Pith. sign in

REVIEW 5 major objections 6 minor 2 cited by

DAVID-XR1: Detecting AI-Generated Videos with Explainable Reasoning

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that fine-tuning a video-language model on hand-labeled defects, with chain-of-thought reasoning supervised by a verdict head, turns it into an explainable AI-video detector that generalizes to unseen generators.

desk verdict Useful dataset and a plausible OOD detection gain, but the explainability claim is undercut by a circular evaluation and an unresolved precision discrepancy. read the letter →

arxiv 2506.14827 v1 pith:Z3YAKG7U submitted 2025-06-13 cs.CV cs.AI

Yifeng Gao , Yifan Ding , Hongyu Su , Juncheng Li , Yunhan Zhao , Lin Luo , Zixing Chen , Li Wang
show 4 more authors
Xin Wang Yixu Wang Xingjun Ma Yu-Gang Jiang
This is my paper
classification cs.CVcs.AI
keywords AI-generatedvideodetectionexplainablereasoningchain-of-thoughtdistillationvideo-languagemodeldefectannotationspatio-temporallocalizationout-of-domaingeneralizationforensics
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 AI-generated video detection should be and can be an explainable diagnostic process rather than a black-box real-versus-fake call. To that end it builds DAVID-X, a corpus of 416 AI-generated clips and 331 real clips in which every visible generation defect is hand-labeled with a defect category, a frame range, a 2D point, and a written rationale. It then trains DAVID-XR1 by distilling five-stage chain-of-thought traces from a commercial reasoning model and fine-tuning an open-source vision-language backbone with an extra binary classifier on the hidden state just before the verdict. The paper reports that this lifts detection accuracy on out-of-domain generators from 26.7% to 76.7% and yields 54.7% explanation precision, positioning the model near commercial systems while keeping every decision auditable.

What carries the argument

The mechanism that carries the argument is CoT-aware supervised fine-tuning. The authors distill chain-of-thought traces from a strong commercial vision-language model using DAVID-X's ground-truth annotations, reformatting detection into five stages: defect discovery, spatio-temporal localization, defect categorization, defect description, and final verdict. They then fine-tune an open-source 7B vision-language backbone on the full <think>...<evidence>...<answer> sequence while adding a lightweight binary classifier that reads the hidden state immediately before the <answer> token and predicts real versus AI-generated. The joint objective $\mathcal{L}=\alpha\mathcal{L}_{\text{LM}}+\beta\mathcal{L}_{\text{CLS}}$ ties the reasoning representation to the verdict, so the explanation is trained to be evidence for the decision rather than mere text.

What would settle it

Have an independent annotation team relabel a random sample of DAVID-X and blindly score DAVID-XR1's explanations; if inter-annotator agreement on defect categories and locations is low, or if blind precision falls well below the reported 54.7%, the explanation-quality claim is an artifact of labeler subjectivity rather than objective defect detection.

Watch

Extended reading notes

Core claim

The central claim is that a general-purpose vision-language model, fine-tuned on a relatively compact set of human-style defect annotations, can detect AI-generated videos from generators it never saw in training and say why. The evidence is the jump from 26.7% to 76.7% accuracy on a 90-video test set that includes Kling 2.0, Pika v2.2, Veo2, MAGI-1, and FramePack outputs, together with an explanation precision of 54.7% measured by manual evaluation. In the model's output, each verdict is backed by a structured chain: reasoning inside <think>, evidence with defect category, timestamp, located frame, 2D point, and explanation inside <evidence>, and the final 'AI generated video' or 'Real video' inside <answer>. This makes the detection transparent and verifiable rather than opaque.

Load-bearing premise

The whole method assumes that the hand-written defect labels, frame ranges, single points, and explanations in DAVID-X are objective and consistent enough to teach a model generalizable evidence, and that scoring the model with the same human labelers who made those labels is a trustworthy measure of explanation quality.

