AEGIS: Authenticity Evaluation Benchmark for AI-Generated Video Sequences

Jieyu Li; Joey Tianyi Zhou; Xin Zhang

REVIEW 3 major objections 6 minor 1 cited by

AEGIS: Authenticity Evaluation Benchmark for AI-Generated Video Sequences

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper builds AEGIS, a 10,000-video authenticity benchmark, and argues that state-of-the-art vision-language models still fail on its hardest AI-generated split even after fine-tuning.

desk verdict Useful benchmark artifact whose headline 'models struggle' is not yet proven because the evaluation lacks ROC-AUC and relies on threshold-dependent accuracy. read the letter →

arxiv 2508.10771 v1 pith:J6KQSQ6J submitted 2025-08-14 cs.CV cs.AI

Jieyu Li , Xin Zhang , Joey Tianyi Zhou This is my paper

classification cs.CVcs.AI

keywords AEGISAI-generatedvideodetectionauthenticitybenchmarkvision-languagemodelsgeneralizationgapSoraKLingmultimodalannotations

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 aims to establish that state-of-the-art vision-language models—systems that answer questions about video frames—cannot yet reliably tell hyper-realistic AI-generated video apart from real footage, and that the reason is a generalization gap invisible in older benchmarks. To show this, the authors assemble AEGIS, a benchmark of over 10,000 curated real and synthetic videos, and construct a Hard Test Set whose synthetic clips come from Sora and KLing under GPT-4o-refined prompts, alongside deliberately complex real footage. On this split, Qwen2.5-VL labels synthetic clips as real about three times out of four, structured reasoning prompts make things worse rather than better, and LoRA fine-tuning on the training split lifts macro-F1 only from 0.52 to 0.55. The paper treats these results as evidence that the Hard Test Set captures forgery realism beyond current model generalization, making it a useful stress instrument for future detectors.

What carries the argument

The load-bearing object is AEGIS's three-stage construction pipeline and the Hard Test Set it produces. The pipeline applies three filters: Reality removes non-photorealistic or heavily edited content; Diversity balances scenes, durations, and resolutions; Difficulty uses Qwen2.5-VL's zero-shot confidence to discard synthetic clips from the TIP-I2V pool that it can flag without effort. The Hard Test Set then combines only high-fidelity synthetic outputs from proprietary generators with independently curated real footage, so a model's score on it is meant to measure detection under distribution shift rather than on easy artifacts. Supporting this design, each video carries multimodal annotati

What would settle it

Run two checks. First, train an independent detector on the synthetic clips that the Difficulty filter discarded from the TIP-I2V pool and on the clips it retained; if the retained clips are no harder for the new detector, the definition of 'hard' is specific to Qwen2.5-VL rather than to forgery detectability. Second, train a detector on Sora and KLing outputs drawn from a different set of prompts and evaluate it on the AEGIS Hard Test Set; high accuracy there would show the gap is a distribution mismatch caused by the benchmark's prompt style rather than a general limitation of current models

Watch

Extended reading notes

Core claim

The central claim, stated as the authors would state it, is that AEGIS is a valid instrument for exposing a real generalization gap. Its Hard Test Set pairs 218 synthetic videos generated by Sora and KLing from GPT-4o-refined prompts with 218 real videos selected for visual complexity. Zero-shot, Qwen2.5-VL 3B and 7B achieve synthetic-video accuracies of 0.23 and 0.22 while real-video accuracies reach 0.80 and 0.89; Video-LLaVA collapses to predicting 'AI' for everything, scoring 0.0 real accuracy. The paper's strongest demonstration is the fine-tuning contrast: low-rank adaptation (LoRA) on the AEGIS training split lifts Qwen2.5-VL 7B's in-domain macro-F1 from 0.43 to 0.82, but the same mod

Load-bearing premise

The load-bearing premise is that the Hard Test Set's difficulty is a property of the forgeries themselves, rather than an artifact of the specific generators, prompts, and filter models used to build the set; if another model or generator family makes the same clips easy, the claimed generalization gap collapses.

