Pith. sign in

REVIEW 3 major objections 4 minor 38 references

Can Your Model Separate Yolks with a Water Bottle? Benchmarking Physical Commonsense Understanding in Video Generation Models

T0 review · 3 major / 4 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 No text-to-video model nails basic physics; best scores 36% on new benchmark

desk verdict A genuinely useful tool-use physical commonsense benchmark for T2V, but the QA-bank label inconsistency makes the headline accuracy numbers un-auditable until the released questions are checked. read the letter →

arxiv 2507.15824 v1 pith:QG2ZCA6C submitted 2025-07-21 cs.CV

Enes Sanli , Baris Sarper Tezcan , Aykut Erdem , Erkut Erdem This is my paper
classification cs.CV
keywords text-to-videogenerationphysicalcommonsensebenchmarktooluseandaffordancecaption-basedevaluationspatialreasoningtemporaldynamics
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 tries to establish that current text-to-video models, despite producing realistic-looking clips, lack robust physical commonsense. It introduces PhysVidBench, a benchmark of 383 everyday tool-use scenarios (derived from PIQA) with 7 physical reasoning dimensions, and evaluates 12 open and closed-source models through a caption-based QA pipeline. The best model, Cosmos-14B, answers only 36.1% of the physics questions correctly, and no model reaches even 40%. Spatial reasoning and temporal dynamics are the weakest dimensions across all models. If right, this means today's video generators cannot be trusted as world simulators for tasks requiring physical causality.

What carries the argument

The caption-based three-stage pipeline carries the argument: physics questions are generated from upsampled prompts, each video is captioned by AuroraCap with eight captions (one general, seven dimension-specific), and an LLM judge (Gemini-2.5-Flash) answers the yes/no questions from captions only. A response counts as correct if any caption supports 'yes'. This indirect setup avoids direct VLM hallucination, but it also is the instrument whose reliability decides the benchmark's meaning.

What would settle it

Run the same prompt set through a human annotation of the generated videos against the same questions; if human-video agreement is substantially higher than the caption-driven LLM scores on the same videos, the gap measures caption loss rather than model physics failure.

Watch

Extended reading notes

Core claim

PhysVidBench is the claim: a benchmark for physical commonsense in text-to-video generation, built from 383 PIQA-derived prompts that require tool use and object affordances. Evaluated on 12 models, the best accuracy on 4,123 yes/no physics questions was 36.1% (Cosmos-14B), followed by Wan2.1-14B at 33.9, Veo-2 at 34.8, Sora at 31.4. Spatial Reasoning and Temporal Dynamics are systematically hardest; Object Properties and Affordances is easiest, likely because models use surface visual heuristics. Prompt upsampling helps most large models but not Sora, which the paper hypothesizes already performs internal prompt upsampling. The paper concludes that scaling alone isn't enough: architecture and physically grounded training are needed.

Load-bearing premise

The whole score depends on the caption model capturing every physics-relevant detail of the video, and on the LLM judge answering correctly from those captions; if subtle physics is missed in captions, the benchmark underestimates model ability.

Editorial extensions

If this is right

  • If PhysVidBench scores reflect true model ability, no current open or closed text-to-video model has robust physical commonsense for tool-mediated actions.
  • Spatial reasoning (object fit, occlusion, geometry) and temporal dynamics (ordering, delays, sequencing) are the binding constraints; fixing them would move average scores more than improving realism.
  • Prompt upsampling with explicit physical detail extracts latent ability in large models, so prompt engineering can partially mask or reveal underlying capability.
  • The easy dimension, object properties and affordances, may reward visual heuristics rather than causal understanding, so high scores there cannot be read as physical reasoning.

