Pith. sign in

REVIEW 3 major objections 3 minor 49 references

Parameter-Efficient Adaptation of a Multi-Stream Vision-Language Framework for Blind Image Quality Assessment

T0 review · 3 major / 3 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read On synthetic image-quality benchmarks, the standard train/test split hides real difficulty; splitting by reference content exposes it, and adapting 0.23% of the backbone's parameters recovers scores in proportion to that exposed difficulty.

desk verdict Solid protocol-inflation study; the adaptation-proportionality claim is a plausible hypothesis but the current experiments don't statistically secure it. read the letter →

arxiv 2606.02002 v3 pith:EA6C4PQ3 submitted 2026-06-01 cs.CV

Bishr Omer Adam , Xu Li This is my paper · ORCID
classification cs.CV
keywords blindimagequalityassessmentvision-languagemodelsparameter-efficientfine-tuningLoRAcontentleakagereference-levelsplittingevaluationprotocolfrozenfeatures
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

Blind image quality assessment predicts how humans rate image quality without access to a pristine reference. This paper builds a three-stream BIQA model from frozen vision-language embeddings plus classical natural-scene statistics, then asks when adapting the backbone is worth the cost. It shows that the conventional way of splitting synthetic datasets—randomly assigning distorted images to train and test—lets distorted versions of the same reference image appear on both sides, inflating frozen-feature rank correlations by up to 0.44 and making all datasets look similarly easy. Under reference-level splitting, frozen features score from 0.91 down to 0.51 across five synthetic datasets; applying Low-Rank Adaptation (LoRA) to just 0.23% of the SigLIP backbone's parameters yields gains as large as +0.357 SROCC on TID2013, where frozen features were weakest, and almost nothing where they were already strong. The paper's central point: the evaluation protocol, not just the model, determines whether backbone adaptation appears valuable.

What carries the argument

The mechanism is a controlled comparison of two evaluation protocols. Image-level splitting randomly assigns distorted images to train and test partitions; reference-level splitting assigns all distorted versions of the same reference image to a single partition, removing content leakage. The model itself is a three-stream fusion of a 138-dimensional NSS descriptor, frozen SigLIP embeddings, and frozen CLIP-H embeddings, standardized, PCA-whitened to 256 dimensions per VLM stream, concatenated into a 650-dimensional vector, and passed to an MLP head. LoRA injects trainable low-rank matrices into the query and value projections of SigLIP's attention layers, training 995,328 parameters (0.23%

What would settle it

Re-run the adaptation study with multiple reference-level splits (for example, 5-fold cross-validation by reference image) and report the distribution of gains. If the +0.357 TID2013 gain collapses or the monotonic relationship between frozen-feature SROCC and LoRA gain disappears across folds, the proportionality claim fails. Also recompute CSIQ and LIVE-MD with fixed epochs instead of selecting the best test-correlation epoch; if their small gains become consistently negative, the 'no benefit where strong' conclusion would still stand but the clean monotonic story would need revision.

Watch

Extended reading notes

Core claim

The paper's central claim is that LoRA adaptation of a frozen SigLIP backbone recovers performance in proportion to the difficulty exposed by reference-level splitting, and that this relationship is invisible under image-level splitting. With content overlap removed, frozen-feature SROCC on synthetic benchmarks spans 0.514 (TID2013) to 0.912 (CSIQ). LoRA with rank 8 and scaling 16, trained jointly with a lightweight regression head, lifts TID2013 to 0.871 (+0.357), PIPAL from 0.576 to 0.707, and KADID-10k from 0.787 to 0.927, while leaving already-strong datasets essentially unchanged (CSIQ +0.015, LIVE-MD -0.011). On KonIQ-10k, an authentic dataset with no reference structure, adaptation ra

Load-bearing premise

The load-bearing premise is that a single held-out reference-level split per dataset represents that dataset's true difficulty; if the chosen splits are unrepresentative, or if selecting the best test-correlation epoch on CSIQ and LIVE-MD inflates adapted scores, the ordering of LoRA gains across datasets could change and the 'gains track difficulty' conclusion would weaken.

