Pith. sign in

REVIEW 3 major objections 4 minor 69 references

The paper claims that a multiscale patch-based CLIP model with learned quality-aware saliency can assess image quality at arbitrary resolutions, preserving fine distortion cues and cutting compute by ranking patches, and demonstrates this a

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

ReLIQS, a CLIP-based multiscale patch model with learned importance-guided patch selection, achieves SOTA or near-SOTA no-reference image quality scores across authentic, synthetic, AIGC, and ultra-high-resolution benchmarks while preserving original-resolution detail and keeping compute adjustable.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Solid architecture for resolution-agnostic IQA with a genuinely new compute-control idea, but the compute-adaptive claim needs an equal-budget baseline and Table 1 has a clear copy-paste error. the 3 major comments →

arxiv 2608.01730 v1 pith:XMRBQFHI submitted 2026-08-03 cs.CV

Learning Where to Look and How to Judge: Resolution-agnostic Image Quality Assessment with Quality-aware Saliency

Hakan Emre Gedik , Shashank Gupta , Alan Bovik This is my paper
classification cs.CV
keywords no-reference image quality assessmentresolution-agnosticmultiscale patch samplingperceptual importance estimationCLIPlatent quality axesmulti-dataset trainingUHD image quality
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 build a no-reference image quality (NR IQA) model that works off the shelf on any input resolution, from small photos to ultra-high-resolution frames, without losing fine distortion cues or demanding excessive compute. Its central move is to avoid feeding whole images at off-distribution sizes to the encoder: fixed-size patches are sampled from the original image and from downscaled variants, so each patch stays close to the backbone's pretraining resolution while native-resolution texture, blur, and noise remain visible. A learned Perceptual Importance Estimator predicts where quality-relevant content lies, and those importance weights are used both to pool patch embeddings and to rank which patches to keep at a given compute budget. The paper reports that the resulting model generalizes across authentic, synthetic, and AI-generated benchmarks, and that on UHD images it outperforms CNN-, CLIP-, and multimodal-LLM baselines while using a fraction of their compute. The broader claim is that these four properties—resolution-agnostic input, preservation of low-level cues, budget-adaptive cost, and multi-dataset training—can be met simultaneously by a patch-driven CLIP architecture.

Core claim

ReLIQS claims that no-reference image quality can be assessed on arbitrary resolutions by decomposing the image into fixed-size patches at several scales and learning both where to look and how to judge. Patches from the original resolution carry low-level cues such as blur, noise, and texture breaks; patches from resized views carry semantic and compositional structure. The Perceptual Importance Estimator produces an IQA-specific saliency field on the smallest scale, upsampled to guide patch weighting, and the Latent Quality Axis Module learns a small set of quality axes whose scores are adaptively fused across scales. On UHD the original-resolution scale is decisive: using only the 224-pix

What carries the argument

The central object is the Perceptual Importance Estimator (PIE) working inside a multiscale patch pipeline. PIE is a TinyCLIP ViT-8M backbone with a convolutional decoder that runs only on the 224-short-side resized image, producing a dense importance field; that field is bilinearly upsampled to the 512 and original scales, and each sampled patch gets a normalized weight. These weights serve two purposes: they pool patch embeddings within each scale, and they rank patches at test time so the model can encode only the top-k. The other load-bearing component is the Latent Quality Axis Module (LQAM), which keeps a few learned positive/negative quality direction pairs and axis queries, computes

Load-bearing premise

The importance field that decides which full-resolution patches to keep is computed on a 224-pixel resized image, so the method assumes the regions that look quality-relevant at small scale are the same regions that contain fine, resolution-dependent distortions at full scale.