Editorial extensions

If this is right

Deployed zero-shot video authenticity checks based on current vision-language models will miss roughly three of every four Sora- or KLing-class synthetic clips, a miss rate large enough to matter for misinformation monitoring.
Structured reasoning prompts—asking a model to examine frequency artifacts, lighting consistency, compression noise, and physical plausibility—do not rescue detection and can push synthetic accuracy lower, from 0.22 to 0.16 for Qwen2.5-VL 7B.
Fine-tuning on an in-domain benchmark set is not sufficient for generalization: large in-domain gains fail to transfer to the Hard Test Set, so detector development should be evaluated on held-out generators and prompt distributions.
AEGIS's annotations and split design give later work a common stress test for measuring whether new approaches are actually improving robustness rather than overfitting one generator family.

Reading between the lines

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

Editorial: The sharp in-domain-to-hard-set gap suggests the Hard Test Set may be tracking domain shift—different generators, prompt styles, and scene distributions—as much as intrinsic forgery difficulty. Training on AEGIS plus Sora/KLing data from unrelated prompts would separate the two.
Editorial: The benchmark's own results leave the multimodal annotations unused; none of the benchmarked models sees optical flow or Fourier features. Feeding those cues to the same backbones is a natural test of whether the hard-set gap is a signal problem or a model-capacity problem.
Editorial: Because the real-video side draws heavily on YouTube, TikTok, and DVF footage, a detector could exploit platform-specific codec or compression fingerprints rather than authenticity. Evaluating on independently captured footage from other sources would clarify whether 'real' is being detected or 'web video' is being detected.
Editorial: If an independent detector confirms the hard-set difficulty, the practical implication is that automated video authenticity screening should assume high miss rates on commercial generators and pair detector output with human review or provenance signals.

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.

Desk Editor's Note

read the letter

Hi — quick take on AEGIS.

What's new and good: they built a real benchmark artifact, not just a paper. AEGIS curates 5,199 synthetic videos across open and proprietary generators, adds a hard test set with Sora/KLing outputs, and ships rich multimodal annotations (semantic, flow, frequency). The dataset is on HuggingFace. That's a concrete contribution for video authenticity detection, and the hard test set does look harder than TIP-I2V alone.

The weak spot is in the evaluation. The only metrics are per-class accuracy and macro-F1. The Qwen2.5-VL zero-shot row on the hard test set is Acc_real=0.89, Acc_ai=0.22 — that pattern is exactly what you'd see if the model's decision threshold is poorly calibrated, not necessarily if the videos are hard. Without ROC-AUC or calibration curves, the headline claim 'current models struggle on AEGIS' is not proven. Video-LLaVA has the opposite extreme (Acc_real=0.0, Acc_ai=1.0). Those two rows scream threshold artifact, not evidence of a generalization gap.

One important correction to the reader's take: the Qwen2.5-VL difficulty filter in Sec 3.2 applies to the TIP-I2V training/validation pools, not to the KLing/Sora hard test set — those were selected by manual realism review and GPT-4o-refined prompts. So the direct circularity concern is smaller than it first appears. That said, the filter is still used to build the training set from which the same model family is fine-tuned, and the hard test set is small (436 videos, 218 synthetic), so any headline number has wide error bars. The LoRA gain on the hard set (macro-F1 0.52 to 0.55) is within noise.

The authors also only evaluate two VLM families. No classical detectors, no AUC, no confidence intervals, no released prompts or eval code. For a benchmark to be adopted, those need to be there.

Bottom line: the artifact deserves attention; the evidence as presented does not. If this is under review, it should be conditional — require ROC-AUC/calibration, more baselines, exact prompts and code. It's a legitimate candidate for peer review, not a desk reject.

Referee Report

3 major / 6 minor

Summary. This paper introduces AEGIS, a benchmark for detecting AI-generated videos, containing 5,199 synthetic and 5,271 real videos, including a Hard Test Set of 436 videos built from Sora, KLing, DVF, and YouTube sources. The authors describe a filtering pipeline (reality, difficulty, diversity), provide multimodal annotations (semantic descriptions, optical flow, frequency features), and evaluate Qwen2.5-VL (3B/7B) and Video-LLaVA under zero-shot, structured-reasoning, and LoRA fine-tuning. The headline finding is that zero-shot and fine-tuned VLMs perform near chance on the Hard Test Set (e.g., Acc_ai≈0.22 and macro-F1≈0.52–0.55), which the paper interprets as a generalization gap.

