Pith. sign in

REVIEW 4 major objections 5 minor 58 references

Towards Holistic Visual Quality Assessment of AI-Generated Videos: A LLM-Based Multi-Dimensional Evaluation Model

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read AIGVEval decomposes AI-generated video quality into technical, motion, and semantic dimensions, feeds the features to a fine-tuned LLM, and reports state-of-the-art results on the NTIRE 2025 AI-generated video quality task.

desk verdict Solid incremental AIGV quality metric with honest ablations and a real leaderboard result, but the 'holistic' title overreaches given the model ignores text-video consistency. read the letter →

arxiv 2506.04715 v2 pith:MCWZQS62 submitted 2025-06-05 cs.CV

classification cs.CV
keywords AI-generatedvideoqualityassessmentno-referencelargelanguagemodelssemanticanchorsLoRAfine-tuningmulti-dimensionalvisualtext-to-videoevaluationNTIRE2025challenge
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes AIGVEval, an automatic no-reference evaluator for AI-generated videos, and sets out to show that splitting visual quality into technical quality, motion quality, and video semantics, then feeding those features to a fine-tuned large language model, yields more accurate scores than existing video quality assessment methods. The authors argue that earlier LLM-based approaches leave the LLM undertrained because their encoders miss motion information and their prompts do not explain what the extracted feature tokens mean. AIGVEval therefore combines a 3D Swin transformer for distortions, SlowFast for dynamics, BLIP for semantics, and a Vicuna-7B LLM that receives the projected tokens through a prompt with semantic anchors naming each dimension. The model placed second in the NTIRE 2025 Quality Assessment of AI-Generated Content Challenge Track 2 and beat seven compared methods on the challenge test set. The practical payoff is a quality signal that could support content moderation for AI-generated video and guide improvements to text-to-video generators.

What carries the argument

The load-bearing mechanism is the triple-encoder front end paired with a semantic-anchor prompt inside the LLM. Technical quality is captured by a 3D Swin Transformer branch from DOVER operating on grid mini-patch sampled frames; motion quality by SlowFast-R50 with dense sampling and learnable positional embeddings; video semantics by BLIP's vision backbone; each encoder's output is projected into the LLM's token space. These tokens are concatenated with a fixed prompt that labels each block ('The key frames of this video are', 'the technical quality features of the video are', 'the motion quality features of the video are', 'Please assess the quality of this video'), and Vicuna-7B with LoRA fine-tuning maps the multimodal sequence to a five-level quality distribution. The final score is the softmax-weighted sum of the five level logits with weights 1 to 5. The semantic anchors are what enable the LLM to associate each feature block with the quality concept it should reason about.

What would settle it

Take a set of AIGVs whose technical and motion quality are matched but whose captions deliberately contradict the content, and have human raters score them; if a model that reads the prompt keeps its correlation with human scores while AIGVEval's rank correlation drops markedly, the claim that the three visual dimensions plus semantic anchors capture holistic quality is refuted. A second check is re-running the published ablation on a different AIGV dataset, since the reported motion-encoder gain comes from a single competition validation split.

Watch

Extended reading notes

Core claim

The central claim is that holistic visual quality of AI-generated video can be predicted from three complementary feature streams once an LLM is told what each stream represents. On the NTIRE 2025 challenge test set, AIGVEval reached a MainScore of 0.631 (PLCC 0.654, SROCC 0.608), ranking second among challenge teams and ahead of DOVER, SimpleVQA, FAST-VQA, Q-Align, T2VQA, Q-Eval-Score, and T2VEval trained on the same data. Validation-set ablations show that removing the technical encoder costs 0.149 MainScore, removing the motion encoder costs 0.085, removing the semantic encoder costs 0.040, and trading semantic anchors for direct token concatenation or cross-attention fusion costs 0.059 and 0.028 respectively. The paper also reports that on T2VQA-DB, where text-video consistency contributes to the ground truth, AIGVEval remains competitive but trails T2VQA and T2VEval, a gap the authors attribute to their deliberate exclusion of text prompts.

Load-bearing premise

The load-bearing premise is the paper's reading of the Q-Eval dataset description: that the MOS labels reflect visual quality only and do not encode text-video consistency; if those labels secretly include alignment with the prompt, then AIGVEval's deliberate omission of the text prompt would handicap it and the reported leaderboard scores would not mean what the paper claims.