Editorial extensions

If this is right

  • A compact corpus of 747 videos with defect-level annotations is enough to move a 7B vision-language backbone from 26.7% to 76.7% accuracy on generators excluded from training.
  • Structured output with defect category, timestamp, point, and explanation makes the verdict auditable by human moderators, which is the stated requirement of social media and video platforms.
  • Jointly training the language generation and a binary classifier on the pre-answer hidden state outperforms language-only fine-tuning on explanation precision across the loss-weight settings in the ablation.
  • Detection transfers across generation paradigms, including next-generation diffusion models, an autoregressive model, and outputs from the FramePack project.
  • Explanation diversity is bounded by training cues: limiting each sample to 1-3 defects keeps the model focused but lowers its recall of possible evidence compared with GPT-4.1.

Reading between the lines

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

  • Editorial extension: because the paper's evaluation is scored by the same annotators who wrote the labels, an independent blind relabeling of DAVID-X and rescoring of DAVID-XR1's explanations would separate model capability from labeler subjectivity.
  • Editorial extension: the same recipe of small curated defect annotations, chain-of-thought distillation, and an auxiliary verdict head is not tied to video-specific features and could plausibly transfer to explainable detection of AI images or manipulated audio.
  • Editorial extension: the dataset's single-point annotations are designed to feed a point-to-mask segmentation model, so an agentic pipeline that lets the detector segment and display the flawed region is a natural deployment the paper does not evaluate.
  • Editorial extension: the paper reports that reinforcement learning added nothing because the fixed output format suppressed new reasoning paths; relaxing the format could be the next lever for explanation diversity, which the paper identifies as a remaining weakness.
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 DAVID-X, a dataset of 416 AI-generated clips from 15 generators and a set of real videos, annotated with defect categories, temporal and spatial localization, and written rationales. On top of this dataset, the authors build DAVID-XR1 by distilling Chain-of-Thought traces from Gemini 2.5 Pro, constraining them to match DAVID-X annotations, and fine-tuning Qwen2.5-VL-7B with an auxiliary binary classification head on the hidden state preceding the answer token. The paper claims that this CoT-aware SFT raises out-of-domain detection accuracy from 26.7% to 76.7% on a 90-video test set spanning unseen generators (Kling 2.0, Pika v2.2, Veo2, MAGI-1, FramePack) and yields explanation precision of 54.7% or 62.5%, bringing the model's reasoning on par with GPT-4.1 and Gemini 2.5 Pro.

Significance. If the claims hold, the work is a useful step toward explainable AI-generated video detection: it contributes a fine-grained annotation dataset, an OOD test set with genuinely unseen generators, and an ablation showing that joint language-modeling and classification training helps over language-only fine-tuning. The paper also makes a concrete, falsifiable prediction that a compact dataset plus CoT distillation can turn a general-purpose VLMM into a competitive explainable detector. However, the current evidence base for the explanation-precision claim is thin, and the detection claim rests on a small test set without uncertainty quantification.

