Pith. sign in

REVIEW 4 major objections 5 minor 48 references

VQ-Insight: Teaching VLMs for AI-Generated Video Quality Understanding via Progressive Visual Reinforcement Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read VQ-Insight claims that a 7B vision-language model, trained with progressive reinforcement learning, can evaluate AI-generated video quality as accurately as specialists that require hundreds of thousands of human labels.

desk verdict A promising progressive RL recipe for AIGC video quality, but the temporal reward and natural-video overclaims need fixing. read the letter →

arxiv 2506.18564 v1 pith:EKFR44JN submitted 2025-06-23 cs.CV

Xuanyu Zhang , Weiqi Li , Shijie Zhao , Junlin Li , Li Zhang , Jian Zhang This is my paper · ORCID
classification cs.CV
keywords AIGCvideoqualityassessmentvision-languagemodelGroupRelativePolicyOptimizationtemporalmodelingrewardpreferencecomparisonmulti-dimensionscoringtext-to-videoalignmentDiffusionDPO
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

VQ-Insight claims that a 7B vision-language model can be turned into an accurate and general evaluator of AI-generated videos using reinforcement learning instead of large supervised datasets. The method trains Qwen-2.5-VL with GRPO through three progressive stages: image-scoring warm-up, task-specific temporal learning with tailored rewards, and alternating optimization with a video generation model. With only a few thousand preference labels, the resulting model outperforms state-of-the-art baselines on preference comparison, multi-dimension scoring, and even natural video scoring. If correct, it offers a data-efficient way to build reward models that can align text-to-video generators with human preferences.

What carries the argument

The engine is Group Relative Policy Optimization (GRPO), an actor-only reinforcement learning algorithm that normalizes rewards within a sampled group and clips policy ratios, removing the need for a critic network. VQ-Insight shapes this with four rule-based rewards: an image scoring reward (1 − ℓ1 error), a temporal modeling reward comparing answer probabilities on sequentially ordered versus randomly shuffled video frame tokens, a length-control reward, and task rewards for multi-dimension scoring and preference comparison. The temporal reward (r_temp = α if w_seq > 0.8·w_rand) is the mechanism meant to instill temporal awareness, and the stage-wise curriculum is what lets small data transfer to natural video.

What would settle it

If on a temporal-quality subset of LGVQ or LSVQ, removing the temporal modeling reward changes scores by less than noise, or if the shuffled-token accuracy gap does not correlate with human temporal MOS, then the central temporal mechanism would fail its purpose.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a curriculum of GRPO reward shaping can elicit video-quality understanding from a general-purpose VLM without large-scale human annotation. The key claim is that combining a multi-dimension ℓ1 scoring reward, a preference-comparison reward, a temporal modeling reward that checks whether shuffling video frame tokens degrades the answer, and a length-control reward lets the model learn spatial, temporal, and alignment quality jointly. The paper reports that VQ-Insight reaches 50.80 tau accuracy on GenAI and 61.20 on MonetBench preference comparison, outperforming VisionReward and UnifiedReward; raises PLCC on LGVQ text-video alignment from 0.660 (Qwen-SFT) to 0.836; and improves T2V-Turbo's VBench overall score from 0.8095 to 0.8185 when used as the reward model for DPO. The progressive scheme — warm-up before temporal learning before joint generation-understanding finetuning — is presented as the reason the model generalizes to out-of-domain natural videos without catastrophic forgetting.

Load-bearing premise

The temporal modeling reward assumes that a large drop in answer probability after shuffling the video's frame tokens means the model truly understands motion and temporal consistency, but this proxy is never validated against human temporal-quality judgments.

Editorial extensions

If this is right

  • With a few thousand labels, a 7B VLM can match or beat specialist video quality models that train on hundreds of thousands of annotations.
  • VQ-Insight can serve as a drop-in reward model for DPO or GRPO alignment of text-to-video generators, improving semantic and quality scores without extra human feedback.
  • The same model covers AIGC preference comparison, multi-dimension scoring, and natural video scoring, suggesting quality judgment transfers across domains.
  • The length-control reward induces shorter, focused reasoning chains, which improves both scoring accuracy and the interpretability of the model's decisions.