Editorial extensions

If this is right

  • Accurate per-dimension scores give video-generator developers a direct diagnostic: a low technical or motion score identifies blur, noise, jitter, or low dynamics as the target for training or post-processing.
  • Because the ablation credits the anchors with a measurable gain, the same labeled-token prompting recipe should transfer to other LLM-based regression tasks that concatenate heterogeneous visual features.
  • The relative size of the ablation drops suggests that on this benchmark, technical and motion distortions dominate perceived quality, so generative models should prioritize temporal stability and sharpness over semantic variety.
  • On any benchmark whose ground truth includes text-video alignment, AIGVEval is expected to underperform models that consume the prompt, matching the paper's own observation on T2VQA-DB.

Reading between the lines

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

  • A natural extension the authors do not pursue is to add a fourth encoder or adapter that ingests the original text prompt, letting the same semantic-anchor framework cover text-video consistency without giving up the visual decomposition.
  • The ablation pattern suggests a testable hypothesis for other multimodal fusion pipelines: semantic anchors matter most when one or more feature streams are noisy, so similar gains should appear in image or audio quality regression with multiple expert encoders.
  • Because the leaderboard is a single test snapshot, the durable claim is the directional gain from motion encoding; re-running the same ablations on datasets from other generator families would show whether that gain is an artifact of the NTIRE test distribution.
  • The five-level softmax weighting is a coarse scoring head; a continuous regression head or question-answering style scoring might close the gap to T2VEval on text-alignment-sensitive data while keeping the three-encoder front end.
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 AIGVEval, a no-reference video quality assessment model for AI-generated videos (AIGVs), decomposing visual quality into semantic, technical, and motion dimensions. The video is encoded by BLIP (semantics), a 3D Swin Transformer (technical quality), and SlowFast-R50 (motion); the resulting tokens are concatenated with specially designed 'semantic anchor' text prompts and fed into a Vicuna-7B LLM, which is fine-tuned with LoRA to regress a five-level quality score. The method reports second place in the NTIRE 2025 Quality Assessment of AI-Generated Content Challenge Track 2 (MainScore 0.631 on the test set), validates each component via ablations on the validation set (MainScore 0.695 with all components), and shows competitive but lower performance on T2VQA-DB (PLCC 0.7494), which the authors attribute to the dataset's inclusion of text-video consistency.

Significance. If the Q-Eval label protocol is exactly as the paper describes, the paper demonstrates a well-engineered architecture and validates its three-encoder decomposition through ablations. The external leaderboard result and the public benchmark comparison provide evidence of effectiveness beyond a single self-built test set, and the released code supports reproducibility. The main significance is limited by the fact that the model deliberately ignores the textual prompts that accompany the videos: the evidence supports a claim about visual-quality-only assessment, not about holistic AIGV quality in the broader sense used by methods such as T2VQA or AIGC-VQA, which explicitly include text-video alignment. The paper is a useful challenge report and architectural study, but its conclusions need to be scoped carefully.