Editorial extensions

If this is right

  • Reported SROCC on synthetic BIQA benchmarks is protocol-dependent; image-level splitting can overstate frozen-feature performance by up to 0.44, so cross-paper comparisons are only meaningful when the split protocol is identical.
  • A frozen VLM backbone can be adequate on some datasets (CSIQ, LIVE-MD, KonIQ) but seriously weak on others (TID2013, PIPAL), so 'frozen versus adapted' conclusions should be drawn per dataset under reference-level evaluation.
  • Lightweight adaptation of just 0.23% of the backbone's parameters can recover most of the performance lost under content-independent evaluation, making parameter-efficient tuning a practical alternative to full fine-tuning for resource-constrained BIQA.
  • Because adaptation gains scale with frozen-feature weakness, the reference-level frozen score can serve as a predictor of when backbone adaptation is worthwhile.
  • The lack of correlation between inflation gap and reference count challenges a simple content-overlap heuristic; the nature of the distortions, rather than how many reference images a dataset has, drives how much frozen features can lean on memorized content.

Reading between the lines

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

  • If the proportionality between frozen-feature difficulty and LoRA gain holds across more datasets, a practical rule follows: measure frozen-feature SROCC under reference-level splitting first, and apply adapter-based tuning only when it falls below a threshold, saving compute on easier benchmarks.
  • The protocol effect likely extends beyond SigLIP and LoRA to other PEFT methods (adapters, visual prompt tuning) and other frozen VLM backbones, though the paper only tests one combination.
  • A natural extension is adapting CLIP-H or both VLM streams jointly; on TID2013 and PIPAL the single-backbone adapter still leaves a gap (e.g., 0.871 versus 0.91+ on CSIQ), so multi-backbone adaptation might close it further.
  • The 0.06–0.44 inflation gap remains without a quantitative predictor, suggesting an open dataset-design question: identifying which distortion properties cause a frozen model to rely on memorized content could turn the gap into a predictable function of benchmark statistics.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper proposes a BIQA framework that combines a 138-dimensional NSS descriptor with frozen SigLIP and CLIP-H embeddings through a lightweight regression head, and then studies parameter-efficient LoRA adaptation of the SigLIP backbone (0.23% of parameters). Across six datasets, the authors compare image-level and reference-level splitting protocols on synthetic benchmarks, report that image-level splitting inflates frozen-feature SROCC by up to 0.44, and claim that LoRA adaptation recovers performance in proportion to the difficulty exposed by reference-level splitting, with the largest gain on TID2013 (+0.357 SROCC) and negligible gains where frozen features are already strong.

Significance. The protocol-inflation finding is practically important and actionable: it adds systematic evidence for reference-level evaluation of synthetic-distortion IQA benchmarks, an issue with known reproducibility consequences. The efficient LoRA adaptation study, with public code and a small trainable-parameter budget, is a useful reference point for practitioners. The main quantitative claim — that adaptation benefit is proportional to frozen-feature difficulty — is plausible and worth publishing, but the current statistical support is insufficient: single held-out splits, best-test-epoch selection on two datasets, and missing error bars leave the central ranking non-secured.