Reading between the lines

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

  • Because the temporal reward penalizes order-insensitive answers, VQ-Insight may be rewarding sensitivity to any frame-order artifact, not human-perceived temporal quality; a direct validation against human temporal MOS would clarify this.
  • The alternating generation-understanding loop is a general recipe: any pairwise preference reward model can be co-trained with the generator it evaluates, which could extend to image generation and editing.
  • One testable extension: apply the same progressive GRPO curriculum to larger VLMs or video-native encoders and measure whether the gains scale, since the current evidence is for a single 7B base.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes VQ-Insight, a 7B vision-language model trained with a three-stage progressive reinforcement learning scheme (image scoring warm-up, general task-specific temporal learning, and unified finetuning with a video generation model) to perform AIGC video preference comparison, multi-dimension scoring, and natural video scoring. Rewards include a multi-dimension ℓ1 scoring reward, a preference comparison reward, a temporal modeling reward based on shuffled frame tokens, and a length control reward. The authors report improvements over existing baselines on GenAI/MonetBench preference accuracy, LGVQ multi-dimension scoring, several natural video quality datasets, and an application to DPO-based finetuning of T2V-Turbo evaluated on VBench.

Significance. If the results hold, the paper would show that a 7B VLM can be turned into a competitive AIGC video quality evaluator and reward model using only a few thousand labeled examples plus outcome-driven GRPO, without large-scale SFT datasets or hand-crafted CoT annotations. The progressive curriculum, the combination of rule-based and temporal rewards, and the application to generator alignment are all interesting and timely. The paper also provides component ablations and an application study, which are strengths. However, the evidence falls short of the abstract's blanket 'consistently outperforms' claim, and the temporal modeling reward is not validated against human temporal-quality judgments, leaving a load-bearing part of the method unsupported.

