Visual-Language Model Knowledge Distillation Method for Image Quality Assessment

Jiarun Song; Yongkang Hou

REVIEW 3 major objections 5 minor 27 references

Visual-Language Model Knowledge Distillation Method for Image Quality Assessment

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A lightweight student trained by CLIP knowledge distillation can beat much larger IQA models.

desk verdict Good method packaging, broken evaluation: the random image-level split on CSIQ/LIVE leaks near-duplicate content, so the headline SOTA numbers are not credible as stated. read the letter →

arxiv 2507.15680 v3 pith:C26FJ66P submitted 2025-07-21 cs.CV

Yongkang Hou , Jiarun Song This is my paper · ORCID

classification cs.CV

keywords imagequalityassessmentvision-languagemodelknowledgedistillationCLIPlightweightstudentprompt-basedscoringcosineannealingno-referenceIQA

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the image-quality knowledge of a large vision-language model such as CLIP can be transferred to a small image encoder that is cheap enough to deploy. The method turns CLIP into a quality scorer with five quality-graded text prompts, fine-tunes only CLIP's image encoder on human quality labels, and then trains a lightweight student under a cosine-annealed mixture of soft teacher-feature alignment and hard label supervision. On four IQA datasets the distilled students, led by a 28M-parameter Swin Transformer, are reported to match or surpass specialized IQA baselines and 151M-parameter CLIP-based methods. If true, the practical consequence is that CLIP-scale semantic quality priors can be compressed into mobile-friendly encoders without giving up accuracy on either synthetic or real-world distortions.

What carries the argument

The load-bearing mechanism is the modality-adaptive knowledge distillation loop, with a fixed five-prompt template set defining the quality semantic space, a fine-tuning stage that updates only the CLIP image encoder, and a training objective $\mathcal{L} = \lambda(t)\mathcal{L}_{soft} + (1-\lambda(t))\mathcal{L}_{hard}$, where the soft loss maximizes cosine similarity between student and teacher image features and the hard loss is MSE against human quality scores. The cosine-annealed weight $\lambda(t)=\frac{1}{2}(1+\cos(\frac{t}{T}\pi))$ starts near 1 so the student first aligns with CLIP's semantic space, then drifts toward 0 so the student can refine local distortions with its own architecture. This scheduling is what the paper credits for transferring CLIP's knowledge while letting the student escape the teacher's constraints.

What would settle it

Re-run the same distillation on CSIQ and LIVE with a content-disjoint split, keeping all distorted versions of each source image entirely in either training or testing; if the near-0.990 PLCC/SROCC values drop materially under that protocol, the reported advantage comes largely from train/test content overlap rather than from the distillation strategy itself.

Watch

Extended reading notes

Core claim

The central claim is that a fixed five-level Likert prompt set gives CLIP a usable IQA scoring head, that fine-tuning only the image encoder sharpens its quality perception without breaking the text side, and that a modality-adaptive distillation stage then lets a student encoder inherit that knowledge and improve on it. The student is built by swapping CLIP's image encoder for a smaller local-feature-oriented encoder (Swin Transformer T, ResNet-18, MobileViT S, or EfficientNet B0) while keeping the frozen teacher text features and the same scoring weights. Training mixes a soft cosine-similarity loss against the teacher's image features with a hard MSE loss against human scores, with the soft weight $\lambda(t)=\frac{1}{2}(1+\cos(\frac{t}{T}\pi))$ decaying from 1 toward 0 across training. On CSIQ and LIVE the 28M-parameter Swin student tops every baseline listed in the paper, reaching PLCC 0.990 and SROCC 0.988 on CSIQ; on KonIQ and SPAQ it is at or near the top of the same table.

Load-bearing premise

The load-bearing assumption is that the random 80/20 image-level split used for training and testing keeps the CSIQ and LIVE results honest, even though many distorted versions of the same source image can land on both sides of that split.