major comments (3)
  1. [Section IV-D, Table IV, Section VI-C] The central claim that LoRA adaptation 'recovers performance in proportion to exposed difficulty' rests on a single held-out split per dataset and on reporting the best test-correlation epoch for CSIQ and LIVE-MD. The manuscript itself concedes (Section VI-C) high epoch-to-epoch variance on these datasets. With this protocol, the CSIQ (+0.015) and LIVE-MD (−0.011) rows are within selection noise, and the TID2013 +0.357 gain may be upwardly selected. Moreover, the frozen baseline is not described as receiving the same best-epoch selection, so the comparison is asymmetric. Please report means and standard deviations over multiple splits or training seeds, or apply a fixed validation-based early-stopping rule to both frozen and adapted models, and scale the proportionality claim to what the intervals support.
  2. [Table IV, KonIQ row; Section IV-D] The KonIQ-10k frozen entry in Table IV (0.887) is not the same as the three-stream frozen SROCC of 0.914 reported in Table II and 0.910 in Section III-C. Section IV-D states that KonIQ results normally use 5-fold cross-validation while the backbone-adaptation experiments use a single held-out split. It must be made explicit whether Table IV reports LoRA on the SigLIP stream only or on the full three-stream model, and whether 0.887 is the correspondingly matched frozen baseline. As written, the +0.064 KonIQ gain and its place in the proportionality argument are not a controlled comparison.
  3. [Section V-D, Table III] The inflation-gap analysis in Table III is based on single image-level and reference-level partitions per dataset, and the statement that the gap is 'not explained by reference count' uses a rank correlation computed over only five datasets with no interval. This is a secondary but prominent claim in the abstract and conclusions. Please provide multiple random splits with intervals, or explicitly label this part as exploratory and rephrase the abstract/conclusion strength accordingly.
minor comments (3)
  1. [Section V-C/V-D] Section V-C and V-D have the same title and the first paragraph of V-D repeats the paragraph of V-C. The duplicate should be removed and the section numbering fixed.
  2. [Table II caption] The caption contains an apparent typo: 'LIVE-itW = ITSC LIVE C COLUMN' should read 'LIVE-itW = its LIVE Challenge column' or similar.
  3. [Table IV] The table caption should state explicitly the model configuration and split protocol for every row (e.g., SigLIP-only versus three-stream, single split versus 5-fold), since the current default of cross-referencing Section IV-D is easy to misread.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: adaptation gains are measured differences and the difficulty-proportionality claim is post-hoc description, not a fitted identity.

full rationale

The paper's central claims are measured differences, not derived identities. The frozen-vs-LoRA comparison (Table IV) uses the identical reference-level split for both conditions, as stated in Section IV-D: 'In each comparison the frozen and adapted models use the identical reference-level split, so that differences reflect adaptation alone.' The gain is therefore an empirical quantity, not a parameter fitted to produce the claimed proportionality. The 'adaptation value tracks frozen-feature difficulty' statement is a post-hoc description of six measured points, not an equation whose output equals its input. The protocol-inflation finding (Table III) is likewise a direct comparison of two splitting protocols on the same frozen model. The only self-citations (refs [18]–[20]) support the NSS/Benford descriptor, and the central protocol/adaptation conclusion does not depend on those features; no uniqueness theorem or ansatz is imported from the authors' prior work. The paper's stated limitations—single held-out split per dataset, best-test-epoch selection on CSIQ and LIVE-MD, and high epoch-to-epoch variance—are statistical robustness concerns, not circularity. No equation in the paper reduces to its own input, and no prediction is a renamed fitted parameter.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central empirical claims rest on evaluation-protocol choices (reference-level splitting, best-epoch selection) and on standard hyperparameters; no new physical or architectural entities are introduced. The free parameters are disclosed but not swept for sensitivity, so the effect-size claims in Tables III/IV are conditional on them.