major comments (4)
  1. [Abstract; Table 3] The abstract claims that VQ-Insight 'consistently outperforms state-of-the-art baselines' in natural video scoring, but Table 3 shows the opposite on the in-domain LSVQ-Test set: VQ-Insight's PLCC is 0.876 and SRCC is 0.875, both below Q-Align's 0.882 and 0.883, and on LSVQ-1080p its PLCC of 0.823 is below Q-Align's 0.833. The body text correctly says 'best or near-best', so the abstract overstates the findings. Please revise the claim and discuss the mixed results on these datasets.
  2. [Eq. (4), 'Temporal Modeling Reward'] The temporal modeling reward is not sufficiently validated for the role it plays in the paper. First, the quantities w_seq and w_rand are never formally defined; for continuous scoring tasks, where the output is a real-valued score and Eq. (6) uses an ℓ1 error, a 'probability of giving the correct answer' is not well-defined. Second, the reward operationalizes temporal understanding as sensitivity of the model's own output probabilities to token order, which is not the same as agreement with human temporal-quality judgments; because the policy controls both w_seq and w_rand, the reward can be increased by making shuffled responses less confident or less parseable without improving temporal quality. The only supporting evidence, Table 4, reports aggregate PLCC/KRCC/SRCC for multi-dimension scoring and does not break out the temporal dimension, and the text explicitly says TMR was not validated on preference tasks. Please either validate Eq. (4) against the human temporal MOS dimensions of LGVQ or LSVQ, or provide a per-dimension ablation showing that TMR specifically improves temporal quality.
  3. [Tables 1–6, 'Experimental Setup'] All reported numbers are point estimates without standard deviations, confidence intervals, or repeated runs. Several of the headline improvements are small relative to the likely run-to-run variation (e.g., GenAI tau of 50.80 vs. UnifiedReward's 49.67; VBench overall score of 0.8185 vs. 0.8167 for VideoDPO). Without multiple seeds or uncertainty estimates, the 'outperforms' claims and the ablation conclusions are not statistically substantiated. Please report at least three seeds or confidence intervals for the main comparisons and the two ablation tables.
  4. [Eq. (8) and 'United Finetuning of Generation and Understanding'] The Stage 3 alternating optimization creates a feedback loop in which VQ-Insight's own preferences select win-lose pairs for DPO, the updated generator produces new samples, and those samples are then used to update VQ-Insight. If VQ-Insight has systematic biases, this loop can amplify them, and the paper does not analyze or test for such bias amplification. Please discuss this risk and, if possible, compare the updated comparison model D'θ with the Stage 2 model Dθ on held-out human preference benchmarks to show that the loop does not degrade preference accuracy.
minor comments (5)
  1. [Eqs. (3) and (6)] The score range used in the ℓ1 rewards is not stated; please clarify whether scores are normalized to [0,1] or [0,100], since the reward scale and the meaning of 'significant' changes accordingly.
  2. [Eqs. (4)–(6)] The hyperparameters α=0.3, μ=0.8, γ=0.1, lmin=320, lmax=512, and λj=1 are given without sensitivity analysis; a short study or a justification of these values would strengthen confidence in the method.
  3. [Throughout] There are several typos: 'manipultation' in the Introduction, 'united/unified' is used inconsistently in headings, 'wieght' in Implementation Details, and 'TWR' should be 'TMR' in the ablation paragraph. Please proofread.
  4. [Figure 2] Figure 2 is very dense and the three training stages are difficult to distinguish; adding clear stage boundaries or numbered annotations would make the pipeline much easier to follow.
  5. [Reproducibility] The paper does not state whether code or trained weights will be released; please include an availability statement, as the preference and scoring results are otherwise hard to reproduce.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reduction found; the training rewards are externally supervised or hand-set, the claimed gains are benchmarked on external datasets, and the temporal-modeling proxy is an unvalidated operational definition rather than a circular derivation.

full rationale

The central quantities used in training are not constructed from the target outputs. Eq. (3) uses image ground-truth scores, Eq. (6) uses per-dimension ground-truth MOS values from LGVQ, Eq. (7) uses ground-truth preference/VQA labels, and the GRPO objective (Eq. (2)) is the standard clipped policy-gradient objective. Hyperparameters (alpha=0.3, mu=0.8, gamma=0.1, lambda_j=1) are declared constants, so no fitted parameter is renamed as a prediction. Eq. (4)'s temporal reward is an operational proxy: it defines 'captured temporal information' as w_seq > 0.8 * w_rand, and the paper itself concedes it is not validated on preference comparisons ('Since performing temporal shuffle on two videos simultaneously can lead to preference label confusion, we do not validate the effectiveness of TWR on the preference comparison task'). That is a construct-validity and Goodharting risk, not a circular derivation; the claimed scoring gains are checked against external human-labeled benchmarks (GenAI, MonetBench, LGVQ, LSVQ, LIVE-VQC, KonViD-1k), and generation gains in Table 6 are checked against independent VBench. Stage 3 is iterative co-training between evaluator and generator, but the final generation claims do not rest on the loop alone. The only self-citation used in the evaluation (Q-Insight) appears as a baseline comparison, not as a load-bearing premise. Thus I find no specific reduction of any claimed result to its inputs and no significant circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a small set of hand-set reward weights and thresholds, on the correctness of existing quality labels, and on the temporal shuffle proxy. No new physical or conceptual entities are introduced, so the invented-entities ledger is empty.