Editorial extensions

If this is right

A 28M-parameter Swin Transformer student outperforms every baseline listed in the paper's comparison on CSIQ (PLCC 0.990, SROCC 0.988) and LIVE (PLCC 0.985, SROCC 0.982), including 151M-parameter CLIP-based methods, which is the paper's headline evidence that a heavy teacher is not needed at inference time.
Students as small as 5M parameters (EfficientNet B0) reach PLCC 0.981 on CSIQ and 0.977 on LIVE, so the reported benefit of the distillation pipeline is not tied to one architecture or parameter budget.
Because the five text features are precomputed and the text encoder stays frozen throughout, a deployed student needs only its small image encoder plus five fixed vectors to produce a quality score.
The ablation table shows that dropping either supervision signal hurts: hard-label-only training for the Swin student on CSIQ falls to PLCC 0.841, and soft-label-only training falls short of the full strategy, which is the paper's evidence that both teacher alignment and human labels contribute.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A content-disjoint split of CSIQ and LIVE, keeping all distorted versions of each source image on one side of the train/test boundary, would clarify how much of the near-0.99 scores depends on the random 80/20 split placing near-duplicate content in both training and testing.
The same prompt-and-distillation recipe could plausibly carry over to aesthetic scoring or other subjective image-perception tasks, because the teacher's quality space is defined by language templates rather than by dataset-specific labels; that is a natural test of whether the mechanism is about CLIP's semantic priors or about IQA labels specifically.
A sweep of annealing schedules (cosine versus linear versus stepwise decay, and different starting soft-weight values) would test whether the gradual teacher-to-student handoff itself, rather than the mere presence of soft labels, produces the reported gains; the ablation table varies components but not the schedule shape.
Because all student models start from ImageNet-1K pretrained weights, training a randomly initialized student under the same distillation would separate the contribution of the teacher's knowledge transfer from the contribution of generic pretraining.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The short version: the method is a reasonable packaging of existing ideas, but the evaluation protocol sinks the headline numbers. The paper reports state-of-the-art on CSIQ and LIVE using a random 80/20 image-level split. On those two datasets, which contain multiple distorted versions of a handful of source images, that split leaks near-duplicate content into both train and test. Published baselines use content-disjoint splits. So Table I is not comparing like for like, and the 0.99 PLCC on CSIQ is not credible as a real gain over QFM-IQM at 0.965. No split indices, code, or error bars are provided, so the numbers are uncheckable. That is the load-bearing flaw, and it is enough to reject.

What the paper does well: the writing is clear, the problem is real (CLIP is too heavy for edge deployment), and the proposed pipeline is coherent: prompt-based scoring, fine-tune only the CLIP image encoder, then distill into a lightweight student with a cosine-annealed soft-to-hard label schedule. The specific combination appears not to have been published before. The ablation, apart from one odd collapse, shows the full strategy beats direct regression and soft-label-only variants. The students are genuinely small (5-28M parameters).

The soft spots: first, the split protocol, as above. Second, the ablation's hard-label-only variant collapses on every dataset (e.g., 0.841 vs 0.987 on CSIQ for the Swin student). That looks like a mis-specified loss or a bug, and the paper does not explain it. Third, no significance tests or variance for the median numbers they report. Fourth, the novelty is incremental; every component is established in the literature (CLIP-IQA, LIQE, CLIP-KD, cosine annealing). The paper also does not compare against other distillation-based IQA methods.

The citation pattern is fine; they cite the relevant CLIP-IQA and distillation work. The math is simple and correct; the issue is the evaluation, not the derivation.

Who is this for? Someone working on efficient IQA might find the approach worth trying if the split issue is fixed, but the paper as submitted does not demonstrate that the method actually works. I would not cite it, and I would not send it to a serious referee in its current form. If the authors re-run with content-disjoint splits, report variance, and release code, a revised version might be worth looking at.

Recommendation: reject.

Referee Report

3 major / 5 minor

Summary. The paper proposes a three-stage knowledge distillation method for image quality assessment (IQA). First, quality-graded text prompts ('a photo of [bad/poor/fair/good/perfect] quality') are used with CLIP to produce quality scores via cosine similarity and softmax. Second, CLIP is fine-tuned on IQA data using MSE loss while only the image encoder is updated, keeping the text encoder and temperature frozen. Third, a modality-adaptive knowledge distillation strategy trains lightweight student image encoders (Swin Transformer T, ResNet-18, MobileViT S, EfficientNet B0) by aligning their image features to the teacher's (soft label) and also regressing to subjective quality scores (hard label), with a cosine-annealing schedule that transitions from soft to hard supervision. Experiments on CSIQ, LIVE, KonIQ, and SPAQ report PLCC/SROCC values that exceed most compared baselines, including CLIP-based methods, while using far fewer parameters.