major comments (5)
  1. [§4.2 / Table 3] The explanation-precision number is reported inconsistently: Section 4.2 states that DAVID-XR1 achieves 62.5% explanation precision, Table 3 reports 54.67%, and the abstract and introduction report 54.7%. These cannot all describe the same evaluation, and the paper gives no reconciliation. The exact protocol and the correct number must be stated before the explainability claim can be assessed.
  2. [§4.1 / §3.1.3 / §3.2] The explanation evaluation is not independent of the training labels. Section 4.1 says 'All manual annotations are performed by our DAVID-X labeler', and Section 3.1.3 indicates this same annotation effort generated the ground-truth defect categories, frame ranges, points, and explanations used for training. Section 3.2 then requires the distilled CoT to match that ground truth exactly. Scoring precision with the same labeler therefore measures agreement with the annotator's own subjective judgments, not the objective validity of the cues; no inter-annotator agreement is reported. The 'on par with GPT-4.1 and Gemini 2.5 Pro' claim needs a blind, independent evaluation, or at minimum IAA statistics together with a held-out set of explanation labels not used in training.
  3. [§4.1 / Table 3] The headline detection improvement from 26.7% to 76.7% is computed on a 90-video test set with 15 videos per condition. With n=15, a single video changes category accuracy by 6.7 points, yet no confidence intervals, bootstrap estimates, or per-video results are reported. Additionally, DAVID-XR1's real-video recall drops from 86.7% (backbone) to 73.3%, a 13.3-point loss that is not discussed; this false-positive trade-off matters for deployment. Please provide uncertainty estimates and a false-positive analysis.
  4. [§4.1 / §4.2] The diversity metric is defined inconsistently. It is first described as 'the fraction of ground-truth evidence items (from all correct model responses) that the tested model retrieves', but the denominator is then said to be 'the total number of correct evidence items identified by all evaluated models'. This is a union of model outputs, not the DAVID-X ground-truth evidence, so the metric is relative to the model set and can reward verbosity. As a result, the comparison with GPT-4.1's higher diversity in Section 4.2 is hard to interpret without also reporting recall against the DAVID-X evidence annotations.
  5. [§4.3] The ablation study varies only the loss weights α/β and the number of fine-tuning iterations. It does not isolate the contributions of the DAVID-X annotations versus the Gemini 2.5 Pro CoT distillation, nor the classification head. For example, there is no condition that trains on DAVID-X with plain captions or no rationales, and no condition that uses Gemini CoT without DAVID-X constraint. The paper's attribution of the 50-point gain to 'fine-tuning on our compact dataset and enhanced with chain-of-thought distillation' is therefore not fully supported by the reported experiments.
minor comments (6)
  1. [§3.1.2 / Table 1] The real-video count is inconsistent: Section 3.1.2 says the collection 'yielded a total of 331 real videos', while Table 1 and the abstract report 337. Also, the stated corpus size of 747 does not match 416 + 337 = 753. Please reconcile these numbers.
  2. [§3.1.2 / References] The 190 InternVid segments are cited as reference [1], but reference [1] is listed as SafeVid-350K. Please correct the citation to the proper InternVid source.
  3. [Table 3] Some cells in Table 3 are typeset without spacing (e.g., '53.366.710080.0'), which makes them difficult to read. Please format all entries clearly.
  4. [§4.3 / Table 3] The final hyperparameters α and β used to produce Table 3 are not stated. Section 4.3 says a 1:10 ratio yields interpretability gains, but it does not identify the configuration used for the main results. Please report the exact values.
  5. [§4.1] The unit of an explanation 'cue' is not defined. It is unclear whether a cue corresponds to a defect category, a defect instance, a temporal span, or a spatial point, which directly affects how precision and diversity are counted.
  6. [Throughout] There are several typographical and copy-editing issues: 'fruther' (Section 3.2), 'explaination' (Figure 5 caption), inconsistent 'DA VID-XR1' spacing, inconsistent 'GPT-o1'/'GPT-o3' naming, and reference [37] listed as 'Piks' rather than 'Pika'. Please proofread the manuscript.

Circularity Check

1 steps flagged · score 5.0 of 10

The explanation-precision claim is scored by the same labeler who produced the training annotations, so it measures fidelity to that labeler's subjective standard rather than independent validity; the OOD detection claim remains externally defined.

  1. self definitional [Sec. 4.1 Evaluation Metrics; Sec. 3.2 Visual-CoT Distillation]
    "All manual annotations are performed by our DAVID-X labeler, as current models lack sufficient defect-understanding capabilities. ... All evidence tags must match the ground truth exactly, except for <explanation>, which may be expanded upon by the model."

