Visible 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. We call this periodic trace VAE Rhythm. Real videos are captured frame by frame and do not have it. To reveal it without training, we propose a phase scan. The scan reconstructs a clip with an off-the-shelf VAE from successive start offsets and measures the periodicity of the reconstruction error across these offsets. We find VAE Rhythm in most of the 39 variants of 31 generators we test, including the closed-source Sora, Kling 3.0 and Seedance 2.0. We use the rhythm as a zero-shot score, as a plug-in for existing detectors and as a data engine that pairs real videos with their VAE reconstructions to train our detector CaDet. On DF26, the training-free rhythm score outperforms all 15 retrained baselines, and as a plug-in it improves all 19 tested detectors by 6.7% to 23.7% AUROC. CaDet achieves the best AUROC on AIGVDBench, DF26 and three 2026 closed-source models, and its variant trained without any generated video still exceeds 90% on AIGVDBench.
Figures & tables
Figure 1: VAE Rhythm. (a) A 4× causal VAE decodes frames in groups, illustrated on Sora frames. (b) As the VAE window shifts, its groups realign with those of the clip every four starts. The residual r(k) of a typical Sora clip therefore oscillates, while that of a typical real clip stays nearly flat.
Figure 2
Figure 4: Barcode view. Residual profiles of up to 100 clips per group with one clip per row. Red is high and blue is low.
Figure 5: Overview. (a) A frozen video VAE has two uses. The periodicity of its phase scan gives the zero-shot score sc , and its round trip creates the VAE pairs, which differ only by VAE decoding. (b) On one frozen DINOv3 backbone, CaDet trains an artifact path with the pairs and a semantic path without them. (c) The plug-in adds positive rhythm evidence to the score of any detector without retraining it. All plots, tables and the difference map are measured.
Open-source
480p
Closed-source
480p
720p
Wan2.1 T2V
99.5
Sora
90.4
84
Wan 2.2 T2V
94.7
Wan 2.6
87.7
96.0
EasyAnimate I2V
94.2
Kling 3.0
87.2
88.2
RepVideo
89.3
Luma
79.3
89
Open-Sora
77.0
Seedance 2.0
73.5
78.5
HunyuanVideo 1.5 T2V
76.7
Grok Imagine
70.0
79.4
Table 1: Zero-shot AUROC of representative generators (%) using the rhythm statistic alone without training. Table S2 in the supplement lists all 39 variants.
AIGVDBench
2026 closed
DF26
Method
venue
20 open
19 unseen
11 closed
3 models
all
open
closed
2AFC
Image forensics, frame-level
CNNSpot [ 84 ]
CVPR’20
82.1 ± 1.2
81.1 ± 1.2
71.4 ± 2.4
78.0
53.5
65.5
35.5
61.0
Gram-Net [ 45 ]
CVPR’20
82.4 ± 0.4
81.4 ± 0.5
67.0 ± 1.4
76.9
52.9
63.3
37.3
59.4
FreDect [ 20 ]
ICML’20
60.9 ± 1.8
58.9 ± 1.9
52.9 ± 1.2
60.8
60.8
59.0
63.4
62.4
Fusing [ 34 ]
ICIP’22
67.0
65.3
51.5
64.9
46.0
53.3
35.0
46.4
Table 2: Main results (AUROC, %). All detectors are trained on AIGVDBench’s Open-Sora and real clips and tested without retraining. ± is the standard deviation over 3 seeds, and bold marks the best per column.
Figure 6: Analysis of the plug-in and the two paths. (a) Each dot is a generator. The plug-in gain averaged over the 19 detectors grows with the zero-shot AUROC of the generator. (b) Each dot is a detector, and weaker detectors gain more on DF26. (c) Each dot is an unseen generator, and the semantic path is better above the diagonal. Colors follow (a), and bands are bootstrap 95% intervals of the fit.
Figure 7: Robustness to post-processing on 250 DF26 clips (AUROC, %). Real and generated clips receive the same processing, and every calibration stays fixed.