Significance. If the Hard Test Set is valid, AEGIS would be a valuable resource: it includes recent proprietary generators, realistic real-video sources, and multimodal annotations that go beyond existing benchmarks. The dataset is released, and the construction pipeline is described in enough detail to be reproduced. However, the current paper does not establish the central claim that current models 'struggle' in a model-independent sense, because the evidence is threshold-dependent and lacks AUC/calibration and uncertainty quantification.

major comments (3)

[Section 4.3, Tables 2a/2b/3] The claim that current VLMs have limited detection capability on the Hard Test Set rests on Acc_ai and macro-F1. With class-balanced data, the observed patterns (Qwen: Acc_real=0.89, Acc_ai=0.22; Video-LLaVA: Acc_real=0.0, Acc_ai=1.0) are exactly what one expects from two miscalibrated decision thresholds, not necessarily from absent discriminative signal. Please report ROC-AUC, balanced accuracy, and calibration curves, and compute confidence intervals or bootstrap estimates given n=436. Without these, the reader cannot distinguish 'models cannot detect Sora/KLing' from 'models have poor default thresholds'.
[Section 3.2 (2)] The Difficulty filter uses Qwen2.5-VL confidence to discard synthetic videos that are 'confidently predicted as AI-generated,' but no confidence threshold is given. Because the same model family is later fine-tuned and evaluated, the in-domain results and the difficulty of the filtered pool are partly defined by one model's failures. The Hard Test Set itself is not Qwen-filtered (Sora/KLing videos are manually selected), so this is not fully circular, but the threshold must be specified for reproducibility and the claimed 'hardness' should be validated with a model-independent difficulty measure (e.g., human realism ratings or agreement across several detectors).
[Section 3.3 / Table 1] The Hard Test Set comprises only 218 synthetic videos (107 Sora, 111 KLing) and 218 real videos. The claimed generalization gap is based on small differences (e.g., macro-F1 0.52 to 0.55) that could be within sampling noise; no per-source breakdown is provided. Report results separately for Sora and KLing and provide uncertainty intervals.

minor comments (6)

[Section 3.2 (2)] The retention rule is ambiguous: confident 'True' predictions are discarded and 'not AI-generated' with lower confidence are retained, but what happens to uncertain 'True' predictions? Define the threshold and the exact rule.
[Section 4.2] Typo in metric definitions: 'ccuracy on synthetic videos' should be 'accuracy'.
[Figure 2 / Section 3.3] It is not clear from Figure 2 which filtering operations apply to the Hard Test Set vs. training/validation; state this explicitly in Section 3.3.
[Table 2] Add the number of videos (n=436) to the Table 2 caption, since the small sample size is relevant to the reported accuracies.
[Reference [19]] The Kling URL is malformed: 'https://https://klingai.kuaishou.com//'.
[Section 3.4 (3)] Define RIO (Radial Integral Operations) at first use.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Hard Test Set is independently constructed and the headline claim is an empirical measurement, not an identity with the construction inputs.

full rationale

The paper's derivation chain is benchmark construction followed by evaluation: Section 3.1 collects real and synthetic videos; Section 3.2 filters them by reality, difficulty, and diversity; Section 3.3 creates the training, validation, and Hard Test Set; Section 4 benchmarks VLMs on the in-domain and Hard Test Sets. The headline claim — that Qwen2.5-VL and Video-LLaVA achieve low synthetic accuracy on the Hard Test Set — is not equivalent to any fitted input. The Hard Test Set synthetic videos come from KLing and Sora, selected via GPT-4o-refined prompts and manual realism review (Sections 3.1.2 and 3.3); they are NOT the TIP-I2V videos filtered by Qwen2.5-VL confidence in Section 3.2. Thus the low Acc_ai on the Hard Test Set is an empirical result on an independently curated test set, not a construction tautology. The Qwen2.5-VL-based Difficulty filter applies only to the TIP-I2V training/validation synthetic pools; the paper does not use that filtered pool as the headline Hard Test Set, so no self-definitional reduction occurs. The LoRA experiments train on the AEGIS training split and evaluate on a held-out Hard Test Set; no fitted parameter is renamed as a prediction, and no external benchmark is used to define the target result. All references are to external prior work with no load-bearing self-citation chain. The skeptic's threshold/calibration concern (Acc_real=0.89, Acc_ai=0.22 may reflect a default decision threshold rather than absent discriminative signal) is a legitimate validity/correctness caveat about the chosen metrics, but it is not a circularity: benchmark hardness is not defined by the same metric that reports it. Overall, the central claim is self-contained and independently testable.

