MSc Thesis · 2026
Generator-Agnostic
AI-Generated Video Detection
WOOF
Marom Sverdlov1
1st supervisor Martin R. Oswald1
2nd supervisor Marc Pollefeys2
1 University of Amsterdam·2 ETH Zürich
WOOF
We train on 4 generators from 2024 and generalize to modern generators, obtaining SOTA performance on benchmarks while only having around 100M parameters.
Clips WOOF labelled correctly of generators not seen during training.
WOOF combines motion and forensic features. Full method details are being held back ahead of an upcoming submission.
Average accuracy (%) = (Real% + Fake%) / 2. Higher is better.
| Method | Source | Avg ACC | Real | Fake |
|---|---|---|---|---|
| WOOF (ours) | this work | 94.2 | 91.1 | 99.0 |
| WaveRep G4 | Cozzolino et al., 2024 | 86.2 | 79.9 | 92.5 |
| BusterX++ | Wen et al., 2025a | 77.5 | 95.3 | 57.9 |
| BusterX | Wen et al., 2025b | 68.3 | 86.4 | 53.1 |
| ReStraV | Salvi et al., 2025 | 66.3 | 72.2 | 60.3 |
| MiMo-VL-7B-RL | Xiaomi, 2025 | 64.4 | 86.1 | 42.7 |
| InternVL3-8B | OpenGVLab, 2025 | 60.7 | 80.1 | 41.3 |
| Qwen2.5-VL-7B | Alibaba, 2025 | 60.2 | 92.6 | 27.7 |
| Qwen2.5-Omni-7B | Alibaba, 2025 | 56.4 | 81.4 | 31.4 |
| MiniCPM-o 2.6 | OpenBMB, 2025 | 54.1 | 78.4 | 29.8 |
| Keye-VL-8B | Kuaishou, 2025 | 53.8 | 95.7 | 11.8 |
Per-split balanced accuracy (%). ID = in-distribution generators, OOD = related but unseen, Wild = 2026-era frontier generators.
| Method | Source | ID | OOD | Wild |
|---|---|---|---|---|
| WOOF (ours) | this work | 95.1 | 94.2 | 88.7 |
| BusterX | Wen et al., 2025b | 85.5 | 84.9 | 81.5 |
| WaveRep G4 | Cozzolino et al., 2024 | 82.5 | 80.1 | 77.8 |
| DeMamba | Chen et al., 2024 | 82.0 | 80.4 | 70.9 |
| VideoMAE | Tong et al., 2022 | 79.1 | 76.4 | 72.0 |
| ViViT | Arnab et al., 2021 | 78.5 | 76.8 | 69.8 |
| 3D ResNeXt | Hara et al., 2018 | 72.6 | 64.6 | 61.9 |
| ReStraV | Salvi et al., 2025 | 71.9 | 57.9 | 53.1 |