free parameters (3)
  • Hybrid loss weights λ_PLCC, λ_rank = 0.5 each
    Set on a held-out KonIQ-10k validation split and fixed across all datasets; affects the trained regression head and downstream SROCC comparisons.
  • LoRA rank r and scale α = r=8, α=16
    Chosen for the SigLIP adaptation study (Section III-G); the adaptation gains and their ranking across datasets depend on this capacity choice.
  • PCA dimensionality per VLM stream = 256 components
    Selected from an ablation on KonIQ-10k (Table V) as a trade-off; affects frozen-feature quality and the entire protocol comparison.
assumptions (3)
  • domain assumption Reference-level splitting yields content-independent partitions and is the correct protocol for measuring generalization on synthetic BIQA benchmarks.
    Section IV-C defines the protocol and verifies zero reference overlap, but does not verify that distortion-type distributions stay balanced across splits; the central difficulty/adaptation conclusions depend on this.
  • ad hoc to paper Reporting the best test correlation over training epochs for CSIQ and LIVE-MD is an acceptable substitute for early stopping.
    Section IV-D states this practice for the two smallest datasets; it assumes the selection does not materially overstate adaptation gains relative to the frozen baseline.
  • domain assumption The 138-dimensional NSS descriptor and Benford-style features provide signal complementary to frozen VLM embeddings.
    Section III-B/III-E motivate this from prior work; Table VI supports complementarity only through KonIQ-10k SROCC, so the fusion gain is not independently verified across all datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Parameter-Efficient Adaptation of a Multi-Stream Vision-Language Framework for Blind Image Quality Assessment." pith.science (2026). https://pith.science/paper/EA6C4PQ3

@misc{pith2026260602002,
  author       = {Pith},
  title        = {Pith review of: Parameter-Efficient Adaptation of a Multi-Stream Vision-Language Framework for Blind Image Quality Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EA6C4PQ3}},
  note         = {Machine review of arXiv:2606.02002}
}
abstract

Blind image quality assessment (BIQA) predicts perceived image quality without access to a pristine reference and is fundamental to applications such as image compression, transmission, and restoration. Recent BIQA methods increasingly rely on large vision-language models (VLMs). Although frozen VLMs provide an efficient alternative to computationally expensive full fine-tuning, it remains unclear how much performance is sacrificed by not adapting the backbone and, more importantly, under what conditions such adaptation is truly beneficial. Answering this question, however, is complicated by the widespread use of image-level splitting on synthetic-distortion benchmarks, where distorted versions of the same reference image can appear in both training and test partitions. This content overlap artificially inflates the apparent performance of frozen representations, masking their true generalization ability and potentially leading to incorrect conclusions about the value of backbone adaptation. We therefore address these two issues jointly. We develop an efficient BIQA framework that fuses a natural-scene-statistics descriptor with frozen SigLIP and CLIP-H embeddings through a lightweight regression head, and then apply parameter-efficient Low-Rank Adaptation (LoRA) to the SigLIP backbone, training only $0.23\%$ of its parameters. Evaluating both frozen and adapted models across six datasets under image-level and reference-level protocols, we find that image-level splitting inflates frozen-feature SROCC by up to $0.44$ and masks wide variation in true difficulty, which reference-level evaluation reveals. Under this content-independent protocol, LoRA adaptation recovers performance in proportion to the exposed difficulty, with the largest gains where frozen features generalize poorly (up to $+0.357$ SROCC on TID2013) and little benefit where they are already strong.

Figures

Figures reproduced from arXiv: 2606.02002 by the authors.

Figure 1
Figure 1. Overall architecture of the proposed framework. Three frozen extractors (NSS, SigLIP, CLIP-H) [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 1
Figure 1. Overall architecture of the proposed framework. Three frozen streams (NSS, SigLIP, CLIP-H) are processed in parallel, dimensionality-reduced, and [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Per-distortion SROCC on KADID-10k for the full NSS + SigLIP + CLIP-H model, sorted ascending. [PITH_FULL_IMAGE:figures/full_fig_p021_2.png] view at source ↗
Figures from the paper (6 more)
Figure 2
Figure 2. Figure 2: Frozen SigLIP SROCC under image-level and reference-level splitting [PITH_FULL_IMAGE:figures/full_fig_p007_2.png]
Figure 3
Figure 3. Figure 3: Per-distortion delta SROCC (NSS+SigLIP+CLIP-H minus SigLIP+CLIP-H) on KADID-10k. [PITH_FULL_IMAGE:figures/full_fig_p022_3.png]
Figure 4
Figure 4. Figure 4: Test SROCC versus adaptation epoch for three representative datasets [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 4
Figure 4. Figure 4: Gate analysis on KADID-10k [PITH_FULL_IMAGE:figures/full_fig_p023_4.png]
Figure 3
Figure 3. Figure 3: Frozen versus LoRA-adapted SigLIP under reference-level evaluation [PITH_FULL_IMAGE:figures/full_fig_p008_3.png]
Figure 5
Figure 5. Figure 5: Gating input ablation on KADID-10k. Left: SROCC is nearly identical across all three options. Right: NSS-only gating loses interpretability (ρ = 0.097, p = 0.642) while SigLIP-only achieves ρ = 0.33. tonomously discovers distortion-stream affinity patterns consistent w…