major comments (4)
  1. [§4.1, Table 4, Abstract] The central claim of 'holistic' AIGV quality assessment is not supported beyond visual-quality-only protocols. The method never consumes the textual prompts that Section 4.1 says are available in all three splits, and the reported leaderboard result depends entirely on the unverified assertion that Q-Eval MOS labels are 'solely based on the visual quality of the AIGVs and do not involve an assessment of text-video consistency.' Table 4 is directly relevant: on T2VQA-DB, which the authors state includes text-video consistency, AIGVEval reaches 0.7494 PLCC against 0.8175 for T2VEval and 0.8066 for T2VQA, and the authors attribute this gap to the fact that text-video consistency is 'not considered in our method.' That is an admission that the method is not holistic for AIGV quality as that term is used in the AIGC-VQA/T2VQA line of work. Please either verify the Q-Eval label protocol with explicit evidence and citations, or rescope the title, abstract, and conclusion to 'visual-quality-only assessment' and present the T2VQA-DB shortfall as a known limitation rather than an afterthought.
  2. [§4.4, Table 1 vs. Table 3] The validation-to-test gap is not addressed. Table 3 reports MainScore 0.695 on the validation set, while Tables 1 and 2 report MainScore 0.631 on the test set; PLCC drops from 0.706 to 0.654. The paper does not report error bars, multiple seeds, or any discussion of how the validation set was used for hyperparameter selection, so the reader cannot tell whether this gap is statistical noise or overfitting. Because the central effectiveness claim is the leaderboard score, this gap should be quantified and discussed. At minimum, report the variance across at least three training runs and state the validation-based model-selection procedure.
  3. [§3.1–§3.3, §4.2] Projector architectures and update rules are underspecified. Equations (1), (3), and (4) define Proj_s, Proj_t, and Proj_m only by input/output shapes, and Figure 3 marks them 'Trainable' without giving layer types, dimensions, or parameter counts. Section 4.2 explicitly freezes only the BLIP component; it does not state whether the 3D-Swin and SlowFast backbones are frozen or fine-tuned, nor whether the positional embeddings added to motion tokens are learned or fixed. Without these details, the reported results cannot be reproduced, and the ablation results cannot be cleanly attributed to the three encoders versus their fine-tuning regimes.
  4. [§3.6, Eq. (7)] Equation (7) is not well defined: the function e(y_i, y_j) never appears in the text. Depending on its intended meaning (e.g., the sign of the MOS difference, or an indicator with an epsilon), the rank loss behaves very differently, and the reader cannot verify the training objective. Similarly, Eq. (5) says 'softmax of each token' but λ_i is presumably the logit of the i-th quality-level token; please define λ_i and the softmax normalization explicitly.
minor comments (5)
  1. [Throughout] Typos and inconsistent notation should be corrected: 'senmatic' (§4.2), 'Tabel 3' (repeated), 'descript' (§3.4), and the dimension mismatch in Eq. (2), where V is written as R^{T×W×W} after the text introduced T×W×H.
  2. [§4.3] MainScore is described as 'ignoring the sign and reporting the average of absolute values (PLCC+SROCC)/2'; please give the explicit formula, since PLCC and SROCC are signed correlation coefficients and the current wording is ambiguous.
  3. [References [34] and [35]] References [34] and [35] appear to describe the same T2VEval method with near-identical titles; please merge them or clarify the distinction between the two entries.
  4. [Figure 3] In Figure 3, the prompt text includes stray '?' characters and the arrow labels are difficult to read; please regenerate the figure with readable fonts and clean text.
  5. [§4.2] Please report the batch size, number of GPUs, and total optimizer steps; the current description gives epochs, learning rate, and warmup epochs but not the effective batch size, which is necessary for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the effectiveness claims rest on held-out NTIRE 2025 Track 2 data and the external T2VQA-DB benchmark, while the self-citations to T2VEval appear only as related work and comparison baselines, not as load-bearing premises.

full rationale

The derivation chain is self-contained: the three encoders (BLIP for semantics, 3D Swin for technical quality, SlowFast for motion) produce tokens that are concatenated with textual semantic anchors and fed to a Vicuna LLM, whose logits over five quality levels are converted to a score by Eq. (5). Nothing in this chain is defined in terms of the target MOS, and no fitted parameter is renamed as a prediction. The central effectiveness claim is supported by the held-out NTIRE 2025 Track 2 test set (Table 1) and by the external T2VQA-DB benchmark (Table 4). The self-citations to T2VEval ([34], [35]) appear in related work and as a baseline; they do not justify the proposed architecture or the semantic-anchor prompt design, so they are not load-bearing. Section 4.1 contains an explicit data-description assumption that Q-Eval MOSs are 'solely based on the visual quality of the AIGVs and do not involve an assessment of text-video consistency.' This is an unsupported premise that, if false, would weaken the 'holistic' claim, but it is not a circular step. The paper also flags its own scope limitation on T2VQA-DB ('This may be because T2VQA-DB considers the consistency of video and textual prompts, which is not considered in our method'), making the limitation explicit rather than hiding it. The equal-interval weights 1-5 in Eq. (5) are a scoring convention, not fitted to the test labels. No equation is equivalent to its own input by construction, and no prediction reduces to a fit, so circularity is absent.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method uses only existing model components and standard hyperparameters; no new entities are introduced. The ledger lists the hand-set constants and domain assumptions that the central claim depends on.