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

The central benchmark results rest on an untested selection assumption (Qwen confidence as difficulty) and on labels inherited from upstream datasets. There are no new physical entities. The benchmark is an assembled resource rather than a fitted model, but the curation choices function as hidden parameters that shape the reported difficulty.

free parameters (2)

k in k-means keyframe selection = 8
Section 3.4(1): CLIP embeddings are clustered with k=8 to pick representative keyframes for GPT-4V description generation. The choice is stated without justification or sensitivity analysis.
Qwen2.5-VL confidence threshold for 'confidently predicted as AI-generated' = not specified
Section 3.2(2): synthetic videos with high zero-shot AI-generated confidence are discarded; the exact threshold is not given, making the filter unreproducible and effectively a hidden parameter.

assumptions (4)

domain assumption Vript videos labeled 'real' and TIP-I2V videos labeled 'synthetic' have correct ground-truth labels.
Section 3.1 uses upstream dataset labels without independent verification of every clip; mislabeled clips would inject label noise into benchmark results.
ad hoc to paper Qwen2.5-VL confidence is a valid proxy for detection difficulty, so filtering out its confident predictions yields a genuinely harder set for all models.
Section 3.2(2) uses this proxy to build the Hard Test Set; this is the main source of circularity in the benchmark's evaluation.
domain assumption Optical flow (RAFT) and frequency-domain (FFT-RIO) features capture authenticity-relevant cues rather than content differences.
Section 3.4(2,3) provides these features as annotations; no analysis shows they distinguish real from generated video independently of content.
domain assumption GPT-4V can produce valid semantic and authenticity descriptions from keyframes.
Section 3.4(1) relies on GPT-4V for all real-video descriptions and authenticity reasoning; these are not validated against human judgments.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AEGIS: Authenticity Evaluation Benchmark for AI-Generated Video Sequences." pith.science (2026). https://pith.science/paper/J6KQSQ6J

@misc{pith2026250810771,
  author       = {Pith},
  title        = {Pith review of: AEGIS: Authenticity Evaluation Benchmark for AI-Generated Video Sequences},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J6KQSQ6J}},
  note         = {Machine review of arXiv:2508.10771}
}

read the original abstract

Recent advances in AI-generated content have fueled the rise of highly realistic synthetic videos, posing severe risks to societal trust and digital integrity. Existing benchmarks for video authenticity detection typically suffer from limited realism, insufficient scale, and inadequate complexity, failing to effectively evaluate modern vision-language models against sophisticated forgeries. To address this critical gap, we introduce AEGIS, a novel large-scale benchmark explicitly targeting the detection of hyper-realistic and semantically nuanced AI-generated videos. AEGIS comprises over 10,000 rigorously curated real and synthetic videos generated by diverse, state-of-the-art generative models, including Stable Video Diffusion, CogVideoX-5B, KLing, and Sora, encompassing open-source and proprietary architectures. In particular, AEGIS features specially constructed challenging subsets enhanced with robustness evaluation. Furthermore, we provide multimodal annotations spanning Semantic-Authenticity Descriptions, Motion Features, and Low-level Visual Features, facilitating authenticity detection and supporting downstream tasks such as multimodal fusion and forgery localization. Extensive experiments using advanced vision-language models demonstrate limited detection capabilities on the most challenging subsets of AEGIS, highlighting the dataset's unique complexity and realism beyond the current generalization capabilities of existing models. In essence, AEGIS establishes an indispensable evaluation benchmark, fundamentally advancing research toward developing genuinely robust, reliable, broadly generalizable video authenticity detection methodologies capable of addressing real-world forgery threats. Our dataset is available on https://huggingface.co/datasets/Clarifiedfish/AEGIS.

