Pith. sign in

REVIEW 5 major objections 5 minor 56 references

A video's aesthetic score is shaped by its salient moments and ending, not its average frames.

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 ↗

T0 review · deepseek-v4-flash

2026-08-02 03:15 UTC pith:G224L5H5

load-bearing objection Solid, useful VAA architecture with a nice cognitive-prior hook, but the SOTA claim is unverified: baseline protocol is underspecified, and the evaluation is single-split with test-set tuning. the 5 major comments →

arxiv 2607.13941 v1 pith:G224L5H5 submitted 2026-07-15 cs.CV

Peak-End-Net: A Peak-End Rule Inspired Framework for Generalizable Video Aesthetic Assessment

Geng Li , Haiwen Li , Rui Chen , Jing Tang , Lei Sun , Xiangxiang Chu This is my paper
classification cs.CV
keywords video aesthetic assessmentpeak-end ruletemporal aggregationimage aesthetic transferzero-shot generalizationgated fusioninterpretable video analysis
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.

Video aesthetic assessment is subjective and underexplored, and the paper argues that human judgments follow the psychological peak-end rule: people rate an experience mainly by its strongest moments and its final moments. Peak-End-Net turns this into a computational mechanism by using frame-level aesthetic scores from a frozen image-aesthetic model as a saliency signal, then pooling video features with attention weights that emphasize peaks, valleys, and the ending. It adds a rhythm encoder to capture how aesthetics evolve over time, and a gated fusion step that blends the video model's overall score with a global image-aesthetic score for better cross-dataset transfer. The paper reports state-of-the-art in-domain results on VADB and zero-shot results on DIVIDE-3K, with only 1.05M trainable parameters on a frozen vision transformer.

Core claim

The paper claims that explicitly implementing the peak-end rule in temporal aggregation improves video aesthetic assessment. Per-frame aesthetic priors, produced by a frozen head trained on the AVA image dataset, act as surrogate saliency labels in the absence of frame-level video annotations. These priors feed three response terms — peak, valley, and end — that produce frame attention weights, so the video representation emphasizes aesthetically salient moments and the ending rather than uniformly averaging frames. The authors also introduce an aesthetic rhythm encoder that extracts trends and fluctuations from the score sequence, and a dynamic gated fusion that adaptively balances the vide

What carries the argument

The central mechanism is the peak-end rule-based temporal aggregation: for each sampled frame, the model computes peak, valley, and end response terms from frame-wise aesthetic scores produced by an AVA-pretrained image aesthetic head, then combines them with a learnable base into softmax attention weights over frames. Aesthetic rhythm pattern encoding, a 1D-convolution encoder on the score sequence, is added to retain temporal evolution, and a dynamic gated fusion module learns to blend the VAA overall prediction with the global IAA estimate.

Load-bearing premise

The paper's mechanism relies on the AVA-pretrained per-frame aesthetic scores being a trustworthy proxy for which moments in a video are aesthetically salient to human viewers; if those scores misidentify peaks and endings, the peak-end aggregation loses its psychological grounding.

What would settle it

Take a held-out set of videos with human keyframe importance annotations and compare them to the frames Peak-End-Net assigns the highest peak and end weights; if the overlap is no better than chance, the method is not actually reflecting the peak-end rule. Alternatively, retrain the published baselines on the identical 8:2 VADB split and re-run the comparison; if the reported margins do not persist, the state-of-the-art claim is protocol-dependent.

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

If this is right

  • Peak-end aggregation outperforms mean, max, and a temporal-transformer baseline on VADB (RMSE 0.4548 vs 0.4654, SRCC 0.8804 vs 0.8683).
  • Replacing the randomly initialized IAA head with an AVA-pretrained head improves all metrics, supporting the transfer of image aesthetic knowledge to video.
  • Adding the rhythm encoder yields further gains, indicating that preserving temporal progression helps beyond isolated salient moments.
  • Dynamic gated fusion improves zero-shot DIVIDE-3K SRCC from 0.4600 to 0.5427 while causing only a small in-domain drop, a favorable trade-off compared to static averaging.
  • The total trainable parameter count is 1.05M (less than 1% of the frozen ViT), and the method outperforms a 7B-parameter LLM-based method on VADB overall score.

