Leave-one-type-out: Construct a quality prior using the remaining 24 types after excluding one source distortion type.

For each configuration, the source quality prior and corresponding source-anchor calibration are independently constructed from the selected KADIS degradations. The frozen backbone, quality prompts, amplification factor, and target evaluation protocol remain unchanged. No target images or MOS annotations are used for source-prior construction.

We evaluated these configurations on three authentic IQA datasets:

DatasetFrozen CLIPFull 25 TypesSingle-Type Variants Improving over Frozen
BID0.75680.787322/25
KonIQ0.64520.756523/25
LIVEC0.71810.777322/25

Notably, 20 of the 25 single-type quality priors improve performance on all three authentic datasets, suggesting that transferable quality-related information is not exclusive to any particular synthetic degradation.

More importantly, removing any single source distortion type produces a direction highly aligned with the complete direction, with cosine similarities of 0.9929–0.9999. The corresponding absolute SRCC changes remain within 0.0066 across the three authentic datasets. These results demonstrate that the source-derived quality prior is not dominated by any single degradation category and remains stable under changes to source distortion composition.

Q1: Source–Target Independence and Potential Data Leakage

We clarify that our quality prior is constructed exclusively from KADIS, without accessing target images, target reference images, or subjective quality annotations. Source anchors are used to estimate the quality direction; they are not employed as a target-image retrieval bank during inference. Furthermore, KADIS-700k and KADID-10k were explicitly constructed using disjoint sets of pristine reference images[1]. Although the two datasets share synthetic distortion-generation methods, this represents shared degradation categories rather than image-level data leakage.

To provide additional empirical verification, we conducted a reference-image overlap audit involving 100 sampled KADIS source references in 15 different seeds(total 1484) and 281 target reference images (81 from KADID and 200 from PIPAL). We examined exact file identity, pixel-level identity, and perceptual-hash similarity. Near-duplicate candidates were identified using a perceptual-hash Hamming-distance threshold of 6. The audit found zero identical-file pairs, zero pixel-identical pairs, and zero perceptual-hash near-duplicate candidates under this threshold. This audit covers the sampled source references rather than an exhaustive search over the entire KADIS collection. Nevertheless, together with the documented dataset construction protocols, it provides additional evidence supporting source–target independence.

Q2: Baseline Protocols, Reproducibility, and Statistical Reliability

We appreciate the opportunity to clarify the evaluation protocols already documented in Appendix I. As specified in Appendix I.1, Table 1 reports established cross-domain and MLLM baselines using results transcribed from published evaluations. In contrast, the broader training-free baselines in Table 2, including CLIP-IQA and MA-CLIP, are reproduced under the same testing protocol as our method, with per-dataset SRCC/PLCC results explicitly reported. For these controlled evaluations, we consistently use frozen CLIP RN50 with aspect-ratio-preserving preprocessing and no center crop. The source-derived direction, quality prompts, and amplification factor are fixed across target datasets. No target-domain information is used for parameter selection or adaptation. We therefore distinguish published cross-method comparisons in Table 1 from controlled reproduced comparisons in Table 2, rather than assuming that all published baselines share identical preprocessing or implementation settings. The training and target-access requirements of the compared methods are explicitly indicated.

Appendix I.3 (Table 27) already reports SRCC/PLCC standard deviations over 15 independently sampled KADIS reference sets, demonstrating that performance is stable with respect to source-reference selection. To further quantify the reliability of performance improvements, we conducted additional image-level paired bootstrap and paired permutation tests, comparing our complete method with frozen CLIP-IQA on identical test images. Specifically, we perform 1,000 paired bootstrap repetitions, sampling image indices with replacement and computing the SRCC difference between the two methods using the same resampled indices. The 95% percentile CI is calculated from the 2.5th and 97.5th percentiles of the resulting SRCC differences. We additionally perform 1,000 paired permutation trials by randomly exchanging the two methods' predictions within paired image units and recomputing the SRCC difference. Two-sided Monte Carlo p-values are obtained by comparing the absolute randomized differences with the observed absolute difference, using the standard plus-one correction. The results are summarized below:

