VAE Rhythm detects AI video from periodic temporal decoding traces
Original title: VAE Rhythm: Temporal Decoding Traces for AI-Generated Video Detection
A training-free phase scan beats 15 baselines on DF26 and lifts 19 detectors by 6.7% to 23.7% AUROC.
Most video generators decode latents with a 3D causal VAE that groups frames, leaving a periodic fidelity trace the authors call VAE Rhythm. Real video lacks it. A training-free phase scan reconstructs clips from successive offsets with an off-the-shelf VAE and scores reconstruction-error periodicity. The authors report the trace in most of 39 variants of 31 generators, including Sora, Kling 3.0 and Seedance 2.0, and train detector CaDet that leads on AIGVDBench, DF26 and three 2026 closed-source models.
Unconfirmed: Detection rates and benchmark wins are the authors' reported results; independent replication is not stated.
- Visible artifacts are unreliable because newer generators learn to remove them.
- The rhythm appears in most of 39 variants across 31 generators.
- Plug-in use improves all 19 tested detectors by 6.7% to 23.7% AUROC.
- A CaDet variant trained without generated video still exceeds 90% on AIGVDBench.
- Closed-source models tested include Sora, Kling 3.0 and Seedance 2.0.
- AI models
- SoraKling 3.0Seedance 2.0CaDet
From the original article
1 min read at huggingface.coVisible artifacts are an unreliable cue for detecting generated videos, because newer generators learn to remove them. We instead detect a trace left by the generation pipeline itself. Most video generators decode latents with a 3D causal VAE that groups consecutive frames in time, so the fidelity of generated frames rises and falls periodically with their position in a group.
By Ke Sun, Shuwei Li, Haiyang Mei · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.