Reading between the lines

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

  • The same caption-based evaluation could be turned around as a diagnostic: asking which physical facts survive captioning, and where caption loss correlates with 'No' answers, would expose whether the bottleneck is generation or measurement.
  • If these numbers generalize, robotics and embodied-AI efforts that treat video generators as world models will need to validate physical plausibility downstream—physically wrong rollouts are near-certain at below-40% averaged QA accuracy.
  • A testable extension: generate paired variants of the same scenario with only the physical constraint changed (e.g., friction, density, order of steps) and measure whether model accuracy tracks the perturbation; invariance to perturbations would confirm heuristic rather than causal generation.
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

3 major / 4 minor

Summary. PhysVidBench is proposed as a benchmark for physical commonsense in text-to-video (T2V) generation, built from 383 PIQA-derived prompts plus upsampled variants. The evaluation pipeline generates yes/no questions from the upsampled prompts (with the stated ground truth 'always yes'), captions each generated video with AuroraCap (one general plus seven dimension-specific captions), and asks an LLM to answer the questions from the captions, scoring a question correct if any caption supports 'Yes'. Accuracies are reported for 12 T2V models across seven dimensions, with the best model (Cosmos-14B) at 36.1% average accuracy; the paper concludes that T2V models lack robust physical commonsense, particularly in spatial and temporal reasoning. The benchmark and code are released, and a small human study is reported.

Significance. If the measurement were valid, the benchmark would be a useful complement to existing physical-commonsense video benchmarks, since it targets tool use and affordance, includes prompt upsampling, and provides an interpretable caption-based QA protocol. Strengths include the breadth of evaluated models, the release of the dataset and code, and a (limited) human-correlation check. However, the central quantitative claims are currently not interpretable because of an internal inconsistency in the QA ground-truth construction and because the scoring rule is not insulated from prompt-caption agreement.

major comments (3)
  1. [§3.3 (Stage 1), Appendix B.3, Appendix C.2] The main text states that 'the ground truth is always yes' and that no-answer questions were omitted, but the Stage-1 generation prompt in Appendix B.3 explicitly asks Gemini 2.5 Pro to 'Include a balanced mix of Yes and No answers' and to return 'A: Yes / No' for every question. No post-processing step that filters out No-answer questions is described. The worked examples in Appendix C.2 contain questions whose ground-truth answer is plainly No (e.g., 'Is the tea towel described as rectangular?' with answer 'No'; 'Is the goggle lens described as hard?' with answer 'No'). Under the scoring rule in §3.3 (Stage 3), which counts a question as correct only if the judge answers Yes for at least one caption, any such No-answer question is scored as wrong even when the judge is right. The reported accuracies in Table 3 therefore depend on an unverified label set and could change substantially (including rankings) after re-scoring with the true per-question labels. The released question labels must be audited and the evaluation re-run, or the absolute accuracy numbers should be withdrawn.
  2. [§3.3 (Stage 1 and Stage 3), Fig. 4, Appendix F] The evaluation loop is circular with respect to prompt language. Questions are generated from the upsampled prompt; the ground truth is always yes; and the judge is instructed to answer Yes if any of the eight captions supports the claim. Since the videos are generated from the same upsampled prompts, and the captions are produced by a general-purpose video captioner that tends to describe objects with typical properties, a correct answer can reflect lexical/paraphrase overlap between the caption and the prompt rather than the presence of the physical interaction in the video. The paper's own Limitations section (Appendix F) acknowledges that deriving questions from upsampled prompts 'could introduce a subtle evaluation bias,' but the magnitude is not quantified. I recommend adding negative controls—e.g., prompts describing physically impossible events, or a subset of questions whose ground truth is No—and showing that scores drop when the video content is physically altered. Without such controls, the absolute accuracies in Table 3 cannot be interpreted as measuring physical commonsense.
  3. [§3.3 (Stage 1), Appendix C.2] The majority of the example questions are of the form 'Is X described as Y?' (e.g., 'Is the toothbrush handle described as plastic?', 'Is the goggle lens described as hard?'). These questions test whether the caption matches the upsampled prompt's incidental descriptive details, not whether the depicted action obeys physical laws. A physically plausible video that renders the same action but changes an incidental material property (e.g., a metal instead of plastic handle) would be marked wrong, even though it does not violate physical commonsense. The claimed seven-dimensional physical ontology is therefore not cleanly operationalized by the question bank; the benchmark conflates prompt-fidelity with physical commonsense. The authors should separate prompt-fidelity questions from physics-plausibility questions, or explicitly re-weight the analysis.
minor comments (4)
  1. [Section 4, Table 3] The text says 'Cosmos-14B exhibits an average gain of over 8 points from prompt enrichment,' but Table 3 reports a change of -7.1 for Cosmos-14B, which corresponds to a gain of 7.1 points; this is also inconsistent with Section 5, which correctly says 'Cosmos-14B gains 7 points.'
  2. [Appendix C.1, Table 4] The difficulty subset is called 'extra hard' in the main text but 'Very Hard' in Appendix C.1 and Table 4; the thresholds (>40%, 20–40%, <20%) appear only in the appendix and should be stated in the main text.
  3. [Appendix A] There is a typo: 'Exmple Prompt 1' should be 'Example Prompt 1.'
  4. [Appendix D] The human evaluation uses only 30 prompts and 15 participants; the reported Pearson correlations (0.45–0.69) are moderate, and the paper should report confidence intervals and statistical significance rather than only the point estimates.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PhysVidBench is an externally grounded benchmark; caption-omission and question-label audit issues are validity concerns, not circular reductions.

