VAE Rhythm: Detecting AI Videos via Periodic 3D Causal VAE Decoding Traces
showlab · hf · 2026-11-19
showlab proposes detecting AI-generated videos not by visible artifacts (which newer generators learn to remove) but via a trace left by the generation pipeline itself. Most video generators decode latents with a 3D causal VAE that groups consecutive frames, so frame fidelity rises and falls periodically within each group — dubbed "VAE Rhythm". Real videos, captured frame by frame, lack it.
A training-free "phase scan" reveals the trace: reconstruct a clip with an off-the-shelf VAE from successive start offsets and measure the periodicity of reconstruction error. The rhythm appears in most of 39 variants across 31 generators, including closed-source Sora, Kling 3.0, and Seedance 2.0.
Results:
- The zero-shot rhythm score beats all 15 retrained baselines on DF26;
- As a plug-in it lifts all 19 tested detectors by 6.7%–23.7% AUROC;
- CaDet, trained on pairs of real videos and their VAE reconstructions, achieves the best AUROC on AIGVDBench, DF26, and three 2026 closed-source models; a variant trained with zero generated videos still exceeds 90% on AIGVDBench.
More from Multimodal
- Scenario open-sources a full AI art team as a Photoshop plugin, MCP-ready — aziz4ai · 2026-11-19
- MAVE reconstructs editable 3D blockouts from film shots, works inside Blender — bdsqlsz · 2026-11-19
- AI-made Dragon Ball parody music video nails the anime look — Top_Brilliant_9587 · 2026-11-19
- GTA 6-style LoRA generates strikingly authentic rain-soaked neon artwork — freshstart2027 · 2026-11-19
- NUS's Jeveration turns a frozen vision-language model into a single-pass image generator — NUS-DSA3101 · 2026-11-19
- KAIST's ReMo: training-free cross-modal token compression trims Omni-LLM inference cost — kaist-ai · 2026-11-19