Review history (3 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references

  1. [1]

    AGIQA-3k: An open database for AI-generated image quality assessment,

    J. Li, T. Wu, J. Wang, P. Zhou, Z. Yang, and Q. Yan, “AGIQA-3k: An open database for AI-generated image quality assessment,”IEEE Trans. Circuits Syst. Video Technol., vol. 34, no. 8, pp. 6833–6846, 2024

    2024

  2. [2]

    AIG- CIQA2023: A large-scale image quality assessment database for AI generated images: From the perspectives of quality, authenticity and correspondence,

    J. Wang, H. Duan, J. Liu, S. Chen, X. Min, and G. Zhai, “AIG- CIQA2023: A large-scale image quality assessment database for AI generated images: From the perspectives of quality, authenticity and correspondence,” inProc. CAAI Int. Conf. Artif. Intell., 2023, pp. 46– 57

    2023

  3. [3]

    No-reference image quality assessment in the spatial domain,

    A. Mittal, A. K. Moorthy, and 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

  4. [4]

    Making a ‘completely blind’ image quality analyzer,

    A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a ‘completely blind’ image quality analyzer,”IEEE Signal Process. Lett., vol. 20, no. 3, pp. 209–212, 2013. 10

    2013

  5. [5]

    A feature-enriched completely blind image quality evaluator,

    L. Zhang, L. Zhang, and A. C. Bovik, “A feature-enriched completely blind image quality evaluator,”IEEE Trans. Image Process., vol. 24, no. 8, pp. 2579–2591, 2015

    2015

  6. [6]

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

    W. Zhang, K. Ma, J. Yan, D. Deng, and Z. Wang, “Blind image quality assessment using a deep bilinear convolutional neural network,”IEEE Trans. Circuits Syst. Video Technol., vol. 30, no. 1, pp. 36–47, 2020

    2020

  7. [7]

    MANIQA: Multi-dimension attention network for no-reference image quality assessment,

    S. Yang, T. Wu, S. Shi, S. Lao, Y . Gong, M. Cao, J. Wang, and Y . Yang, “MANIQA: Multi-dimension attention network for no-reference image quality assessment,” inProc. IEEE/CVF CVPRW, 2022, pp. 1191–1200

    2022

  8. [8]

    Exploring CLIP for assessing the look and feel of images,

    J. Wang, K. C. K. Chan, and C. C. Loy, “Exploring CLIP for assessing the look and feel of images,” inProc. AAAI, vol. 37, no. 2, 2023, pp. 2555–2563

    2023

Show all 49 references
  1. [9]

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

    W. Zhang, G. Zhai, Y . Wei, X. Yang, and K. Ma, “Blind image quality assessment via vision-language correspondence: A multitask learning perspective,” inProc. IEEE/CVF CVPR, 2023, pp. 14071–14081

    2023

  2. [10]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” inProc. ICLR, 2022

    2022

  3. [11]