free parameters (5)
  • Grid size G_f = 32
    Chosen for GMS sampling to balance local detail and global coverage.
  • SlowFast alpha = 4
    Frame sampling ratio for fast branch.
  • LoRA rank = 8
    Low-rank adaptation matrix rank.
  • Loss weight lambda = 0.3
    Balance between PLCC and rank loss.
  • Quality level weights = 1,2,3,4,5
    Fixed linear weights for softmax probabilities; chosen by hand.
assumptions (4)
  • domain assumption Pretrained encoders (BLIP, 3D Swin-T, SlowFast) provide features that are informative for AI-generated video quality.
    The method freezes these encoders and relies on their transferable features; see Sections 3.1-3.3.
  • domain assumption Q-Eval dataset MOS labels are reliable and reflect solely visual quality, not text-video alignment.
    The paper states this in Section 4.1; if false, the model's objective is mis-specified.
  • domain assumption Vicuna v1.5 7B with LoRA can map concatenated visual tokens to quality scores via the designed prompt.
    The quality regression module uses the LLM as a regressor; the paper provides no theoretical guarantee.
  • domain assumption The softmax of the five quality-level tokens followed by weighted sum with fixed weights produces scores comparable to MOS.
    Section 3.6, Eq. (5); assumes equal-interval scale.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Holistic Visual Quality Assessment of AI-Generated Videos: A LLM-Based Multi-Dimensional Evaluation Model." pith.science (2026). https://pith.science/paper/MCWZQS62

@misc{pith2026250604715,
  author       = {Pith},
  title        = {Pith review of: Towards Holistic Visual Quality Assessment of AI-Generated Videos: A LLM-Based Multi-Dimensional Evaluation Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MCWZQS62}},
  note         = {Machine review of arXiv:2506.04715}
}
read the original abstract

The development of AI-Generated Video (AIGV) technology has been remarkable in recent years, significantly transforming the paradigm of video content production. However, AIGVs still suffer from noticeable visual quality defects, such as noise, blurriness, frame jitter and low dynamic degree, which severely impact the user's viewing experience. Therefore, an effective automatic visual quality assessment is of great importance for AIGV content regulation and generative model improvement. In this work, we decompose the visual quality of AIGVs into three dimensions: technical quality, motion quality, and video semantics. For each dimension, we design corresponding encoder to achieve effective feature representation. Moreover, considering the outstanding performance of large language models (LLMs) in various vision and language tasks, we introduce a LLM as the quality regression module. To better enable the LLM to establish reasoning associations between multi-dimensional features and visual quality, we propose a specially designed multi-modal prompt engineering framework. Additionally, we incorporate LoRA fine-tuning technology during the training phase, allowing the LLM to better adapt to specific tasks. Our proposed method achieved \textbf{second place} in the NTIRE 2025 Quality Assessment of AI-Generated Content Challenge: Track 2 AI Generated video, demonstrating its effectiveness. Codes can be obtained at https://github.com/QiZelu/AIGVEval.

Figures

Figures reproduced from arXiv: 2506.04715 by the authors.

Figure 1
Figure 1. Typical distortion types of AI-Generated videos. (a) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework of our method. LLMs understand the specific meaning of the features extracted by the encoders, making it quite a challenge to establish reasoning associations between features and quality. To address the aforementioned issues, in this work, we propose a LLM-based AIGV visual quality evaluation model called AIGVEval, which consists of three encoders and a quality regression module incorporating the LLM.… view at source ↗
Figure 3
Figure 3. (a) The overall architecture of our proposed AIGVEval, which includes three encoders and an LLM-based quality regression [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 24 canonical work pages

  1. [1]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

  2. [2]

    Paligemma: A versatile 3b vlm for trans- fer.arXiv preprint arXiv:2407.07726, 2024

    Lucas Beyer, Andreas Steiner, Andr ´e Susano Pinto, Alexan- der Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for trans- fer.arXiv preprint arXiv:2407.07726, 2024. 3

  3. [3]

    Video generation models as world simulators.OpenAI Blog, 1:8, 2024

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luh- man, Eric Luhman, et al. Video generation models as world simulators.OpenAI Blog, 1:8, 2024. 2

  4. [4]

    Gaia: Rethinking action quality assessment for ai-generated videos.Advances in Neural Information Processing Systems, 37:40111–40144, 2024

    Zijian Chen, Wei Sun, Yuan Tian, Jun Jia, Zicheng Zhang, Wang Jiarui, Ru Huang, Xiongkuo Min, Guangtao Zhai, and Wenjun Zhang. Gaia: Rethinking action quality assessment for ai-generated videos.Advances in Neural Information Processing Systems, 37:40111–40144, 2024. 3

  5. [5]