free parameters (5)
  • alpha (temporal reward magnitude) = 0.3
    Set by hand in Eq. (4) to weight the temporal shuffle reward.
  • mu (shuffle threshold) = 0.8
    Hand-set threshold in Eq. (4) that decides when a shuffled answer counts as temporally aware.
  • gamma (length reward magnitude) = 0.1
    Hand-set bonus in Eq. (5) for answers whose length is inside the target interval.
  • length interval (lmin, lmax) = 320, 512
    Empirically chosen answer length bounds in Eq. (5); the authors note an 'aha moment' at these settings.
  • dimension weights lambda_j = 1
    Equal weights for spatial, temporal, and alignment dimensions in Eq. (6); the authors set lambda_j = 1.
assumptions (5)
  • standard math GRPO objective in Eq. (2) is a working optimization method and does not require a critic or external reward model.
    The training pipeline relies on the prior GRPO formulation from Guo et al. (2025).
  • domain assumption Dataset labels (c_gt, v_gt) used in Eqs. (3), (6), and (7) are correct, unbiased, and transferable to AIGC video quality.
    The paper acknowledges annotator bias but treats the aggregated labels as ground truth.
  • ad hoc to paper A large difference between sequential and shuffled-answer probabilities indicates genuine temporal understanding.
    Eq. (4) defines the temporal modeling reward on this proxy, with no independent validation or calibration against human temporal quality scores.
  • standard math DPO in Eq. (9) improves T2V-Turbo given reliable win-lose pairs.
    Stage 3 applies the DiffusionDPO loss to align the generation model.
  • domain assumption Qwen2.5-VL-7B-Instruct is a suitable base VLM for video quality reasoning.
    All experiments start from this checkpoint, and the method is not analyzed across other base models.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VQ-Insight: Teaching VLMs for AI-Generated Video Quality Understanding via Progressive Visual Reinforcement Learning." pith.science (2026). https://pith.science/paper/EKFR44JN

@misc{pith2026250618564,
  author       = {Pith},
  title        = {Pith review of: VQ-Insight: Teaching VLMs for AI-Generated Video Quality Understanding via Progressive Visual Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EKFR44JN}},
  note         = {Machine review of arXiv:2506.18564}
}
read the original abstract

Recent advances in AI-generated content (AIGC) have led to the emergence of powerful text-to-video generation models. Despite these successes, evaluating the quality of AIGC-generated videos remains challenging due to limited generalization, lack of temporal awareness, heavy reliance on large-scale annotated datasets, and the lack of effective interaction with generation models. Most current approaches rely on supervised finetuning of vision-language models (VLMs), which often require large-scale annotated datasets and tend to decouple understanding and generation. To address these shortcomings, we propose VQ-Insight, a novel reasoning-style VLM framework for AIGC video quality assessment. Our approach features: (1) a progressive video quality learning scheme that combines image quality warm-up, general task-specific temporal learning, and joint optimization with the video generation model; (2) the design of multi-dimension scoring rewards, preference comparison rewards, and temporal modeling rewards to enhance both generalization and specialization in video quality evaluation. Extensive experiments demonstrate that VQ-Insight consistently outperforms state-of-the-art baselines in preference comparison, multi-dimension scoring, and natural video scoring, bringing significant improvements for video generation tasks.

Figures

Figures reproduced from arXiv: 2506.18564 by the authors.

Figure 1
Figure 1. We propose a reasoning-style vision-language model [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the proposed VQ-Insight and our progressive visual reinforcement learning framework. In stage 1, we [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Reasoning process comparison between Q-Insight and the proposed VQ-Insight on preference comparison tasks. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Generation result comparisons between our method and other competitive methods. The video generation model [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 14 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bai, S.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Song, S.; Dang, K.; Wang, P.; Wang, S.; Tang, J.; et al. 2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923

    arXiv 2025

  4. [4]

    Chen, H.; Zhang, Y.; Cun, X.; Xia, M.; Wang, X.; Weng, C.; and Shan, Y. 2024. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7310--7320

    2024

  5. [5]

    Deutsch, D.; Foster, G.; and Freitag, M. 2023. Ties Matter: Meta-Evaluating Modern Metrics with Pairwise Accuracy and Tie Calibration. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 12914--12929

    work page 2023

  6. [6]

