生成式影片模型近年快速發展，文字生成影片（Text-to-Video, T2V）與影像生成影片（Image-to-Video, I2V）已能產生具高度擬真性與時間連貫性的影片。本文聚焦於 I2V 與 T2V 兩類生成影片的分類，藉此觀察不同生成條件是否會在影片視覺特徵上留下可辨識的統計差異。本研究從統計視覺特徵角度出發，透過探索性資料分析等工具，探討 I2V 與 T2V 生成影片是否存在可辨識之視覺統計差異，用以建立兼具公平性與可解釋性的辨識方法。

本研究參考 TIP-I2V 與 GenVideo-100K 兩個資料庫，並以 OpenSora 生成模型所產生之影片為研究對象，樣本包含 I2V 與 T2V 各 500 部影片，每部影片均勻抽取 8 幀影格，合計 1,000 部影片與 8,000 幀影格。為避免解析度、長寬比與有效內容範圍差異造成不公平比較，本研究採用公平有效內容區域（Fair Region of Interest, Fair ROI）作為前處理基準，並以影片層級切分降低資料洩漏風險。在特徵設計方面，本研究建構單幀視覺特徵與跨幀穩定性特徵，前者涵蓋色彩、亮度、紋理與邊緣等視覺資訊，後者則量化同一部影片中多張影格的時間波動與延續性。本文將上述特徵輸入 Logistic Regression、Random Forest 與 XGBoost 進行分類，並與 Xception 及 TimeSformer 等深度學習基準模型比較。

實證分析顯示，以單幀與跨幀穩定性完整特徵組（Full81 + TS54）搭配 Logistic Regression 之準確率可達 92.99%，高於僅使用單幀視覺特徵組（Full81）之 89.44%，與影片序列深度學習模型 TimeSformer 之 92.60% 相近，顯示以統計思維整理視覺特徵，亦可取得接近深度學習模型的分類表現。相較於深度學習模型，統計方法本身的模型結構較簡潔、計算成本較低，以及可攫取具有詮釋性的視覺特徵等優勢。整體而言，本研究提出之方法可作為兼具可解釋性、可重現性之 I2V 與 T2V 生成影片來源的辨識工具。

Generative video models have advanced rapidly in recent years, enabling both Text-to-Video (T2V) and Image-to-Video (I2V) systems to generate highly realistic and temporally coherent videos. This thesis investigates whether these two generation paradigms exhibit distinguishable statistical visual characteristics and proposes a fair and interpretable framework for distinguishing between I2V- and T2V-generated videos.

The experiments were conducted using OpenSora-generated videos from the TIP-I2V and GenVideo-100K datasets. The balanced dataset comprised 500 I2V and 500 T2V videos, with eight frames uniformly sampled from each video, yielding 1,000 videos and 8,000 frames. To ensure fair comparison, a Fair Region of Interest (Fair ROI) was adopted to reduce the influence of differences in resolution, aspect ratio, and effective content area. Single-frame visual features and temporal stability features were extracted and used to train Logistic Regression, Random Forest, and XGBoost models, with Xception and TimeSformer serving as deep learning baselines.

The results showed that Logistic Regression with the complete feature set (Full81 + TS54) achieved an average classification accuracy of 92.99%, outperforming the model using only single-frame features (89.44%) and achieving performance comparable to TimeSformer (92.60%). These findings indicate that statistical visual features can effectively distinguish between I2V- and T2V-generated videos while providing greater interpretability, lower computational cost, and better reproducibility than deep learning models. Overall, the proposed framework offers an effective and interpretable approach for identifying the source of AI-generated videos.

一、中文文獻

高崇哲（2025）。強化深偽偵測：以統計方法辨識影像的圖像特徵［碩士論文，國立政治大學商學院統計學系］。

黃政嘉（2024）。Deepfake 與 GAN 真偽人臉圖像統計分析［碩士論文，國立政治大學商學院統計學系］。

陳怡穎（2026 年 2 月 17 日）。是 AI 啦！豬哥亮「重磅回歸」民視除夕節目，網掀兩極討論。Yahoo 奇摩新聞。https://tw.news.yahoo.com/%E6%98%AFai%E5%95%A6-%E8%B1%AC%E5%93%A5%E4%BA%AE-%E9%87%8D%E7%A3%85%E5%9B%9E%E6%AD%B8-%E6%B0%91%E8%A6%96%E9%99%A4%E5%A4%95%E7%AF%80%E7%9B%AE-%E7%B6%B2%E6%8E%80%E5%85%A9%E6%A5%B5%E8%A8%8E%E8%AB%96-091121516.html