What would settle it

Take a UHD image whose only visible defect (e.g., a thin blur streak or localized noise) lies outside the top regions of the PIE map, and compare PIE-guided top-k selection against uniform random selection at the same patch budget: if uniform selection scores as well or better, the saliency-based compute savings are not coming from where to look.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • NR IQA can be deployed on arbitrary-resolution inputs without global resizing, so fine-grained artifacts that resizing hides remain available to the scorer.
  • Compute becomes a dial: selecting top-k patches per scale lets the same model run in a few GMACs on UHD content or with full coverage when budget allows.
  • Joint training across datasets with incompatible MOS scales is feasible using within-dataset ranking and correlation losses, so future models can absorb more subjective studies without scale alignment.
  • Multimodal-LLM-based scoring is not a prerequisite for top generalization; a CLIP patch pipeline can beat or match it on most evaluated benchmarks at lower cost.
  • UHD-specific assessment no longer requires specially tailored challenge models; the general-purpose multiscale design outperforms them at comparable compute.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the importance field itself may be resolution-sensitive; testing PIE computed at 512 instead of 224 would reveal whether top-k selection changes, and whether the saliency is genuinely scale-invariant.
  • Beyond the paper: the learned quality axes could be inspected against human-interpretable attributes (sharpness, noise, color, composition) to see whether the latent decomposition has semantic meaning.
  • Beyond the paper: the same patch-and-saliency recipe could transfer to other CLIP-based regression or ranking tasks—such as video quality, aesthetic scoring, or artifact localization—where resolution robustness and budget adaptivity are valuable.
  • Beyond the paper: an equal-budget comparison of PIE-guided selection versus uniform or grid selection would quantify how much of the compute savings comes from 'where to look' as opposed to mere patch count.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. ReLIQS is a no-reference image quality assessment (NR IQA) model built from a CLIP ViT-B/16 patch encoder, a Perceptual Importance Estimator (PIE) that produces dense IQA-specific saliency maps from a short-side-224 view, and a Latent Quality Axis Module (LQAM) that fuses axis-aware scale representations into a single quality score. Patches are sampled from the original resolution and from resized variants with short sides 512 and 224; PIE weights are used both for within-scale pooling and, at test time, for selecting the top-k most important patches under a compute budget. Training combines within-dataset margin-ranking and PLCC losses with uncertainty-based weighting, enabling joint training on multiple IQA datasets. The paper evaluates ReLIQS on authentic, synthetic, and AIGC benchmarks, reports a new state of the art on the UHD dataset, and presents ablations showing that original-resolution scales are critical on UHD and that PIE gives small but consistent gains.

Significance. If the claims are substantiated, ReLIQS would be a practically valuable general-purpose NR IQA model: it is resolution-agnostic by construction, preserves original-resolution quality cues, is trainable on multiple heterogeneous datasets, and has a compute-adaptive inference mode. The paper provides a broad evaluation (nine datasets), multi-dataset training comparisons against strong baselines, extensive ablations (Tables 4-6), and explicit GMACs reporting. The UHD results in particular are a strength. However, two load-bearing issues currently prevent the claims from being fully established: the AGIQA-3K column in Table 1 is identical to the LIVE column, and the central 'learning where to look' claim lacks an equal-budget comparison against uniform/random patch selection.

major comments (3)
  1. [Table 1] The AGIQA-3K column is identical to the LIVE column for every method (e.g., NIQE 0.560/0.533, DeQA 0.809/0.729, ReLIQS 0.768/0.705). This cannot be correct and invalidates the claimed AIGC comparison in the abstract and in Sec. 4.2 ('on AGIQA-3K, ReLIQS trails Q-Align and DeQA'). Please provide the actual AGIQA-3K numbers or explicitly remove/relabel the column, and re-check all statements about AIGC generalization.
  2. [Sec. 4.4, Eqs. (2)-(5)] PIE is computed only on the short-side-224 resized image and bilinearly upsampled to all other scales (Eq. 3). The top-k patch selection at the original resolution therefore assumes that importance at 224 px is a reliable proxy for the location of full-resolution-only distortions. The paper reports performance versus k under PIE selection (Fig. 2, Fig. 4) but never compares PIE-based selection with uniform or random selection at the same patch budget. Without such an equal-budget baseline, the compute-adaptive claim for ReLIQS* (47 GMACs, Table 3) and the central 'learning where to look' contribution are not established. A full-resolution artifact in a low-importance region would be discarded by this procedure. Please add an ablation comparing PIE-ranked top-k, uniform top-k, and random top-k at identical patch counts, and ideally also an importance field computed at the original resolut
  3. [Sec. 4.3 and Supp. Sec. 5.1] The patch budget for the UHD experiments is inconsistent. Main-text Sec. 4.3 says ReLIQS* reduces cost by 'limiting the number of sampled patches to 4 instead of 48', while Supp. Sec. 5.1 says that for UHD the model uses 'a total of 36 + 11 + 1 = 48 patches per image'. Table 3 lists ReLIQS* at 47 GMACs and ReLIQS at 543 GMACs. Please specify exactly how many patches at each scale are used for ReLIQS* and for full ReLIQS on UHD, and recompute/report the GMACs consistently. This is necessary to verify the claimed compute reduction.
minor comments (4)
  1. [Sec. 3.1] The notation R is used both for the number of resized variants and as an index for the smallest scale (x^(R), Eq. (2)). This is confusing; please use separate symbols or clarify the indexing convention.
  2. [Fig. 2] The caption says 'GMACs percentage', but the axes are not fully labeled. Please state explicitly what is plotted on each axis and how the GMACs percentage is computed.
  3. [Tables 4 and 5] The abbreviation 'o.' in the 'Scales' rows should be defined in the captions as 'original-resolution scale'.
  4. [Sec. 4.2] The text says ReLIQS is 'on par with, and in most cases superior to' Q-Align and DeQA, but Table 1 shows ReLIQS trailing both on AGIQA-3K. After correcting Table 1, please qualify this statement to match the actual cross-dataset results.