Significance. If the reported results hold under standard evaluation protocols, the paper would make a useful contribution to deploying IQA on resource-constrained devices: the three-stage pipeline is clearly described, the ablation study covers multiple student architectures, and the proposed combination of quality-graded prompts, CLIP fine-tuning, and feature-level distillation is a sensible design. The paper also claims parameter reductions of up to 30x relative to CLIP-based teachers, which is practically relevant. However, the significance is currently undermined by the evaluation protocol: the random image-level 80/20 split on CSIQ and LIVE can leak near-duplicate content across folds, and the absence of error bars or significance tests makes the headline numbers unverifiable. The ablation results additionally show an unexplained collapse for hard-label-only training, which needs investigation before the distillation mechanism can be fully trusted.

major comments (3)

[Section III-B and Table I] The implementation details state that "Each dataset is randomly split into 80% for training and 20% for testing." On CSIQ and LIVE, which contain many distorted versions of the same source images, an image-level random split places near-duplicate content in both training and test sets. This can inflate PLCC and SROCC by allowing the model to recognize source content or distortion pattern rather than generalize to unseen images. Published IQA methods typically use content-disjoint splits (by source image), so the comparison in Table I is not on equal footing. The large reported gains on CSIQ (0.990 vs 0.965 for QFM-IQM) could be partially or wholly explained by this leakage. The authors must re-evaluate under content-disjoint splits (e.g., splitting by source image) and report those results, or justify why their random split is content-disjoint.
[Section III-C and Table I] The paper states "We conduct multiple experiments and take the median as the final performance result," but no error bars, standard deviations, or significance tests are reported anywhere. Given the large margins over baselines (e.g., 0.025 on CSIQ PLCC), the reader cannot assess whether these differences are statistically meaningful. The authors should provide per-dataset mean and standard deviation across runs (at least 5 runs) and, ideally, a paired significance test (e.g., Wilcoxon signed-rank) against the strongest baselines to support the claim of "excellent evaluation metrics."
[Table II and Section III-D] The ablation results show that hard-label-only distillation (e.g., A+hard label) collapses to PLCC 0.841 on CSIQ, far below the direct regression baseline A (0.987). This is counterintuitive because hard labels are the ground-truth quality scores, and it raises serious questions about what the soft-label supervision actually contributes. The authors should analyze why training with quality scores alone performs so poorly: is it due to loss scale, optimization instability, the student's architecture, or the way the quality score is computed (weighted sum over prompt similarities)? Without a clear explanation, the claim that the distillation strategy effectively combines teacher knowledge with student local-feature learning is not fully substantiated. The current numbers suggest that the method relies almost entirely on aligning with the teacher's features, which may not be quality-specific.

minor comments (5)

[Section II-B heading] The heading "CLIP IQA Capability Enhance" is a grammatical fragment and should be "CLIP IQA Capability Enhancement."
[Section II-C (Student Model Design)] The description of the student model is ambiguous: "keeping the rest of the structure unchanged" could imply the student retains the CLIP text encoder, which would contradict the parameter counts in Table I. Clarify that the text features are precomputed and stored, so the student only runs the image encoder at inference, and that the reported parameter counts correspond to the student image encoder alone.
[Section III-A (Experimental Steps)] The paper does not specify whether the CLIP teacher is fine-tuned separately on each dataset's training split or shared across datasets. State this explicitly, as it affects reproducibility and the interpretation of cross-dataset generalization.
[Table I] Several baselines have missing entries (e.g., TIQA and MetaIQA on SPAQ, LoDa on CSIQ). Indicate whether these are not applicable or not reported, to avoid ambiguity.
[Conclusion] The conclusion states the method achieves "excellent evaluation metrics" without acknowledging the split-protocol caveat. The wording should be tempered until content-disjoint results are available.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the distillation pipeline is trained on external human quality scores and standard teacher-student losses.

full rationale

The paper's derivation chain is self-contained relative to its inputs. The teacher CLIP is fine-tuned on the IQA dataset using MSE against subjective quality scores (Eq. 5), and the student is trained with a cosine-annealed combination of a feature-alignment soft loss (Eq. 6) and a hard-label MSE loss (Eq. 7) against the same subjective scores. The 'prediction' reported is test-set PLCC/SROCC after training; it is not obtained by construction from any fitted parameter renamed as a prediction. The student's reuse of the teacher's fixed text features is an explicit design choice in standard knowledge distillation, not a definitional identification of output with input. The loss functions and scheduling formula (Eq. 10) are stated openly and depend on external ground-truth labels, so the central improvement claim is not forced by the method's own equations. There are no load-bearing self-citations: the cited CLIP, LIQE, CLIP-KD, and student-architecture works are independent, and no uniqueness theorem or ansatz is imported from the present authors' prior work. Concerns about the random 80/20 image-level split and possible content leakage on CSIQ/LIVE pertain to evaluation comparability and result validity, not to circularity of the derivation, and therefore do not raise the circularity score under the review rules.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper relies on standard distillation and prompt-engineering practices. The hand-chosen score weights and the annealing schedule are the main free choices. No new physical or conceptual entities are postulated beyond the distillation strategy itself.