陳慧霜（2023）。影像分析與深偽影片的偵測［碩士論文，國立政治大學商學院統計學系］。

二、英文文獻

Bertasius, G., Wang, H., & Torresani, L. (2021). Is space-time attention all you need for video understanding? [Preprint]. arXiv. https://arxiv.org/abs/2102.05095

Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32. https://doi.org/10.1023/A:1010933404324

Chen, H., Hong, Y., Huang, Z., Xu, Z., Gu, Z., Li, Y., Lan, J., Zhu, H., Zhang, J., Wang, W., & Li, H. (2024). DeMamba: AI-generated video detection on million-scale GenVideo benchmark [Preprint]. arXiv. https://arxiv.org/abs/2405.19707

Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 785–794). Association for Computing Machinery. https://doi.org/10.1145/2939672.2939785

Chollet, F. (2017). Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 1251–1258). IEEE. https://doi.org/10.1109/CVPR.2017.195

Cox, D. R. (1958). The regression analysis of binary sequences. Journal of the Royal Statistical Society: Series B (Methodological), 20(2), 215–232. https://doi.org/10.1111/j.2517-6161.1958.tb00292.x

Dalal, N., & Triggs, B. (2005). Histograms of oriented gradients for human detection. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Vol. 1, pp. 886–893). IEEE. https://doi.org/10.1109/CVPR.2005.177

European Commission. (2026). Code of practice on transparency of AI-generated content. https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content

Geary, R. C. (1954). The contiguity ratio and statistical mapping. The Incorporated Statistician, 5(3), 115–145. https://doi.org/10.2307/2986645

Gu, Z., Chen, Y., Yao, T., Ding, S., Li, J., Huang, F., & Ma, L. (2021). Spatiotemporal inconsistency learning for DeepFake video detection [Preprint]. arXiv. https://arxiv.org/abs/2109.01860

Haralick, R. M., Shanmugam, K., & Dinstein, I. (1973). Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics, SMC-3(6), 610–621. https://doi.org/10.1109/TSMC.1973.4309314

Houben, M., Van Den Noortgate, W., & Kuppens, P. (2015). The relation between short-term emotion dynamics and psychological well-being: A meta-analysis. Psychological Bulletin, 141(4), 901–930. https://doi.org/10.1037/a0038822

Hu, B., Lei, C., Wang, D., Zhang, S., & Chen, Z. (2019). A preliminary study on data augmentation of deep learning for image classification [Preprint]. arXiv. https://arxiv.org/abs/1906.11887

Hunter, T., & Harwell, D. (2025, October 11). AI videos of dead celebrities are horrifying many of their families. The Washington Post. https://www.washingtonpost.com/technology/2025/10/11/openai-sora-dead-celebrities-ai/

Koval, P., Pe, M. L., Meers, K., & Kuppens, P. (2013). Affect dynamics in relation to depressive symptoms: Variable, unstable or inert? Emotion, 13(6), 1132–1141. https://doi.org/10.1037/a0033579

Kumar, N., Bhandari, P., & Maragatham, G. (2025). Bridging text and video generation: A survey [Preprint]. arXiv. https://arxiv.org/abs/2510.04999

Li, T., Zheng, G., Jiang, R., Zhan, S., Wu, T., Lu, Y., Lin, Y., Deng, C., Xiong, Y., Chen, M., Cheng, L., & Li, X. (2025). RealCam-I2V: Real-world image-to-video generation with interactive complex camera control [Preprint]. arXiv. https://arxiv.org/abs/2502.10059

Lowe, D. G. (2004). Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2), 91–110. https://doi.org/10.1023/B:VISI.0000029664.99615.94

Ni, Z., Yan, Q., Huang, M., Yuan, T., Tang, Y., Hu, H., Chen, X., & Wang, Y. (2025). GenVidBench: A 6-million benchmark for AI-generated video detection [Preprint]. arXiv. https://arxiv.org/abs/2501.11340

Ortutay, B. (2026, February 15). Hollywood groups condemn ByteDance’s AI video generator, claiming copyright infringement. AP News. https://apnews.com/article/7e445388401d172c6bf51d0d42aa4f24

Pardesi, R. S. (2026, February 16). ByteDance pledges to prevent unauthorised IP use on AI video tool after Disney threat. Reuters. https://www.reuters.com/world/china/disney-sends-cease-and-desist-bytedance-over-ai-generated-videos-2026-02-16/

Powers, D. M. W. (2011). Evaluation: From precision, recall and F-measure to ROC, informedness, markedness & correlation. Journal of Machine Learning Technologies, 2(1), 37–63.