    Feichtenhofer, C.; Fan, H.; Malik, J.; and He, K. 2019. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, 6202--6211

    2019

  7. [7]

    Feng, K.; Gong, K.; Li, B.; Guo, Z.; Wang, Y.; Peng, T.; Wang, B.; and Yue, X. 2025. Video-r1: Reinforcing video reasoning in mllms. arXiv preprint arXiv:2503.21776

    arXiv 2025

  8. [8]

    Guo, D.; Yang, D.; Zhang, H.; Song, J.; Zhang, R.; Xu, R.; Zhu, Q.; Ma, S.; Wang, P.; Bi, X.; et al. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    arXiv 2025

Show all 48 references
  1. [9]

    He, X.; Jiang, D.; Zhang, G.; Ku, M.; Soni, A.; Siu, S.; Chen, H.; Chandra, A.; Jiang, Z.; Arulraj, A.; et al. 2024. VideoScore: Building Automatic Metrics to Simulate Fine-grained Human Feedback for Video Generation. In Proceedings of the 2024 Conference on Empirical Methods ...

    2024

  2. [10]

    L.; and Choi, Y

    Hessel, J.; Holtzman, A.; Forbes, M.; Bras, R. L.; and Choi, Y. 2021. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718

    2021 arXiv

  3. [11]

    ???? CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers

    Hong, W.; Ding, M.; Zheng, W.; Liu, X.; and Tang, J. ???? CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers. In The Eleventh International Conference on Learning Representations

  4. [12]

    Hosu, V.; Hahn, F.; Jenadeleh, M.; Lin, H.; Men, H.; Szir \'a nyi, T.; Li, S.; and Saupe, D. 2017. The Konstanz natural video database (KoNViD-1k). In 2017 Ninth international conference on quality of multimedia experience (QoMEX), 1--6. IEEE

    2017

  5. [13]

    Hosu, V.; Lin, H.; Sziranyi, T.; and Saupe, D. 2020. KonIQ-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29: 4041--4056

    2020

  6. [14]

    Huang, Z.; He, Y.; Yu, J.; Zhang, F.; Si, C.; Jiang, Y.; Zhang, Y.; Wu, T.; Jin, Q.; Chanpaisit, N.; et al. 2024. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21807--21818

    2024

  7. [15]

    Jia, Z.; Zhang, Z.; Qian, J.; Wu, H.; Sun, W.; Li, C.; Liu, X.; Lin, W.; Zhai, G.; and Min, X. 2024. VQA2: Visual Question Answering for Video Quality Assessment. arXiv preprint arXiv:2411.03795

    2024 arXiv

  8. [16]

    Jiang, D.; Guo, Z.; Zhang, R.; Zong, Z.; Li, H.; Zhuo, L.; Yan, S.; Heng, P.-A.; and Li, H. 2025. T2i-r1: Reinforcing image generation with collaborative semantic-level and token-level cot. arXiv preprint arXiv:2505.00703

    2025 arXiv

  9. [17]

    Jiang, D.; Ku, M.; Li, T.; Ni, Y.; Sun, S.; Fan, R.; and Chen, W. 2024. Genai arena: An open evaluation platform for generative models. Advances in Neural Information Processing Systems, 37: 79889--79908

    2024

  10. [18]

    Li, J.; Feng, W.; Fu, T.-J.; Wang, X.; Basu, S.; Chen, W.; and Wang, W. Y. 2024 a . T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback. arXiv preprint arXiv:2405.18750

    2024 arXiv

  11. [19]

    Li, W.; Zhang, X.; Zhao, S.; Zhang, Y.; Li, J.; Zhang, L.; and Zhang, J. 2025. Q-insight: Understanding image quality via visual reinforcement learning. arXiv preprint arXiv:2503.22679

    2025 arXiv

  12. [20]