DatasetFrozen SRCCOurs SRCCΔSRCC95% Bootstrap CIp-value
KonIQ0.64520.7565+0.1113[0.1033, 0.1191]0.001
LIVEC0.71810.7773+0.0592[0.0422, 0.0764]0.005
AGIQA-3K0.65170.7235+0.0718[0.0584, 0.0847]0.001
KADID0.47390.6737+0.1998[0.1890, 0.2099]0.001

All four datasets exhibit positive performance differences, with image-level paired bootstrap intervals excluding zero and permutation (p)-values below 0.01. These intervals quantify image-level resampling variability. Together with the 15-source-sampling analysis already reported in Table 27, these findings provide complementary evidence for both source-construction stability and prediction-level performance reliability. Furthermore, Table 3 already separates the image-amplification and text-calibration contributions, confirming that image-side directional amplification accounts for most of the improvement.

[1] Lin, H., Hosu, V., & Saupe, D. (2019). KADID-10k: A Large-scale Artificially Distorted IQA Database.

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

This manuscript investigates whether a frozen CLIP can give a quality score without training on target dataset. They compute high/low-quality centroids on several IQA datasets, and find that the so-called quality directions are partly shared and low-dimensional. Basing on this observation, the authors propose a training-free, target-free readout: the source direction v_q amplifies the v_q part of the image feature, the good/bad text contrast is calibrated to source anchors, and the score is a cosine similarity. On synthetic, authentic, AIGC and aesthetic benchmarks the proposed method beats the frozen CLIP baseline and is close to trained methods.

Critical strengths
  1. The idea is interesting and is supported by several different probes: direction alignment, energy diagnostics, removal tests and random/noise controls.

  2. The gain is stable on 12+ targets and several backbones, the source-sampling variance is small (Table 27), and App. H discusses the limitations honestly.

  3. A comprehensive set of ablation studies and analysis are provided, covering nearly each part and aspect of the method. Hyperparameter sensitivity and injection are provided, computational cost and storage are provided, and reproducibility and evaluation protocal are detailed.

  4. The uploaded code matches the stated hyperparameters (alpha=2, tau_a=0.5, K=50, R=100, KADID-MOS K=2000) with no train/test leakage: when KADID is the source, KADID-10K is excluded.

Critical weaknesses
  1. Sec. 3.2 says each KADIS pristine image has 'M=125 distorted variants'. But maybe this is wrong: public KADIS-700K has only 5 per image. I think the 125 variants come from the authors' own KADID-pipeline code, and the paper never says so.

  2. Sec. 4.1 says 15 repeated samples with the median, but according to the uploaded code, it only uses one seed (42) plus a cached prior; App. G.2's 'repeated splits' is also one split.

Questions for authors
  1. Please state clearly that the 125 KADIS distortions come from your own KADID-pipeline code, and release the 15-sample median code.
4 comments

We sincerely thank the reviewer for the positive assessment of our work and for carefully examining both the manuscript and the submitted code package. We particularly appreciate the precise observations regarding distortion generation and the reproducibility of repeated experiments. We address these concerns below.

Q1 & W2: Reproducibility of Repeated Evaluations

We appreciate the reviewer identifying the discrepancy between the reported repeated evaluations and the submitted code package. The code package prepared for conference review provides a single-run implementation of our core method, whereas the reported experimental results were obtained through a separate multi-run evaluation procedure.

(1) KADIS source sampling. The reported experiments were conducted over 15 independently randomized source samplings. Specifically, we included a fixed seed of 42 and generated the remaining 14 seeds by randomly sampling distinct integer offsets from [1, 10000] without replacement and adding 42 to each offset. For each seed, we independently sample 100 pristine references from KADIS-700K and construct the corresponding quality prior using the same fixed hyperparameters ((K=50), (α=2), and (τa=0.5\tau_a=0.5)). No target-domain information is involved in source sampling or quality-prior construction. The main tables report the median SRCC/PLCC across the 15 runs, while Appendix I.3 (Table 27) reports the corresponding standard deviations and observed half-ranges.

