Introduction
As video generators learn to suppress obvious distortions, flicker, and texture defects, detectors that rely on visible flaws have a moving target. A study from Show Lab proposes looking elsewhere: at a temporal trace that may arise from how generated video is decoded. The researchers call this signal VAE Rhythm.
A rhythm in the decoding pipeline
Many video generators create latent representations and then decode them into frames with a 3D causal VAE. Rather than treating every frame independently, these decoders often process consecutive frames in groups. The paper notes group sizes of four frames are common, while models such as LTX and Kling 3.0 may use groups of eight. The fidelity of reconstructed frames can vary with each frame’s position within a group, producing a periodic pattern. Frame-by-frame camera footage does not have the same generator-defined grouping and is expected to lack this particular trace.
The distinction is not simply that generated video has higher or lower reconstruction error. The proposed clue is how the error changes periodically as the alignment of frame groups shifts.
Scanning for the signal
The team’s training-free phase scan uses an off-the-shelf video VAE to reconstruct a clip from successive starting-frame offsets. It then measures the periodicity of reconstruction error across those offsets. When the scan’s frame grouping aligns with the original decoding groups, the residual falls; shifting the offset changes that alignment and produces an oscillation. Real clips, in the reported experiments, tend to show a flatter pattern. The method therefore evaluates periodic structure, not the absolute size of the error.
Across 39 variants of 31 generators, the researchers report finding VAE Rhythm in most tested systems, including the closed-source Sora, Kling 3.0, and Seedance 2.0. On DF26, the training-free rhythm score reaches 76.2% AUROC, outperforming the 15 retrained baselines included in the comparison. Adding rhythm evidence as a plug-in improves all 19 tested detectors on the same benchmark, by 6.7 to 23.7 AUROC points.
From a signal to a detector
The paper also uses real clips and their VAE reconstructions to create paired training data under shared post-processing, then trains CaDet. The authors report strong AUROC results for CaDet on AIGVDBench and DF26, as well as tests involving three 2026 closed-source models. A version called CaDet-GF, trained without any generated video clips, also exceeds 90% in both reported AIGVDBench results. This suggests that reconstruction traces may serve not only as a stand-alone score but also as complementary evidence for learned detectors.
Why it matters—and what remains open
VAE Rhythm shifts attention from appearance defects, which model developers can deliberately reduce, to a structural property of the generation pipeline. Its phase scan does not require training on examples from a target generator, and its plug-in form offers a way to combine the signal with existing detectors. But the findings do not establish that every generator leaves the same trace, or that one signal can settle every authenticity question. Further evaluation across codecs, editing operations, and future model designs will matter for real-world use.
Source: Hugging Face Daily Papers
Comments
Checking sign-in status...
Loading comments...