Most AI-generated video detectors analyze decoded pixels and overlook how videos are produced. We show that causal video autoencoders leave an architecture-specific rhythm in compressed bitstreams: their temporal stride concentrates unpredictable content at latent-chunk boundaries, so rate-distortion optimized encoders allocate bits periodically. Without decoding pixels, we observe four-frame pulses in CogVideoX and Wan2.2 and eight-frame pulses in LTX-Video, which are absent from content-matched camera footage. Passing camera footage through a generator's autoencoder alone induces the pulse, supporting its proposed origin. A training-free peak-ratio statistic separates generated from camera video (AUC 0.84-0.94) and persists across tested quantization settings and H.264/H.265 encoding. Frame-size traces also identify four commercial generators, with 89.2%, 84.1%, and 83.0% accuracy under H.264, H.265, and AV1, beating the pixel-based UnivFD baseline (72.7-73.6%). Our results directly link generative architecture to measurable coding demand.
Figures & tables
Figure 2: A causal video autoencoder lets new information in once per chunk. The decoder expands z0 into frame 0 and each later zk into s frames; only frame 0 and the first frame of each chunk (dark; t≡1(mods) ) carry much that the past cannot predict. Camera footage carries surprise at a roughly constant rate.
Figure 3: At fixed QP, a P-frame’s bits measure its surprise. The encoder predicts each frame from the previous one and pays only for the residual. On generated video the bits per frame bt pulse once per chunk and their spectrum peaks at 1/s , with harmonics; on camera video they are flat. Measured: median peak ratio 27–29 for the stride-4 generators, 1.1–1.4 for camera clips ( Tab. 1 ).
Figure 4: CoDe reads frame and row sizes from the bitstream and never decodes pixels. The three detectors share the same normalized coding-demand signals: (a) needs no training, (b) and (c) are learned.
Median
Share of clips
Generator
s
PRs
PRs>3
PRs>10
AUC
CogVideoX
4
27.3 / 1.4
0.80 / 0.25
0.65 / 0.05
0.84
Wan2.2
4
29.0 / 1.1
0.85 / 0.15
0.65 / 0.00
0.94
LTX-Video
8
6.1 / 1.1
0.75 / 0.10
0.40 / 0.00
0.89
Table 1: Generated clips pulse at their own stride; camera clips do not. x265, QP 27. Pairs are generated / camera: 20 generated clips per row vs. their 20 content- and resolution-matched camera references. PRs : peak ratio of the bit spectrum at the generator’s own stride.
Figure 5: The peak follows the stride. Mean ( ± standard error, s.e.) power spectrum of normalized, detrended per-frame bits, x265 QP 27, P-frames 6–45. Dashed: camera clips frame-dropped to the generator’s rate, which acquire a spurious peak at 1/4 ; all camera references in this paper are therefore used at their native rate.
Autoencoder
s
Median PRs
AUC [95% CI]
CogVideoX
4
87.0 / 1.4
0.99 [0.96, 1.00]
Wan2.2
4
12.9 / 1.1
0.86 [0.72, 0.96]
LTX-Video
8
10.4 / 1.1
0.86 [0.74, 0.96]
Table 2: An autoencoder round trip alone puts the rhythm into camera clips. x265 QP 27, 20 clips, no frames dropped ( k=0 ). Median PRs : round trip / original. CI: bootstrap confidence interval. Median PSNR of the round trip: 30.8, 31.4 and 27.7 dB.
Condition
AUC
Quantization: constant QP 22 / 27 / 32 / 37
x265
0.92 / 0.89 / 0.85 / 0.81
x264
0.92 / 0.89 / 0.82 / 0.80
Edits of generated clips † (unedited: 0.72–0.90)
Temporal smoothing of every frame
0.70–0.93
Blending chunk-start frames with neighbours
0.93–0.99
Table 3: The rhythm propagates into the encoded video bitstream and mostly survives quantization, edits and grain. Training-free detector ( PRs ), stride corpus, camera vs. generated, 20 vs. 20 clips per generator. Quantization: mean over the three generators; other rows: range over generators, x265 QP 27 unless stated. † , ‡ : two regenerations of the corpus, with their own clean AUC; motion terciles have 3–10 clips per class. Learned detectors on this corpus: Tab. 4 .
Bitstream (ours)
Pixels
Condition
PRs
GBDT
Transf.