full rationale

PhysVidBench's central measurement is an external benchmark score, not a derived prediction. The 383 prompts come from PIQA's correct solutions (Section 3.2), the videos are generated by independent T2V models, and the QA bank is fixed before scoring; no parameter is fitted to the models being ranked. The scoring rule ('count as correct if the LLM answers Yes for at least one caption', Section 3.3, Figure 4) is a noisy proxy for physical commonsense, but it is not equal by construction to the conclusion that models lack physical commonsense—a model that faithfully renders the prompt's physics could in principle score high, and Section D reports a human-correlation check (r = 0.45-0.69), providing independent evidence. The acknowledged caption-omission failure (Section F) is a validity threat that could underestimate scores, not a circular reduction. The apparent inconsistency between Section 3.3 ('ground truth is always yes') and the Appendix B.3 generation instruction ('Include a balanced mix of Yes and No answers') affects the auditability of the released labels and would be a correctness/scoring issue if unresolved, but it does not make the reported accuracy identical to an input by construction. There are no load-bearing self-citations or imported uniqueness theorems: PIQA, PhyT2V, AuroraCap, and the model references are external to the authors' prior work. Therefore no circular step meets the evidentiary bar.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The benchmark's load rests on three measurement assumptions: captions faithfully describe videos, prompt-derived questions have fixed yes answers, and the LLM judge is faithful. The difficulty thresholds are the only fitted quantities.

free parameters (1)
  • Difficulty subset thresholds = 40% and 20% average top-4 model accuracy
    The medium/hard/very-hard boundaries are set from the accuracy of the four best open-source models on the same questions being reported, making the difficulty labels dependent on the models under test.
assumptions (4)
  • domain assumption AuroraCap captions preserve all physics-relevant visual details of each generated video.
    Invoked throughout Section 3.3 Stage 2 and Section E; the paper's own Limitations (Section F) concedes captions may omit critical details, which would penalize correct videos.
  • domain assumption A yes/no question derived from the upsampled prompt has ground truth yes, and answering it via captions measures video physical correctness.
    Stated in Section 3.3 Stage 1: 'the ground truth is always yes.' This assumes the prompt describes the only physically correct outcome.
  • domain assumption Gemini-2.5-Flash answers based only on the provided captions and does not leak prompt or prior knowledge.
    Assumed in Section 3.3 Stage 3; the authors note LLM hallucination as a limitation in Section 6.
  • domain assumption PIQA correct solutions transferred into video prompts remain physically valid and visually demonstrable.
    Assumed in Section 3.2 Stage 2; the reliance on Gemini filtering could introduce selection bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Your Model Separate Yolks with a Water Bottle? Benchmarking Physical Commonsense Understanding in Video Generation Models." pith.science (2026). https://pith.science/paper/QG2ZCA6C

@misc{pith2026250715824,
  author       = {Pith},
  title        = {Pith review of: Can Your Model Separate Yolks with a Water Bottle? Benchmarking Physical Commonsense Understanding in Video Generation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QG2ZCA6C}},
  note         = {Machine review of arXiv:2507.15824}
}
read the original abstract