Circularity Check

1 steps flagged

UHD patch budget is tuned on the UHD test set and then reported as ReLIQS*'s predicted accuracy; the compute-adaptive claim is partially circular.

specific steps
  1. fitted input called prediction [Supplementary Sec. 5.1 (Fig. 4) and main Sec. 4.4 / Table 3]
    "Figure 4. SRCC on UHD test set over patch counts at original resolution and short image dimension 512. We select the (11, 36) patch configuration as a good trade-off between performance and computational cost. ... Utilizing this saturation, we choose a total of 36 + 11 + 1 = 48 patches per image, allocated from higher to lower scales, respectively, which attains near-maximum performance at substantially reduced computational cost."

    The number of patches k that defines the compute-adaptive variant ReLIQS* (Sec. 4.4, Table 3) is chosen by reading SRCC on the UHD test set (Fig. 4), and the same UHD test set is then used to report ReLIQS*'s PLCC/SRCC in Table 3. Thus the claim that ReLIQS* (47 GMACs) 'preserves accuracy' or 'attains near-maximum performance' is not an independent prediction; it is the result of selecting the evaluation hyperparameter to maximize performance on the very data being scored. The 'near-maximum' wording is self-referential: the maximum is read off the same test-set curve used for selection.

full rationale

ReLIQS's core derivation is not circular: the pipeline equations (1)-(10) are feed-forward, PIE weights are learned from MOS supervision rather than hand-set, and the main benchmark results in Tables 1 and 2 are evaluated on external datasets with standard SRCC/PLCC metrics. There is no self-definition, no load-bearing self-citation chain, and no ansatz smuggled in via citation. The one genuine circularity concern is confined to the compute-adaptive UHD evaluation: the patch budget k for ReLIQS* is selected using UHD test-set performance curves, and the resulting UHD score is then presented as the model's accuracy. This makes the specific claim 'ReLIQS* preserves accuracy at 47 GMACs' partially circular, because the hyperparameter was fitted to the very test set used for scoring. The full-resolution ReLIQS row and the cross-dataset generalization claims remain independent of this tuning, so the overall circularity is partial rather than pervasive.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central claims rest on four assumptions the text does not independently verify: CLIP embeddings preserve low-level distortion sensitivity (argued qualitatively in Sec. 2, not measured), patch sizes near CLIP pretraining resolution avoid off-distribution degradation (Sec. 1 design claim), the 224px-scale importance field transfers to original-resolution patch selection (Eqs. 2-3), and ranking+PLCC losses suffice for cross-dataset training despite incommensurate MOS scales (Sec. 3.2). Free parameters are concentrated in the evaluation protocol: the (36,11,1) UHD patch budget is fitted to test-set curves, A=4 axes are chosen by ablation with small effect, and the margin delta=0.01 is stipulated. The invented components (PIE saliency, latent quality axes) carry no independent evidence outside the paper; the axis-count ablation shows they add little (0.858 to 0.865 SRCC on UHD).

free parameters (5)
  • UHD evaluation patch budget (36 original + 11 at 512-scale + 1 at 224-scale) = 48 patches; ReLIQS* uses 4
    Selected from SRCC-vs-patch-count curves on the UHD test set (supp. Fig. 4) and KonIQ test set (supp. Fig. 5), then used in Table 3; the trade-off point is a hand choice.
  • Number of latent quality axes A = 4
    Ablation Table 5: SRCC 0.858 (1 axis) to 0.865 (4 axes); 8 axes slightly worse; negligible effect on low/mid-resolution datasets.
  • Loss margin delta = 0.01
    Eq. (11) margin hyperparameter; no sensitivity analysis reported.
  • Train-time patches per scale (original / 512 / 224) = 6 / 5 / 1
    Sec. 4.1 implementation detail; hand-chosen sampling schedule, no ablation reported.
  • Uncertainty loss weights sigma1, sigma2 (initial values) = 1.0 each (learned thereafter)
    Sec. 3.2 Eq. (14), supp. Sec. 5.5; init value stated, learned during training, early stopping applied.
axioms (5)
  • domain assumption CLIP ViT-B/16 (OpenAI) embeddings are discriminative for low-level image quality without distortion-style pretraining augmentations.
    Sec. 2 states CLIP 'does not rely on heavy distortion-style augmentations, which helps preserve sensitivity to image degradations'; the model's accuracy inherits this unverified property.
  • domain assumption Fixed-size patches near CLIP pretraining resolution avoid the off-distribution performance drop that full-image resizing causes.
    Sec. 1 design claim: 'By choosing patch sizes close to the CLIP pretraining resolution, we avoid degradations that occur when feeding off-distribution resolutions into the backbone'; no direct measurement is provided.
  • ad hoc to paper The perceptual importance field computed at short-side-224 resolution transfers to all scales via bilinear upsampling.
    Eqs. (2)-(3); all scale-importance maps are upsampled from S(x^(R)) without computing any original-resolution importance evidence.
  • domain assumption Within-dataset ranking and PLCC losses can be jointly optimized across datasets with incompatible MOS scales.
    Sec. 3.2, Eqs. (11)-(14); validated only indirectly by Table 2 results, not by a controlled analysis of scale mismatch.
  • domain assumption MOS labels faithfully represent human quality perception and can supervise patch-importance learning.
    Standard IQA assumption; used to train PIE end-to-end through Eqs. (4)-(6).
invented entities (2)
  • IQA-specific saliency fields (PIE output) no independent evidence
    purpose: Predict patch importance for weighting (Eq. 5) and for test-time top-k patch selection (Sec. 4.4).
    Trained only with indirect MOS supervision through the aggregation loss; no ground-truth importance labels exist. Fig. 3 shows maps 'strongly biased toward semantically salient regions', and the PIE ablation gain is small (KonIQ PLCC 0.951 to 0.958), so its role is mostly compute control, not a validated saliency model.
  • Latent quality axes U_A and axis queries (LQAM) no independent evidence
    purpose: Aggregate scale representations into a single score via softmax over + and - directions per axis (Eqs. 7-10).
    Latent directions with no external validation; Table 5 shows 1 axis achieves 0.858 vs 0.865 SRCC for 4 axes on UHD, so the axis interpretation contributes little beyond the architecture capacity.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Where to Look and How to Judge: Resolution-agnostic Image Quality Assessment with Quality-aware Saliency." pith.science (2026). https://pith.science/paper/XMRBQFHI

