Research map / Video / Video generation
Real Data Lies: Unveiling and Closing the Quality Shortcut in Generalizable AI-Generated Video Detection
Ziyuan Fang, Tianyi Wei, Guanjie Wang, Weiming Zhang, Nenghai Yu, Wenbo Zhou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Kinetics-400 + Pika training pair with quality-matched InternVid-AES + Pika (ablation) | VideoMAEv2-B · GenVideo revised protocol, Overall Mean | AUC (%): 81.70 -> 87.78 | Improved |
| quality-aligned training data (InternVid-AES + Pika) instead of Kinetics-400 + Pika (ablation) | DINOv3-L · GenVideo revised protocol, Overall Mean | AUC (%): 87.09 -> 86.59 | Hurt |
| quality-aligned training data (InternVid-AES + Pika) instead of Kinetics-400 + Pika (ablation) | SAFE · GenVideo revised protocol, Overall Mean | AUC (%): 77.24 -> 77.33 | No clear effect |
| add quality-aware data expansion (LQ and UHQ real/fake pairs) on top of aligned data | DINOv2-L · GenVideo revised protocol, Overall Mean | AUC (%): 89.81 -> 94.83 | Improved |
| add quality-aware data expansion on top of aligned data | SAFE · GenVideo revised protocol, Overall Mean | AUC (%): 77.33 -> 85.15 | Improved |
| add LQ expansion (Kinetics-400 + OpenSora) only (ablation) | VideoMAEv2-B Aligned · GenVideo revised protocol, Overall Mean | AUC (%): 87.78 -> 92.56 | Improved |
| add UHQ expansion (OpenVidHD + DynamiCrafter) only (ablation) | VideoMAEv2-B Aligned · GenVideo revised protocol, Overall Mean | AUC (%): 87.78 -> 90.86 | Improved |
| add both LQ and UHQ expansion (ablation) | VideoMAEv2-B Aligned · GenVideo revised protocol, Overall Mean | AUC (%): 87.78 -> 93.39 | Improved |
| quality-matched aligned set (1x data) versus 3x larger biased set (ablation) | VideoMAEv2-B · 9 real datasets, Mean* (unseen datasets) | AUC (%): 84.88 -> 86.15 | Improved |
| 3x scale up of biased training data (Biased++) (ablation) | VideoMAEv2-B · 9 real datasets, Overall Mean | AUC (%): 81.88 -> 89.40 | Improved |
| expanded quality-aware data versus Biased++ at equivalent scale (ablation) | VideoMAEv2-B · 9 real datasets, Overall Mean | AUC (%): 89.40 -> 94.23 | Improved |
| train on high-quality real data matching Pika (InternVid-AES) instead of low-quality (Youku-mPLUG) or ultra-HQ (OpenVidHD) (ablation) | VideoMAEv2-B · Overall Mean over 9 real test datasets | AUC (%): 81.88 (Kinetics-400), 78.48 (Youku-mPLUG), 81.23 (OpenVidHD) -> 88.35 (InternVid-AES) | Improved |
Closest papers by meaning
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.88)
- Seeing What Matters: Generalizable AI-generated Video Detection with Forensic-Oriented Augmentation (NeurIPS 2025 · similarity 0.87)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.87)
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · similarity 0.86)
- AI-Generated Video Detection via Perceptual Straightening (NeurIPS 2025 · similarity 0.86)
- Provably Improving Generalization of Few-shot models with Synthetic Data (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