Figures

Figures reproduced from arXiv: 2508.10771 by the authors.

**Figure 1.** Overview of the AEGIS dataset. AEGIS comprises a diverse collection of synthetic videos covering a broad range of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** The AEGIS Dataset Construction Pipeline. Step 1: Data Collection – Collecting real and synthetic videos from [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

RobustSora: De-Watermarked Benchmark for Robust AI-Generated Video Detection
cs.CV 2025-12 conditional novelty 8.0 of 10

RobustSora benchmark demonstrates that current AI video detectors rely heavily on visible watermarks, with average accuracy drops of 6.6 percentage points when watermarks are erased and increased false alarms when wat...

Reference graph

Works this paper leans on

53 extracted references · 28 canonical work pages · cited by 1 Pith paper

[1]

Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al . 2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

arXiv 2025
[2]

Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, Kai-Wei Chang, and Aditya Grover. 2024. Videophy: Evaluating physical commonsense for video generation.arXiv preprint arXiv:2406.03520 (2024)

arXiv 2024
[3]

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

arXiv 2023
[4]

Nuria Alina Chandra, Ryan Murtfeldt, Lin Qiu, Arnab Karmakar, Hannah Lee, Emmanuel Tanumihardja, Kevin Farhat, Ben Caffee, Sejin Paik, Changyeon Lee, et al. 2025. Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024. arXiv preprint arXiv:2503.02857 (2025)

arXiv 2025
[5]

Chirui Chang, Zhengzhe Liu, Xiaoyang Lyu, and Xiaojuan Qi. 2024. What Matters in Detecting AI-Generated Videos like Sora? arXiv preprint arXiv:2406.19568 (2024)

arXiv 2024
[6]

Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, et al. 2024. Demamba: Ai- generated video detection on million-scale genvideo benchmark. arXiv preprint arXiv:2405.19707 (2024)

arXiv 2024
[7]

Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah
[8]

Fanda Fan, Chunjie Luo, Wanling Gao, and Jianfeng Zhan. 2023. Aigcbench: Com- prehensive evaluation of image-to-video content generated by ai. BenchCouncil Transactions on Benchmarks, Standards and Evaluations (2023)

work page 2023

Show all 53 references

[9]

Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky TQ Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. 2024. Discrete flow matching. In NeurIPS

2024
[10]

Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets

2014
[11]

Xinan He, Yue Zhou, Bing Fan, Bin Li, Guopu Zhu, and Feng Ding. 2025. VL- Forgery Face Triad: Detection, Localization and Attribution via Multimodal Large Language Models. arXiv preprint arXiv:2503.06142 (2025)

2025 arXiv
[12]

Vlad Hondru, Eduard Hogea, Darian Onchis, and Radu Tudor Ionescu. 2025. ExDDV: A New Dataset for Explainable Deepfake Detection in Video. arXiv preprint arXiv:2503.14421 (2025)

2025
[13]

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

2022
[14]

Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. 2024. Vbench: Comprehensive benchmark suite for video generative models. In CVPR

2024
[15]

Yoori Hwang, Ji Youn Ryu, and Se-Hoon Jeong. 2021. Effects of disinformation us- ing deepfake: The protective effect of media literacy education. Cyberpsychology, Behavior, and Social Networking (2021)

2021
[16]

Xinyi Jin, Zhuoyue Zhang, Bowen Gao, Shuqing Gao, Wenbo Zhou, Nenghai Yu, and Guoyan Wang. 2025. Assessing the perceived credibility of deepfakes: The impact of system-generated cues and video characteristics. new media & society (2025)

2025
[17]

Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. 2020. Analyzing and improving the image quality of stylegan. In CVPR

2020
[18]

Leon Klein, Andreas Krämer, and Frank Noé. 2023. Equivariant flow matching. In NeurIPS