@misc{pith2026260801730,
  author       = {Pith},
  title        = {Pith review of: Learning Where to Look and How to Judge: Resolution-agnostic Image Quality Assessment with Quality-aware Saliency},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XMRBQFHI}},
  note         = {Machine review of arXiv:2608.01730}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

No-reference image quality assessment (NR IQA) has recently benefited from deep and multimodal models, yet many SOTA systems still violate at least one basic requirement: they either discard critical quality cues via aggressive resizing, fail to generalize across resolutions, cannot be jointly trained on heterogeneous IQA datasets with mismatched MOS scales, or require prohibitive computation. We present \textbf{ReLIQS}, a model for \textbf{Re}solution-agnostic \textbf{L}earning for \textbf{I}mage \textbf{Q}uality with \textbf{S}aliency, which is resolution-agnostic, preserves original-resolution quality cues, learns from multiple subjective studies, and remains computationally efficient and budget-adaptive. ReLIQS is a CLIP-based multiscale patch-driven architecture that learns both \emph{where to look} and \emph{how to judge} quality. Fixed-size patches are sampled across multiple resolutions, including the original resolution, and encoded with a CLIP vision backbone. A lightweight Perceptual Importance Estimator then predicts IQA-specific importance maps to select a small set of informative patches, and a Latent Quality Axis Module aggregates their embeddings into a single image-level score. Across authentic, synthetic, and AIGC benchmarks spanning diverse resolutions and distortions, ReLIQS generalizes better than strong CNN-, CLIP-, and MLLM-based baselines with matching or reduced computational cost.

Figures

Figures reproduced from arXiv: 2608.01730 by Alan Bovik, Hakan Emre Gedik, Shashank Gupta.

Figure 1
Figure 1. Figure 1: ReLIQS pipeline: patches from three scales (including original resolution). A lightweight [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: PLCC and SRCC with GMACs percentage on UHD with [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: SRCC on UHD test set over patch counts at original [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: SRCC on KonIQ-10K test set over patch counts at orig [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Learning progression of IQA-specific saliency maps for models trained on KonIQ-10K, CLIVE, BID, KADID, CSIQ, and LIVE. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Learned IQA-specific saliency maps (middle) with input images (left) and overlays (right). Fine-tuning purely on MOS supervi [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Learned IQA-specific saliency maps (middle) with input images (left) and overlays (right). Fine-tuning purely on MOS supervi [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Input images (left), learned IQA-specific saliency maps (middle), and UNISAL visual saliency maps (right). In these examples, [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

69 extracted references · 66 canonical work pages

  1. [1]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e Je- gou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. InIEEE/CVF International Conference on Computer Vision (ICCV), pages 9630–9640, 2021. 2

    work page 2021

  2. [2]

    Topiq: A top-down approach from semantics to distortions for image quality assessment.IEEE Transactions on Image Processing, 33:2404–2418, 2024

    Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment.IEEE Transactions on Image Processing, 33:2404–2418, 2024. 2, 7

    work page 2024

  3. [3]

    Girshick, and Kaiming He

    Xinlei Chen, Haoqi Fan, Ross B. Girshick, and Kaiming He. Improved baselines with momentum contrastive learn- ing.ArXiv, abs/2003.04297, 2020. 2

    Pith/arXiv arXiv 2003

  4. [4]