Recent progress in text-to-video (T2V) generation has enabled the synthesis of visually compelling and temporally coherent videos from natural language. However, these models often fall short in basic physical commonsense, producing outputs that violate intuitive expectations around causality, object behavior, and tool use. Addressing this gap, we present PhysVidBench, a benchmark designed to evaluate the physical reasoning capabilities of T2V systems. The benchmark includes 383 carefully curated prompts, emphasizing tool use, material properties, and procedural interactions, and domains where physical plausibility is crucial. For each prompt, we generate videos using diverse state-of-the-art models and adopt a three-stage evaluation pipeline: (1) formulate grounded physics questions from the prompt, (2) caption the generated video with a vision-language model, and (3) task a language model to answer several physics-involved questions using only the caption. This indirect strategy circumvents common hallucination issues in direct video-based evaluation. By highlighting affordances and tool-mediated actions, areas overlooked in current T2V evaluations, PhysVidBench provides a structured, interpretable framework for assessing physical commonsense in generative video models.

Figures

Figures reproduced from arXiv: 2507.15824 by the authors.

Figure 1
Figure 1. Understanding Physical Commonsense in Video Generation Models. Humans intuitively understand how physical interactions unfold—how objects respond to forces, how tools function, and how materials behave under manipulation. To assess whether text-to-video (T2V) models share this understanding, we introduce PhysVidBench, a benchmark designed to evaluate core dimensions of physical commonsense. The figure shows two exam… view at source ↗
Figure 2
Figure 2. Physical commonsense dimensions tested in PhysVidBench, each illustrated with a video generated by Cosmos-14B. Prompts are designed to probe specific categories such as force and motion, object affordance, spatial containment, temporal progression, and material interaction. Each row shows sampled frames from one generated video paired with its corresponding prompt. 3 Benchmark 3.1 Tested Physical Commonsense Concept… view at source ↗
Figure 3
Figure 3. Video Generation Pipeline. An overview of the four-stage pipeline used to construct PhysVidBench. We begin with correct solutions from the PIQA dataset and filter them using a large language model (LLM) to identify instances of secondary tool use and object affordances (Stage 1). These filtered solutions are converted into short, physically grounded video generation prompts via another LLM pass (Stage 2). Each base … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Evaluation Pipeline. Overview of our three-stage evaluation framework designed to assess physical commonsense in generated videos. Using the upsampled prompts, we generate yes/no questions aligned with one or more of the seven physical commonsense dimensions in our ont…
Figure 5
Figure 5. Figure 5: Comparison of performances of video generation models on PhysVidBench. Each violin plot corresponds to a distinct model and shows the probability distribution of accuracy scores across all evaluated prompts. involving materials (MT), object properties (OP), and force d…
Figure 6
Figure 6. Figure 6: User Evaluation Interface. Our web-based interface for the human study mimicked the auto-eval QA structure. Participants watched a video, then answered five Yes/No questions drawn from the same QA bank used in automatic evaluation. The design enforced randomization acr…
Figure 7
Figure 7. Figure 7: Pearson Correlation between Human Evaluations and Auto-Evaluations. Plots com￾paring auto-eval scores with human judgment scores across 30 prompts for four open-source top￾performing models. Each subplot shows a linear fit and corresponding Pearson correlation coeffici…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 20 canonical work pages

  1. [1]

    Agarwal, A

    N. Agarwal, A. Ali, M. Bala, Y . Balaji, E. Barker, T. Cai, P. Chattopadhyay, Y . Chen, Y . Cui, Y . Ding, et al. Cosmos world foundation model platform for physical AI. arXiv preprint arXiv:2501.03575, 2025

    arXiv 2025

  2. [2]

    Z. Bai, H. Ci, and M. Z. Shou. Impossible videos. arXiv preprint arXiv:2503.14378, 2025

    arXiv 2025

  3. [3]

    Bansal, Z

    H. Bansal, Z. Lin, T. Xie, Z. Zong, M. Yarom, Y . Bitton, C. Jiang, Y . Sun, K.-W. Chang, and A. Grover. VideoPhy: Evaluating physical commonsense for video generation. arXiv preprint arXiv:2406.03520, 2024

    arXiv 2024

  4. [4]