2023
[19]

Kuaishou. 2024. Kling: AI Video Generation Model. https://https://klingai. kuaishou.com//

2024
[20]

Chunyi Li, Tengchuan Kou, Yixuan Gao, Yuqin Cao, Wei Sun, Zicheng Zhang, Yingjie Zhou, Zhichao Zhang, Weixia Zhang, Haoning Wu, et al. 2024. Aigiqa-20k: A large database for ai-generated image quality assessment. In CVPR

2024
[21]

Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. 2023. Agiqa-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology (2023)

2023
[22]

Xiaoming Li, Xinyu Hou, and Chen Change Loy. 2024. When stylegan meets stable diffusion: a w+ adapter for personalized image generation. In CVPR

2024
[23]

Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. 2023. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 (2023)

2023 arXiv
[24]

Li Lin, Neeraj Gupta, Yue Zhang, Hainan Ren, Chun-Hao Liu, Feng Ding, Xin Wang, Xin Li, Luisa Verdoliva, and Shu Hu. 2024. Detecting multimedia generated by large ai models: A survey. arXiv preprint arXiv:2402.00045 (2024)

2024 arXiv
[25]

Xiao Liu, Xinhao Xiang, Zizhong Li, Yongheng Wang, Zhuoheng Li, Zhuosheng Liu, Weidi Zhang, Weiqi Ye, and Jiawei Zhang. 2024. A survey of ai-generated MM ’25, October 27–31, 2025, Dublin, Ireland Jieyu Li, Xin Zhang, and Joey Tianyi Zhou video evaluation. arXiv preprint arXiv:...

2024
[26]

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

2024
[27]

Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al . 2024. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177 (2024)

2024 arXiv
[28]

Yibo Miao, Yifan Zhu, Lijia Yu, Jun Zhu, Xiao-Shan Gao, and Yinpeng Dong
[29]

Zhenliang Ni, Qiangyu Yan, Mouxiao Huang, Tianning Yuan, Yehui Tang, Hailin Hu, Xinghao Chen, and Yunhe Wang. 2025. GenVidBench: A Challenging Bench- mark for Detecting AI-Generated Video. arXiv preprint arXiv:2501.11340 (2025)

2025
[30]

OpenAI. 2023. GPT-4V. https://openai.com/index/gpt-4v-system-card/

2023
[31]

OpenAI. 2024. GPT-4o. https://platform.openai.com/docs/models/gpt-4o

2024
[32]

OpenAI. 2024. Sora: AI Video Generation Model. https://openai.com/sora

2024
[33]

Pika Labs. 2024. Pika: AI Video Generation Platform. https://www.pika.art/

2024
[34]

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 learning

2021
[35]

Xiufeng Song, Xiao Guo, Jiache Zhang, Qirui Li, Lei Bai, Xiaoming Liu, Guangtao Zhai, and Xiaohong Liu. 2024. On learning multi-modal forgery representation for diffusion generated video detection. arXiv preprint arXiv:2410.23623 (2024)

2024 arXiv
[36]

Zhen Sun, Ziyi Zhang, Zeren Luo, Zeyang Sha, Tianshuo Cong, Zheng Li, Shi- wen Cui, Weiqiang Wang, Jiaheng Wei, Xinlei He, Qi Li, and Qian Wang. 2025. FragFake: A Dataset for Fine-Grained Detection of Edited Images with Vision Language Models. arXiv preprint arXiv:2505.15644 (2025)

2025
[37]

Zachary Teed and Jia Deng. 2020. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16 . Springer, 402–419

2020
[38]

Cristian Vaccari and Andrew Chadwick. 2020. Deepfakes and disinformation: Exploring the impact of synthetic political video on deception, uncertainty, and trust in news. Social media+ society (2020)

2020
[39]

Jiarui Wang, Huiyu Duan, Jing Liu, Shi Chen, Xiongkuo Min, and Guangtao Zhai
[40]

Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. 2023. VideoFactory: Swap Attention in Spatiotemporal Diffusions for Text-to-Video Generation. arXiv preprint arXiv:2305.10874 (2023)