    KADID-10k: A large-scale artificially distorted IQA database,

    H. Lin, V . Hosu, and D. Saupe, “KADID-10k: A large-scale artificially distorted IQA database,” inProc. QoMEX, 2019, pp. 1–3

    2019

  4. [12]

    Image database TID2013: Peculiarities, results and perspectives,

    N. Ponomarenko, L. Jin, O. Ieremeiev, V . Lukin, K. Egiazarian, J. Astola, B. V ozel, K. Chehdi, M. Carli, F. Battisti, and C.-C. J. Kuo, “Image database TID2013: Peculiarities, results and perspectives,”Signal Pro- cess. Image Commun., vol. 30, pp. 57–77, 2015

    2015

  5. [13]

    Critical analysis on the repro- ducibility of visual quality assessment using deep features,

    F. G ¨otz-Hahn, V . Hosu, and D. Saupe, “Critical analysis on the repro- ducibility of visual quality assessment using deep features,”PLOS ONE, vol. 17, no. 8, p. e0269715, 2022

    2022

  6. [14]

    Adaptive image quality assessment via teaching large multimodal model to compare,

    H. Zhu, X. Wu, Q. Li, H. Zhang, Y . Xu, X. Yang, and W. Zhou, “Adaptive image quality assessment via teaching large multimodal model to compare,” inAdv. Neural Inf. Process. Syst. (NeurIPS), 2024

    2024

  7. [15]

    Sigmoid loss for language image pre-training,

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer, “Sigmoid loss for language image pre-training,” inProc. IEEE/CVF ICCV, 2023, pp. 11975–11986

    2023

  8. [16]

    OpenCLIP,

    G. Ilharco, M. Wortsman, R. Wightman, C. Gordon, N. Carlini, R. Taori, A. Dave, V . Shankar, H. Namkoong, J. Miller, H. Hajishirzi, A. Farhadi, and L. Schmidt, “OpenCLIP,” Software, 2021. [Online]. Available: https: //github.com/mlfoundations/open clip

    2021

  9. [17]

    Blind image quality assessment: From natural scene statistics to perceptual quality,

    A. K. Moorthy and A. C. Bovik, “Blind image quality assessment: From natural scene statistics to perceptual quality,”IEEE Trans. Image Process., vol. 20, no. 12, pp. 3350–3364, 2011

    2011

  10. [18]

    A no-reference quality assessment method for hyperspectral sharpened images via Benford’s law,

    X. Hao, X. Li, J. Wu, B. Wei, Y . Song, and B. Li, “A no-reference quality assessment method for hyperspectral sharpened images via Benford’s law,”Remote Sens., vol. 16, no. 8, p. 1167, 2024

    2024

  11. [19]

    A no-reference spectral quality assessment method for multispectral pansharpening,

    J. Wu, X. Li, B. Wei and L. Li, “A no-reference spectral quality assessment method for multispectral pansharpening,” inProc. IEEE Int. Geosci. Remote Sens. Symp. (IGARSS), Pasadena, CA, USA, Jul. 2023

    2023

  12. [20]

    A no-reference multivariate Gaussian-based spectral distortion index for pansharpened images,

    B. O. Adam, X. Li, J. Wu, and X. Hao, “A no-reference multivariate Gaussian-based spectral distortion index for pansharpened images,” Sensors, vol. 26, 2026

    2026

  13. [21]

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

    S. Su, Q. Yan, Y . Zhu, C. Zhang, X. Ge, J. Sun, and Y . Zhang, “Blindly assess image quality in the wild guided by a self-adaptive hyper network,” inProc. IEEE/CVF CVPR, 2020, pp. 3667–3676

    2020

  14. [22]

    SGDR: Stochastic gradient descent with warm restarts,

    I. Loshchilov and F. Hutter, “SGDR: Stochastic gradient descent with warm restarts,” inProc. ICLR, 2017

    2017

  15. [23]

    Uncertainty-aware blind image quality assessment in the laboratory and in the wild,