    Bansal, C

    H. Bansal, C. Peng, Y . Bitton, R. Goldenberg, A. Grover, and K.-W. Chang. VideoPhy-2: A challenging action-centric physical commonsense evaluation in video generation.arXiv preprint arXiv:2503.06800, 2025

    arXiv 2025

  5. [5]

    Y . Bisk, R. Zellers, R. L. Bras, J. Gao, and Y . Choi. Piqa: Reasoning about physical common- sense in natural language, 2019. URL https://arxiv.org/abs/1911.11641

    arXiv 2019

  6. [6]

    W. Chai, E. Song, Y . Du, C. Meng, V . Madhavan, O. Bar-Tal, J.-N. Hwang, S. Xie, and C. D. Manning. AuroraCap: Efficient, performant video detailed captioning and a new benchmark. In Proceedings of the International Conference on Learning Representations (ICLR), 2025

    work page 2025

  7. [7]

    H. Chen, Y . Zhang, X. Cun, M. Xia, X. Wang, C. Weng, and Y . Shan. VideoCrafter2: Overcom- ing data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310–7320, 2024

    work page 2024

  8. [8]

    W. Chow, J. Mao, B. Li, D. Seita, V . Guizilini, and Y . Wang. PhysBench: Benchmarking and enhancing vision-language models for physical world understanding. In Proceedings of the International Conference on Learning Representations (ICLR), 2025

    work page 2025

Show all 38 references
  1. [9]

    Gemini 2.5 Pro

    Google DeepMind. Gemini 2.5 Pro. https://deepmind.google/technologies/gemini/ pro/, 2025. Accessed May 15, 2025

    2025

  2. [10]

    Gemini 2.5 Flash

    Google DeepMind. Gemini 2.5 Flash. https://deepmind.google/technologies/ gemini/flash/, 2025. Accessed May 15, 2025

    2025

  3. [11]

    Google DeepMind. Veo-2. https://deepmind.google/technologies/veo/veo-2/,

  4. [12]

    HaCohen, N

    Y . HaCohen, N. Chiprut, B. Brazowski, D. Shalem, D. Moshe, E. Richardson, E. Levin, G. Shiran, N. Zabari, O. Gordon, P. Panet, S. Weissbuch, V . Kulikov, Y . Bitterman, Z. Melumian, and O. Bibi. LTX-Video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024

    2024 arXiv

  5. [13]

    H. He, Y . Zhang, L. Lin, Z. Xu, and L. Pan. Pre-trained video generative models as world simulators. arXiv preprint arXiv:2502.07825, 2025

    2025 arXiv

  6. [14]

    Hessel, A

    J. Hessel, A. Holtzman, M. Forbes, R. Le Bras, and Y . Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514–7528, 2021

    2021

  7. [15]

    A. Hu, L. Russell, H. Yeo, Z. Murez, G. Fedoseev, A. Kendall, J. Shotton, and G. Corrado. GAIA-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023

    2023 arXiv

  8. [16]

    Y . Hu, L. Wang, X. Liu, L.-H. Chen, Y . Guo, Y . Shi, C. Liu, A. Rao, Z. Wang, and H. Xiong. Simulating the real world: A unified survey of multimodal generative models. arXiv preprint arXiv:2503.04641, 2025

    2025

  9. [17]

    Huang, H

    W. Huang, H. Liu, M. Guo, and N. Z. Gong. Visual hallucinations of multi-modal large language models. In Findings of the Association for Computational Linguistics (ACL), 2024

    2024

  10. [18]

    Huang, Y

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

    2024

  11. [19]

    W. Kong, Q. Tian, Z. Zhang, R. Min, Z. Dai, J. Zhou, J. Xiong, X. Li, B. Wu, J. Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024

    2024 arXiv

  12. [20]

    Kuaishou Team. Kling. https://klingai.kuaishou.com/, 2024. Accessed May 12, 2025

    2024

  13. [21]

    Y . Liu, L. Li, S. Ren, R. Gao, S. Li, S. Chen, X. Sun, and L. Hou. FETV: A benchmark for fine- grained evaluation of open-domain text-to-video generation. Advances in Neural Information Processing Systems, 36:62352–62387, 2023

    2023

  14. [22]

