Research map / Video / Video segmentation and tracking
Seeing What Matters: Generalizable AI-generated Video Detection with Forensic-Oriented Augmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes WaveRep, a wavelet-band replacement augmentation (plus VAE-reconstructed fakes) that trains a DINOv2 detector on a single generator (Pyramid Flow) to generalize to many unseen AI video generators.
Problem
AI-generated video detectors generalize poorly to unseen generators, and existing detectors are often trained on biased datasets where compression and semantic cues leak into decisions.
Approach
Identify diagonal mid-high frequency wavelet bands as discriminative and compression-robust forensic cues shared across generators, then train with a forensic augmentation that replaces low-frequency and horizontal/vertical wavelet bands of fake frames with those of the paired real video, forcing the detector to rely on diagonal mid-high bands; combined with VAE-reconstructed fakes from paired rea
Key results
- Average balanced accuracy of 94.3% across 16 generators when baselines are trained on their original datasets, about 10 points above DeMamba (second best).
- With WaveRep augmentation, a single-generator (Pyramid Flow) DINOv2 detector reaches 98.7 AUC and 91.8 bAcc on four recent generators, versus 94.7 AUC and 71.0 bAcc without the reconstruction and augmentation.
- Ranked first on the SAFE Synthetic Video Detection Challenge 2025 (per the paper).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add reconstructed fake videos to training (ablation) | DINOv2 detector with standard augmentation · Recent generators average | balanced accuracy: 71.0 -> 88.2 | Helped |
| Add WaveRep wavelet-band replacement (ablation) | DINOv2 detector trained with reconstructed fakes · Recent generators average | balanced accuracy: 88.2 -> 91.8 | Helped |
| Add MixUp (ablation) | DINOv2 detector trained with reconstructed fakes · Recent generators average | balanced accuracy: 88.2 -> 86.5 | Hurt |
| Add CutMix (ablation) | DINOv2 detector trained with reconstructed fakes · Recent generators average | balanced accuracy: 88.2 -> 88.7 | No clear effect |
| End-to-end fine-tune the backbone instead of linear probing (ablation) | DINOv2 detector with reconstructed fakes and WaveRep · Recent generators average | balanced accuracy: 78.8 -> 91.8 | Helped |
| Use 3 wavelet decomposition levels instead of 2 (ablation) | WaveRep detector · Recent generators average | balanced accuracy: 89.8 -> 91.8 | Helped |
| Use 3 wavelet decomposition levels instead of 4 (ablation) | WaveRep detector · Recent generators average | balanced accuracy: 88.5 -> 91.8 | Helped |
| Increase training generators from one to two (ablation) | DINOv2 detector with WaveRep · Recent generators average | balanced accuracy: 93.8 -> 96.6 | Helped |
| Increase training generators from two to four (ablation) | DINOv2 detector with WaveRep · Recent generators average | balanced accuracy: 96.6 -> 98.6 | Helped |
| Use WaveRep during training (ablation) | DINOv2 detector trained with one generator · Recent generators average | balanced accuracy: 73.3 -> 93.8 | Helped |
| Add mid-high diagonal frequencies from synthetic videos to real test videos | Ours · Modified real-video test set | TNR: 98.7 to 0.2 | Hurt |
| Train on four generators instead of one (ablation) | DINOv2 detector without WaveRep · Recent generators average | balanced accuracy: 73.3 -> 93.2 | Helped |
Surprising findings
- Image-oriented detectors (e.g., UnivFD, RINE) retrained on the same data outperform video-oriented detectors (AIGVDet, DeMamba, MM-Det), suggesting video detectors do not exploit 3D structure well.
- MM-Det scores below random guessing (about 42% average bAcc) when using its original training data, which the authors attribute to the biased DVF training set.
Limitations
- Not explicitly designed to capture temporal artifacts; relies on spatial per-frame cues.
- Trained only on Pyramid Flow fakes; may need adaptation when new generators use different synthesis paradigms.
- Accuracy drops substantially under severe compression (about 60% bAcc at CRF 32), requiring threshold calibration.
Experimental setup
Datasets: Panda70M WebVid-10M GenVideo DVF
Benchmarks: GenVideo custom recent-generator test set (Allegro, CogVideoX1.5, Mochi-1, Open-Sora Plan, Sora, NOVA, FLUX)
Metrics: balanced accuracy AUC Pd@5% FPR NLL ECE
Baselines: DMID UnivFD RINE FreqNet FIRE AIGVDet DeMamba MM-Det
Builds on: DINOv2 Pyramid Flow MixUp CutMix
Closest papers by meaning
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.90)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.89)
- BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation (NeurIPS 2026 · similarity 0.88)
- Forensic Self-Descriptions Are All You Need for Zero-Shot Detection, Open-Set Source Attribution, and Clustering of AI-generated Images (CVPR 2025 · similarity 0.87)
- Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.85)
- Locate-Then-Examine: Grounded Region Reasoning Improves Detection of AI-Generated Images (CVPR 2026 · 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.