To further verify reproducibility, we independently repeated the complete evaluation using a newly generated set of 15 random seeds, with all hyperparameters and evaluation protocols unchanged. The resulting median performance matched or even exceeded the values reported in the original manuscript. This additional replication provides further evidence that our results are stable across independently sampled KADIS reference sets and are not dependent on a particular random seed configuration.Remarkably, the median SRCC/PLCC differences between the original and replicated experiments were on the order of 10−310^{-3}, with the replicated results matching or slightly exceeding those reported in the manuscript. The single-seed implementation (seed=42) included in the submitted code package also produces results closely matching the reported medians. We acknowledge that the submitted code package demonstrates the core single-run implementation rather than the complete multi-run orchestration procedure. We will clarify the random-seed generation and repeated-sampling protocols in the revised manuscript and provide the corresponding multi-run evaluation procedure in the final reproducibility materials.

(2) Appendix G.2: Repeated splits. The within-domain experiments were conducted using 15 independent random 1:9 reference/test splits. In each repetition, the quality direction is constructed exclusively from the reference split using the corresponding MOS annotations, while evaluation is performed on the disjoint test split. To further clarify the repeated evaluation protocol and its statistical reliability, we additionally computed the standard deviations across all 15 splits. The updated results are reported below.

Table R1. Within-domain direction estimation under 15 independent 1:9 reference/test splits. Results are SRCC / PLCC, reported as mean ± standard deviation.

DatasetFrozen CLIPIn-domain Direction
KonIQ-10K.6455 ± .0019 / .7211 ± .0018.7720 ± .0133 / .8135 ± .0093
KADID-10K.4716 ± .0053 / .5013 ± .0041.6878 ± .0145 / .6934 ± .0132
PIPAL.3682 ± .0034 / .3729 ± .0041.3922 ± .0141 / .4196 ± .0120
AGIQA-3K.6514 ± .0026 / .7157 ± .0026.7845 ± .0075 / .8534 ± .0035
AVA.4085 ± .0038 / .4308 ± .0038.5128 ± .0222 / .5285 ± .0191

For KADID-10K and PIPAL, the splits are performed at the pristine-reference level, ensuring that images derived from the same reference do not appear in both the reference and test partitions. For the other datasets, we employ random image-level splits. The same evaluation procedure is applied across repetitions, and PLCC is computed following the five-parameter logistic fitting protocol used in the main experiments. The results demonstrate consistent improvements across all five datasets, while the standard deviations quantify the variability introduced by different reference/test partitions. We will update Table 22 to report these mean ± standard deviation results and explicitly document the number of repetitions and splitting procedure.

W1: Clarification of the 125 Distorted Variants

We thank the reviewer for this careful observation. The reviewer is correct that the original KADIS-700K dataset provides only five pre-generated distorted images per pristine reference. As the reviewer correctly inferred from our submitted code, we reimplemented the original MATLAB distortion-generation pipeline of KADID-10K in Python, and this implementation is already included in the submitted code package. Specifically, for each pristine reference sampled from KADIS-700K, we generate 125 distorted variants (25 distortion types × 5 severity levels) on the fly, rather than using the five pre-generated KADIS distortions. Therefore, these variants are generated using our Python reproduction of the established KADID distortion pipeline, not a newly designed degradation process. This allows us to construct high- and low-quality source anchors across diverse distortion types without requiring subjective quality annotations.

We acknowledge that the distinction between the original KADIS data and our on-the-fly distortion generation was insufficiently explained in the manuscript. We will explicitly clarify this implementation detail in Sec. 3.2 and Appendix I.3.

Remarks

We sincerely thank the reviewer again for the careful assessment and constructive suggestions. We agree that explicitly distinguishing our Python reproduction of the established distortion pipeline from the original KADIS data, and providing a clearer description of the repeated evaluation procedures, will improve the precision and reproducibility of our work.

I have carefully read the author's responses to me and the other reviewers. I'll retain my score.

We would like to express our sincere gratitude for your continued support and encouraging assessment of our work. We especially appreciate the time and effort you devoted to carefully reviewing our responses and reconsidering the additional evidence.