free parameters (2)

Quality-level score weights w_i = i for i = 1..5 = 1, 2, 3, 4, 5
Chosen by hand in Eq. (4) to convert the five prompt probabilities into a scalar quality score. The final score is a linear function of these weights, so the choice affects every reported metric.
Cosine annealing schedule horizon T = 100 epochs; lambda(t) = 0.5 * (1 + cos(t/T * pi))
The soft-to-hard schedule in Eqs. (9)-(10) is set by hand and tied to the 100-epoch training setting. The claimed benefit of the method depends on this schedule.

assumptions (4)

domain assumption The five natural-language templates 'a photo of [level] quality' with levels {bad, poor, fair, good, perfect} carry subjective quality semantics that CLIP can rank by cosine similarity.
Eqs. (1)-(4) rely on CLIP's text encoder having meaningful representations of these quality adjectives; no validation of prompt quality is provided.
ad hoc to paper Freezing the CLIP text encoder and temperature while fine-tuning only the image encoder preserves cross-modal alignment and improves IQA perception.
Section II-B states this design choice but provides no evidence that it is better than updating the text encoder.
domain assumption Student encoders pre-trained on ImageNet1K provide a good initialization for the IQA distillation, and their local feature modeling is an advantage for IQA.
Section III-A introduces all students as ImageNet1K pre-trained, but no analysis links local feature modeling to the observed gains.
ad hoc to paper Maximizing cosine similarity between student and teacher image features transfers quality-relevant knowledge from teacher to student.
Eq. (6) is the core distillation assumption; it is standard practice but not derived or empirically validated in this paper beyond the headline results.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Visual-Language Model Knowledge Distillation Method for Image Quality Assessment." pith.science (2026). https://pith.science/paper/C26FJ66P

@misc{pith2026250715680,
  author       = {Pith},
  title        = {Pith review of: Visual-Language Model Knowledge Distillation Method for Image Quality Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C26FJ66P}},
  note         = {Machine review of arXiv:2507.15680}
}

read the original abstract

Image Quality Assessment (IQA) is a core task in computer vision. Multimodal methods based on vision-language models, such as CLIP, have demonstrated exceptional generalization capabilities in IQA tasks. To address the issues of excessive parameter burden and insufficient ability to identify local distorted features in CLIP for IQA, this study proposes a visual-language model knowledge distillation method aimed at guiding the training of models with architectural advantages using CLIP's IQA knowledge. First, quality-graded prompt templates were designed to guide CLIP to output quality scores. Then, CLIP is fine-tuned to enhance its capabilities in IQA tasks. Finally, a modality-adaptive knowledge distillation strategy is proposed to achieve guidance from the CLIP teacher model to the student model. Our experiments were conducted on multiple IQA datasets, and the results show that the proposed method significantly reduces model complexity while outperforming existing IQA methods, demonstrating strong potential for practical deployment.

Figures

Figures reproduced from arXiv: 2507.15680 by the authors.

**Figure 1.** Overall Architecture II. METHODOLOGY A visual-language model knowledge distillation method is proposed for IQA, with the overall architecture shown in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 25 canonical work pages

[1]

Learning transferable visual models from natural language supervision,

A. Radford, J. W. Kim, C. Hallacy, et al., "Learning transferable visual models from natural language supervision," in Proc. Int. Conf. Mach. Learn., 2021, pp. 8748–8763

work page 2021
[2]

Exploring clip for assessing the look and feel of images,

J. Wang, K. C. K. Chan, C. C. Loy, "Exploring clip for assessing the look and feel of images," in Proc. AAAI Conf. Artif. Intell., vol. 37, no. 2, pp. 2555–2563, 2023

work page 2023
[3]

Blind image quality assessment via vision-language correspondence: A multitask learning perspective,