    W. Zhang, K. Ma, G. Zhai, and X. Yang, “Uncertainty-aware blind image quality assessment in the laboratory and in the wild,”IEEE Trans. Image Process., vol. 30, pp. 3474–3486, 2021

    2021

  16. [24]

    MUSIQ: Multi-scale image quality transformer,

    J. Ke, Q. Wang, Y . Wang, P. Milanfar, and F. Yang, “MUSIQ: Multi-scale image quality transformer,” inProc. IEEE/CVF ICCV, 2021, pp. 5148– 5157

    2021

  17. [25]

    IQA vision transformed: A survey of transformer architectures in perceptual image quality assessment,

    M. U. Rehman, I. F. Nizami, F. Ullah, and I. Hussain, “IQA vision transformed: A survey of transformer architectures in perceptual image quality assessment,”IEEE Access, vol. 12, pp. 183369–183393, 2024

    2024

  18. [26]

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

    A. Saha, S. Mishra, and A. C. Bovik, “Re-IQA: Unsupervised learning for image quality assessment in the wild,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 5846–5855

    2023

  19. [27]

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

    G. Qin, R. Hu, Y . Liu, X. Zheng, H. Liu, X. Li, and Y . Zhang, “Data- efficient image quality assessment with attention-panel decoder,” in Proc. AAAI Conf. Artif. Intell., vol. 37, no. 2, 2023, pp. 2091–2100

    2023

  20. [28]

    Learning transferable visual models from natural language supervi- sion,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervi- sion,” inProc. ICML, 2021, pp. 8748–8763

    2021

  21. [29]

    Quality-aware pre- trained models for blind image quality assessment,

    S. Zhao, H. Zhou, X. Li, Y . Fang, and H. Zhu, “Quality-aware pre- trained models for blind image quality assessment,” inProc. IEEE/CVF CVPR, 2023, pp. 22–31

    2023

  22. [30]

    Q-Align: Teaching LMMs for visual scoring via discrete text-defined levels,

    H. Wu, Z. Zhang, W. Zhang, C. Chen, L. Liao, C. Li, Y . Gao, A. Wang, E. Zhang, W. Sun, Q. Yan, X. Min, G. Zhai, and W. Lin, “Q-Align: Teaching LMMs for visual scoring via discrete text-defined levels,” in Proc. ICML, 2024, pp. 54015–54029

    2024

  23. [31]

    Teaching large language models to regress accurate image quality scores using score distribution,

    Z. You, X. Cai, J. Gu, T. Xue, and C. Dong, “Teaching large language models to regress accurate image quality scores using score distribution,” inProc. IEEE/CVF CVPR, 2025, pp. 14483–14494

    2025

  24. [32]

    TOPIQ: A top-down perspective for image quality assessment,

    X. Chen, H. Zhang, R. Lu, L. Gu, X. Chu, J. Qiao, and C. Dong, “TOPIQ: A top-down perspective for image quality assessment,”IEEE Trans. Image Process., vol. 33, pp. 2940–2955, 2024

    2024

  25. [33]

    ARNIQA: Learning distortion manifold for image quality assessment,

    L. Agnolucci, L. Galteri, M. Bertini, and A. Del Bimbo, “ARNIQA: Learning distortion manifold for image quality assessment,” inProc. IEEE/CVF WACV, 2024, pp. 3344–3353

    2024

  26. [34]

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

    K. Xu, L. Liao, J. Xiao, C. Chen, H. Wu, Q. Yan, and W. Lin, “Boosting image quality assessment through efficient transformer adaptation with local feature enhancement,” inProc. IEEE/CVF CVPR, 2024, pp. 2662– 2672

    2024

  27. [35]

    Parameter-efficient transfer learning for NLP,

    N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly, “Parameter-efficient transfer learning for NLP,” inProc. Int. Conf. Mach. Learn. (ICML), 2019, pp. 2790–2799