    The ground truth for explanation validity on the test set is the same DAVID-X labeler who created the defect categories, frame ranges, points, and explanations used as CoT-SFT training targets. The language-model loss forces the student to reproduce those labels, and the precision metric then asks the same labeler whether the model's cues 'appear in the video and are a valid indicator of AI generation.' A high precision score therefore measures agreement with the labeler's own annotation standard, not an independent property of AI-generated video. No inter-annotator agreement is reported, so the 'explanation precision' claim (54.7% in Table 3, 62.5% in Sec. 4.2) cannot distinguish model skill from mimicry of the labeler's subjective viewing style.

full rationale

The detection half of the central claim is not circular: the 90-video OOD test set has fixed binary labels (known generators vs. real recordings), and DAVID-XR1's 76.7% accuracy is computed against those labels with no fitting on the test set. The backbone improvement from 26.7% to 76.7% is an externally falsifiable result. The explanation half is circular in its evaluation: DAVID-X's training annotations are hand-produced by the DAVID-X labeler, the CoT distillation prompt forces Gemini to reproduce these tags exactly, and the explanation-precision metric is then scored by the same DAVID-X labeler, checking 'whether each cue the model provides actually appears in the video and is a valid indicator of AI generation.' Because 'valid indicator' is operationalized as the labeler's own judgment, the precision number measures self-consistency with the training-label source rather than independent correctness. This is compounded by no reported inter-annotator agreement and by an internal numeric conflict (54.67% in Table 3 vs. 62.5% in Sec. 4.2), making the explanation claim unreliable as reported. No self-citation chain or imported uniqueness theorem is involved; the score reflects partial circularity of the explainability metric while the OOD detection claim remains independent.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim depends on a small hand-annotated dataset, a teacher-model distillation pipeline, and an internal evaluation protocol. The free parameters are the SFT loss weights and the video-retrieval threshold; the axioms are the assumptions that visible defects are decisive evidence, that the anchor-type framework is valid, that Gemini 2.5 Pro's traces are faithful, and that the authors' labeler provides a reliable yardstick. No new physical or theoretical entities are introduced.

free parameters (2)
  • Loss weights alpha and beta = alpha:beta = 1:10, with beta=0 as language-only control
    Selected from ablation in Section 4.3; the reported interpretability gains depend on this ratio and it is not validated on a separate hold-out.
  • VideoCLIP cosine similarity threshold for real video selection = 0.22
    Hand-set retention threshold for semantic matching of real videos in Section 3.1.2; it changes the distribution of real videos and hence the difficulty of the task.
assumptions (4)
  • domain assumption AI-generated videos can be reliably identified from human-recognizable spatiotemporal defects.
    The dataset and model only label visible flaws (Section 3.1.3); if future generators produce clean outputs, the method's premise fails.
  • domain assumption The anchor-type dichotomy (natural recorded vs handcrafted) is a valid basis for adjusting artifact tolerance.
    The annotation pipeline assigns each synthetic clip an anchor and applies looser standards for handcrafted footage; the model inherits this assumption in Section 3.1.3.
  • domain assumption Gemini 2.5 Pro's distilled CoT traces faithfully reflect the ground-truth defect evidence.
    CoT distillation instructs Gemini to derive, not invent, the given cues, but there is no measurement of trace fidelity in Section 3.2.
  • domain assumption Manual evaluation by the DAVID-X labeler is a valid measure of explanation precision.
    The same team that annotated training data judges whether model cues are valid; no external auditors or inter-rater agreement are reported in Section 4.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DAVID-XR1: Detecting AI-Generated Videos with Explainable Reasoning." pith.science (2026). https://pith.science/paper/Z3YAKG7U

@misc{pith2026250614827,
  author       = {Pith},
  title        = {Pith review of: DAVID-XR1: Detecting AI-Generated Videos with Explainable Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z3YAKG7U}},
  note         = {Machine review of arXiv:2506.14827}
}
read the original abstract