    Li, W.; Zhao, S.; Mou, C.; Sheng, X.; Zhang, Z.; Wang, Q.; Li, J.; Zhang, L.; and Zhang, J. 2024 b . OmniDrag: Enabling Motion Control for Omnidirectional Image-to-Video Generation. arXiv preprint arXiv:2412.09623

    2024 arXiv

  13. [21]

    Lin, Z.; Pathak, D.; Li, B.; Li, J.; Xia, X.; Neubig, G.; Zhang, P.; and Ramanan, D. 2024. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, 366--384. Springer

    2024

  14. [22]

    Liu, J.; Liu, G.; Liang, J.; Li, Y.; Liu, J.; Wang, X.; Wan, P.; Zhang, D.; and Ouyang, W. 2025 a . Flow-grpo: Training flow matching models via online rl. arXiv preprint arXiv:2505.05470

    2025 arXiv

  15. [23]

    Liu, J.; Liu, G.; Liang, J.; Yuan, Z.; Liu, X.; Zheng, M.; Wu, X.; Wang, Q.; Qin, W.; Xia, M.; et al. 2025 b . Improving Video Generation with Human Feedback. arXiv preprint arXiv:2501.13918

    2025 arXiv

  16. [24]

    Liu, R.; Wu, H.; Zheng, Z.; Wei, C.; He, Y.; Pi, R.; and Chen, Q. 2025 c . Videodpo: Omni-preference alignment for video diffusion generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, 8009--8019

    2025

  17. [25]

    Prabhudesai, M.; Mendonca, R.; Qin, Z.; Fragkiadaki, K.; and Pathak, D. 2024. Video diffusion alignment via reward gradients. arXiv preprint arXiv:2407.08737

    2024 arXiv

  18. [26]

    D.; Ermon, S.; and Finn, C

    Rafailov, R.; Sharma, A.; Mitchell, E.; Manning, C. D.; Ermon, S.; and Finn, C. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36: 53728--53741

    2023

  19. [27]

    Shen, H.; Liu, P.; Li, J.; Fang, C.; Ma, Y.; Liao, J.; Shen, Q.; Zhang, Z.; Zhao, K.; Zhang, Q.; et al. 2025. Vlm-r1: A stable and generalizable r1-style large vision-language model. arXiv preprint arXiv:2504.07615

    2025 arXiv

  20. [28]

    Sinno, Z.; and Bovik, A. C. 2018. Large-scale study of perceptual video quality. IEEE Transactions on Image Processing, 28(2): 612--627

    2018

  21. [29]

    Sun, W.; Wen, W.; Min, X.; Lan, L.; Zhai, G.; and Ma, K. 2024. Analysis of video quality datasets via design of minimalistic video quality models. IEEE Transactions on Pattern Analysis and Machine Intelligence

    2024

  22. [30]

    Wallace, B.; Dang, M.; Rafailov, R.; Zhou, L.; Lou, A.; Purushwalkam, S.; Ermon, S.; Xiong, C.; Joty, S.; and Naik, N. 2024. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8228--8238

    2024

  23. [31]

    C.; and Loy, C

    Wang, J.; Chan, K. C.; and Loy, C. C. 2023. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, volume 37, 2555--2563

    2023

  24. [32]

    Wang, W.; and Yang, Y. 2024. Vidprom: A million-scale real prompt-gallery dataset for text-to-video diffusion models. arXiv preprint arXiv:2403.06098

    2024 arXiv

  25. [33]

    Wang, Y.; Li, Z.; Zang, Y.; Wang, C.; Lu, Q.; Jin, C.; and Wang, J. 2025 a . Unified multimodal chain-of-thought reward model through reinforcement fine-tuning. arXiv preprint arXiv:2505.03318

    2025

  26. [34]

    Wang, Y.; Tan, Z.; Wang, J.; Yang, X.; Jin, C.; and Li, H. 2024. Lift: Leveraging human feedback for text-to-video model alignment. arXiv preprint arXiv:2412.04814

    2024 arXiv

  27. [35]

