CMTA: Toward Generalizable AI-Generated Video Detection via Cross-Modal Temporal Artifacts

The researchers propose the CMTA framework, which detects AI-generated videos by capturing cross-modal temporal artifacts. Its starting point is that the cross-modal semantic alignment of real videos exhibits natural temporal fluctuations as semantics change, whereas AI-generated videos are governed by input prompts and their semantic trajectories are abnormally stable. Methodologically, CMTA uses BLIP to generate frame-level image captions and CLIP to extract visual-text representations, uses a GRU branch to characterize the temporal fluctuations of cross-modal alignment, and uses a Transformer encoder to capture fine-grained inter-frame changes from fused features. The authors claim new SOTA on 40 subsets of four datasets—GenVideo, EvalCrafter, VideoPhy, and VidProM—and stronger cross-generator generalization, with code and models to be released at https://github.com/hwang-cs-ime/CMTA。