    Finevq: Fine-grained user generated content video quality assessment.arXiv preprint arXiv:2412.19238,

    Huiyu Duan, Qiang Hu, Jiarui Wang, Liu Yang, Zitong Xu, Lu Liu, Xiongkuo Min, Chunlei Cai, Tianxiao Ye, Xiaoyun Zhang, et al. Finevq: Fine-grained user generated content video quality assessment.arXiv preprint arXiv:2412.19238,

  6. [6]

    Slowfast networks for video recognition

    Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202–6211, 2019. 3, 4, 5

  7. [7]

    Lmm-vqa: Advancing video quality assessment with large multimodal models.arXiv preprint arXiv:2408.14008,

    Qihang Ge, Wei Sun, Yu Zhang, Yunhao Li, Zhongpeng Ji, Fengyu Sun, Shangling Jui, Xiongkuo Min, and Guangtao Zhai. Lmm-vqa: Advancing video quality assessment with large multimodal models.arXiv preprint arXiv:2408.14008,

  8. [8]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Ab- hinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024. 3

Show all 58 references
  1. [9]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025. 3

  2. [10]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 2

  3. [11]

    Videoscore: Building automatic metrics to simulate fine-grained human feedback for video genera- tion.arXiv preprint arXiv:2406.15252, 2024

    Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, et al. Videoscore: Building automatic metrics to simulate fine-grained human feedback for video genera- tion.arXiv preprint arXiv:2406.15252, 2024. 3

  4. [12]

    Cogvideo: Large-scale pretraining for text-to-video generation via transformers.arXiv preprint arXiv:2205.15868, 2022

    Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers.arXiv preprint arXiv:2205.15868, 2022. 2

  5. [13]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022. 2, 5

  6. [14]

    Vbench++: Comprehensive and ver- satile benchmark suite for video generative models.arXiv preprint arXiv:2411.13503, 2024

    Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, et al. Vbench++: Comprehensive and ver- satile benchmark suite for video generative models.arXiv preprint arXiv:2411.13503, 2024. 2

  7. [15]

    T2vbench: Benchmarking temporal dynamics for text-to- video generation

    Pengliang Ji, Chuyang Xiao, Huilin Tai, and Mingxiao Huo. T2vbench: Benchmarking temporal dynamics for text-to- video generation. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 5325–5335, 2024. 2, 3

  8. [16]

    Vqa 2: Visual question answering for video quality assessment.arXiv preprint arXiv:2411.03795,

    Ziheng Jia, Zicheng Zhang, Jiaying Qian, Haoning Wu, Wei Sun, Chunyi Li, Xiaohong Liu, Weisi Lin, Guangtao Zhai, and Xiongkuo Min. Vqa 2: Visual question answering for video quality assessment.arXiv preprint arXiv:2411.03795,

  9. [17]

    Mantis: Interleaved multi-image instruction tuning.arXiv preprint arXiv:2405.01483, 2024

    Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning.arXiv preprint arXiv:2405.01483, 2024. 3

  10. [18]

    The kinetics hu- man action video dataset.arXiv preprint arXiv:1705.06950,

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

  11. [19]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024. 2

  12. [20]

    Subjective-aligned dataset and metric for text-to-video qual- ity assessment

    Tengchuan Kou, Xiaohong Liu, Zicheng Zhang, Chunyi Li, Haoning Wu, Xiongkuo Min, Guangtao Zhai, and Ning Liu. Subjective-aligned dataset and metric for text-to-video qual- ity assessment. InProceedings of the 32nd ACM Interna- tional Conference on Multimedia, pages 7793–7802, ...

  13. [21]

    Bowen Li, Weixia Zhang, Meng Tian, Guangtao Zhai, and Xianpei Wang. Blindly assess quality of in-the-wild videos via quality-aware pre-training and motion perception.IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5944–5958, 2022. 7

  14. [22]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InInterna- tional conference on machine learning, pages 12888–12900. PMLR, 2022. 2, 3, 5, 7

  15. [23]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InIn- ternational conference on machine learning, pages 19730– 19742. PMLR, 2023. 3

  16. [24]

    Unmasked teacher: Towards training-efficient video foundation models

    Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 19948–19960, 2023. 2

  17. [25]

    Evaluating text-to-visual generation with image-to-text gen- eration

    Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text gen- eration. InEuropean Conference on Computer Vision, pages 366–384. Springer, 2024. 2

  18. [26]

    Evalcrafter: Benchmarking and eval- uating large video generation models

    Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and eval- uating large video generation models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Re...

  19. [27]

    Fetv: A bench- mark for fine-grained evaluation of open-domain text-to- video generation.Advances in Neural Information Process- ing Systems, 36, 2024

    Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A bench- mark for fine-grained evaluation of open-domain text-to- video generation.Advances in Neural Information Process- ing Systems, 36, 2024. 7

  20. [28]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 11976–11986,

  21. [29]

    Video swin transformer

    Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202–3211, 2022. 5

  22. [30]

    Aigc- vqa: A holistic perception metric for aigc video quality assessment

    Yiting Lu, Xin Li, Bingchen Li, Zihao Yu, Fengbin Guan, Xinrui Wang, Ruling Liao, Yan Ye, and Zhibo Chen. Aigc- vqa: A holistic perception metric for aigc video quality assessment. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6384–...

  23. [31]

    Step-video-t2v technical re- port: The practice, challenges, and future of video founda- tion model.arXiv preprint arXiv:2502.10248, 2025

    Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xi- aoniu Song, Xing Chen, et al. Step-video-t2v technical re- port: The practice, challenges, and future of video founda- tion model.arXiv preprint arXiv:2502.10248, 2025. 2

  24. [32]

    Open-sora 2.0: Training a commercial-level video generation model in 200k.arXiv preprint arXiv:2503.09642, 2025

    Xiangyu Peng, Zangwei Zheng, Chenhui Shen, Tom Young, Xinying Guo, Binluo Wang, Hang Xu, Hongxin Liu, Mingyan Jiang, Wenjun Li, et al. Open-sora 2.0: Training a commercial-level video generation model in 200k.arXiv preprint arXiv:2503.09642, 2025. 2

  25. [33]

    Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023. 2

  26. [34]

    T2veval: Benchmark dataset and objective evaluation method for t2v-generated videos, 2025

    Zelu Qi, Ping Shi, Shuqi Wang, Chaoyang Zhang, Fei Zhao, Zefeng Ying, Da Pan, Xi Yang, Zheqi He, and Teng Dai. T2veval: Benchmark dataset and objective evaluation method for t2v-generated videos, 2025. 6, 7

  27. [35]

    Comprehensive subjective and objective evaluation method for text-generated video.arXiv preprint arXiv:2501.08545, 2025

    Zelu Qi, Ping Shi, Shuqi Wang, Zhaoyang Zhang, Zefeng Ying, and Da Pan. Comprehensive subjective and objective evaluation method for text-generated video.arXiv preprint arXiv:2501.08545, 2025. 3

  28. [36]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, p...

  29. [37]

    Make-a-video: Text-to-video generation without text-video data.arXiv preprint arXiv:2209.14792,

    Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data.arXiv preprint arXiv:2209.14792,

  30. [38]

    A deep learning based no-reference quality assessment model for ugc videos

    Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai. A deep learning based no-reference quality assessment model for ugc videos. InProceedings of the 30th ACM Interna- tional Conference on Multimedia, pages 856–865, 2022. 6, 7

  31. [39]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of con- text.arXiv preprint arXiv:2403.05530, 2024

    Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of con- text.arXiv preprint arXiv:2403.05530, 2024. 3

  32. [40]

    Aigv-assessor: Benchmarking and eval- uating the perceptual quality of text-to-video generation with lmm.arXiv preprint arXiv:2411.17221, 2024

    Jiarui Wang, Huiyu Duan, Guangtao Zhai, Juntong Wang, and Xiongkuo Min. Aigv-assessor: Benchmarking and eval- uating the perceptual quality of text-to-video generation with lmm.arXiv preprint arXiv:2411.17221, 2024. 3

  33. [41]

    Videomae v2: Scaling video masked autoencoders with dual masking

    Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi- nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 14549–14560, 2023. 2

  34. [42]

    Internvid: A large-scale video-text dataset for multimodal understanding and generation.arXiv preprint arXiv:2307.06942, 2023

    Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation.arXiv preprint arXiv:2307.06942, 2023. 7

  35. [43]

    An ensemble approach to short-form video quality assess- ment using multimodal llm

    Wen Wen, Yilin Wang, Neil Birkbeck, and Balu Adsumilli. An ensemble approach to short-form video quality assess- ment using multimodal llm. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2025. 3

  36. [44]

    Godiva: Gen- erating open-domain videos from natural descriptions.arXiv preprint arXiv:2104.14806, 2021

    Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Gen- erating open-domain videos from natural descriptions.arXiv preprint arXiv:2104.14806, 2021. 7

  37. [45]

    Fast- vqa: Efficient end-to-end video quality assessment with frag- ment sampling

    Haoning Wu, Chaofeng Chen, Jingwen Hou, Liang Liao, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Fast- vqa: Efficient end-to-end video quality assessment with frag- ment sampling. InEuropean conference on computer vision, pages 538–554. Springer, 2022. 4, 6, 7

  38. [46]

    Exploring video quality assessment on user gener- ated contents from aesthetic and technical perspectives

    Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jing- wen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user gener- ated contents from aesthetic and technical perspectives. In Proceedings of the IEEE/CVF International Conferenc...

  39. [47]

    Q-align: Teaching lmms for visual scoring via discrete text-defined levels.arXiv preprint arXiv:2312.17090, 2023

    Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching lmms for visual scoring via discrete text-defined levels.arXiv preprint arXiv:2312.17090, 2023. 3, 6

  40. [48]

    Q-instruct: Improving low-level visual abilities 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, et al. Q-instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF conference on computer v...

  41. [49]

    Grit: A gener- ative region-to-text transformer for object understanding

    Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, and Lijuan Wang. Grit: A gener- ative region-to-text transformer for object understanding. In European Conference on Computer Vision, pages 207–224. Springer, 2024. 2

  42. [50]

    Ntire 2025 xgc quality assessment challenge: Methods and results.CVPR Workshop, 2025

    Qiang Hu Xiaohong Liu, Xiongkuo Min. Ntire 2025 xgc quality assessment challenge: Methods and results.CVPR Workshop, 2025. 2

  43. [51]

    Imagere- ward: Learning and evaluating human preferences for text- to-image generation.Advances in Neural Information Pro- cessing Systems, 36:15903–15935, 2023

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere- ward: Learning and evaluating human preferences for text- to-image generation.Advances in Neural Information Pro- cessing Systems, 36:15903–15935, 2023. 2

  44. [52]

    Imagere- ward: Learning and evaluating human preferences for text- to-image generation.Advances in Neural Information Pro- cessing Systems, 36, 2024

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere- ward: Learning and evaluating human preferences for text- to-image generation.Advances in Neural Information Pro- cessing Systems, 36, 2024. 7

  45. [53]

    Qwen2.5 technical report.arXiv preprint arXiv:2412.15115, 2024

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report.arXiv preprint arXiv:2412.15115, 2024. 3

  46. [54]

    Chronomagic-bench: A bench- mark for metamorphic evaluation of text-to-time-lapse video generation.Advances in Neural Information Processing Sys- tems, 37:21236–21270, 2024

    Shenghai Yuan, Jinfa Huang, Yongqi Xu, Yaoyang Liu, Shaofeng Zhang, Yujun Shi, Rui-Jie Zhu, Xinhua Cheng, Jiebo Luo, and Li Yuan. Chronomagic-bench: A bench- mark for metamorphic evaluation of text-to-time-lapse video generation.Advances in Neural Information Processing Sys- t...

  47. [55]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. InProceedings of the IEEE/CVF international conference on computer vision, pages 11975–11986, 2023. 3

  48. [56]

    Q-eval-100k: Evaluating visual quality and alignment level for text-to-vision content.arXiv preprint arXiv:2503.02357, 2025

    Zicheng Zhang, Tengchuan Kou, Shushi Wang, Chunyi Li, Wei Sun, Wei Wang, Xiaoyu Li, Zongyu Wang, Xuezhi Cao, Xiongkuo Min, et al. Q-eval-100k: Evaluating visual quality and alignment level for text-to-vision content.arXiv preprint arXiv:2503.02357, 2025. 1, 6

  49. [57]

    Judging llm-as-a-judge with mt-bench and chatbot arena.Advances in Neural Information Processing Systems, 36:46595–46623, 2023

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena.Advances in Neural Information Processing Systems, 36:46595–46623, 2023. 3, 5

  50. [58]

    Open-sora: Democratizing efficient video production for all, march 2024.URL https://github

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, march 2024.URL https://github. com/hpcaitech/Open-Sora, 1(3):4, 2024. 2

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.