    Alexandre Ciancio, Andr ´e Luiz N Targino Targino da Costa, Eduardo A. B. da Silva, Amir Said, Ramin Samadani, and Pere Obrador. No-reference blur assessment of digital pic- tures based on multifeature classifiers.IEEE Transactions on Image Processing, 20(1):64–75, 2011. 5

    work page 2011

  5. [5]

    Multi-task learning using uncertainty to weigh losses for scene geom- etry and semantics

    Roberto Cipolla, Yarin Gal, and Alex Kendall. Multi-task learning using uncertainty to weigh losses for scene geom- etry and semantics. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7482–7491, 2018. 5

    work page 2018

  6. [6]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. InIEEE Conference on Computer Vision and Pat- tern Recognition, pages 248–255, 2009. 2

    work page 2009

  7. [7]

    Alison Noble

    Richard Droste, Jianbo Jiao, and J. Alison Noble. Unified image and video saliency modeling. InComputer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V, page 419–435, Berlin, Heidelberg, 2020. Springer-Verlag. 2

    work page 2020

  8. [8]

    Perceptual quality assessment of smartphone photog- raphy

    Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photog- raphy. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3674–3683, 2020. 3, 5

    work page 2020

  9. [9]

    Deepti Ghadiyaram and Alan C. Bovik. Massive online crowdsourced study of subjective and objective picture qual- ity.IEEE Transactions on Image Processing, 25(1):372–387,

    work page

  10. [10]

    Alireza Golestaneh, Saba Dadsetan, and Kris M

    S. Alireza Golestaneh, Saba Dadsetan, and Kris M. Kitani. No-reference image quality assessment via transformers, rel- ative ranking, and self-consistency. InIEEE/CVF Win- ter Conference on Applications of Computer Vision, pages 3989–3999, 2022. 2

    work page 2022

  11. [11]

    Saliency-guided quality assessment of screen content images.IEEE Transactions on Multimedia, 18(6):1098–1110, 2016

    Ke Gu, Shiqi Wang, Huan Yang, Weisi Lin, Guangtao Zhai, Xiaokang Yang, and Wenjun Zhang. Saliency-guided quality assessment of screen content images.IEEE Transactions on Multimedia, 18(6):1098–1110, 2016. 3

    work page 2016

  12. [12]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016. 7

    work page 2016

  13. [13]

    Momentum contrast for unsupervised visual rep- resentation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual rep- resentation learning. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9726–9735, 2020. 2

    work page 2020

  14. [14]

    Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment.IEEE Transactions on Image Processing, 29:4041–4056, 2020

    Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment.IEEE Transactions on Image Processing, 29:4041–4056, 2020. 1, 3, 5

    work page 2020

  15. [15]

    Uhd-iqa benchmark database: Push- ing the boundaries of blind photo quality assessment

    Vlad Hosu, Lorenzo Agnolucci, Oliver Wiedemann, Daisuke Iso, and Dietmar Saupe. Uhd-iqa benchmark database: Push- ing the boundaries of blind photo quality assessment. In Computer Vision – ECCV 2024 Workshops, pages 467–482, Cham, 2025. Springer Nature Switzerland. 5, 7

    work page 2024

  16. [16]

    Vlad Hosu, Marcos V . Conde, Lorenzo Agnolucci, Naba- jeet Barman, Saman Zadtootaghaj, Radu Timofte, Wei Sun, Weixia Zhang, Yuqin Cao, Linhan Cao, Jun Jia, Zijian Chen, Zicheng Zhang, Xiongkuo Min, Guangtao Zhai, Songbai Tan, Lixin Zhang, Guanghui Yue, Daekyu Kwon, Dongy- oung Kim, Seon Joo Kim, Yunchen Zhang, Xiangkai Xu, Hong Gao, Yiming Bao, Ji Shi, Xi...

    work page 2024

  17. [17]

    Convolu- tional neural networks for no-reference image quality assess- ment

    Le Kang, Peng Ye, Yi Li, and David Doermann. Convolu- tional neural networks for no-reference image quality assess- ment. InIEEE Conference on Computer Vision and Pattern Recognition, pages 1733–1740, 2014. 2

    work page 2014

  18. [18]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. InIEEE/CVF International Conference on Computer Vision, pages 5128–5137, 2021. 1, 2, 5, 6

    work page 2021

  19. [19]

    On large- batch training for deep learning: Generalization gap and sharp minima

    Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large- batch training for deep learning: Generalization gap and sharp minima. InInternational Conference on Learning Rep- resentations, 2017. 5

    work page 2017

  20. [20]

    Attiqa: General- izable image quality feature extractor using attribute-aware pretraining

    Daekyu Kwon, Dongyoung Kim, Sehwan Ki, Younghyun Jo, Hyong-Euk Lee, and Seon Joo Kim. Attiqa: General- izable image quality feature extractor using attribute-aware pretraining. InComputer Vision – ACCV 2024, pages 284– 300, Singapore, 2025. Springer Nature Singapore. 3

    work page 2024

  21. [21]