As AI-generated video becomes increasingly pervasive across media platforms, the ability to reliably distinguish synthetic content from authentic footage has become both urgent and essential. Existing approaches have primarily treated this challenge as a binary classification task, offering limited insight into where or why a model identifies a video as AI-generated. However, the core challenge extends beyond simply detecting subtle artifacts; it requires providing fine-grained, persuasive evidence that can convince auditors and end-users alike. To address this critical gap, we introduce DAVID-X, the first dataset to pair AI-generated videos with detailed defect-level, temporal-spatial annotations and written rationales. Leveraging these rich annotations, we present DAVID-XR1, a video-language model designed to deliver an interpretable chain of visual reasoning-including defect categorization, temporal-spatial localization, and natural language explanations. This approach fundamentally transforms AI-generated video detection from an opaque black-box decision into a transparent and verifiable diagnostic process. We demonstrate that a general-purpose backbone, fine-tuned on our compact dataset and enhanced with chain-of-thought distillation, achieves strong generalization across a variety of generators and generation modes. Our results highlight the promise of explainable detection methods for trustworthy identification of AI-generated video content.

Figures

Figures reproduced from arXiv: 2506.14827 by the authors.

Figure 1
Figure 1. An illustration of DAVID-XR1’s explain￾able reasoning capability. Inspired by the visual reasoning principle of “extract evidence first, then reason step-by-step” [18, 42, 48, 50], we begin by distilling Chain￾of-Thought (CoT) traces from the proprietary Gemini 2.5 Pro, restructuring video forensics into five explicit stages: defect discovery, spa￾tial–temporal localization, defect categorization, defect description… view at source ↗
Figure 2
Figure 2. Overview of the data construction and annotation workflow for our [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Statistics of DAVID-X dataset. For real videos, we first manually-curated 147 clips from major video platforms using the same prompts as for the generated videos. To further increase diversity, we include 190 segments from the scene-diverse subset of InternVid [1]. These videos were non-overlappingly cropped into 5s, 10s, 20s, and 30s chunks, and only those with a VideoCLIP [51] embedding cosine similarity of ≥ 0.22… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Illustration of CoT-aware SFT. CoT-Aware SFT We apply supervised fine-tuning (SFT) to adapt a pretrained VLMM to our task: given a video, the model generates a coherent reasoning chain within <think></think>, summarizes the evidence in <evidence></evidence>, and conclu…
Figure 5
Figure 5. Figure 5: The detection and explaination performances of DAVId-XR1 under different hyperparame [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Example detection of an AI-generated video (by Veo2) using our [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Keywords of top-30 prompt clusters We cluster the real-world generation prompts from VidProM [47] into 80 groups and retain the top 30, which together account for over 89 % of all queries—thereby capturing the predominant user preferences. We then apply TF-IDF to each …
Figure 8
Figure 8. Figure 8: DAVID-X Annotation Tool D DAVID-XR1 Demo We present a demonstration of DAVID-XR1 in real-world deployment with Chainlit and include the demo video in the supplementary materials. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: DAVID-XR1 Demo 17 [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Detecting AI-Generated Video: A Vision-Language Dual-View Survey

    cs.CV 2026-07 conditional novelty 6.0 of 10

    AIGC-V detection should be treated as factual fidelity verification and organized by a four-layer vision-language dual-view taxonomy spanning cues, motion, cross-modal consistency, and world-level reasoning.

  2. SafeGuard: A Multi-Agent Perception-Reasoning Framework for Social-Risk AI-Generated Video Detection

    cs.CV 2026-07 conditional novelty 6.0 of 10

    SafeGuard bridges low-level forensic perception and high-level semantic reasoning via multi-agent collaboration, lifting accuracy +18.7% on a new social-risk AI-video benchmark.