Where Pith is reading between the lines

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

  • The paper's saliency signal is a single frozen image model; retraining or swapping this prior (e.g., with a video-specific saliency model or human keyframe annotations) could either strengthen or weaken the peak-end benefit, so the mechanism's ceiling is not yet mapped.
  • The end-span hyperparameter is set to the last one-third of frames; a natural extension is to test whether the optimal end span scales with video duration or content genre.
  • The same saliency-weighted pooling recipe could transfer to other subjective sequential judgments, such as video quality or user experience, where a proxy saliency score is available.
  • The paper reports an intentional in-domain drop from gated fusion; a finer analysis of the gate weight lambda could reveal whether the gate is selecting on interpretable properties like motion or content category.

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

5 major / 5 minor

Summary. The manuscript proposes Peak-End-Net, a video aesthetic assessment (VAA) framework built on a frozen CLIP ViT-L/14 backbone with only 1.05M trainable parameters. Frame-level features are passed through an AVA-pretrained IAA head to produce frame-wise aesthetic scores. These scores are used to compute peak, valley, and end responses that are combined into attention weights (Eqs. 5–8), yielding a holistic video representation. A separate rhythm encoder models the temporal progression of the aesthetic score sequence, and a VAA head predicts 11 aesthetic dimensions. In a second stage, a dynamic gated fusion (Eq. 12) combines the VAA overall prediction with a global IAA score. The model is trained on the public 7,881-video VADB subset with an 8:2 split and evaluated in-domain on VADB and zero-shot on DIVIDE-3K. The paper claims state-of-the-art performance on both benchmarks, with Stage 1 best on VADB (RMSE 0.4344, SRCC 0.8875, PLCC 0.8853, KRCC 0.7014) and Stage 2 best on DIVIDE-3K (RMSE 0.4905, SRCC 0.5427, PLCC 0.5773, KRCC 0.3745).

Significance. If the reported comparisons are reproducible under an identical protocol, the work makes a useful contribution: it introduces a psychologically motivated, interpretable, and parameter-efficient approach to VAA, and the cross-dataset zero-shot evaluation is a meaningful stress test. The ablations in Tables 3–5 provide reasonable evidence for the contribution of the IAA prior, the rhythm encoder, and the gated fusion. The explicit statement of code/model release is a further strength. However, the central state-of-the-art claim is not currently established: baseline acquisition and target normalization are unstated, all results come from a single split without variance or significance testing, and key hyperparameters were selected on the test set. These are fixable within the manuscript's scope, but they are load-bearing for the paper's headline claim.