    Larson and Damon M

    Eric C. Larson and Damon M. Chandler. Most apparent dis- tortion: full-reference image quality assessment and the role of strategy.Journal Electronic Imaging, 19:011006, 2010. 3, 5

    work page 2010

  22. [22]

    Agiqa-3k: An open database for ai-generated image quality assessment.IEEE Transactions on Circuits and Sys- tems for Video Technology, 34(8):6833–6846, 2024

    Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiqa-3k: An open database for ai-generated image quality assessment.IEEE Transactions on Circuits and Sys- tems for Video Technology, 34(8):6833–6846, 2024. 3, 5, 6

    work page 2024

  23. [23]

    Distilling spatially- heterogeneous distortion perception for blind image quality assessment

    Xudong Li, Wenjie Nie, Yan Zhang, Runze Hu, Ke Li, Xiawu Zheng, and Liujuan Cao. Distilling spatially- heterogeneous distortion perception for blind image quality assessment. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2344–2354, 2025. 3

    work page 2025

  24. [24]

    Kadid-10k: A large-scale artificially distorted iqa database

    Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. InInter- national Conference on Quality of Multimedia Experience, pages 1–3, 2019. 3, 5

    work page 2019

  25. [25]

    Visual attention in objective image quality assessment: Based on eye-tracking data.IEEE Transactions on Circuits and Systems for Video Technology, 21(7):971–982, 2011

    Hantao Liu and Ingrid Heynderickx. Visual attention in objective image quality assessment: Based on eye-tracking data.IEEE Transactions on Circuits and Systems for Video Technology, 21(7):971–982, 2011. 3, 7, 8

    work page 2011

  26. [26]

    SGDR: Stochastic gradi- ent descent with warm restarts

    Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradi- ent descent with warm restarts. InInternational Conference on Learning Representations, 2017. 5

    work page 2017

  27. [27]

    Decoupled weight de- cay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight de- cay regularization. InInternational Conference on Learning Representations. OpenReview.net, 2019. 5

    work page 2019

  28. [28]

    Madhusudana, Neil Birkbeck, Yilin Wang, Balu Adsumilli, and Alan C

    Pavan C. Madhusudana, Neil Birkbeck, Yilin Wang, Balu Adsumilli, and Alan C. Bovik. Image quality assessment us- ing contrastive learning.IEEE Transactions on Image Pro- cessing, 31:4149–4161, 2022. 2, 6

    work page 2022

  29. [29]

    No-reference image quality assessment in the spa- tial domain.IEEE Transactions on Image Processing, 21 (12):4695–4708, 2012

    Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spa- tial domain.IEEE Transactions on Image Processing, 21 (12):4695–4708, 2012. 1, 2, 6

    work page 2012

  30. [30]

    completely blind

    Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Mak- ing a “completely blind” image quality analyzer.IEEE Sig- nal Processing Letters, 20(3):209–212, 2013. 1, 2, 6

    work page 2013

  31. [31]

    Visual im- portance pooling for image quality assessment.IEEE Jour- nal of Selected Topics in Signal Processing, 3(2):193–201,

    Anush Krishna Moorthy and Alan Conrad Bovik. Visual im- portance pooling for image quality assessment.IEEE Jour- nal of Selected Topics in Signal Processing, 3(2):193–201,

    work page

  32. [32]

    Blind im- age quality assessment: From natural scene statistics to per- ceptual quality.IEEE Transactions on Image Processing, 20 (12):3350–3364, 2011

    Anush Krishna Moorthy and Alan Conrad Bovik. Blind im- age quality assessment: From natural scene statistics to per- ceptual quality.IEEE Transactions on Image Processing, 20 (12):3350–3364, 2011. 1, 2

    work page 2011

  33. [33]

    Maxime Oquab, Timoth ´ee Darcet, Th´eo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Je- gou, Julien Mairal, Patr...

    work page 2024

  34. [34]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InInternational Conference on Machine Learning, pages 8748–8763. PMLR,

  35. [35]

    Avinab Saha, Sandeep Mishra, and Alan C. Bovik. Re-iqa: Unsupervised learning for image quality assessment in the wild. InIEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 5846–5855, 2023. 1, 2

    work page 2023

  36. [36]

    A novel just-noticeable-difference-based saliency-channel attention residual network for full-reference image quality predictions

    Soomin Seo, Sehwan Ki, and Munchurl Kim. A novel just-noticeable-difference-based saliency-channel attention residual network for full-reference image quality predictions. IEEE Transactions on Circuits and Systems for Video Tech- nology, 31(7):2602–2616, 2021. 3

    work page 2021

  37. [37]