Reference graph

Works this paper leans on

59 extracted references · 17 canonical work pages · cited by 2 Pith papers

  1. [1]

    Safevid-350k.https://huggingface.co/datasets/yxwang/SafeVid-350K, 2025. 9

    work page 2025

  2. [2]

    Gradio: Hassle-free sharing and testing of ml models in the wild.arXiv preprint arXiv:1906.02569, 2019

    Abubakar Abid, Ali Abdalla, Ali Abid, Dawood Khan, Abdulrahman Alfozan, and James Zou. Gradio: Hassle-free sharing and testing of ml models in the wild.arXiv preprint arXiv:1906.02569, 2019

    arXiv 1906

  3. [3]

    Jimeng.https://jimeng.jianying.com/, 2024

    Jimeng AI. Jimeng.https://jimeng.jianying.com/, 2024

    work page 2024

  4. [4]

    Kling 1.5.https://www.klingai.com/, 2024

    KLING AI. Kling 1.5.https://www.klingai.com/, 2024

    work page 2024

  5. [5]

    Bream machine.https://lumalabs.ai/dream-machine, 2024

    Luma AI. Bream machine.https://lumalabs.ai/dream-machine, 2024

    work page 2024

  6. [6]

    Vidu 1.5.https://www.vidu.com/, 2024

    Vidu AI. Vidu 1.5.https://www.vidu.com/, 2024

    work page 2024

  7. [7]

    Ai-generated video detection via spatial-temporal anomaly learning

    Jianfa Bai, Man Lin, Gang Cao, and Zijie Lou. Ai-generated video detection via spatial-temporal anomaly learning. InChinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 460–470. Springer, 2024

    2024

  8. [8]

    Qwen2.5-vl technical report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report. a...

    arXiv 2025

Show all 59 references
  1. [9]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Do- minik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    2023 arXiv

  2. [10]

    Align your latents: High-resolution video synthesis with latent diffusion models

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pa...

    2023

  3. [11]

    Video generation models as world simulators

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. URL https://openai.com/research/ video-generation-models...

    2024

  4. [12]

    K-means clustering.Virginia Tech, Advanced Research Computing, Interdisci- plinary Center for Applied Mathematics, 2009

    John Burkardt. K-means clustering.Virginia Tech, Advanced Research Computing, Interdisci- plinary Center for Applied Mathematics, 2009

    2009

  5. [13]

    Videocrafter1: Open diffusion models for high-quality video generation.arXiv preprint arXiv:2310.19512, 2023

    Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation.arXiv preprint arXiv:2310.19512, 2023

    arXiv

  6. [14]

    Videocrafter2: Overcoming data limitations for high-quality video diffusion models

    Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming 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

    2024

  7. [15]

    Demamba: Ai-generated video detection on million-scale genvideo benchmark, 2024

    Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, and Huaxiong Li. Demamba: Ai-generated video detection on million-scale genvideo benchmark, 2024. URL https://arxiv.org/abs/2405. 19707

    2024

  8. [16]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. InCVPR, 2024

    2024

  9. [17]

    Video-of-thought: Step-by-step video reasoning from perception to cognition.arXiv preprint arXiv:2501.03230, 2024

    Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition.arXiv preprint arXiv:2501.03230, 2024

    2024 arXiv

  10. [18]

    Videoespresso: A large-scale chain-of-thought dataset for fine-grained video reasoning via core frame selection.arXiv preprint arXiv:2411.14794, 2024

    Songhao Han, Wei Huang, Hairong Shi, Le Zhuo, Xiu Su, Shifeng Zhang, Xu Zhou, Xiaojuan Qi, Yue Liao, and Si Liu. Videoespresso: A large-scale chain-of-thought dataset for fine-grained video reasoning via core frame selection.arXiv preprint arXiv:2411.14794, 2024

    2024 arXiv

  11. [19]

    Exposing ai-generated videos: A benchmark dataset and a local-and-global temporal defect based detection method.arXiv preprint arXiv:2405.04133, 2024

    Peisong He, Leyao Zhu, Jiaxing Li, Shiqi Wang, and Haoliang Li. Exposing ai-generated videos: A benchmark dataset and a local-and-global temporal defect based detection method.arXiv preprint arXiv:2405.04133, 2024

    2024 arXiv

  12. [20]

    Latent video diffusion models for high-fidelity long video generation.arXiv preprint arXiv:2211.13221, 2022

    Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation.arXiv preprint arXiv:2211.13221, 2022. 10

    2022 arXiv

  13. [21]

    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

    2022 arXiv

  14. [22]

    Vbench: Comprehensive benchmark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. InProceedings of the IEEE/CVF Conference on Computer Vision and Patt...