Your constructive feedback has been genuinely valuable in helping us strengthen the clarity, rigor, and reproducibility of our paper. We are particularly grateful for your recognition of our contributions and for maintaining your positive assessment.

Thank you once again for your thoughtful review, generosity, and support.

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

It is a well written paper in cross-domain IQA. It proposes a training and target-data free framework that amplifies the latent quality components already encoded within frozen contrastive VLM like CLIP. The authors show that perceptual quality is latent but underweighted in CLIP's default (semantic) geometry, by showing that independently extracted quality directions across diverse IQA datasets have non-random alignment and occupy a common low-dimensional subspace. They construct a source-derived quality direction by contrasting high- and low-quality feature centroids from a synthetic source. At inference, both the image and text embeddings are amplified along this quality direction achieving good performance with low cost overhead.

Critical strengths
  1. The ability to construct an effective quality direction without human annotations by using CLIP affinity ranking is a good contribution.
  2. The authors provide good representational analysis to validate that quality directions concentrate in a common low-dimensional space.
  3. It demonstrates that feature amplification boosts quality sensitivity while preserving CLIP’s high-level semantic representation.
Critical weaknesses
  1. The framework relies on a fixed, manually chosen amplification strength across all heterogeneous target domains. Enforcing a global constant fails to adapt to dataset-specific semantic densities, which can lead to a sub-optimal trade-off between quality sensitivity and semantic structure on highly complex domains.
  2. The linear projection update is mathematically symmetric and amplifies both correct and incorrect projection signs. If an image's default latent representation has an initially incorrect projection sign relative to the ground-truth quality, the amplification step directly magnifies this error, resulting in severe quality under- or over-estimation (like in figure 16).
  3. Because the quality direction is constructed from global synthetic distortions and computed over globally pooled CLIP features, the model lacks the spatial granularity required to detect sparse, localized, or highly non-additive artifacts
  4. Representing quality as a single 1D vector assumes a monotonic, unidimensional axis. This formulation is structurally incapable of generalizing to multidimensional attributes, failing to represent cases where objective visual fidelity and subjective aesthetic preferences diverge (like in table 23).
Questions for authors
  1. Have you considered any simple, image-adaptive heuristics (such as feature norm or entropy) to dynamically scale the amplification factor on a per-image basis, rather than using a fixed global constant?
  2. How difficult would it be to apply this directional amplification to the spatial tokens of CLIP's visual transformer prior to attention pooling, in order to capture localized or patch-level distortions?
  3. Since the first three principal components capture over 90% of the cross-dataset quality energy, does extending vq to a 2D or 3D projection subspace yield noticeable accuracy gains over the 1D direction?
8 comments

We sincerely thank the reviewer for recognizing the contributions of our work, particularly the MOS-free quality prior construction, cross-domain representation analysis, and semantic preservation. We appreciate the constructive questions concerning the robustness and expressiveness of our framework. To address these concerns, we provide additional experiments on image-adaptive amplification, pairwise ordering reliability, multidimensional quality guidance, and spatial-level intervention, together with clarifications of our existing analyses and acknowledged limitations.

W1 & Q1: Fixed Amplification Strength and Image-Adaptive Scaling

We appreciate the reviewer's constructive suggestion regarding image-adaptive amplification. The fixed amplification factor is a deliberate design choice for target-data-free generalization, rather than an assumption that a single value is optimal for every domain.

As shown in Appendix A.8 (Fig. 9), moderate amplification consistently improves quality prediction across heterogeneous datasets, although the optimal strength differs across domains. We deliberately fix (α=2) for all targets without target-specific hyperparameter selection. Furthermore, Appendix D (Table 18) demonstrates that this setting retains 93.4% ImageNet prediction consistency and 0.984 feature cosine similarity with the original CLIP representations, indicating a favorable trade-off between quality sensitivity and semantic preservation.

Additional experiment: Fixed vs. image-adaptive amplification. To directly investigate Q1, we evaluate two training-free, image-adaptive strategies:

  • Text-prediction entropy: We estimate confidence from the binary entropy of positive/negative quality-text predictions, using a fixed temperature of 0.05. Lower entropy corresponds to higher confidence.
  • Source-anchor margin: We estimate confidence from the normalized absolute distance between an image's quality-direction coordinate and the midpoint of the source high-/low-quality anchors.