    Sheikh and A.C

    H.R. Sheikh and A.C. Bovik. Image information and visual quality.IEEE Transactions on Image Processing, 15(2):430– 444, 2006. 2

    work page 2006

  38. [38]

    Sheikh, M.F

    H.R. Sheikh, M.F. Sabir, and A.C. Bovik. A statistical eval- uation of recent full reference image quality assessment al- gorithms.IEEE Transactions on Image Processing, 15(11): 3440–3451, 2006. 3, 5

    work page 2006

  39. [39]

    Blind image quality assessment based on geometric order learning

    Nyeong-Ho Shin, Seon-Ho Lee, and Chang-Su Kim. Blind image quality assessment based on geometric order learning. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12799–12808, 2024. 3

    work page 2024

  40. [40]

    Oriane Sim ´eoni, Huy V . V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha ¨el Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timoth´ee Darcet, Th´eo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie,...

    Pith/arXiv arXiv 2025

  41. [41]

    Leslie N. Smith. Cyclical learning rates for training neural networks. InIEEE Winter Conference on Applications of Computer Vision (WACV), pages 464–472, 2017. 5

    work page 2017

  42. [42]

    Rajiv Soundararajan and Alan C. Bovik. Rred indices: Re- duced reference entropic differencing for image quality as- sessment.IEEE Transactions on Image Processing, 21(2): 517–526, 2012. 2

    work page 2012

  43. [43]

    Learning generalizable perceptual represen- tations for data-efficient no-reference image quality assess- ment

    Suhas Srinath, Shankhanil Mitra, Shika Rao, and Rajiv Soundararajan. Learning generalizable perceptual represen- tations for data-efficient no-reference image quality assess- ment. InIEEE/CVF Winter Conference on Applications of Computer Vision, pages 22–31, 2024. 3

    work page 2024

  44. [44]

    Blindly assess image qual- ity in the wild guided by a self-adaptive hyper network

    Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image qual- ity in the wild guided by a self-adaptive hyper network. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3664–3673, 2020. 2, 5, 6

    work page 2020

  45. [45]

    Assessing uhd image quality from aesthetics, dis- tortions, and saliency

    Wei Sun, Weixia Zhang, Yuqin Cao, Linhan Cao, Jun Jia, Zijian Chen, Zicheng Zhang, Xiongkuo Min, and Guang- tao Zhai. Assessing uhd image quality from aesthetics, dis- tortions, and saliency. InComputer Vision – ECCV 2024 Workshops, pages 109–126, Cham, 2025. Springer Nature Switzerland. 6, 7

    work page 2024

  46. [46]

    Nima: Neural image assessment.IEEE Transactions on Image Processing, 27(8): 3998–4011, 2018

    Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment.IEEE Transactions on Image Processing, 27(8): 3998–4011, 2018. 2, 6

    work page 2018

  47. [47]

    Chan, and Chen Change Loy

    Jianyi Wang, Kelvin C.K. Chan, and Chen Change Loy. Ex- ploring clip for assessing the look and feel of images. In Conference on Artificial Intelligence. AAAI Press, 2023. 2, 6

    work page 2023

  48. [48]

    Zhou Wang and Alan C. Bovik. Reduced- and no-reference image quality assessment.IEEE Signal Processing Maga- zine, 28(6):29–40, 2011. 2

    work page 2011

  49. [49]

    Wang, E.P

    Z. Wang, E.P. Simoncelli, and A.C. Bovik. Multiscale struc- tural similarity for image quality assessment. InAsilomar Conference on Signals, pages 1398–1402 V ol.2, 2003. 2

    work page 2003

  50. [50]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Processing, 13(4): 600–612, 2004. 2

    work page 2004

  51. [51]

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

    Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, and Weisi Lin. Q-bench: A benchmark for general-purpose foundation models on low-level vision. InInternational Conference on Learning Representations,

    work page

  52. [52]

    Q-instruct: Improving low-level visual abil- ities for multi-modality foundation models

    Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, Geng Xue, Wenxiu Sun, Qiong Yan, and Weisi Lin. Q-instruct: Improving low-level visual abil- ities for multi-modality foundation models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25490–25500, 2024. 2

    work page 2024

  53. [53]

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

    Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, Qiong Yan, Xiongkuo Min, Guang- tao Zhai, and Weisi Lin. Q-align: Teaching LMMs for visual scoring via discrete text-defined levels. InProceedings of the 41st International Conference on Machine Learning, pages 54015–54029. PMLR, ...

    work page 2024

  54. [54]

    Tinyclip: Clip distillation via affinity mimicking and weight inheritance

    Kan Wu, Houwen Peng, Zhenghong Zhou, Bin Xiao, Mengchen Liu, Lu Yuan, Hong Xuan, Michael Valenzuela, Xi Stephen Chen, Xinggang Wang, Hongyang Chao, and Han Hu. Tinyclip: Clip distillation via affinity mimicking and weight inheritance. InIEEE/CVF International Confer- ence on Computer Vision, pages 21913–21923, 2023. 5

    work page 2023

  55. [55]