    2024

  15. [23]

    Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    2024 arXiv

  16. [24]

    Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    2024 arXiv

  17. [25]

    Text-to-Video, 2025

    Kling AI. Text-to-Video, 2025. URL https://app.klingai.com/global/ text-to-video/new. Accessed: 2025-05-12

    2025

  18. [26]

    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

    2024 arXiv

  19. [27]

    Towards a universal synthetic video detector: From face or background manipulations to fully ai-generated content.arXiv preprint arXiv:2412.12278, 2024

    Rohit Kundu, Hao Xiong, Vishal Mohanty, Athula Balachandran, and Amit K Roy-Chowdhury. Towards a universal synthetic video detector: From face or background manipulations to fully ai-generated content.arXiv preprint arXiv:2412.12278, 2024

    2024 arXiv

  20. [28]

    Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326, 2024

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326, 2024

    2024 arXiv

  21. [29]

    Open-sora plan: Open-source large video generation model.arXiv preprint arXiv:2412.00131, 2024

    Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model.arXiv preprint arXiv:2412.00131, 2024

    2024 arXiv

  22. [30]

    Turns out i’m not real: Towards robust detection of ai-generated videos.arXiv preprint arXiv:2406.09601, 2024

    Qingyuan Liu, Pengyuan Shi, Yun-Yun Tsai, Chengzhi Mao, and Junfeng Yang. Turns out i’m not real: Towards robust detection of ai-generated videos.arXiv preprint arXiv:2406.09601, 2024

    2024 arXiv

  23. [31]

    Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation

    Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu 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

  24. [32]

    Detect- ing ai-generated video via frame consistency, 2025

    Long Ma, Zhiyuan Yan, Qinglang Guo, Yong Liao, Haiyang Yu, and Pengyuan Zhou. Detect- ing ai-generated video via frame consistency, 2025. URL https://arxiv.org/abs/2402. 02085

    2025

  25. [33]

    Umap: Uniform manifold approximation and projection for dimension reduction.arXiv preprint arXiv:1802.03426, 2018

    Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction.arXiv preprint arXiv:1802.03426, 2018

    2018 arXiv

  26. [34]

    Compositional chain-of- thought prompting for large multimodal models

    Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of- thought prompting for large multimodal models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420–14431, 2024

    2024

  27. [35]

    Sora.https://openai.com/sora/, 2024

    OpenAI. Sora.https://openai.com/sora/, 2024

    2024

  28. [36]

    Gpt-o3.https://openai.com/index/introducing-o3-and-o4-mini/, 2025

    OpenAI. Gpt-o3.https://openai.com/index/introducing-o3-and-o4-mini/, 2025

    2025

  29. [37]

    Piks.https://pika.art/about, 2024

    Pika. Piks.https://pika.art/about, 2024

    2024

  30. [38]

    Pixverse v3.https://app.pixverse.ai/onboard, 2024

    PixVerse. Pixverse v3.https://app.pixverse.ai/onboard, 2024

    2024

  31. [39]

    Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    2024 arXiv

  32. [40]

    Gen-3 alpha

    Runway. Gen-3 alpha. https://runwayml.com/research/introducing-gen-3-alpha , 2024

    2024

  33. [41]

    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

  34. [42]

    Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Advancing multi-modal language models with a comprehen- sive dataset and benchmark for chain-of-thought reasoning.Advances in Neural Information Processing Systems,...