W. Zhang, G. Zhai, Y. Wei, et al., "Blind image quality assessment via vision-language correspondence: A multitask learning perspective," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2023, pp. 14071– 14081

work page 2023
[4]

An image is worth 16x16 words: Transformers for image recognition at scale,

A. Dosovitskiy, L. Beyer, A. Kolesnikov, et al., "An image is worth 16x16 words: Transformers for image recognition at scale," arXiv preprint arXiv:2010.11929, 2020

arXiv 2010
[5]

Boosting image quality assessment through efficient transformer adaptation with local feature enhancement,

K. Xu, L. Liao, J. Xiao, et al., "Boosting image quality assessment through efficient transformer adaptation with local feature enhancement," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. , 2024, pp. 2662–2672

work page 2024
[6]

Clip -kd: An empirical study of clip model distillation,

C. Yang, Z. An, L. Huang, et al., "Clip -kd: An empirical study of clip model distillation," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2024, pp. 15952–15962

work page 2024
[7]

Most apparent distortion: Full - reference image quality assessment and the role of strategy,

E. C. Larson and D. M. Chandler, "Most apparent distortion: Full - reference image quality assessment and the role of strategy," J. Electron. Imaging, vol. 19, no. 1, Mar. 2010

work page 2010
[8]

A statistical evaluation of recent full reference image quality assessment algorithms,

H. R. Sheikh, M. F. Sabir, A. C. Bovik, "A statistical evaluation of recent full reference image quality assessment algorithms," IEEE Trans. Image Process., vol. 15, no. 11, pp. 3440–3451, Nov. 2006

work page 2006

Show all 27 references

[9]

KonIQ -10k: An ecologically valid database for deep learning of blind image quality assessment,

V. Hosu, H. Lin, T. Sziranyi, et al., "KonIQ -10k: An ecologically valid database for deep learning of blind image quality assessment," IEEE Trans. Image Process., vol. 29, pp. 4041–4056, 2020

2020
[10]

Perceptual quality assessment of smartphone photography,

Y. Fang, H. Zhu, Y. Zeng, et al., "Perceptual quality assessment of smartphone photography," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2020, pp. 3677–3686

2020
[11]

Swin transformer: Hierarchical vision transformer using shifted windows,

Z. Liu, Y. Lin, Y. Cao, et al., "Swin transformer: Hierarchical vision transformer using shifted windows," in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2021, pp. 10012–10022

2021
[12]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, et al., "Deep residual learning for image recognition," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2016, pp. 770–778

2016
[13]

Mobilevit: Light -weight, general-purpose, and mobile-friendly vision transformer,

S. Mehta, M. Rastegari, "Mobilevit: Light -weight, general-purpose, and mobile-friendly vision transformer," arXiv preprint arXiv:2110.02178 , 2021

2021 arXiv
[14]

Efficientnet: Rethinking model scaling for convolutional neural networks,

M. Tan, Q. Le, "Efficientnet: Rethinking model scaling for convolutional neural networks," in Proc. Int. Conf. Mach. Learn., 2019, pp. 6105–6114

2019
[15]

Imagenet: A large-scale hierarchical image database,

J. Deng, W. Dong, R. Socher, et al., "Imagenet: A large-scale hierarchical image database," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 2009, pp. 248–255

2009
[16]

No -reference image quality assessment in the spatial domain,

A. Mittal, A. K. Moorthy, A. C. Bovik, "No -reference image quality assessment in the spatial domain," IEEE Trans. Image Process., vol. 21, no. 12, pp. 4695–4708, 2012

2012
[17]

Making a 'completely blind' image quality analyzer,

A. Mittal, R. Soundararajan, A. C. Bovik, "Making a 'completely blind' image quality analyzer," IEEE Signal Process. Lett. , vol. 20, no. 3, pp. 209–212, 2012

2012
[18]

Blind image quality assessment using a deep bilinear convolutional neural network,

W. Zhang, K. Ma, J. Yan, et al., "Blind image quality assessment using a deep bilinear convolutional neural network," IEEE Trans. Circuits Syst. Video Technol., vol. 30, no. 1, pp. 36–47, 2018

2018
[19]

Transformer for image quality assessment,

J. You, J. Korhonen, "Transformer for image quality assessment," in Proc. IEEE Int. Conf. Image Process., 2021, pp. 1389–1393

2021
[20]

