Skip to content

VAE Rhythm: temporal decoding traces for AI-generated video detection

VAE Rhythm: Temporal Decoding Traces for AI-Generated Video Detection

Hugging Face Daily Papers· arXiv:2611.21114· Ke Sun, Shuwei Li, Haiyang Mei, Mike Zheng Shou·Indexed by this site · 14 hr ago· Originally publishedDefenses· Its place in the scan record →
AI summary

The paper proposes VAE Rhythm, which uses periodic decoding traces left by the generation pipeline to detect AI-generated video. Most video generators use a 3D causal VAE to decode consecutive frames in temporal groups, and frame fidelity fluctuates periodically with position within the group; real videos shot frame by frame do not have this trace. The accompanying phase scan requires no training: it uses an off-the-shelf VAE to reconstruct clips from different starting offsets and measures the periodicity of reconstruction error. The trace was detected on most of the 39 variants of the 31 generators tested, including the closed-source Sora, Kling 3.0, and Seedance 2.0. As a zero-shot score, the trace surpasses all 15 retrained baselines on DF26; as a plug-in, it raises the AUROC of the 19 tested detectors by 6.7% to 23.7%; CaDet, trained on pairs of real videos and their VAE reconstructions, achieves the best AUROC on AIGVDBench, DF26, and three 2026 closed-source models, and the variant trained without any generated video still exceeds 90% on AIGVDBench.

Read originalhuggingface.co