    Wang, Y.; Zang, Y.; Li, H.; Jin, C.; and Wang, J. 2025 b . Unified reward model for multimodal understanding and generation. arXiv preprint arXiv:2503.05236

    2025 arXiv

  28. [36]

    Wu, H.; Chen, C.; Hou, J.; Liao, L.; Wang, A.; Sun, W.; Yan, Q.; and Lin, W. 2022. Fast-vqa: Efficient end-to-end video quality assessment with fragment sampling. In European conference on computer vision, 538--554. Springer

    2022

  29. [37]

    Wu, H.; Zhang, Z.; Zhang, E.; Chen, C.; Liao, L.; Wang, A.; Xu, K.; Li, C.; Hou, J.; Zhai, G.; et al. 2024 a . Q-instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogn...

    2024

  30. [38]

    Wu, H.; Zhang, Z.; Zhang, W.; Chen, C.; Liao, L.; Li, C.; Gao, Y.; Wang, A.; Zhang, E.; Sun, W.; et al. 2024 b . Q-ALIGN: teaching LMMs for visual scoring via discrete text-defined levels. In Proceedings of the 41st International Conference on Machine Learning, 54015--54029

    2024

  31. [39]

    Xu, J.; Huang, Y.; Cheng, J.; Yang, Y.; Xu, J.; Wang, Y.; Duan, W.; Yang, S.; Jin, Q.; Li, S.; et al. 2024 a . Visionreward: Fine-grained multi-dimensional human preference learning for image and video generation. arXiv preprint arXiv:2412.21059

    2024 arXiv

  32. [40]

    Xu, Z.; Zhang, X.; Li, R.; Tang, Z.; Huang, Q.; and Zhang, J. 2024 b . Fakeshield: Explainable image forgery detection and localization via multi-modal large language models. arXiv preprint arXiv:2410.02761

    2024 arXiv

  33. [41]

    Xu, Z.; Zhang, X.; Zhou, X.; and Zhang, J. 2025. AvatarShield: Visual Reinforcement Learning for Human-Centric Video Forgery Detection. arXiv preprint arXiv:2505.15173

    2025

  34. [42]

    Yang, Z.; Teng, J.; Zheng, W.; Ding, M.; Huang, S.; Xu, J.; Yang, Y.; Hong, W.; Zhang, X.; Feng, G.; et al. 2024. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072

    2024 arXiv

  35. [43]

    Ye, Q.; Xu, H.; Ye, J.; Yan, M.; Hu, A.; Liu, H.; Qian, Q.; Zhang, J.; and Huang, F. 2024. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the ieee/cvf conference on computer vision and pattern recognition, 13040--13051

    2024

  36. [44]

    Ying, Z.; Mandal, M.; Ghadiyaram, D.; and Bovik, A. 2021. Patch-vq:'patching up'the video quality problem. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 14019--14029

    2021

  37. [45]

    You, Z.; Cai, X.; Gu, J.; Xue, T.; and Dong, C. 2025. Teaching Large Language Models to Regress Accurate Image Quality Scores using Score Distribution. arXiv preprint arXiv:2501.11561

    2025

  38. [46]

    You, Z.; Li, Z.; Gu, J.; Yin, Z.; Xue, T.; and Dong, C. 2024. Depicting beyond scores: Advancing image quality assessment through multi-modal language models. In European Conference on Computer Vision, 259--276

    2024

  39. [47]

    Zhang, Z.; Li, X.; Sun, W.; Jia, J.; Min, X.; Zhang, Z.; Li, C.; Chen, Z.; Wang, P.; Ji, Z.; et al. 2024. Benchmarking AIGC Video Quality Assessment: A Dataset and Unified Model. arXiv preprint arXiv:2407.21408

    2024 arXiv

  40. [48]

    Zheng, Z.; Peng, X.; Yang, T.; Shen, C.; Li, S.; Liu, H.; Zhou, Y.; Li, T.; and You, Y. 2024. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404

    2024 arXiv

Pith tools

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