    Y . Liu, X. Cun, X. Liu, X. Wang, Y . Zhang, H. Chen, Y . Liu, T. Zeng, R. Chan, and Y . Shan. EvalCrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139–22149, 2024

    2024

  15. [23]

    Y . Liu, K. Zhang, Y . Li, Z. Yan, C. Gao, R. Chen, Z. Yuan, Y . Huang, H. Sun, J. Gao, L. He, and L. Sun. Sora: A review on background, technology, limitations, and opportunities of large vision models, 2024. URL https://arxiv.org/abs/2402.17177

    2024 arXiv

  16. [24]

    F. Meng, J. Liao, X. Tan, W. Shao, Q. Lu, K. Zhang, Y . Cheng, D. Li, Y . Qiao, and P. Luo. To- wards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv preprint arXiv:2410.05363, 2024

    2024 arXiv

  17. [25]

    Motamed, L

    S. Motamed, L. Culp, K. Swersky, P. Jaini, and R. Geirhos. Do generative video models learn physical principles from watching videos? arXiv preprint arXiv:2501.09038, 2025

    2025 arXiv

  18. [26]

    Sora: Generating videos from text, 2025

    OpenAI. Sora: Generating videos from text, 2025. URL https://openai.com/sora. Ac- cessed May 14, 2025

    2025

  19. [27]

    Gen-3 Alpha

    Runway Team. Gen-3 Alpha. https://runwayml.com/research/ introducing-gen-3-alpha , 2024. Accessed May 10, 2025

    2024

  20. [28]

    Salimans, I

    T. Salimans, I. Goodfellow, W. Zaremba, V . Cheung, A. Radford, and X. Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016

    2016

  21. [29]

    Magi-1: Autoregressive video generation at scale, 2025

    Sand-AI. Magi-1: Autoregressive video generation at scale, 2025. URL https://static. magi.world/static/files/MAGI_1.pdf. 11

    2025

  22. [30]

    Unterthiner, S

    T. Unterthiner, S. Van Steenkiste, K. Kurach, R. Marinier, M. Michalski, and S. Gelly. To- wards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018

    2018 arXiv

  23. [31]

    A. Wang, B. Ai, B. Wen, C. Mao, C.-W. Xie, D. Chen, F. Yu, H. Zhao, J. Yang, J. Zeng, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    2025 arXiv

  24. [32]

    X. Wang, Z. Zhu, G. Huang, B. Wang, X. Chen, and J. Lu. Worlddreamer: Towards general world models for video generation via predicting masked tokens. arXiv preprint arXiv:2401.09985, 2024

    2024 arXiv

  25. [33]

    Q. Xue, X. Yin, B. Yang, and W. Gao. PhyT2V: LLM-guided iterative self-refinement for physics-grounded text-to-video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  26. [34]

    Z. Yang, J. Teng, W. Zheng, M. Ding, S. Huang, J. Xu, Y . Yang, W. Hong, X. Zhang, G. Feng, D. Yin, Y . Zhang, W. Wang, Y . Cheng, B. Xu, X. Gu, Y . Dong, and J. Tang. CogVideoX: Text-to-video diffusion models with an expert transformer. In Proceedings of the International Con...

    2025

  27. [35]

    L. Yu, Y . Cheng, K. Sohn, J. Lezama, H. Zhang, H. Chang, A. G. Hauptmann, M.-H. Yang, Y . Hao, I. Essa, and L. Jiang. Magvit: Masked generative video transformer, 2023. URL https://arxiv.org/abs/2212.05199

    2023 arXiv

  28. [36]

    Zheng, Z

    D. Zheng, Z. Huang, H. Liu, K. Zou, Y . He, F. Zhang, Y . Zhang, J. He, W.-S. Zheng, Y . Qiao, et al. VBench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness. arXiv preprint arXiv:2503.21755, 2025

    2025 arXiv

  29. [37]

    are” rather than what agents can “do

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. Xing, et al. Judging LLM-as-a-judge with MT-bench and Chatbot Arena. Advances in Neural Information Processing Systems, 36:46595–46623, 2023. 12 Supplementary Material In this supplementa...

    2023

  30. [2025]

    Accessed May 11, 2025. 10

    2025

Pith tools

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