UnivFD
Wan2.2
QP 27
0.935
0.960
0.895
0.540
QP 37
0.902
0.858
0.810
0.487
Smoothing
0.858
0.848
0.828
0.535
Chunk-start blend
0.963
0.978
0.818
0.545
Table 4: On the stride clips, bitstream detectors beat the pixel detector: the GBDT reaches AUC 0.85–0.98 where UnivFD sits at 0.34–0.70. Camera vs. generated, 20 vs. 20 clips, x265 QP 27 unless stated; learned detectors trained on clean clips and tested on the edited ones, grouped 5-fold × 3 by source clip. 95% group-bootstrap CI widths at QP 27: 0.10–0.18 (Wan2.2), 0.11–0.29 (LTX-Video). LR on normalized / raw bits at QP 27: 0.863 / 0.955 (Wan2.2), 0.983 normalized (LTX-Video). CogVideoX’s learned-detector run was incomplete at submission.
Detector
H.264
H.265
AV1
Same codec
Bitstream (ours)
Transformer, frame/block bits
80.6
78.4/68.4
60.6/59.9
GBDT, frame/block bits
87.9
83.1/82.0
60.0/65.0
LR, normalized/raw bits
89.2 /86.5
84.1 /83.7
83.0 /82.4
MLP, normalized/raw bits
85.4/82.9
81.5/81.6
77.6/76.2
Table 5: Bitstream detectors identify the generator at equal size and length, and a linear model on per-frame bits beats the pixel baseline by 10–17 points. Frontier set, 4 commercial generators, 4-way accuracy (%, chance 25), mean over 3×5 prompt-grouped folds (s.d. 3.9–7.5 for LR, 5.1–6.5 for MLP, 5.3–7.5 for UnivFD). Block bits: per WPP row (H.265) or superblock (AV1). Transformer: 150 epochs. LR, MLP: per-frame bits normalized by the clip mean, or raw. UnivFD: CLIP ViT-L/14 probe on 16 decoded frames. Size probe: all files are 832×480 . Across codecs: trained on the other H.26x codec; AV1 column: all directions to or from AV1.
Classes
Input
(chance)
Transformer
GBDT
Native
7 (19.2)
98.7
91.6 §
Crop, first 16 frames
7 (19.2)
65.6
61.7 §
Crop, first 48 frames
5 (31.2)
73.3
71.5
Crop, first 192 frames
3 (72.1)
74.4
86.2
camera vs. generated
2 (50.0)
51.1
88.8
Table 6: On DeepAction, source identification falls once frame size and clip length are equalized. Source identification, accuracy (%; chance: majority class). Native: original frame size and length. Crop: 512×448 ; fewer classes: subsets of the sources (192 frames: 24–30 fps only); camera vs. generated is balanced. Transformer: 3 seeds, 70/15/15 split; § 5-fold protocol of Tab. 7 .
Crop,
Detector
Native
16 frames
Bitstream
CoDe, Transformer ¶
94.5/75.6
74.1/70.0
CoDe, GBDT
94.9/87.0
69.9/67.3
Motion vectors [ 26 ]
84.9/84.4
71.4/68.8
Pixels
Table 7: Camera vs. generated on DeepAction: on 16-frame crops, where the size probe fails, the pixel baseline beats the bitstream detectors. Balanced accuracy (%), known / unseen generator, 5 prompt-grouped folds; unseen: mean over six held-out generators. ¶ Two-head Transformer, 18,985 parameters. Motion vectors: gradient-boosted statistics; native reads the delivered H.264 files.
∣IG∣ , train / test
Rollout
Class
I-frame share
top residue mod 4
I-frame share
Grok
0.44 / 0.32
0
0.06
HeyGen
0.34 / 0.33
0
0.06
MiniMax
0.24 / 0.21
0
0.06
Wan ( s=4 )
0.50 / 0.57
0 (predicted: 1)
0.06
Uniform
0.06
—
0.06
Table 8: Attributions follow the encoder’s keyframes, not the stride grid. Frontier set, H.265 WPP Transformer. IG: integrated gradients; rollout: attention rollout. I-frames: t≡0(mod16) . Among the other frames, the share on residue 0 mod 4 is 0.21–0.24 (uniform 0.20).
Luddy School of Informatics, Computing, and Engineering, Indiana University Bloomington, Bloomington, IN, USA · College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar
School of Cyber Science and Engineering Sichuan University Chengdu, Sichuan, China · School of Information Science and Technology Xizang University Lhasa, Xizang, China