major comments (5)
  1. [§4.1, Tables 1–2] The central claim 'state-of-the-art performance' depends entirely on the baseline numbers, but the paper does not state whether FastVQA, SimpleVQA, ModularBVQA, CLIPVQA, Q-Align, DOVER, and VADB-Net were retrained on the exact 8:2 split of the 7,881-video VADB subset, evaluated using publicly released checkpoints, or cited from papers using different splits or settings. Without this information, margins such as SRCC 0.8875 vs VADB-Net 0.8576 could be protocol artifacts. Please specify for every baseline the training data, split, checkpoint, and evaluation protocol; if published numbers are used, justify why they are comparable.
  2. [§4.1, Table 2] RMSE comparisons on DIVIDE-3K are only meaningful if the target aesthetic scores are normalized to the same scale as the model's sigmoid output in [0,1]. The paper states that VADB scores are normalized to [0,1] during training, but never states the target transformation for the DIVIDE-3K aesthetic dimension. If baselines were evaluated with differently scaled targets (e.g., raw MOS on a different scale), the reported RMSE values are not comparable. The exact target preprocessing for DIVIDE-3K must be reported.
  3. [§4.2, Tables 1–2] All results are reported for a single random 8:2 split, with no standard deviations, confidence intervals, or significance tests. Some reported advantages over strong baselines are small (e.g., VADB SRCC 0.8875 vs Q-Align 0.8789; KRCC 0.7014 vs 0.6903). A single split cannot establish a statistically reliable ranking. The authors should report results over multiple random splits (mean ± std) and, where feasible, bootstrap or paired significance tests.
  4. [§4.4, Fig. 4] The default hyperparameters K=4 and T=12 are selected based on the VADB test set: the sensitivity analysis reports that K/T = 1/3 'gives the best performance' and that T=12 is chosen as a balance, with both curves computed on the test set. This is test-set-based model selection and can inflate the reported numbers. Hyperparameters should be selected on a held-out validation split or via nested cross-validation, and the test results should be reported for the selected configuration.
  5. [§4.1–4.2, Tables 1–2] The paper conflates two checkpoints when claiming state-of-the-art performance. The best VADB results are from Stage 1 (e.g., SRCC 0.8875), while the best DIVIDE-3K results are from Stage 2. Table 1 reports 'Stage 1/Stage 2' in the Overall column, and Table 2 reports two separate rows. Notably, if Stage 2 is the final model, it is not state-of-the-art on VADB SRCC: its 0.8784 is below Q-Align's 0.8789, contradicting the text that Stage 2 'remains the state-of-the-art'. The paper must clearly state which checkpoint is used for each claim and, if a single deployed model is intended, evaluate that one model on both benchmarks.
minor comments (5)
  1. [§4.4] In the paragraph 'Effectiveness of Peak-End Aggregation', the phrase 'explicitly textitasiz' appears to be a typo; likely 'explicitly emphasizing' was intended.
  2. [Appendix B.5] The text says the temporal transformer baseline appears in 'Table 4 of the main paper', but it is actually reported in Table 3.
  3. [Table 1 caption] The 'Overall' column shows two numbers separated by a slash (0.4344/0.4548 etc.). The caption says these are 'Stage 1/Stage 2', but this convention should also be stated in the main text where Table 1 is discussed, to avoid ambiguity about which model is being compared.
  4. [§3.3, Eq. (14)] Equation (14) uses BCE(λ, λ*) where both arguments are continuous scalars in [0,1]. The precise form of the binary cross-entropy term for scalar values should be defined, since BCE is usually applied to probabilities.
  5. [§4.1] The text says DIVIDE-3K 'contains 3,590 in-the-wild videos' and Table 2 is described as reporting the 'test set'. Please clarify whether the full dataset or a designated test split is used for zero-shot evaluation.

Circularity Check

0 steps flagged

No circularity: the claimed predictions are supervised on external VADB/DIVIDE-3K targets and the IAA prior is reused as a feature, not as the target by construction.

full rationale

The derivation chain is self-contained against external benchmarks. Frame-wise IAA scores s_t are produced by a frozen AVA-pretrained head (Eqs. 2-3) and are used twice: to define peak/valley/end responses in Eqs. (4)-(7), and as the global IAA estimate y_IAA in Eq. (12). This double use is feature reuse, not circularity: no equation defines s_t in terms of the VAA target y, and the final score in Eq. (12) is a convex combination of the VAA-head estimate and the IAA prior, both ultimately supervised by VADB ground truth. The gate lambda in Eq. (13) is trained against a soft target derived from relative errors; this is ordinary supervised training, not a fitted parameter renamed as a prediction. Ablations test the pretrained IAA head against a randomly initialized head and compare peak-end aggregation against mean/max pooling and a temporal transformer on the test split. The reported SOTA claim depends on unverified baseline protocols and a single split with no variance; that is a correctness/robustness concern, not circularity. The manuscript's self-citations are not load-bearing for the central derivation, and the peak-end rule is anchored to an external psychology source [13]. No uniqueness theorem is imported from the authors. Therefore no circular step can be exhibited, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 1 invented entities

The method introduces no physical entities. It relies on standard learning assumptions, the transferability of AVA image-aesthetic priors to video frames, and the applicability of the peak-end rule to aesthetic judgment; both psychological premises enter as domain assumptions and are validated only indirectly through benchmark performance.

