Research map / Video / Video segmentation and tracking
Training-free Detection of Generated Videos via Spatial-Temporal Likelihoods
Omer Ben Hayun, Roy Betser, Meir Yossef Levi, Levi Kassel, Guy Gilboa
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 |
|---|---|---|---|
| use balanced real/synthetic test sets instead of unbalanced (D3 with DinoV3, downsampled real) | D3 evaluation · GenVideo | mean AP: 0.83 -> 0.74 | Hurt |
| duplicate real-video frames 3->8 FPS instead of downsampling (D3 with DinoV3, balanced) | D3 evaluation · GenVideo | mean AP: 0.74 -> 0.94 | Hurt |
| use X-CLIP16 instead of DinoV3 as D3 embedder | D3 · GenVideo (downsample, balanced) | mean AP: 0.74 -> 0.78 | Improved |
| apply l2 normalization to temporal differences (ablation) | STALL temporal features · VATEX calibration set (AD normality test) | normal features (%): 0.0% -> 98.4% | Improved |
| add whitening to l2-normalized temporal differences (ablation) | STALL temporal features · VATEX calibration set (AD normality test) | normal features (%): 98.4% -> 99.6% | Improved |
| add whitening to raw frame embeddings (ablation) | STALL spatial features · VATEX calibration set (AD normality test) | normal features (%): 96.3% -> 98.2% | Improved |
| estimate temporal whitening from one random frame difference per video instead of all differences (ablation) | STALL · GenVideo/VideoFeedback/ComGenVid | average AUC: 0.8110 -> 0.8105 | No clear effect |
| raise temporal derivative order D from 1 to 4 (ablation) | STALL · three benchmarks | AUC | No clear effect |
| insert a flash frame (black or white) into real videos | STALL temporal likelihood · MSR-VTT (400 real videos) | temporal score | Hurt |
| reverse or shuffle frame order | STALL temporal likelihood · MSR-VTT (400 real videos) | temporal score | No clear effect |
| vary frame rate in {2, 4, 8, 12, 24} FPS at inference | STALL · GenVideo 24 FPS subset | detection performance | No clear effect |
| change calibration set from VATEX to DiDeMo (ablation) | STALL · VideoFeedback | average AUC: 0.82 -> 0.86 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- RIGID — RIGID: Recurrent GAN Inversion and Editing of Real Face Videos (ICCV 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Physics-Driven Spatiotemporal Modeling for AI-Generated Video Detection (NeurIPS 2025 · similarity 0.87)
- Taming generative video models for zero-shot optical flow extraction (NeurIPS 2025 · similarity 0.86)
- Manifold Induced Biases for Zero-shot and Few-shot Detection of Generated Images (ICLR 2025 · similarity 0.86)
- AI-Generated Video Detection via Perceptual Straightening (NeurIPS 2025 · similarity 0.85)
- GenRec: Unifying Video Generation and Recognition with Diffusion Models (NeurIPS 2024 · similarity 0.85)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 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.