Metaiqa: Deep meta-learning for no-reference image quality assessment,

H. Zhu, L. Li, J. Wu, et al., "Metaiqa: Deep meta-learning for no-reference image quality assessment," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2020, pp. 14143–14152

2020
[21]

Blindly assess image quality in the wild guided by a self -adaptive hyper network,

S. Su, Q. Yan, Y. Zhu, et al., "Blindly assess image quality in the wild guided by a self -adaptive hyper network," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2020, pp. 3667–3676

2020
[22]

No-reference image quality assessment via transformers, relative ranking, and self -consistency,

S. A. Golestaneh, S. Dadsetan, K. M. Kitani, "No-reference image quality assessment via transformers, relative ranking, and self -consistency," in Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis., 2022, pp. 1220–1230

2022
[23]

MUSIQ: Multi -scale image quality transformer,

J. Ke, Q. Wang, Y. Wang, et al., "MUSIQ: Multi -scale image quality transformer," in Proc. IEEE/CVF Int. Conf. Comput. Vis. , 2021, pp. 5148–5157

2021
[24]

Re -iqa: Unsupervised learning for image quality assessment in the wild,

A. Saha, S. Mishra, A. C. Bovik, "Re -iqa: Unsupervised learning for image quality assessment in the wild," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2023, pp. 5846–5855

2023
[25]

Data -efficient image quality assessment with attention-panel decoder,

G. Qin, R. Hu, Y. Liu, et al., "Data -efficient image quality assessment with attention-panel decoder," in Proc. AAAI Conf. Artif. Intell., 2023

2023
[26]

Adaptive feature selection for no -reference image quality assessment by mitigating semantic noise sensitivity,

X. Li, T. Gao, R. Hu, et al., "Adaptive feature selection for no -reference image quality assessment by mitigating semantic noise sensitivity," in Int. Conf. Mach. Learn., 2024, pp. 27808–27821

2024
[27]

Local distortion aware efficient transformer adaptation for image quality assessment,

K. Xu, L. Liao, J. Xiao, et al., "Local distortion aware efficient transformer adaptation for image quality assessment," arXiv preprint arXiv:2308.12001, 2023

2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Learning transferable visual models from natural language supervision,

A. Radford, J. W. Kim, C. Hallacy, et al., "Learning transferable visual models from natural language supervision," in Proc. Int. Conf. Mach. Learn., 2021, pp. 8748–8763

work page 2021

[2] [2]

Exploring clip for assessing the look and feel of images,

J. Wang, K. C. K. Chan, C. C. Loy, "Exploring clip for assessing the look and feel of images," in Proc. AAAI Conf. Artif. Intell., vol. 37, no. 2, pp. 2555–2563, 2023

work page 2023

[3] [3]

Blind image quality assessment via vision-language correspondence: A multitask learning perspective,

W. Zhang, G. Zhai, Y. Wei, et al., "Blind image quality assessment via vision-language correspondence: A multitask learning perspective," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2023, pp. 14071– 14081

work page 2023

[4] [4]

An image is worth 16x16 words: Transformers for image recognition at scale,

A. Dosovitskiy, L. Beyer, A. Kolesnikov, et al., "An image is worth 16x16 words: Transformers for image recognition at scale," arXiv preprint arXiv:2010.11929, 2020

arXiv 2010

[5] [5]

Boosting image quality assessment through efficient transformer adaptation with local feature enhancement,

K. Xu, L. Liao, J. Xiao, et al., "Boosting image quality assessment through efficient transformer adaptation with local feature enhancement," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. , 2024, pp. 2662–2672

work page 2024

[6] [6]

Clip -kd: An empirical study of clip model distillation,

C. Yang, Z. An, L. Huang, et al., "Clip -kd: An empirical study of clip model distillation," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2024, pp. 15952–15962

work page 2024

[7] [7]

Most apparent distortion: Full - reference image quality assessment and the role of strategy,

E. C. Larson and D. M. Chandler, "Most apparent distortion: Full - reference image quality assessment and the role of strategy," J. Electron. Imaging, vol. 19, no. 1, Mar. 2010

work page 2010

[8] [8]

A statistical evaluation of recent full reference image quality assessment algorithms,

H. R. Sheikh, M. F. Sabir, A. C. Bovik, "A statistical evaluation of recent full reference image quality assessment algorithms," IEEE Trans. Image Process., vol. 15, no. 11, pp. 3440–3451, Nov. 2006

work page 2006