free parameters (6)
  • alpha, beta, gamma in Eq. (7) = not reported
    Learnable coefficients balancing peak, valley, and end terms in the temporal attention logits; fitted on VADB Stage 1.
  • r_base in Eq. (7) = not reported
    Learnable base response applied to all frames before softmax weighting.
  • End-span hyperparameter K (lambda = T/K) = 4 (last one-third of frames)
    Controls exponential end emphasis in Eq. (6); selected from VADB test-set sensitivity analysis (Fig. 4a).
  • Number of sampled frames T = 12
    Default frame count chosen from sensitivity analysis (Fig. 4b) balancing accuracy and efficiency.
  • Weights of AVA-pretrained IAA head Phi_IAA = not specified in paper
    External MLP (768->512->10) pretrained on AVA and frozen; provides the frame-wise aesthetic prior used throughout.
  • Weights of Phi_VAA, Phi_RE, Phi_GF = 1.05M trainable parameters total
    Learned on VADB in two stages; standard supervised regression and gating parameters.
axioms (6)
  • domain assumption The peak-end rule (Kahneman et al. 1993) transfers to video aesthetic judgment: salient moments and the ending dominate overall ratings.
    Invoked in Sec. 1 and encoded in Eqs. (5)-(8); no human-subject test in this paper validates the mechanism for video aesthetics.
  • domain assumption IAA scores from an AVA-pretrained head are a valid frame-wise saliency signal for videos across domains.
    Used to define peaks, valleys, and end emphasis in Sec. 3.2; if miscalibrated for video frames, the aggregation weights become arbitrary.
  • domain assumption Uniformly sampling 12 frames preserves the aesthetic trajectory of a 5-20 second video.
    Implementation choice in Sec. 4.1; Appendix D admits sparse sampling can miss narrative or semantic cues.
  • domain assumption CLIP ViT-L/14 [CLS] embeddings, passed through a small MLP, contain sufficient aesthetic signal for frame-level scoring.
    Frozen backbone used throughout; no fine-tuning of visual features.
  • domain assumption VADB professional scores are reliable ground truth for VAA.
    Dataset assumption inherited from [32]; labels provided by 37 professional annotators.
  • standard math Standard deep-learning machinery (softmax, AdamW, MSE/BCE losses, LayerNorm) behaves as expected.
    Used in Eqs. (3), (11), (14); no formal guarantees required.
invented entities (1)
  • Aesthetic rhythm pattern no independent evidence
    purpose: Conceptual representation of temporal variation in frame-wise aesthetic scores, encoded by 1D convolutions in Phi_RE.
    A modeling construct with no direct observable outside the system; its usefulness is measured only through downstream VAA metrics.

pith-pipeline@v1.3.0-alltime-deepseek · 22788 in / 14129 out tokens · 132244 ms · 2026-08-02T03:15:17.070873+00:00 · methodology

0 comments
read the original abstract

Video aesthetic assessment (VAA) aims to predict how aesthetically pleasing a video is, yet remains far less explored than other visual assessment tasks. Its progress is hindered not only by the scarcity of large-scale benchmarks, but also by the intrinsic subjectivity of aesthetic judgment, which is shaped by human perception. In this paper, we revisit VAA from a psychological perspective and propose \textit{Peak-End-Net}, a lightweight and interpretable framework inspired by the \textit{peak-end rule}, which suggests that people tend to judge a temporal experience mainly according to its salient moments and the ending. Building on this intuition, we first transfer knowledge from image aesthetic assessment (IAA) to VAA by introducing a pretrained IAA head to produce frame-wise aesthetic priors, which serve as surrogate signals for identifying aesthetically salient moments and guiding \textit{peak-end rule}-based temporal aggregation. To further capture how a video evolves aesthetically over time, we design an aesthetic rhythm encoder that models temporal progression beyond isolated moments. Additionally, we refine the overall assessment through a dynamic gated fusion mechanism to improve robustness under distribution shift. Our method is built on a frozen vision transformer (ViT) and requires only a small number of trainable parameters, making it scalable and parameter-efficient. Extensive experiments on two existing VAA benchmarks, including in-domain evaluation on VADB and cross-domain testing on DIVIDE-3K, demonstrate that our approach achieves state-of-the-art performance, affirming the value of psychologically grounded modeling for VAA. Our code and models are available at https://github.com/AMAP-ML/Peak-End-Net.

