i.
Socially Fragile Content
- Crime
- Natural Disaster
- CCTV
- Traffic
- War
sec 01
The reality gap
Reality Perception Gap (RPG) exposes the deployment gap in AI-video forensics: detectors tuned on clean clips break when videos are platform re-encoded, post-processed, captured with camera artifacts, locally edited, or generated by frontier models. RPG makes those shifts visible, so reviewers can see where synthetic-video detection fails outside the lab.
sec 02
What is new in Reality Perception Gap
ii.
Coupled Distribution Shifts
- Social Media Compression
- Retiming
- Center Crop
- Camera Optical Artifacts
iii.
Hybrid Manipulation
- Character Replacement
- Video Extension
- Relighting
sec 03
Video demonstrations
Group 01
Socially fragile categories
High-impact domains where manipulated video can cause social harm: combat, traffic, natural disasters, and crime.
Group 02
Localized AI edits in real footage
Mixed-source clips where surrounding scene and camera statistics remain real while a region, attribute, or temporal continuation is synthesized.
Group 03
Camera optical artifacts
Camera-side and camera-style artifact cases, including generated high-ISO noise, that reveal whether detectors confuse optical imperfections, sensor noise, and focus shifts with synthesis residuals.
sec 04
Detector performance on RPG
| Detector | ACC | Macro F1 | AUC up | TPR fake | TNR real | TPR at 1% FPR |
|---|---|---|---|---|---|---|
| D3 | 0.685 | 0.637 | 0.689 | 0.748 | 0.537 | 0.012 |
| WaveRep | 0.591 | 0.591 | 0.803 | 0.438 | 0.945 | 0.237 |
| GenD | 0.330 | 0.287 | 0.583 | 0.060 | 0.955 | 0.016 |
| VideoFACT | 0.562 | 0.479 | 0.583 | 0.562 | 0.561 | 0.025 |
| FreqNet | 0.404 | 0.404 | 0.548 | 0.285 | 0.679 | 0.025 |
| AIGVDet | 0.179 | 0.167 | 0.546 | 0.035 | 0.956 | 0.008 |
| RINE | 0.318 | 0.269 | 0.516 | 0.042 | 0.958 | 0.013 |
| UFD | 0.351 | 0.326 | 0.510 | 0.113 | 0.903 | 0.016 |
sec 05
Key findings
Compression as adversary
Platform encoding erases many frequency cues detectors are trained to read.
Deployment-grade compression removes fragile high-frequency residuals, exposing detectors that perform well only under clean laboratory conditions.
Generator coverage
Frontier generators, not data volume alone, set the binding constraint.
Per-generator behavior varies substantially, suggesting that future benchmarks need deliberate coverage of modern synthesis pipelines.
Threshold calibration
Default thresholds are unstable across realistic deployment shifts.
Detector operating points can drift toward always-real or always-fake behavior when clean evaluation assumptions are removed.
Optical decoys
Camera-style artifacts can look like generative artifacts.
Rolling shutter, chromatic aberration, generated high ISO noise, and autofocus hunting stress whether models distinguish capture physics from synthesis traces.