    Boosting image quality assessment through efficient transformer adaptation with lo- cal feature enhancement

    Kangmin Xu, Liang Liao, Jing Xiao, Chaofeng Chen, Haon- ing Wu, Qiong Yan, and Weisi Lin. Boosting image quality assessment through efficient transformer adaptation with lo- cal feature enhancement. InIEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 2662–2672,

    work page

  56. [56]

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

    Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa: Multi-dimension attention network for no-reference image quality assessment. InIEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1190–1199, 2022. 2, 6

    work page 2022

  57. [57]

    From patches to pic- tures (paq-2-piq): Mapping the perceptual space of picture quality

    Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pic- tures (paq-2-piq): Mapping the perceptual space of picture quality. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3572–3582, 2020. 2, 3, 5

    work page 2020

  58. [58]

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

    Zhiyuan You, Xin Cai, Jinjin Gu, Tianfan Xue, and Chao Dong. Teaching large language models to regress accurate image quality scores using score distribution. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14483–14494, 2025. 2, 3, 6

    work page 2025

  59. [59]

    Depicting beyond scores: Advanc- ing image quality assessment through multi-modal language models

    Zhiyuan You, Zheyuan Li, Jinjin Gu, Zhenfei Yin, Tianfan Xue, and Chao Dong. Depicting beyond scores: Advanc- ing image quality assessment through multi-modal language models. InComputer Vision – ECCV 2024, pages 259–276, Cham, 2025. Springer Nature Switzerland. 2

    work page 2024

  60. [60]

    Vsi: A visual saliency-induced index for perceptual image quality assess- ment.IEEE Transactions on Image Processing, 23(10): 4270–4281, 2014

    Lin Zhang, Ying Shen, and Hongyu Li. Vsi: A visual saliency-induced index for perceptual image quality assess- ment.IEEE Transactions on Image Processing, 23(10): 4270–4281, 2014. 3

    work page 2014

  61. [61]

    Lin Zhang, Lei Zhang, and Alan C. Bovik. A feature- enriched completely blind image quality evaluator.IEEE Transactions on Image Processing, 24(8):2579–2591, 2015. 1, 2

    work page 2015

  62. [62]

    The application of visual saliency models in objective image quality assessment: A statistical evaluation

    Wei Zhang, Ali Borji, Zhou Wang, Patrick Le Callet, and Hantao Liu. The application of visual saliency models in objective image quality assessment: A statistical evaluation. IEEE Transactions on Neural Networks and Learning Sys- tems, 27(6):1266–1278, 2016. 3

    work page 2016

  63. [63]

    Blind image quality assessment using a deep bilinear convolutional neural network.IEEE Transactions on Cir- cuits and Systems for Video Technology, 30(1):36–47, 2020

    Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network.IEEE Transactions on Cir- cuits and Systems for Video Technology, 30(1):36–47, 2020. 2, 6

    work page 2020

  64. [64]

    Uncertainty-aware blind image quality assessment in the laboratory and wild.IEEE Transactions on Image Pro- cessing, 30:3474–3486, 2021

    Weixia Zhang, Kede Ma, Guangtao Zhai, and Xiaokang Yang. Uncertainty-aware blind image quality assessment in the laboratory and wild.IEEE Transactions on Image Pro- cessing, 30:3474–3486, 2021. 1, 6

    work page 2021

  65. [65]

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

    Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision- language correspondence: A multitask learning perspective. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14071–14081, 2023. 1, 2, 6

    work page 2023

  66. [66]

    Zicheng Zhang, Haoning Wu, Erli Zhang, Guangtao Zhai, and Weisi Lin. Q-bench ++: A benchmark for multi-modal foundation models on low-level vision from single images to pairs.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10404–10418, 2024. 2

    work page 2024

  67. [67]

    Quality-aware pretrained models for blind image quality as- sessment

    Kai Zhao, Kun Yuan, Ming Sun, Mading Li, and Xing Wen. Quality-aware pretrained models for blind image quality as- sessment. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22302–22313, 2023. 1, 2, 3

    work page 2023

  68. [68]

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

    Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiqa: Deep meta-learning for no- reference image quality assessment. InIEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 14131–14140, 2020. 2

    work page 2020

  69. [69]

    Adaptive image quality assess- ment via teaching large multimodal model to compare

    Hanwei Zhu, Haoning Wu, Yixuan Li, Zicheng Zhang, Bao- liang Chen, Lingyu Zhu, Yuming Fang, Guangtao Zhai, Weisi Lin, and Shiqi Wang. Adaptive image quality assess- ment via teaching large multimodal model to compare. In Neural Information Processing Systems, 2024. 2, 3, 6 Learning Where to Look and How to Judge: Resolution-agnostic Image Quality Assessme...

    work page 2024

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.