Probst, P., Wright, M. N., & Boulesteix, A.-L. (2019). Hyperparameters and tuning strategies for random forest. WIREs Data Mining and Knowledge Discovery, 9(3), e1301. https://doi.org/10.1002/widm.1301

Qi, Z., Shi, P., Wang, S., Zhang, C., Zhao, F., Ying, Z., Pan, D., Yang, X., He, Z., & Dai, T. (2025). T2VEval: Benchmark dataset and objective evaluation method for T2V-generated videos [Preprint]. arXiv. https://arxiv.org/abs/2501.08545

Reed, G. F., Lynn, F., & Meade, B. D. (2002). Use of coefficient of variation in assessing variability of quantitative assays. Clinical and Diagnostic Laboratory Immunology, 9(6), 1235–1239. https://doi.org/10.1128/CDLI.9.6.1235-1239.2002

Ren, W., Yang, H., Zhang, G., Wei, C., Du, X., Huang, W., & Chen, W. (2024). ConsistI2V: Enhancing visual consistency for image-to-video generation [Preprint]. arXiv. https://arxiv.org/abs/2402.04324

Rowe, N. (2025, October 17). ‘Legacies condensed to AI slop’: OpenAI Sora videos of the dead raise alarm with legal experts. The Guardian. https://www.theguardian.com/technology/2025/oct/17/openai-sora-ai-videos-deepfake

Shorten, C., & Khoshgoftaar, T. M. (2019). A survey on image data augmentation for deep learning. Journal of Big Data, 6, Article 60. https://doi.org/10.1186/s40537-019-0197-0

Singh, J. (2025, June 2). Meta aims to fully automate advertising with AI by 2026, WSJ reports. Reuters. https://www.reuters.com/business/media-telecom/meta-aims-fully-automate-advertising-with-ai-by-2026-wsj-reports-2025-06-02/

Vahdati, D. S., Nguyen, T. D., Azizpour, A., & Stamm, M. C. (2024). Beyond deepfake images: Detecting AI-generated videos [Preprint]. arXiv. https://arxiv.org/abs/2404.15955

Wang, W., & Yang, Y. (2024). TIP-I2V: A million-scale real text and image prompt dataset for image-to-video generation [Preprint]. arXiv. https://arxiv.org/abs/2411.04709

West, A. (2026, February 16). After Tom Cruise’s AI fight video goes viral, SAG-AFTRA condemns Seedance 2.0: “This is unacceptable and undercuts the ability of human talent.” GamesRadar+. https://www.gamesradar.com/entertainment/movies/after-tom-cruises-ai-fight-video-goes-viral-sag-aftra-condemns-seedance-2-0-this-is-unacceptable-and-undercuts-the-ability-of-human-talent/

Xing, Z., Feng, Q., Chen, H., Dai, Q., Hu, H., Xu, H., Wu, Z., & Jiang, Y.-G. (2023). A survey on video diffusion models [Preprint]. arXiv. https://arxiv.org/abs/2310.10647

Xu, Y., Liang, J., Jia, G., Yang, Z., Zhang, Y., & He, R. (2023). TALL: Thumbnail layout for deepfake video detection [Preprint]. arXiv. https://arxiv.org/abs/2307.07494

Zhang, Y., Yang, H., Zhang, Y., Hu, Y., Zhu, F., Lin, C., Mei, X., Jiang, Y., Peng, B., & Yuan, Z. (2025). Waver: Wave your way to lifelike video generation [Preprint]. arXiv. https://arxiv.org/abs/2508.15761

Zheng, Z., Peng, X., Yang, T., Shen, C., Li, S., Liu, H., Zhou, Y., Li, T., & You, Y. (2024). Open-Sora: Democratizing efficient video production for all [Preprint]. arXiv. https://arxiv.org/abs/2412.20404

Zubair, A. R., & Alo, O. A. (2024). Grey level co-occurrence matrix (GLCM) based second order statistics for image texture analysis [Preprint]. arXiv. https://arxiv.org/abs/2403.04038