Both confidence scores are normalized to ([0,1]) and mapped to image-specific amplification factors through (α_i=2(0.5+c_i)), yielding a range of ([1,3]). The same factor is applied to both image and text branches. Neither strategy requires target annotations, target-domain calibration, or additional training.

Table R1. Fixed vs. image-adaptive amplification. Each entry reports SRCC/PLCC.

MethodKonIQLIVECTID2013AGIQA-3KAvg.
Fixed α=2 (Ours).7565/.8090.7773/.8134.6610/.7185.7235/.7623.7296/.7758
Fixed α=3.7546/.8045.7735/.8123.6634/.7197.7206/.7585.7280/.7737
Adaptive (Entropy).7339/.7954.7666/.8005.6460/.7084.7132/.7578.7149/.7655
Adaptive (Source Margin).7546/.8049.7770/.8073.6584/.7111.7137/.7536.7259/.7692

Avg. denotes the arithmetic mean across the four datasets.

The fixed (α=2) achieves the highest average SRCC (0.7296) and PLCC (0.7758) and outperforms both adaptive strategies in SRCC on all four datasets. Compared with entropy-based and margin-based adaptation, it improves average SRCC by 0.0147 and 0.0037, respectively. These findings suggest that simple image-level confidence heuristics do not necessarily provide reliable guidance for quality amplification. Although more sophisticated adaptive strategies remain worth exploring, our results support a simple, target-independent intervention without additional optimization or calibration. We will include the exploration of image-level adaptive scaling coefficients in our future work.

W2: Amplification of Incorrect Quality Projections

We appreciate the reviewer's concern regarding the potential amplification of incorrect quality projections. We explicitly acknowledge this limitation in Appendix H (Fig. 16), where individual images can be overestimated or underestimated when their quality coordinates are incorrectly signed.

Our method does not assume that every quality projection is perfectly aligned with human perception. Rather, it aims to amplify a statistically transferable quality-related component already encoded in frozen CLIP. The main experiments demonstrate substantial improvements in overall quality discrimination. For example, our MOS-free method improves SRCC from 0.474 to 0.674 on KADID (+0.200), 0.645 to 0.757 on KonIQ (+0.112), and 0.718 to 0.777 on LIVEC (+0.059).

Additional experiment: Pairwise ordering accuracy. To investigate whether amplification improves overall quality comparisons despite potential individual errors, we perform a pairwise ordering analysis before and after amplification. For each dataset, we select up to 5,000 images using a fixed random seed and evaluate all valid unordered image pairs. Pairs with identical MOS values or tied predictions are excluded. Pairwise accuracy measures the proportion of pairs whose predicted quality ordering agrees with the corresponding MOS ordering. We use exactly the same image pairs for both methods. MOS annotations are used exclusively for evaluation, not for constructing quality directions or adjusting predictions.

Table R2. Pairwise ordering accuracy before and after amplification.

DatasetBeforeAfterImprovement
KADID66.55%74.47%+7.93 pp
KonIQ73.37%77.97%+4.60 pp
LIVEC76.45%79.25%+2.81 pp
AGIQA-1K64.23%69.94%+5.70 pp
Average70.15%75.41%+5.26 pp

pp denotes percentage points.

The results demonstrate consistent improvements across all four datasets, with average pairwise accuracy increasing by 5.26 percentage points. These findings provide complementary evidence that directional amplification improves overall quality ordering despite potential errors in individual projections.The failure cases illustrated in Fig. 16 remain a limitation of fixed linear amplification. Nevertheless, consistent gains in dataset-level correlations and pairwise comparisons demonstrate that such errors do not negate the overall benefits of our approach. Identifying uncertain quality projections and selectively mitigating erroneous amplification remains a meaningful direction for future work.