Figure S1 : VAE Rhythm across all 39 generator variants. Mean residuals r(k) with one standard error. Orange curves are generated clips, teal curves are real clips of the same benchmark, and n is the number of clips or source segments.
variant
n
stat.
median
AUROC
p
above
R
AIGVDBench, open-source
Wan2.1 T2V
50
C4
0.96 / 0.11
99.5
< .001
96
0.99
EasyAnimate I2V
50
C4
0.81 / 0.11
94.2
< .001
82
0.86
EasyAnimate T2V
50
C4
0.74 / 0.11
89.8
< .001
80
0.59
RepVideo
50
C4
0.61 / 0.11
89.3
< .001
70
0.71
CogVideoX1.5 V2V
50
C4
0.43 / 0.11
77.4
< .001
50
0.49
Table S1 : Real vs. generated statistics of VAE Rhythm for the 36 window-matched variants. Section S1.1 defines the columns.
AIGVDBench, open-source
AIGVDBench, closed-source
DF26 and 2026 models
generator
VAE
480p
generator
VAE
480p
720p
generator
VAE
480p
720p
Wan2.1 T2V
Wan 4 ×
99.5
Sora
Unknown
90.4
84
Wan 2.2 T2V / I2V
Wan 4 ×
94.7 / 88.4
98.8 / 96.7
EasyAnimate I2V / T2V
own 4 ×
94.2 / 89.8
Luma
Unknown
79.3
89
HunyuanVideo 1.5 T2V / I2V
own 4 ×
76.7 / 73.6
84.3 / 81.9
RepVideo
CogX 4 ×
89.3
Gen-3
Unknown
78.6
80
LTX 2.3 T2V / I2V
own 8 ×
76.5 / 60.6
74.3 / 60.0
CogVideoX1.5 T2V / V2V
CogX 4 ×
77.8 ‡ / 77.4
CausVid
Wan 4 ×
80.0
77
Kling 3.0
Unknown, 8-frame
87.2
88.2
Open-Sora
own 4 ×
77.0
Wan (web)
Wan 4 ×
58.7
78
Wan 2.6, 30 fps
Unknown, period 5
87.7
96.0
Table S2 : Zero-shot AUROC (%) per generator using the rhythm statistic alone without training. The AIGVDBench 720p column uses smaller subsets, and ‡ marks short-window scans without a matched real reference.
Figure S2 : Phase locking. Phase histograms of the period-4 component, with the concentration R and the amplitude ratio of each group. The indicated groups use period 8.
Figure S3 : Spectral summaries. Mean normalized periodograms of the sampled profiles. Sample counts are beside each group.
Figure S4 : Dominant period of the residual curve. Bars show the share of clips whose strongest period is 4 in blue, 8 in orange or another period in gray. The black line marks the real period-4 share of the same benchmark.
Figure S5 : The scale law. (a) Zero-shot C4 AUROC of 12 open-source 4× -VAE generators at 256p in light blue and at native resolution capped at 480p in dark blue. Orange shows the controlled real-through-VAE experiment. (b) Closed-source models at three scan scales, labelled with their maximum AUROC across scales.
Figure S6 : Phase scan in detail. This figure expands Figure 1 b. (a) A W -frame window shifts by one frame per pass through the same frozen VAE. (b) Subtracting the same-frame mean along the orange diagonal from the stacked errors and averaging over j>0 gives r(k) . Values are schematic.
Figure S7 : Zero-shot AUROC versus scan budget. Points are group means with bootstrap standard errors. (a) Offsets K at W=13 . (b) Window length W at K=24 . (c) All 42 settings against their cost. The default setting is circled.
time
AIGVDBench
DF26
K
W
KW
(s)
10 closed
1 open 4×
3 open 8×
4 closed
6 open
GB++
Window length W , with K=24
24
5
120
11.6
72.1
99.6
72.2
78.5
89.2
65.7
24
9
216
20.7
69.7
99.3
71.9
82.3
84.7
63.9
24
13
312
29.9
68.8
99.1
62.0
85.7
86.0
68.5
24
17
408
39.0
65.6
98.8
62.4
79.9
83.5
69.0
Table S3 : Scan budget (zero-shot AUROC, %). The table gives accuracy and per-clip time for different numbers of offsets K and window lengths W . Bold marks the best per column, and ∗ marks a time from the fit.
zero-shot
CaDet
Effort
TimeS-
generator
VAE / period
scan
+ plug-in
+ plug-in
former
DF26, closed-source
Veo 3.1
Unknown / 4
73.4
77.9
78.2
77.1
77.8
83.6
Kling 3.0
Unknown / 8
88.2
73.2
77.9
57.9
66.5
52.8
Grok Imagine
Unknown / 4
79.4
72.7
79.7
76.1
81.6
80.5
Wan 2.6, 30 fps
Unknown / 5
96.0
92.5
97.9
82.6
96.9
86.4
Table S4 : Generalization to newer generators (AUROC, %) against the 271 DF26 reals. Bold marks the best per row, and the zero-shot column scans all clips at native ≤ 720p as in the plug-in.