Figures

Figures reproduced from arXiv: 2607.13941 by Geng Li, Haiwen Li, Jing Tang, Lei Sun, Rui Chen, Xiangxiang Chu.

Figure 1
Figure 1. Figure 1: Illustration of our motivation. VAA is inherently [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our method. Given sampled video frames, the system transfers frame-wise aesthetic priors from [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative results on two VADB examples. We [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sensitivity analysis of key hyperparameters. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Some representative samples: the two above are [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results on VADB. We present the tags, language comments, ground-truth annotations, and the predicted [PITH_FULL_IMAGE:figures/full_fig_p012_6.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

56 extracted references · 11 linked inside Pith

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report.arXiv preprint arXiv:2303.08774 (2023)

    Pith/arXiv arXiv 2023

  2. [2]

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al . 2025. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631(2025)

    Pith/arXiv arXiv 2025

  3. [3]

    Zoya Bylinskii, Ellen M DeGennaro, Rishi Rajalingham, Harald Ruda, Jinxia Zhang, and John K Tsotsos. 2015. Towards the quantitative evaluation of visual attention models.Vision research116 (2015), 258–268

    2015

  4. [4]

    Rui Chen, Lei Sun, Jing Tang, Geng Li, and Xiangxiang Chu. 2025. Finger: Content aware fine-grained evaluation with reasoning for ai-generated videos. InProceedings of the 33rd ACM International Conference on Multimedia. 3517– 3526

    2025

  5. [5]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. InInter- national conference on machine learning. PmLR, 1597–1607

    2020

  6. [6]

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. 2024. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 24185–24198

    2024

  7. [7]

    Ritendra Datta, Dhiraj Joshi, Jia Li, and James Z Wang. 2006. Studying aesthetics in photographic images using a computational approach. InEuropean conference on computer vision. Springer, 288–301

    2006

  8. [8]

    Yubin Deng, Chen Change Loy, and Xiaoou Tang. 2017. Image aesthetic assess- ment: An experimental survey.IEEE Signal Processing Magazine34, 4 (2017), 80–106

    2017

  9. [9]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020)

    Pith/arXiv arXiv 2020

  10. [10]

    Xiaokun Feng, Haiming Yu, Meiqi Wu, Shiyu Hu, Jintao Chen, Chen Zhu, Jiahong Wu, Xiangxiang Chu, and Kaiqi Huang. 2025. NarrLV: Towards a Comprehensive Narrative-Centric Evaluation for Long Video Generation Models.arXiv preprint arXiv:2507.11245(2025)

    arXiv 2025

  11. [11]

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. Mo- mentum contrast for unsupervised visual representation learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9729–9738

    2020

  12. [12]

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. 2021. Scaling up visual and vision- language representation learning with noisy text supervision. InInternational conference on machine learning. PMLR, 4904–4916

    2021

  13. [13]

    Daniel Kahneman, Barbara L Fredrickson, Charles A Schreiber, and Donald A Redelmeier. 1993. When more pain is preferred to less: Adding a better end. Psychological science4, 6 (1993), 401–405

    1993

  14. [14]

    Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. 2017. The kinetics human action video dataset.arXiv preprint arXiv:1705.06950(2017)

    Pith/arXiv arXiv 2017

  15. [15]

    Qi Kuang, Xin Jin, Qinping Zhao, and Bin Zhou. 2019. Deep multimodality learning for UAV video aesthetic quality assessment.IEEE Transactions on Multimedia22, 10 (2019), 2623–2634

    2019

  16. [16]

    Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. 2011. HMDB: a large video database for human motion recognition. In2011 International conference on computer vision. IEEE, 2556–2563

    2011

  17. [17]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InInternational conference on machine learning. PMLR, 12888–12900

    2022

  18. [18]

    Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. 2024. Mvbench: A comprehensive multi-modal video understanding benchmark. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 22195–22206

    2024

  19. [19]

    Mingxing Li, Rui Wang, Lei Sun, Yancheng Bai, and Xiangxiang Chu. 2025. Next Token Is Enough: Realistic Image Quality and Aesthetic Scoring with Multimodal Large Language Model.arXiv preprint arXiv:2503.06141(2025)

    Pith/arXiv arXiv 2025

  20. [20]