W3 & Q2: Spatial Granularity and Token-Level Amplification

We appreciate the reviewer's suggestion to investigate finer-grained quality information. Our original submission already investigates spatial-level directional amplification at different stages of frozen CLIP RN50 (Appendix A.8, Fig. 9(c)). These interventions operate on intermediate spatial feature maps before aggregation, rather than ViT patch tokens. Earlier-stage amplification consistently underperforms the final readout, with performance generally deteriorating as injection moves toward earlier layers. Moreover, Appendix F.3 (Fig. 15) presents gradient-based spatial responses and image–text matching maps covering blur, noise, compression artifacts, and localized generation defects. These visualizations suggest that localized degradation cues can contribute to the pooled representation, providing qualitative evidence that local degradation information is preserved to some extent.

To directly address the reviewer's token-level question, we conduct a controlled experiment using frozen CLIP ViT-B/16 with original full-resolution images, without cropping, resizing, or letterboxing, preserving native-resolution distortion details.

We compare six configurations: (1) Frozen ViT, without amplification; (2) Text-only, using source-derived text calibration; (3) Final CLS (Ours), enhancing the final image embedding; and three interventions immediately before the final Transformer block: (4) Intermediate CLS, (5) Intermediate Patch, and (6) Intermediate CLS+Patch. This placement allows enhanced patch tokens to influence the final CLS representation through subsequent self-attention.

All quality directions are constructed exclusively from MOS-free KADIS using the original pristine-affinity-based high-/low-quality anchor selection. Intermediate CLS and patch directions are independently estimated from their corresponding frozen token representations, with the patch direction obtained by aggregating source high-minus-low contrasts across spatial positions. All variants share the same frozen backbone, full-resolution preprocessing, and fixed amplification factor α=2, with consistent source-derived text calibration for enhanced variants. No target images, annotations, or domain-specific tuning are involved.

Table R3. Full-resolution spatial amplification with frozen CLIP ViT-B/16 (SRCC/PLCC).

MethodKonIQLIVECTID2013AGIQA-3KAvg.
Frozen ViT.5677/.5990.6231/.6299.4429/.5057.5821/.6309.5540/.5914
Text-only.5565/.5877.6196/.6219.4375/.5028.5655/.6102.5448/.5806
Final CLS (Ours).6164/.6488.6719/.6750.5382/.5895.6183/.6623.6112/.6439
Intermediate CLS.6163/.6506.6644/.6707.4882/.5444.6037/.6349.5931/.6251
Intermediate Patch.5880/.6156.6094/.6064.4247/.4935.5801/.6245.5506/.5850
Intermediate CLS+Patch.6406/.6687.6524/.6516.4545/.5077.6005/.6295.5870/.6143

All variants use identical full-resolution evaluation settings. Avg. denotes the arithmetic mean across four datasets.

The results reveal three important observations. First, final-feature amplification achieves the highest average SRCC (0.6112) and PLCC (0.6439), leading in SRCC on three of the four datasets. This extends the RN50 injection-location findings to a different architecture, supporting the final readout as a stable intervention point for cross-domain quality prediction.

Second, patch-only amplification does not consistently improve performance. Although it improves KonIQ over frozen ViT, it underperforms the frozen baseline on LIVEC, TID2013, and AGIQA-3K in SRCC. This suggests that amplifying intermediate spatial representations alone does not necessarily yield better quality discrimination at the final readout.

Interestingly, joint CLS+Patch amplification improves KonIQ from 0.6164/0.6488 to 0.6406/0.6687 (ΔSRCC +0.0242; ΔPLCC +0.0198), suggesting potential complementarity between global and spatial quality information. However, this benefit does not consistently transfer to the other datasets.

Together, the RN50 experiments examine injection depth, while the additional full-resolution ViT experiment explicitly investigates CLS- and patch-token-level interventions. Both favor final-feature amplification on average, without implying that spatial modeling is inherently ineffective. We therefore retain the simple final-readout design while acknowledging that globally pooled features may underrepresent highly localized defects (Appendix H), and that spatial–global interactions remain worth exploring.

Comments on the paper

2 comments

Dear Reviewers,

We sincerely thank all reviewers for their constructive comments and suggestions. We have carefully addressed the concerns raised during the discussion and updated our manuscript with additional experiments, statistical analyses, and implementation clarifications. We would like to briefly summarize the key revisions to facilitate further evaluation.

