Slop TVNewsLatest
Explainers

AI video leaves a periodic pulse in its compressed file

A Tsinghua-led preprint reads the pulse straight from the bitstream, with no pixels decoded, and names the generator up to 89.2% of the time.

Illustration: AI video leaves a periodic pulse in its compressed file
Illustration: SLOP TV News. Logos are trademarks of their owners.

Key takeaways

  • The preprint The Rhythm of Generated Video, submitted to arXiv on November 17, 2026, reports four-frame pulses in the bitstreams of CogVideoX and Wan2.2 and eight-frame pulses in LTX-Video, absent from content-matched camera footage.
  • A training-free peak-ratio statistic separates generated from camera video at an AUC of 0.84 to 0.94 with x265 and 0.85 to 0.93 with x264, and holds from quantization parameter 22 to 37.
  • Frame-size traces identify four commercial generators (Grok, HeyGen, MiniMax and Wan) at 89.2%, 84.1% and 83.0% accuracy under H.264, H.265 and AV1, beating the pixel-based UnivFD detector at 72.7% to 73.6%.
  • Passing camera footage through a generator's own autoencoder alone induces the pulse, but delivery-style encoding with rate control and B-frames erased the per-frame pulse in 14 of 15 tested cases.

Most AI-generated video detectors decode a clip to pixels and study the frames. A preprint posted to arXiv on November 17, 2026 argues there is a cheaper tell the file already carries, and that reading it needs no pixels at all.

The paper is The Rhythm of Generated Video: Latent Chunking Leaves Periodic Traces in Compressed Bitstreams, by Chen Ye Gan and Yuxing Han of Tsinghua University, Bryan Zhen Zhang of the University of California, San Diego, and Gene Wen of New York University Shanghai. It is a nine-page manuscript and a preprint. It has not been peer-reviewed, and no one has replicated it.

The finding is that AI-generated video leaves a periodic pulse in its compressed bitstream, a rhythm tied to the architecture of the generator that made it. The pulse shows up in how many bits the video encoder spends on each frame, a number a parser can read from the file without reconstructing a single frame. The authors call their method CoDe.

How a generator leaves a rhythm in the file

Nearly all current video generators are latent generators. A diffusion or flow model samples a compact latent video, and a video autoencoder decodes that latent into pixels. To compress time, those autoencoders are causal three-dimensional networks with a fixed temporal stride, which the paper writes as s. The first frame is decoded on its own, and every later latent frame is decoded into s pixel frames. CogVideoX and Wan2.2 use a stride of 4; LTX-Video uses 8, according to the paper.

The argument rests on a quantity the paper calls surprise: the part of a frame that cannot be predicted from the frames before it. In camera footage, surprise arrives continuously, at roughly the same rate in every frame. In a latent video, the paper proposes, new information can enter only where a new latent frame enters, once every s frames, because the frames in between are decoded from the same latent. Surprise should then arrive on a period of s frames, locked to the latent-chunk grid.

The second half of the pipeline turns that hidden rhythm into a measurable one. Conventional encoders such as H.264 and H.265 predict each frame from frames already coded and are rate-distortion optimized, so they spend bits only on what their temporal prediction cannot anticipate, which is the surprise. The bits spent on each frame, readable from the bitstream, should trace the generator's rhythm. Figure 1 in the paper shows that they do.

The stride is an architectural constant, the same for every clip a model produces, whatever the prompt or the sampler.

What the detector actually reads

CoDe never reconstructs a pixel. A clip is coded at constant quantization with one I-frame followed only by P-frames, and the parser reads the packet sizes and the WPP entry points from the bitstream. Divide each frame's size by the clip's mean and you get frame-size ratios, which the paper feeds to a logistic regression or to a small transformer with 26,000 parameters. A delivered file is re-encoded first.

The simplest check is training-free: a peak-ratio statistic, the strength of the bit spectrum at the generator's own stride. Generated clips concentrate the temporal variation of their bits at that frequency, and camera clips do not.

What the tests found

Re-encoded under identical settings, clips from CogVideoX and Wan2.2 pulse every fourth frame and clips from LTX-Video every eighth, while content-matched camera clips show no pulse. The median peak ratio at the generator's own stride is 27 to 29 for CogVideoX and Wan2.2 and 6.1 for LTX-Video, against 1.1 to 1.4 for their camera references. The peak-ratio statistic separates generated from camera video at an AUC of 0.84 to 0.94 with x265 and 0.85 to 0.93 with x264, and it holds across quantization parameters 22 to 37, falling from an AUC of 0.92 to about 0.80 as compression gets harsher.

The control experiment supports the proposed origin. Passing camera footage through a generator's own autoencoder, with no diffusion sampler, is enough to put the rhythm into it. For CogVideoX the median peak ratio rises from 1.4 to 87.0, an AUC of 0.99; Wan2.2 and LTX-Video reach 12.9 and 10.4. The autoencoder alone, the paper concludes, is sufficient.

The paper also identifies which generator made a clip. On its own Frontier dataset of four commercial generators, Grok, HeyGen, MiniMax and Wan, sharing 40 prompts, a logistic regression on frame-size ratios names the generator with 89.2%, 84.1% and 83.0% accuracy under H.264, H.265 and AV1. The 26K transformer reaches 80.6% and 78.4% under H.264 and H.265. Chance is 25%. On the same clips, the pixel-based UnivFD detector reaches 72.7%, 73.6% and 73.4%, so the bitstream traces beat it by 10 to 17 points.

What the paper does not claim

The method has limits the authors state themselves. Delivery-style encodes, which use rate control and B-frames, erased the per-frame pulse in 14 of 15 tested cases, though x265 still placed its reference frames on the stride-4 grid. LTX-Video is the fragile family: its rhythm needs motion, dropping to an AUC of 0.46 on near-static clips, and strong or film grain largely removes it.

The bitstream advantage is specific to naming the generator. On the harder task of separating camera from generated video with frame size and clip length controlled, the paper reports that UnivFD is stronger, at 90.5% against 69.9% to 74.1% balanced accuracy on 16-frame crops of known generators.

It is a research method, and no one has shipped it. No news outlet has reported on the preprint as of November 19, 2026, so the figures here come from the paper itself. The paper does not claim the rhythm is unbeatable or that it catches every generator; it tested specific model families.

Why it matters for anyone checking a clip

Pixel-based detection has to decode a clip, and its cues often weaken under re-compression. A compressed-domain check runs on the file as distributed, and the paper's statistics are cheap. That idea is not unique to this paper. STREAM, published in the ICML 2026 proceedings, mines motion vectors and residuals from the bitstream and reports an mAP of 0.965; a separate July 2026 preprint, Detect Early, Escalate Rarely, scores the codec's motion field on a CPU.

For a creator, the practical read is conditional. If the traces hold up, a watermark is not the only way to know a clip was generated: the generator's own encoding habit is visible in the delivered file. That cuts both ways. A re-encode, a crop or a delivery-style export can blur the trace, and the paper shows the pulse is easiest to read from the models it tested.

Read the preprint at arxiv.org/abs/2611.22996. For anyone checking a clip today, it adds a second signal to the visual checks: the file's own bit budget alongside the picture.

Paper card

  • Paper: The Rhythm of Generated Video: Latent Chunking Leaves Periodic Traces in Compressed Bitstreams
  • Authors: Chen Ye Gan, Bryan Zhen Zhang, Gene Wen, Yuxing Han
  • Submitted: November 17, 2026
  • Venue: arXiv preprint (cs.CV), nine pages, not peer-reviewed
  • Method: CoDe, reads frame and row sizes from the bitstream, no pixel decoding
  • Reported separation: AUC 0.84 to 0.94 (x265), 0.85 to 0.93 (x264)
  • Generator identification: 89.2%, 84.1%, 83.0% (H.264, H.265, AV1), chance 25%
  • Read it: https://arxiv.org/abs/2611.22996

Sources

  1. arxiv.org - abstract, submission date, authors, figures
  2. arxiv.org - paper body: method, datasets, results, limitations
  3. proceedings.mlr.press - independent compressed-domain detection (STREAM, ICML 2026)
  4. arxiv.org - independent bitstream detection (Detect Early, Escalate Rarely)
  5. sloptvnews.com - SLOP TV's own related fake-clip coverage