    2019

  28. [36]

    The power of scale for parameter-efficient prompt tuning,

    B. Lester, R. Al-Rfou, and N. Constant, “The power of scale for parameter-efficient prompt tuning,” inProc. Conf. Empir. Methods Nat. Lang. Process. (EMNLP), 2021

    2021

  29. [37]

    Visual prompt tuning,

    M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim, “Visual prompt tuning,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2022, pp. 709–727

    2022

  30. [38]

    Adaptive budget allocation for parameter-efficient fine-tuning,

    Q. Zhang, M. Chen, A. Bukharin, P. He, Y . Cheng, W. Chen, and T. Zhao, “Adaptive budget allocation for parameter-efficient fine-tuning,” inProc. Int. Conf. Learn. Represent. (ICLR), 2023

    2023

  31. [39]

    DoRA: Weight-decomposed low-rank adaptation,

    S.-Y . Liu, C.-Y . Wang, H. Yin, P. Molchanov, Y .-C. F. Wang, K.- T. Cheng, and M.-H. Chen, “DoRA: Weight-decomposed low-rank adaptation,” inProc. Int. Conf. Mach. Learn. (ICML), 2024

    2024

  32. [40]

    From patches to pictures (PaQ-2-PiQ): Mapping the perceptual space of picture quality,

    Z. Ying, H. Niu, P. Gupta, D. Mahajan, D. Ghadiyaram, and A. C. Bovik, “From patches to pictures (PaQ-2-PiQ): Mapping the perceptual space of picture quality,” inProc. IEEE/CVF CVPR, 2020, pp. 3575–3585

    2020

  33. [41]

    Q-Bench: A benchmark for general- purpose foundation models on low-level vision,

    H. Wu, Z. Zhang, E. Zhang, C. Chen, L. Liao, A. Wang, C. Li, W. Sun, Q. Yan, G. Zhai, and W. Lin, “Q-Bench: A benchmark for general- purpose foundation models on low-level vision,” inProc. ICLR, 2024

    2024

  34. [42]

    Leakage and the reproducibility crisis in machine-learning-based science,

    S. Kapoor and A. Narayanan, “Leakage and the reproducibility crisis in machine-learning-based science,”Patterns, vol. 4, no. 9, p. 100804, 2023

    2023

  35. [43]

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

    V . Hosu, H. Lin, T. Sziranyi, and D. Saupe, “KonIQ-10k: An ecologi- cally valid database for deep learning of blind image quality assessment,” IEEE Trans. Image Process., vol. 29, pp. 4041–4056, 2020

    2020

  36. [44]

    Massive online crowdsourced study of subjective and objective picture quality,

    D. Ghadiyaram and A. C. Bovik, “Massive online crowdsourced study of subjective and objective picture quality,”IEEE Trans. Image Process., vol. 25, no. 1, pp. 372–387, 2016

    2016

  37. [45]

    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, p. 011006, 2010

    2010

  38. [46]

    Objective quality assessment of multiply distorted images,

    D. Jayaraman, A. Mittal, A. K. Moorthy, and A. C. Bovik, “Objective quality assessment of multiply distorted images,” inProc. Asilomar Conf. Signals, Syst. Comput., 2012, pp. 1693–1697

    2012

  39. [47]

    PIPAL: a large- scale image quality assessment dataset for perceptual image restoration,

    J. Gu, H. Cai, H. Chen, X. Ye, J. S. Ren, and C. Dong, “PIPAL: a large- scale image quality assessment dataset for perceptual image restoration,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2020, pp. 633–651

    2020

  40. [48]

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

    S. A. Golestaneh, S. Dadsetan, and K. M. Kitani, “No-reference im- age quality assessment via transformers, relative ranking, and self- consistency,” inProc. IEEE/CVF WACV, 2022, pp. 3989–3999

    2022

  41. [49]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inProc. ICLR, 2019

    2019

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.