1. Independent validation of the shared quality structure (Sec. 4.3).

We added a seven-dataset Leave-One-Dataset-Out (LODO) analysis to investigate whether the observed low-dimensional quality structure generalizes beyond in-sample SVD. The held-out directions achieve a mean PC1 cosine similarity of 0.6968 (95% CI: [0.6737, 0.7011]), while the top three components retain 75.22% of their energy, substantially exceeding the MOS-permutation null of 5.07% (p ≈ 0.005). A complete per-dataset table and bootstrap/permutation protocols are provided. These findings strengthen our central observation of a partially shared, transferable quality structure in frozen CLIP.

2. Distortion-level robustness and source-direction stability (Appendix A).

We extended the distortion analysis to all 25 KADID-10K distortion categories. The fixed MOS-free KADIS direction improves SRCC over frozen CLIP across all 25 types. Additional single-type and leave-one-type-out experiments show that the extracted direction is not dominated by any particular source distortion: 20/25 single-type directions improve all three evaluated authentic datasets, and leave-one-type-out directions maintain cosine similarities of 0.9929–0.9999 with the full-source direction.

3. Adaptive, multidimensional, and spatial amplification (Appendix A).

We conducted three additional controlled studies:

  • Adaptive amplification: Compared fixed gains with text-entropy-based and source-margin-based adaptive strategies. The original fixed α=2 achieves the best average SRCC/PLCC across four datasets.
  • Multidimensional guidance: Compared our single quality direction with 2D/3D reference-derived and distortion-type-derived subspaces. The original 1D formulation achieves the highest average SRCC, supporting its effectiveness as a compact cross-domain approximation.
  • Spatial granularity: Evaluated six amplification configurations using frozen CLIP ViT-B/16 at original full resolution, without cropping, resizing, or letterboxing. Final-feature amplification achieves the highest average SRCC/PLCC. Joint CLS+Patch amplification offers complementary gains on KonIQ, but these do not consistently generalize across domains.

These experiments clarify both the effectiveness and the scope of our deliberately simple final-feature amplification design.

4. Statistical reliability and reproducibility (Appendices G and I).

We added paired bootstrap confidence intervals and paired permutation tests, confirming statistically significant SRCC improvements on KonIQ, LIVEC, AGIQA-3K, and KADID-10K. We also updated the within-domain experiments to report mean ± standard deviation over 15 independent reference/test splits, with pristine-reference-level separation for KADID-10K and PIPAL.

Furthermore, we clarified that the 125 source distortions are generated on the fly through our Python reimplementation of the established KADID distortion pipeline, rather than taken directly from the five pre-generated KADIS variants. We supplemented the source-sampling reproducibility description and reported an image-overlap audit showing no exact or perceptual-hash matches among the examined KADIS, KADID, and PIPAL references.

5. Additional clarifications.

We refined the related-work discussion to better contextualize DepictQA and other reasoning-based IQA approaches, while maintaining a clear distinction between their evaluation protocols and our training-free, target-data-free setting.

Overall, these revisions provide additional support for our central claim: quality-related information is already present in frozen CLIP, exhibits partially shared cross-domain structure, and can be effectively exploited through a simple source-derived amplification mechanism without training or target-domain adaptation.

We greatly appreciate the reviewers' time and feedback. We hope the revised manuscript and additional evidence address your concerns, and we would be grateful if you could take another look at the updated results when convenient. We welcome any further questions or discussion.

Thank you again for your consideration.

AuthorsOpenReview ↗

General Response: Consolidated Clarifications and Additional Evidence

Dear Area Chair and Reviewers,

We thank reviewers for their thoughtful feedback. We have completed additional experiments, statistical analyses, and manuscript revisions to address the concerns. We summarize the evidence below to facilitate the AC's assessment and further discussion. All target-MOS-derived directions are used only for geometric diagnosis; the deployed prior comes exclusively from MOS-free KADIS, without target data, training, or adaptation.