    2024

  35. [43]

    Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023

    2023 arXiv

  36. [44]

    Beyond deepfake images: Detecting ai-generated videos

    Danial Samadi Vahdati, Tai D Nguyen, Aref Azizpour, and Matthew C Stamm. Beyond deepfake images: Detecting ai-generated videos. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4397–4408, 2024

    2024

  37. [45]

    Veo-Team, :, Agrim Gupta, Ali Razavi, Andeep Toor, Ankush Gupta, Dumitru Erhan, Eleni Shaw, Eric Lau, Frank Belletti, Gabe Barth-Maron, Gregory Shaw, Hakan Erdogan, Hakim Sidahmed, Henna Nandwani, Hernan Moraldo, Hyunjik Kim, Irina Blok, Jeff Donahue, José Lezama, Kory Mathews...

    2024

  38. [46]

    Modelscope text-to-video technical report.arXiv preprint arXiv:2308.06571, 2023

    Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report.arXiv preprint arXiv:2308.06571, 2023

    2023 arXiv

  39. [47]

    Vidprom: A million-scale real prompt-gallery dataset for text-to- video diffusion models.arXiv preprint arXiv:2403.06098, 2024

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

    2024 arXiv

  40. [48]

    Videocot: A video chain-of-thought dataset with active annotation tool.arXiv preprint arXiv:2407.05355, 2024

    Yan Wang, Yawen Zeng, Jingsheng Zheng, Xiaofen Xing, Jin Xu, and Xiangmin Xu. Videocot: A video chain-of-thought dataset with active annotation tool.arXiv preprint arXiv:2407.05355, 2024

    2024 arXiv

  41. [49]

    Internvideo2: Scaling foundation models for multimodal video understanding

    Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Internvideo2: Scaling foundation models for multimodal video understanding. InECCV, 2024

    2024

  42. [50]

    V?: Guided visual search as a core mechanism in multimodal llms

    Penghao Wu and Saining Xie. V?: Guided visual search as a core mechanism in multimodal llms. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13084–13094, 2024

    2024

  43. [51]

    Videoclip: Contrastive pre-training for zero- shot video-text understanding.arXiv preprint arXiv:2109.14084, 2021

    Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero- shot video-text understanding.arXiv preprint arXiv:2109.14084, 2021

    2021 arXiv

  44. [52]

    Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    2024 arXiv

  45. [53]

    Minicpm-v: A gpt-4v level mllm on your phone.arXiv preprint arXiv:2408.01800, 2024

    Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone.arXiv preprint arXiv:2408.01800, 2024

    2024 arXiv

  46. [54]

    Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model? arXiv preprint arXiv:2504.13837, 2025

    Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang, Shiji Song, and Gao Huang. Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model? arXiv preprint arXiv:2504.13837, 2025

    2025 arXiv

  47. [55]

    Packing input frame contexts in next-frame prediction models for video generation.Arxiv, 2025

    Lvmin Zhang and Maneesh Agrawala. Packing input frame contexts in next-frame prediction models for video generation.Arxiv, 2025

    2025

  48. [56]

    Llava-next: A strong zero-shot video understanding model

    Y Zhang, B Li, H Liu, Y Lee, L Gui, D Fu, J Feng, Z Liu, and C Li. Llava-next: A strong zero-shot video understanding model. 2024

    2024

  49. [57]

    Mul- timodal chain-of-thought reasoning in language models.arXiv preprint arXiv:2302.00923, 2023

    Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Mul- timodal chain-of-thought reasoning in language models.arXiv preprint arXiv:2302.00923, 2023

    2023 arXiv

  50. [58]

    Llamafactory: Unified efficient fine-tuning of 100+ language models

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of 100+ language models. InACL, 2024. 12

    2024

  51. [59]

    heat haze

    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. arXiv preprint arXiv:2412.20404, 2024. 13 A Keywords of Top-30 Prompt Clusters Figure 7: Keywords...

    2024 arXiv

Pith tools

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