AIGVD o/c
DF26 c
(a) training data, DINOv3-L LoRA
benchmark set only
68.9 / 81.2
65.2
+ post-processed reals, 3 seeds
73.8 / 86.2
64.3
+ VAE pairs, single head, 3 seeds
93.0 / 92.2
59.3
real + pairs, CaDet-GF, 3 seeds
90.9 / 91.9
51.3
(b) leave one VAE out, single head
Table S5 : Extended ablations (AUROC, %) on AIGVDBench’s open and closed averages and DF26’s closed-source generators.
real side
CaDet
artifact path
semantic path
CaDet-GF
no pairs
Effort
CNNSpot
D3
OpenVid, protocol
93.9 / 95.8
94.1 / 92.0
82.9 / 93.0
90.9 / 91.9
68.9 / 81.2
84.3 / 92.5
82.1 / 71.4
45.6 / 62.7
same clips, re-encoded 360p
32.3 / 28.8
30.8 / 22.6
45.4 / 58.1
65.8 / 59.8
18.7 / 11.8
31.0 / 28.2
34.3 / 17.5
51.7 / 67.9
Youku, GenVideo-100K
17.4 / 11.9
17.9 / 10.7
24.8 / 30.8
45.2 / 34.5
12.2 / 5.4
14.6 / 8.5
19.9 / 8.5
78.4 / 86.6
ViF-Bench reals
28.9 / 25.4
27.9 / 21.0
35.2 / 42.6
54.9 / 46.4
20.8 / 16.8
21.7 / 17.7
28.0 / 15.7
–
Youku, engine + Youku reals
–
–
–
71.2 / 57.5
–
–
–
–
Table S6 : Swapping the real side (open / closed AUROC, %). Generated clips and models are fixed, and only the real videos change.
AIGVDBench
2026 closed
ViF
DF26
Detector
open
closed
all
open
closed
2AFC
zero-shot rhythm alone ( Ppk )
69.0
63.0
62.5
56.9
76.2
74.4
79.0
76.1
CNNSpot
82.9
71.6
78.0
47.0
53.5
65.5
35.5
61.0
+ VAE Rhythm plug-in
86.4 +3.5
76.9 +5.3
80.3 +2.3
51.4 +4.4
69.8 +16.3
79.5 +14.0
55.3 +19.8
75.1 +14.2
Gram-Net
83.3
67.8
76.9
52.4
52.9
63.3
37.3
59.3
+ VAE Rhythm plug-in
86.1 +2.8
74.7 +6.9
79.2 +2.3
56.8 +4.4
71.0 +18.1
79.2 +15.9
58.8 +21.5
74.5 +15.2
Table S7 : The VAE Rhythm plug-in, full table (AUROC, %). Each detector appears with and without the plug-in, and the table includes extra training variants.
AIGVDBench closed
2026 closed
DF26 closed
Detector
none
reported
original
rank
none
reported
original
rank
none
reported
original
rank
Image forensics, frame-level
CNNSpot
71.6
76.9 +5.3
77.0 +5.3
76.9 +5.3
78.0
80.3 +2.3
80.2 +2.2
80.3 +2.3
35.5
55.3 +19.8
55.2 +19.7
55.3 +19.8
Gram-Net
67.8
74.7 +6.9
73.9 +6.1
74.7 +6.9
76.9
79.2 +2.3
77.2 +0.3
79.2 +2.3
37.3
58.8 +21.5
60.9 +23.6
58.8 +21.5
FreDect
53.6
65.2 +11.7
64.3 +10.8
65.2 +11.7
60.8
66.4 +5.6
67.2 +6.5
66.4 +5.6
63.4
75.3 +11.9
78.8 +15.4
75.3 +11.9
Fusing
51.3
63.7 +12.4
65.5 +14.2
63.7 +12.4
64.8
70.2 +5.4
71.5 +6.6
70.2 +5.4
35.0
59.5 +24.5
69.7 +34.7
59.5 +24.5
Table S8 : Plug-in sensitivity (AUROC, %). The reported rule is compared with the original and pure-rank standardizations on the three columns where fusion can hurt a host.
Luddy School of Informatics, Computing, and Engineering, Indiana University Bloomington, Bloomington, IN, USA · College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar