Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Are Synthetic Videos Useful? A Benchmark for Retrieval-Centric Evaluation of Synthetic Videos

T0 review · 5 major / 6 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 Synthetic videos with strong object-action alignment boost retrieval

desk verdict Solid new benchmark, but the augmentation claim is not actually tested — needs a real-data mixing experiment before it can be trusted. read the letter →

arxiv 2507.02316 v1 pith:HOIJNFAL submitted 2025-07-03 cs.CV

Zecheng Zhao , Selena Song , Tong Chen , Zhi Chen , Shazia Sadiq , Yadan Luo This is my paper · ORCID
classification cs.CV
keywords text-to-videoretrievalsyntheticvideobenchmarksemanticalignmentdatasetaugmentationqualityassessmenthumanannotationgenerationAuto-Evaluator
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

The paper argues that current quality metrics for text-to-video models are not enough to tell whether synthetic videos are useful for building retrieval systems. It introduces SynTVA, a set of 2,400 synthetic video-text pairs annotated by humans on four alignment dimensions, and uses it to show that semantic alignment, especially Object & Scene and Actions, predicts downstream text-to-video retrieval gains. Training retrieval models on high-alignment synthetic clips, even just twenty samples, improves recall over a pretrained baseline, making the curated HQ-SynTVA subset a practical augmentation resource. An Auto-Evaluator trained on seven standard video-quality metrics approximates the two critical human alignment scores, reducing the need for costly annotation. This matters because it converts the vague question of whether synthetic videos are useful into a measurable, task-specific one.

What carries the argument

The load-bearing object is SynTVA, a dataset of 800 diverse text queries, 2,400 synthetic video-text pairs generated from three text-to-video models, and human ratings on four semantic alignment dimensions: Object & Scene, Actions, Attributes, and Prompt Fidelity. The argument is carried by the Auto-Evaluator, two linear regression models that map seven standard VQA metrics from VBench to human scores on the two dimensions found most retrieval-relevant, and by the augmentation protocol that trains the X-Pool retrieval model on high- and low-alignment subsets of nine sizes and measures Recall@1 on MSRVTT 1K-A. The mechanism linking all parts is the claimed correlation between alignment quality and downstream retrieval utility, with Object & Scene and Actions identified as the dimensions that matter.

What would settle it

Re-score a random sample of SynTVA pairs with a fresh set of judges and compute inter-annotator agreement; if agreement is low, or if training retrieval models on subsets ranked high by the Auto-Evaluator no longer beats training on random subsets in MSRVTT 1K-A Recall@1, the claimed utility signal would not hold.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the utility of synthetic video for text-to-video retrieval is governed by semantic alignment between the prompt and the generated footage, specifically how faithfully the video depicts the requested objects and scenes and the requested actions, rather than by generic visual quality alone. The evidence is direct: models trained on human-ranked high-quality subsets for Object & Scene and Actions outperform those trained on low-quality subsets by 0.8 and 0.6 in Recall@1, and even 20 high-quality samples beat the pretrained baseline across all alignment dimensions. In contrast, temporal consistency and motion smoothness show weak or negative correlations with human alignment judgments, suggesting that appearance-focused metrics alone mislead when used to select training data. The paper therefore proposes SynTVA as a retrieval-centric evaluation benchmark and the Auto-Evaluator as a way to score new synthetic videos without fresh human annotation.

Load-bearing premise

The entire quality-utility chain treats the average of five human judges' rubric scores as ground truth for how well a video matches its text, and the paper does not report how much the judges agreed, so if those labels are noisy, the high/low split, the Auto-Evaluator, and the utility conclusions all inherit that noise.

Editorial extensions

If this is right

  • Training text-to-video retrieval models on high-alignment synthetic clips improves Recall@1 over the pretrained baseline, with gains generally growing as more high-quality samples are added.
  • Object & Scene and Actions are the alignment dimensions that determine utility, so selecting or generating synthetic videos with these two dimensions in mind is the path to effective augmentation.
  • The Auto-Evaluator shows that existing appearance-focused metrics, when combined in the right way, can approximate human alignment judgment and pick out high-utility synthetic data without fresh human annotation.
  • The SynTVA benchmark protocol can be reused to evaluate future text-to-video generators on retrieval-centric utility instead of only perceptual quality.

Reading between the lines

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

  • Editorial inference: The same quality-versus-utility separation likely applies to other downstream tasks such as action recognition or embodied agents, but the four rubric dimensions would need to be re-targeted to the semantics that matter for each task.
  • Editorial inference: The Auto-Evaluator could be inverted into a reward model for prompt refinement, generating high-utility training videos on demand rather than merely filtering already-generated ones.
  • Editorial inference: Because the paper reports no inter-annotator agreement, the reliability of the benchmark and the transfer of the Auto-Evaluator to new generators remain open; re-annotating a small subset with fresh judges would settle both questions.
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

5 major / 6 minor

Summary. The paper introduces SynTVA, a benchmark dataset of 2,400 synthetic video-text pairs generated from 800 GPT-4o-derived queries spanning 20 MSRVTT-inspired categories, using three text-to-video models (Cosmos, Mochi, Wan2.1). Each pair is annotated by five human judges along four semantic alignment dimensions (Object & Scene, Action, Attributes, Prompt Fidelity). The authors analyze correlations between standard VQA metrics and these human alignment scores, train two linear Auto-Evaluators to predict Object & Scene and Action scores from VBench features, and conduct text-to-video retrieval (TVR) experiments with X-Pool trained on synthetic-only subsets of varying alignment quality and quantity. They report that high-quality synthetic subsets improve R@1 on MSRVTT 1K-A and conclude that SynTVA is a valuable resource for data augmentation and that VQA metrics can predict TVR utility.

Significance. If the central claims held, SynTVA would be a useful community resource: it offers a relatively large set of synthetic video-text pairs with multi-dimensional human alignment annotations, and the attempt to connect appearance-based VQA metrics to downstream retrieval utility is timely. The dataset release and the Auto-Evaluator idea are genuine contributions, and the paper is honest about the cost of human annotation. However, the paper's headline practical claim — that SynTVA enables dataset augmentation that measurably improves TVR — is not actually evaluated, because all TVR experiments train on synthetic-only subsets and never combine SynTVA with real training data. The statistical reporting is also too thin to support the quantitative conclusions. With additional experiments and stronger reporting, the benchmark could still make a solid contribution.

major comments (5)
  1. [Section 3.2, Figure 10] The abstract and conclusion claim that SynTVA is 'a valuable asset for dataset augmentation' and that high-quality samples 'measurably improve TVR outcomes,' but no experiment in Section 3.2 adds HQ-SynTVA to real MSRVTT training data. All conditions train X-Pool on synthetic-only subsets, with Random and Low-Quality synthetic controls compared against a pre-trained baseline. Augmentation, by definition, adds data to an existing training set, so a synthetic-only training result does not establish that adding HQ-SynTVA to a real-data-trained model will help; it may even hurt due to distribution shift. The paper should include conditions that train on real MSRVTT data alone and on real data plus HQ-SynTVA, with matched data quantities.
  2. [Section 3.2, Figure 10] The central TVR result is reported only as a figure with no numeric table, no error bars, no significance tests, no number of runs, and no full training hyperparameters (epochs, batch size, learning rate, optimizer, random seeds). The claims that 'models trained on high-quality synthetic data consistently outperform the pre-trained baseline across all alignment dimensions' and that performance gaps are '0.8 and 0.6' cannot be quantitatively assessed or reproduced from the text. A table of R@1 values with standard deviations across multiple seeds and significance tests is needed for these load-bearing claims.
  3. [Section 3.1.2 and Section 3.2] There is a circularity and extrapolation problem in the Auto-Evaluator narrative. The human alignment scores are used both to define the High/Low quality subsets for the TVR experiments and as the training targets for the Auto-Evaluator, and the paper then concludes that VQA metrics 'have a positive correlation with TVR utility' (Section 1) and that the Auto-Evaluator framework 'suggests a potential direction for leveraging appearance-based VQA metrics.' However, no experiment ever selects synthetic videos with the Auto-Evaluator and trains a TVR model on that selection. The claimed VQA-to-utility link is therefore not directly tested; an experiment comparing TVR performance when training data are selected by (a) human scores, (b) Auto-Evaluator scores, and (c) random selection is required.
  4. [Section 2.2 and Section 2.3] The human annotations are the ground truth for every downstream conclusion, but the paper reports no inter-annotator agreement statistics (e.g., Fleiss' kappa or Krippendorff's alpha), no details on how the five judges' scores were aggregated (mean, median, majority), and no information about judge qualifications, training, or quality control beyond 'cleaning and aggregating.' Given that the High/Low subsets, the Auto-Evaluator training targets, and the central quality-utility correlations all inherit these labels, the missing agreement analysis is a load-bearing gap and should be reported.
  5. [Section 3.1.1 and Section 3.1.2] The correlational and Auto-Evaluator results lack statistical rigor. Figure 8 and Figure 9 report Pearson and Spearman correlations without confidence intervals, p-values, or the underlying sample sizes per model, and the Auto-Evaluator is a linear regression fit on an 80/20 split with no cross-validation, no confidence intervals on the coefficients, and no regularization. The claims that 'Temporal Flickering shows 0.00% contribution' and that Motion Smoothness is unimportant for Actions are not supported without uncertainty estimates or a more robust evaluation protocol.
minor comments (6)
  1. [Section 3.1.1] Inception Score (IS) is described as assessing 'visual quality,' but IS is primarily a measure of label diversity and does not evaluate perceptual quality; the choice and interpretation should be justified or replaced with a perceptual quality metric.
  2. [Section 2.2] The Prompt Fidelity rubric asks whether the video 'introduce[s] any extraneous elements that were not mentioned in the text query,' but a higher score is said to indicate a 'perfect match.' This apparent inversion should be clarified, because a video that introduces extraneous elements would intuitively receive a lower fidelity score.
  3. [Section 2.3, Figure 4] The T-SNE visualization is used to claim that SynTVA has 'significantly wider semantic coverage' than MSRVTT-0.8K, but no T-SNE hyperparameters, quantitative overlap measure, or statistical test are provided; the claim should be supported with a quantitative diversity metric.
  4. [Table 1] The 'Weight (%)' column is not clearly defined: it is unclear whether the reported numbers are normalized absolute regression coefficients, relative importances, or something else, and the table would benefit from a note explaining how the percentages were computed.
  5. [References and formatting] The ACM reference format contains the template placeholder year '2018' and a placeholder DOI; these should be corrected before submission.
  6. [Section 2.2] The paper states that five judges evaluate 'all three synthetic videos generated by the different models,' resulting in fifteen annotations per prompt, but does not describe how disagreements are resolved or whether each judge saw the videos in a randomized order; these details matter for annotation reliability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the utility claims rest on external MSRVTT 1K-A evaluation of human-ranked synthetic subsets, while the Auto-Evaluator is standard supervised regression on held-out human scores.

full rationale

The paper's derivation chain is not circular. SynTVA is constructed by generating videos from GPT-4o-derived prompts and collecting human alignment scores; the Auto-Evaluator is a linear regression trained on VBench metrics to predict those human scores and is evaluated on a held-out 20% split, which is normal supervised modeling rather than a prediction forced by construction. The central TVR utility results in Section 3.2 do not use Auto-Evaluator outputs: they train X-Pool on human-annotated High/Low/Random subsets and evaluate on the external MSRVTT 1K-A test set, so the positive correlation between human-judged alignment and downstream R@1 is an independent empirical finding, not a self-consistency check. The abstract's inference that VQA-based Auto-Evaluators 'enable' high-utility selection is an extrapolation from the Auto-Evaluator's held-out accuracy plus the measured alignment--utility correlation; it is not demonstrated by directly testing Auto-Evaluator-selected data, but that is a scope gap rather than a circular reduction. There are no load-bearing self-citations, no imported uniqueness theorems, and no known result renamed as new organization. The augmentation claim is somewhat overstated because no experiment mixes synthetic data with real MSRVTT training data, but this concerns experimental coverage, not circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The paper does not derive a mathematical law; it constructs a benchmark. The main fitted quantities are the Auto-Evaluator regression weights and the unspecified HQ/LQ thresholds. The central claims rely on human annotations being reliable, on standard VQA metrics being valid proxies, and on X-Pool representing TVR models.

free parameters (2)
  • Auto-Evaluator linear regression coefficients for Object & Scene and Actions = not disclosed (7-dimensional per dimension)
    Fitted on 80% of the 2,400 SynTVA pairs to predict human alignment scores from seven VQA metrics. These coefficients are the learned core of the Auto-Evaluator claim.
  • High/Low quality subset thresholds = not specified
    The paper creates high-quality and low-quality subsets based on human annotation scores for each alignment dimension, but the exact cutoffs are not reported. These thresholds determine all TVR comparisons in Figure 10.
assumptions (4)
  • domain assumption Human annotations on the four rubric dimensions are reliable ground truth for text-video semantic alignment.
    Section 2.2 uses five judges per pair with no inter-annotator agreement metric. These scores define HQ/LQ subsets and train the Auto-Evaluator.
  • domain assumption MSRVTT training split categories and captions represent realistic user search intents.
    Section 2.1 uses GPT-4o on MSRVTT categories and captions to generate 800 queries; the utility test is then measured on MSRVTT 1K-A.
  • domain assumption Standard VQA metrics (IS, CLIP-Score, SD Score, VBench dimensions) measure the constructs they are named after in synthetic videos.
    Section 3.1 builds correlations and the Auto-Evaluator on these metrics without validating them separately for synthetic video content.
  • domain assumption X-Pool with CLIP ViT-B/32 is representative of TVR models for measuring synthetic data utility.
    Section 3.2 tests only X-Pool; no other retrieval architectures or CLIP variants are used to confirm the utility conclusions.
invented entities (1)
  • None
    purpose: No new physical or hypothetical entities are introduced.
    The paper introduces a dataset and an evaluation framework, not new particles, forces, or conserved quantities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Are Synthetic Videos Useful? A Benchmark for Retrieval-Centric Evaluation of Synthetic Videos." pith.science (2026). https://pith.science/paper/HOIJNFAL

@misc{pith2026250702316,
  author       = {Pith},
  title        = {Pith review of: Are Synthetic Videos Useful? A Benchmark for Retrieval-Centric Evaluation of Synthetic Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HOIJNFAL}},
  note         = {Machine review of arXiv:2507.02316}
}
read the original abstract

Text-to-video (T2V) synthesis has advanced rapidly, yet current evaluation metrics primarily capture visual quality and temporal consistency, offering limited insight into how synthetic videos perform in downstream tasks such as text-to-video retrieval (TVR). In this work, we introduce SynTVA, a new dataset and benchmark designed to evaluate the utility of synthetic videos for building retrieval models. Based on 800 diverse user queries derived from MSRVTT training split, we generate synthetic videos using state-of-the-art T2V models and annotate each video-text pair along four key semantic alignment dimensions: Object \& Scene, Action, Attribute, and Prompt Fidelity. Our evaluation framework correlates general video quality assessment (VQA) metrics with these alignment scores, and examines their predictive power for downstream TVR performance. To explore pathways of scaling up, we further develop an Auto-Evaluator to estimate alignment quality from existing metrics. Beyond benchmarking, our results show that SynTVA is a valuable asset for dataset augmentation, enabling the selection of high-utility synthetic samples that measurably improve TVR outcomes. Project page and dataset can be found at https://jasoncodemaker.github.io/SynTVA/.

Figures

Figures reproduced from arXiv: 2507.02316 by the authors.

Figure 1
Figure 1. Overview of the proposed SynTVA dataset, designed [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the SynTVA pipeline and its application. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 5
Figure 5. Case Study of T2V Model Outputs. Comparison of Cosmos [1], Mochi [6], and Wan2.1 [33] video generations for the same text query. 1 2 3 4 5 Objects & Scene 1 2 3 4 5 Actions s = 0.768 1 2 3 4 5 Objects & Scene 1 2 3 4 5 Attributes s = 0.399 1 2 3 4 5 Actions 1 2 3 4 5 Attributes s = 0.371 1 2 3 4 5 Objects & Scene 1 2 3 4 5 Prompt Fidelity s = 0.309 1 2 3 4 5 Actions 1 2 3 4 5 Prompt Fidelity s = 0.233 1 2 3 4 5 Attr… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: T-SNE visualisation comparing the text embed [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 8
Figure 8. Figure 8: Correlations between VQA metrics and human [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 7
Figure 7. Figure 7: Comparison of the performance of T2V models [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 10
Figure 10. Figure 10: Impact of synthetic data quality and quantity on TVR R@1 performance across our alignment aspects. [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Quantifying and Narrowing the Unknown: Interactive Text-to-Video Retrieval via Uncertainty Minimization

    cs.CV 2025-07 conditional novelty 5.0 of 10

    By measuring text ambiguity, mapping uncertainty, and frame quality, UMIVR asks targeted clarifying questions and raises Recall@1 on MSR-VTT-1k to 69.2% after 10 interaction rounds.

Reference graph

Works this paper leans on

37 extracted references · 21 canonical work pages · cited by 1 Pith paper

  1. [1]

    Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. 2025. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575 (2025)

    arXiv 2025

  2. [2]

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

    work page 2024

  3. [3]

    Chaorui Deng, Qi Chen, Pengda Qin, Da Chen, and Qi Wu. 2023. Prompt switch: Efficient clip adaptation for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 15648–15658

    work page 2023

  4. [4]

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. 2017. CARLA: An open urban driving simulator. In Conference on robot learning. PMLR, 1–16

    2017

  5. [5]

    Qihang Ge, Wei Sun, Yu Zhang, Yunhao Li, Zhongpeng Ji, Fengyu Sun, Shangling Jui, Xiongkuo Min, and Guangtao Zhai. 2024. LMM-VQA: Advancing Video Quality Assessment with Large Multimodal Models. CoRR abs/2408.14008 (2024)

    arXiv 2024

  6. [6]

    genmo. [n. d.]. Mochi 1: A new SOTA in open-source video generation models. https://www.genmo.ai/blog

    work page

  7. [7]

    Satya Krishna Gorti, Noël Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. 2022. X-pool: Cross-modal language- video attention for text-video retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 5006–5015

    work page 2022

  8. [8]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. In NIPS. 6626–6637

    work page 2017

Show all 37 references
  1. [9]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems 33 (2020), 6840–6851

    2020

  2. [10]

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. 2022. Video diffusion models. Advances in Neu- ral Information Processing Systems 35 (2022), 8633–8646

    2022

  3. [11]

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

    2022 arXiv

  4. [12]

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. 2024. VBench: Comprehensive Benchmark Suite for Video Generative Mod...

