Try our hybrid model that combines temporal low-level artifacts with CNN features to detect AI-generated video — upload your own video to see it in action.
This study analyzes temporal low-level artifacts in AI-generated video and proposes a generalizable detection model that leverages them. The analysis found that AI-generated videos exhibit a temporal low-level artifact distribution different from real videos, and a hybrid model combining these artifacts with CNN features achieved excellent detection performance. This confirms that temporal low-level artifacts can act complementarily with CNN features to improve detection performance.
Spatial artifacts such as RGB, FFT, and Wavelet are low-level statistics that appear within each individual frame. In contrast, temporal-native artifacts such as frame difference, optical flow, and temporal FFT/Wavelet only emerge from the changes between frames, capturing information that spatial artifacts alone cannot.
Adding spatial artifacts to ResNet alone (AUC 0.9455) improves it slightly to 0.9515, and adding temporal-native artifacts improves it to 0.9516; combining both in a hybrid model raises it to 0.9603. You can compare the detection results of all four models directly on this site.