  1. Shared quality geometry (MqJC: W1/Q3). We added seven-dataset leave-one-dataset-out (LODO) validation (Sec. 4.3), evaluating each excluded direction against a subspace estimated from the other six. All held-out directions align positively with PC1 (cosine 0.6210–0.7470; mean 0.6968, 95% CI [0.6737, 0.7011]). Top-3 retained energy averages 75.22%, versus 5.07% under MOS-label permutation (p≈0.005). The mean pairwise cosine is 0.5726 (95% CI [0.5448, 0.5762]). This supports a partially shared structure beyond in-sample SVD, not a universal quality axis.
  2. Single-direction and adaptive design (iUgM: W4/Q3, W1/Q1). We compared our 1D prior with MOS-free KADIS 2D/3D subspaces from reference-specific (refPCA) or distortion-type (typePCA) contrasts. 1D achieves the best four-dataset mean SRCC (0.7296) versus 0.7266 for 2D-refPCA. Fixed α=2 also outperforms fixed α=3 and text-entropy/source-margin adaptive gains on average (SRCC/PLCC 0.7296/0.7758). These results favor a compact, fixed intervention under cross-domain transfer, without claiming quality is intrinsically one-dimensional.
  3. Reliability and significance (iUgM: W2; MqJC: Q2). We acknowledge individual incorrect projection signs (Fig. 16), but pairwise ordering improves on KADID 66.55%→74.47%, KonIQ 73.37%→77.97%, LIVEC 76.45%→79.25%, AGIQA-1K 64.23%→69.94% (mean +5.26 pp). Paired bootstrap CIs and permutation tests (1,000 repetitions) support significant SRCC gains: KonIQ +0.1113 (CI [.1033,.1191]), LIVEC +0.0592 ([.0422,.0764]), AGIQA-3K +0.0718 ([.0584,.0847]), and KADID +0.1998 ([.1890,.2099]); all p≤0.005.
  4. Distortion robustness (MqJC: W2). The unchanged source prior improves SRCC on all 25 KADID distortion types, with overall SRCC 0.4739→0.6737 (e.g., color quantization −0.2132→0.4343). In source-type ablations, 20/25 single-type directions improve all three authentic targets; leave-one-type-out directions retain cosine 0.9929–0.9999 with the full prior, with maximum absolute ΔSRCC 0.0066. This complements the original 25×25 sensitivity analysis.
  5. Spatial granularity (iUgM: W3/Q2, MqJC: W2). Our original RN50 injection-depth analysis favors final-readout amplification. Additional six-configuration frozen ViT-B/16 experiments preserve the original full resolution (no crop/resize/letterbox). Final-feature amplification attains the highest average SRCC/PLCC 0.6112/0.6439, leading three of four datasets in SRCC. Patch-only gains are inconsistent; joint CLS+Patch helps KonIQ (0.6164→0.6406 SRCC) but not consistently elsewhere.
  6. Provenance, leakage, reproducibility (CdKU: W1/W2; MqJC: Q1). We clarified that 125 variants per KADIS reference (25 types × 5 levels) are generated on the fly using our Python reimplementation of the established KADID distortion pipeline, rather than KADIS's five pre-generated variants; the code was already submitted. Auditing 100 KADIS against 81 KADID and 200 PIPAL pristine references found no exact-file/pixel or pHash near-duplicate matches in these examined sets. Main results are medians over 15 source samplings; an independent 15-run replication agrees within approximately 10⁻³. We clarified the seed-42 code configuration and updated in-domain experiments to mean±std over 15 independent 1:9 splits, grouping KADID/PIPAL by pristine reference.
  7. Baselines and manuscript updates (MqJC: W3/Q4). We clarified published-versus-controlled baseline protocols and incorporated DepictQA into Related Work; its original descriptive/comparative evaluation does not provide a directly matched continuous NR-IQA baseline under our cross-domain protocol. Revisions appear in Sec. 4.3 (LODO), Appendix A (ablations), Appendix G (repeated splits), and Appendix I (statistics/reproducibility).

Overall, the new evidence strengthens our finding of partially shared quality structure in frozen CLIP and supports its simple, training-free, target-data-free exploitation. With the discussion deadline approaching, we would greatly appreciate any remaining concerns so we can provide timely clarifications or targeted experiments.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.