    2024

  5. [13]

    Pengliang Ji, Chuyang Xiao, Huilin Tai, and Mingxiao Huo. 2024. T2VBench: Benchmarking Temporal Dynamics for Text-to-Video Generation. InCVPR Work- shops. IEEE, 5325–5335

    2024

  6. [14]

    Peng Jin, Hao Li, Zesen Cheng, Kehan Li, Xiangyang Ji, Chang Liu, Li Yuan, and Jie Chen. 2023. Diffusionret: Generative text-video retrieval with diffusion model. In Proceedings of the IEEE/CVF international conference on computer vision . 2470–2481

    2023

  7. [15]

    Xiaojie Jin, Bowen Zhang, Weibo Gong, Kai Xu, Xueqing Deng, Peng Wang, Zhao Zhang, Xiaohui Shen, and Jiashi Feng. 2024. Mv-adapter: Multimodal video transfer learning for video text retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2024

  8. [16]

    Pandeng Li, Chen-Wei Xie, Liming Zhao, Hongtao Xie, Jiannan Ge, Yun Zheng, Deli Zhao, and Yongdong Zhang. 2023. Progressive spatio-temporal prototype matching for text-video retrieval. In Proceedings of the IEEE/CVF international conference on computer vision . 4100–4110

    2023

  9. [17]

    Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. 2024. EvalCrafter: Benchmarking and Evaluating Large Video Generation Models. In CVPR. IEEE, 22139–22149

    2024

  10. [18]

    Yuqi Liu, Pengfei Xiong, Luhui Xu, Shengming Cao, and Qin Jin. 2022. Ts2- net: Token shift and selection transformer for text-video retrieval. In European conference on computer vision . Springer, 319–335

    2022

  11. [19]

    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. Neurocomputing 508 (2022), 293–304

    2022

  12. [20]

    OpenAI. [n. d.]. Hello GPT-4o. https://openai.com/index/hello-gpt-4o/

  13. [21]

    William Peebles and Saining Xie. 2023. Scalable diffusion models with transform- ers. In Proceedings of the IEEE/CVF international conference on computer vision . 4195–4205

    2023

  14. [22]

    Mathis Petrovich, Michael J Black, and Gül Varol. 2022. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision. Springer, 480–497

    2022

  15. [23]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)

    2023 arXiv

  16. [24]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  17. [25]

    Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. 2016. Play- ing for data: Ground truth from computer games. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14 . Springer, 102–118

    2016

  18. [26]

    Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen

    Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. 2016. Improved Techniques for Training GANs. InNIPS. 2226–2234

    2016

  19. [27]

    Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. 2022. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)

    2022 arXiv

  20. [28]

    Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. 2018. Towards Accurate Generative Models of Video: A New Metric & Challenges. CoRR abs/1812.01717 (2018)

    2018 arXiv

  21. [29]

    Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. 2019. FVD: A new Metric for Video Gener- ation. In DGS@ICLR. OpenReview.net

    2019

  22. [30]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research 9, 11 (2008)

    2008

  23. [31]

    Gul Varol, Javier Romero, Xavier Martin, Naureen Mahmood, Michael J Black, Ivan Laptev, and Cordelia Schmid. 2017. Learning from synthetic humans. In Proceedings of the IEEE conference on computer vision and pattern recognition . 109–117

    2017

  24. [32]

    Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. 2022. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399 (2022)

    2022 arXiv

  25. [33]

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. 2025. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314 (2025)

    2025 arXiv

  26. [34]

    Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, Tianxiang Jiang, Songze Li, Jilan Xu, Hongjie Zhang, Yifei Huang, Yu Qiao, Yali Wang, and Limin Wang. 2024. InternVideo2: Scaling Foundation Models for Multimodal ...

    2024

  27. [35]

    Jun Xu, Tao Mei, Ting Yao, and Yong Rui. 2016. MSR-VTT: A Large Video Description Dataset for Bridging Video and Language. In CVPR. IEEE Computer Society, 5288–5296

    2016

  28. [36]

    Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. 2022. Clip-vip: Adapting pre-trained image-text model to video- language representation alignment. arXiv preprint arXiv:2209.06430 (2022)

    2022 arXiv

  29. [37]

    Youngjae Yu, Jongseok Kim, and Gunhee Kim. 2018. A Joint Sequence Fusion Model for Video Question Answering and Retrieval. In ECCV (7) (Lecture Notes in Computer Science, Vol. 11211). Springer, 487–503

    2018

Pith tools

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