Spiking Neural Networks Detect AI Video Temporal Smoothness
Researchers have developed a new method for detecting AI-generated videos using Spiking Neural Networks (SNNs). The approach, detailed in a preprint on arXiv (2605.05895), targets temporal inconsistencies that remain even in photorealistic single-frame AI video. The team identified two key signatures of AI-generated content: smoother frame-to-frame pixel-level temporal residuals and more compact trajectories in semantic feature space, indicating a temporal smoothness gap. Existing detectors that use full-frame sequences, fixed video-level descriptors, or comparison metrics degrade under cross-generator evaluation. The SNN-based detector jointly exploits these signatures on the GenVidBench benchmark, offering improved robustness across different generators.
Key facts
- arXiv preprint 2605.05895 proposes SNN-based AI video detection.
- AI videos have smoother pixel-level temporal residuals.
- AI videos show more compact semantic feature trajectories.
- Existing detectors degrade under cross-generator evaluation.
- GenVidBench is the benchmark used for evaluation.
- The method exploits a temporal smoothness gap at two levels.
- Single-frame AI video is photorealistic.
- Inter-frame dynamics are the main detection axis.
Entities
Institutions
- arXiv