    Weiqi Li, Xuanyu Zhang, Shijie Zhao, Yabin Zhang, Junlin Li, Li Zhang, and Jian Zhang. 2025. Q-insight: Understanding image quality via visual reinforcement learning.arXiv preprint arXiv:2503.22679(2025)

    Pith/arXiv arXiv 2025

  21. [21]

    Xinran Ling, Chen Zhu, Meiqi Wu, Hangyu Li, Xiaokun Feng, Cundian Yang, Aiming Hao, Jiashu Zhu, Jiahong Wu, and Xiangxiang Chu. 2025. Vmbench: A benchmark for perception-aligned video motion generation. InProceedings of the IEEE/CVF International Conference on Computer Vision. 13087–13098

    2025

  22. [22]

    Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101(2017)

    Pith/arXiv arXiv 2017

  23. [23]

    Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. 2022. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning.Neurocomputing508 (2022), 293–304

    2022

  24. [24]

    Yiwen Luo and Xiaoou Tang. 2008. Photo and video quality evaluation: Focusing on the subject. InEuropean conference on computer vision. Springer, 386–399

    2008

  25. [25]

    Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. 2022. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM international conference on multimedia. 638–647

    2022

  26. [26]

    Anne-Sofie Maerten, Li-Wei Chen, Stefanie De Winter, Christophe Bossens, and Johan Wagemans. 2025. LAPIS: A novel dataset for personalized image aes- thetic assessment. InProceedings of the Computer Vision and Pattern Recognition Conference. 6302–6311

    2025

  27. [27]

    Yachun Mi, Yu Li, Yanting Li, Chen Hui, Tong Zhang, Zhixuan Li, Chenyue Song, Wei Yang Bryan Lim, and Shaohui Liu. 2025. Q-CLIP: Unleashing the Power of Vision-Language Models for Video Quality Assessment through Unified Cross- Modal Adaptation.arXiv preprint arXiv:2508.06092(2025)

    arXiv 2025

  28. [28]

    Xiongkuo Min, Huiyu Duan, Wei Sun, Yucheng Zhu, and Guangtao Zhai. 2024. Perceptual video quality assessment: A survey.Science China Information Sciences 67, 11 (2024), 211301

    2024

  29. [29]

    Anush K Moorthy, Pere Obrador, and Nuria Oliver. 2010. Towards computational models of the visual aesthetic appeal of consumer videos. InEuropean conference on computer vision. Springer, 1–14

    2010

  30. [30]

    Naila Murray, Luca Marchesotti, and Florent Perronnin. 2012. AVA: A large-scale database for aesthetic visual analysis. In2012 IEEE conference on computer vision and pattern recognition. IEEE, 2408–2415

    2012

  31. [31]

    Yuzhen Niu and Feng Liu. 2012. What makes a professional video? A computa- tional aesthetics approach.IEEE Transactions on Circuits and Systems for Video Technology22, 7 (2012), 1037–1049

    2012

  32. [32]

    Qianqian Qiao, DanDan Zheng, Yihang Bo, Bao Peng, Heng Huang, Longteng Jiang, Huaye Wang, Jingdong Chen, Jun Zhou, and Xin Jin. 2025. VADB: A Large-Scale Video Aesthetic Database with Professional and Multi-Dimensional Annotations.arXiv preprint arXiv:2510.25238(2025)

    arXiv 2025

  33. [33]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sand- hini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al

  34. [34]

    Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. 2012. Ucf101: A dataset of 101 human actions classes from videos in the wild.arXiv preprint arXiv:1212.0402(2012)

    Pith/arXiv arXiv 2012

  35. [35]

    Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai. 2022. A deep learning based no-reference quality assessment model for ugc videos. InProceedings of the 30th ACM International Conference on Multimedia. 856–865

    2022

  36. [36]

    Yunlong Tang, Jing Bi, Siting Xu, Luchuan Song, Susan Liang, Teng Wang, Daoan Zhang, Jie An, Jingyang Lin, Rongyi Zhu, et al. 2025. Video understanding with large language models: A survey.IEEE Transactions on Circuits and Systems for Video Technology(2025)

    2025

  37. [37]

    Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. 2016. Yfcc100m: The new data in multimedia research.Commun. ACM59, 2 (2016), 64–73

    2016

  38. [38]

    Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. 2025. Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features.arXiv preprint arXiv:2502.14786(2025)

    Pith/arXiv arXiv 2025

  39. [39]

    Christos Tzelepis, Eftichia Mavridaki, Vasileios Mezaris, and Ioannis Patras. 2016. Video aesthetic quality assessment using kernel Support Vector Machine with isotropic Gaussian sample uncertainty (KSVM-IGSU). In2016 IEEE International Conference on Image Processing (ICIP). IEEE, 2410–2414

    2016

  40. [40]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  41. [41]

    Mengmeng Wang, Jiazheng Xing, and Yong Liu. 2021. Actionclip: A new para- digm for video action recognition.arXiv preprint arXiv:2109.08472(2021)

    Pith/arXiv arXiv 2021

  42. [42]

    Wen Wen, Mu Li, Yabin Zhang, Yiting Liao, Junlin Li, Li Zhang, and Kede Ma

  43. [43]

    Haoning Wu, Chaofeng Chen, Jingwen Hou, Liang Liao, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. 2022. Fast-vqa: Efficient end-to-end video quality Geng Li et al. assessment with fragment sampling. InEuropean conference on computer vision. Springer, 538–554

    2022

  44. [44]

    H. Wu, E. Zhang, L. Liao, et al . 2023. Towards explainable in-the-wild video quality assessment: a database and a language-prompted approach.Proceedings of the 31st acm international conference on multimedia(2023), 1045–1054

    2023

  45. [45]

    Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. 2023. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. InProceedings of the IEEE/CVF international conference on computer vision. 20144– 20154

    2023

  46. [46]

    Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al . 2023. Q-align: Teaching lmms for visual scoring via discrete text-defined levels.arXiv preprint arXiv:2312.17090(2023)

    Pith/arXiv arXiv 2023

  47. [47]

    Tianhe Wu, Jian Zou, Jie Liang, Lei Zhang, and Kede Ma. 2025. Visualquality-r1: Reasoning-induced image quality assessment via reinforcement learning to rank. arXiv preprint arXiv:2505.14460(2025)

    arXiv 2025

  48. [48]

    Wulin Xie, Rui Dai, Ruidong Ding, Kaikui Liu, Xiangxiang Chu, Xinwen Hou, and Jie Wen. 2026. Q-Hawkeye: Reliable Visual Policy Optimization for Image Quality Assessment.arXiv preprint arXiv:2601.22920(2026)

    arXiv 2026

  49. [49]

    Fengchuang Xing, Mingjie Li, Yuan-Gen Wang, Guopu Zhu, and Xiaochun Cao

  50. [50]

    Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. 2021. Videoclip: Contrastive pre-training for zero-shot video-text understanding. InProceedings of the 2021 conference on empirical methods in natural language processing. 6787– 6800

    2021

  51. [51]

    Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. 2024. mplug-owl2: Revolutionizing multi-modal large lan- guage model with modality collaboration. InProceedings of the ieee/cvf conference on computer vision and pattern recognition. 13040–13051

    2024

  52. [52]

    Clipvqa: Video quality assessment via clip.IEEE Transactions on Broadcast- ing71, 1 (2024), 291–306

    2024

  53. [53]

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. 2023. Sigmoid loss for language image pre-training. InProceedings of the IEEE/CVF international conference on computer vision. 11975–11986. Peak-End-Net: A Peak-End Rule Inspired Framework for Generalizable Video Aesthetic Assessment This appendix mainly includes a detailed introduction t...

    2023

  54. [55]

    Kun Yuan, Hongbo Liu, Mading Li, Muyi Sun, Ming Sun, Jiachao Gong, Jinhua Hao, Chao Zhou, and Yansong Tang. 2024. Ptm-vqa: efficient video quality assessment leveraging diverse pretrained models from the wild. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2835–2845

    2024

  55. [2021]

    In International conference on machine learning

    Learning transferable visual models from natural language supervision. In International conference on machine learning. PMLR, 8748–8763

  56. [2024]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Modular blind video quality assessment. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2763–2772