2023 arXiv
[41]

Wenhao Wang and Yi Yang. 2024. TIP-I2V: A Million-Scale Real Text and Image Prompt Dataset for Image-to-Video Generation. arXiv preprint arXiv:2411.04709 (2024)

2024 arXiv
[42]

In CAAI ICAI

Aigciqa2023: A large-scale image quality assessment database for ai gener- ated images: from the perspectives of quality, authenticity and correspondence. In CAAI ICAI
[43]

Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. 2024. Vript: A video is worth thousands of words. Advances in Neural Information Processing Systems 37 (2024), 57240–57261

2024
[44]

Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al . 2024. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer. arXiv preprint arXiv:2408.06072 (2024)

2024 arXiv
[45]

Zhiyuan Yan, Taiping Yao, Shen Chen, Yandan Zhao, Xinghe Fu, Junwei Zhu, Donghao Luo, Chengjie Wang, Shouhong Ding, Yunsheng Wu, et al. 2024. Df40: Toward next-generation deepfake detection. arXiv preprint arXiv:2406.13495 (2024)

2024 arXiv
[46]

Ruihan Zhang, Borou Yu, Jiajian Min, Yetong Xin, Zheng Wei, Juncheng Nemo Shi, Mingzhen Huang, Xianghao Kong, Nix Liu Xin, Shanshan Jiang, et al. 2025. Generative AI for Film Creation: A Survey of Recent Advances. arXiv preprint arXiv:2504.08296 (2025)

2025 arXiv
[47]

Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. 2023. I2vgen-xl: High-quality image- to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145 (2023)

2023 arXiv
[48]

Jiquan Yuan, Fanyi Yang, Jihe Li, Xinyan Cao, Jinming Che, Jinlong Lin, and Xixin Cao. 2024. PKU-AIGIQA-4K: A Perceptual Quality Assessment Database for Both Text-to-Image and Image-to-Image AI-Generated Images. arXiv preprint arXiv:2404.18409 (2024)

2024 arXiv
[49]

Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. 2024. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404 (2024)

2024 arXiv
[50]

Yufan Zhou, Bingchen Liu, Yizhe Zhu, Xiao Yang, Changyou Chen, and Jinhui Xu. 2023. Shifted diffusion for text-to-image generation. In CVPR. Received 30 May 2025; accepted 1 Aug 2025

2023
[51]

Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024. Llamafactory: Unified efficient fine-tuning of 100+ language models. arXiv preprint arXiv:2403.13372 (2024)

2024 arXiv
[2023]

IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)

Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)

2023
[2024]

In NeurIPS

T2vsafetybench: Evaluating the safety of text-to-video generative models. In NeurIPS

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al . 2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

arXiv 2025

[2] [2]

Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, Kai-Wei Chang, and Aditya Grover. 2024. Videophy: Evaluating physical commonsense for video generation.arXiv preprint arXiv:2406.03520 (2024)

arXiv 2024

[3] [3]

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

arXiv 2023

[4] [4]

Nuria Alina Chandra, Ryan Murtfeldt, Lin Qiu, Arnab Karmakar, Hannah Lee, Emmanuel Tanumihardja, Kevin Farhat, Ben Caffee, Sejin Paik, Changyeon Lee, et al. 2025. Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024. arXiv preprint arXiv:2503.02857 (2025)

arXiv 2025

[5] [5]

Chirui Chang, Zhengzhe Liu, Xiaoyang Lyu, and Xiaojuan Qi. 2024. What Matters in Detecting AI-Generated Videos like Sora? arXiv preprint arXiv:2406.19568 (2024)

arXiv 2024

[6] [6]

Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, et al. 2024. Demamba: Ai- generated video detection on million-scale genvideo benchmark. arXiv preprint arXiv:2405.19707 (2024)

arXiv 2024

[7] [7]

Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah

[8] [8]

Fanda Fan, Chunjie Luo, Wanling Gao, and Jianfeng Zhan. 2023. Aigcbench: Com- prehensive evaluation of image-to-video content generated by ai. BenchCouncil Transactions on Benchmarks, Standards and Evaluations (2023)

work page 2023