cs.CVMar 16, 2026

Training-free Detection of Generated Videos via Spatial-Temporal Likelihoods

Authors: Omer Ben Hayun, Roy Betser, Meir Yossef Levi, Levi Kassel, Guy Gilboa

Organizations: Technion – Israel Institute of Technology, Haifa, Israel

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Following major advances in text and image generation, the video domain has surged, producing highly realistic and controllable sequences. Along with this progress, these models also raise serious concerns about misinformation, making reliable detection of synthetic videos increasingly crucial. Image-based detectors are fundamentally limited because they operate per frame and ignore temporal dynamics, while supervised video detectors generalize poorly to unseen generators, a critical drawback given the rapid emergence of new models. These challenges motivate zero-shot approaches, which avoid synthetic data and instead score content against real-data statistics, enabling training-free, model-agnostic detection. We introduce STALL, a simple, training-free, theoretically justified detector that provides likelihood-based scoring for videos, jointly modeling spatial and temporal evidence within a probabilistic framework. We evaluate STALL on two public benchmarks and introduce ComGenVid, a new benchmark with state-of-the-art generative models. STALL consistently outperforms prior image- and video-based baselines. Code and data are available at https://omerbenhayun.github.io/stall-video.

Figures & tables

Figure 2: Qualitative comparison of ZED, D3, and our method (STALL). Each row shows a video clip with natural or unnatural spatial and temporal behavior, together with the corresponding predictions. ZED (spatial-only) misses in cases dominated by temporal inconsistency; D3 (temporal-only) fails when spatial realism is misleading. STALL fuses spatial and temporal likelihoods, yielding robust detection when either modality alone is insufficient. Additional examples with more details are given in Supp. D.7.
Figure 3: Method overview. A video is split into frames and encoded into embeddings. The spatial branch scores the likelihood of each frame embedding; the temporal branch normalizes inter-frame differences and scores their likelihood analogously. The two scores are then fused into a unified measure that separates AI-generated from real videos.
Figure 4: Correlations among spatial and temporal aggregation methods. Values computed on VATEX [ 82 ] , which is not used in our evaluations. When all individual likelihood detectors perform reasonably well on the evaluated benchmarks (see Supp. Section D.2), lower correlations are desirable.
Figure 5: Temporal embedding coordinates. Raw coordinates of temporal embeddings (frame differences) are not Gaussian; after normalization, each coordinate is approximately Gaussian (full histogram comparison in Supp. B.2.1).
Figure 6: Comparison of detectors and calibration data. We compare our method against six detectors: three image-based [ 68 , 42 , 28 ] and three video-based [ 93 , 1 , 5 ] , for performance and efficiency. (a) Inference latency per video. (b) Average AUC across all three benchmarks; our method is both high-performing and efficient. (c) GenVideo [ 24 ] results using different datasets as the calibration set, showing that same-distribution calibration is only slightly better, indicating robustness to the calibration choice (see Supp. D.6.6).
Figure 7: Robustness to calibration set size and image perturbations. (a) Varying calibration-set size between 1k to 34k; each size is resampled 5 times and we report mean AUC ±\pm standard deviation. (b) Robustness to four common image perturbations applied randomly to frames at five severity levels; our method maintains high separation across perturbation type and intensity. Both experiments are performed on GenVideo [ 24 ] .
Figure 8: Temporal ablations. Our method remains robust under all temporal variations: (a) temporal step size for likelihood computation; (b) video length; (c) frame rate (FPS).
RepresentationTestAvg ScoreThresholdNormal Features (↑)(\uparrow)approximately gaussian?
Raw embeddings xtx_{t}AD0.4750<0.752<0.75296.3%✓
DP0.3931>0.05>0.0599.2%✓
Whitened embeddings yty_{t}AD0.5034<0.752<0.75298.2%✓
DP0.3207>0.05>0.0599.3%✓
Transition vector Δt=xt+1−xt\Delta_{t}=x_{t+1}-x_{t}AD3.3992<0.752<0.7520.0%✗
DP0.0093>0.05>0.050.0%✗
Table 4: Normality tests. Results of Anderson Darling (AD) and D’Agostino Pearson (DP) normality tests for different representations on the VATEX [ 82 ] calibration set. “Avg Score” is the mean test statistic across embedding coordinates and groups, “Threshold” specifies the acceptance condition for normality, and “Normal Features” is the percentage of coordinates satisfying that condition.
Figure 9: Raw vs. Normalized Temporal Difference Histograms. Histogram comparison of raw temporal differences (Raw Δt\Delta_{t} , left) versus normalized temporal differences (Normalized Δt\Delta_{t} , right) for the first four dimensions of DINOv3 [ 71 ] embeddings, computed from all adjacent frame pairs in the VATEX [ 82 ] calibration set. Red curves show Gaussian distributions fitted using the empirical mean and variance of the data (i.e., a moment-matched Gaussian). Even under this moment-matched fit, raw differences exhibit clear deviations from Gaussianity, while normalized differences closely match Gaussian distributions across all dimensions.
Figure 10: Real and AI-generated videos embeddings. Histograms of selected embedding dimensions from DINOv3 [ 71 ] features extracted from randomly sampled frames in the GenVideo [ 24 ] dataset. The left column corresponds to real videos, and the right column corresponds to generated videos. Red curves show Gaussian distributions fitted using the empirical mean and variance of each feature dimension.
Figure 11: Uniformity on the sphere and Gaussian distribution. Histograms of the first coordinate of UdU_{d} , where UdU_{d} is uniformly distributed on the dd -dimensional unit sphere Sd−1\mathbb{S}^{d-1} , shown for several values of dd and overlaid with the N(0,1)\mathcal{N}(0,1) density.
Figure 12: Uniformity on the sphere. Cosine similarity distributions over all unordered pairs of 3k3k randomly sampled normalized temporal differences Δ~t\tilde{\Delta}_{t} from the VATEX [ 82 ] calibration set and 3k3k points drawn uniformly from the unit sphere.
Figure 13: Ablations on temporal derivative order. STALL performance for different temporal derivative orders D∈{1,2,3,4}D\in\{1,2,3,4\} : (a) AUC across orders, (b) Pearson correlation between scores from different orders, and (c) Spearman correlation between scores from different orders.
Figure 14: Components ablation. Bar plots of spatial-only, temporal-only, and combined detectors, using both raw scores and percentile-ranked scores. Spatial and Temporal denote single-branch detectors, and Combined refers to the fused Spatial and Temporal detector. Configurations labeled with “P” use percentile-ranked scores, where “P” stands for percentile. “Avg” stands for average of the two scores and “Prod” stands for the product of the two scores.
Figure 15: Frame-level aggregation ablation. Heatmaps of AP and AUC, for different combinations of spatial and temporal frame-level aggregation operators. Rows correspond to the spatial aggregation and columns to the temporal aggregation, while all other parts of the pipeline follow the default configuration described in the main paper.
Figure 16: Temporal scores under temporal perturbations. Each condition applies a different perturbation to 400 real MSR-VTT videos: Original (no perturbation), Reversed (frames in reverse order), Shuffle (consecutive frames shuffled), Black flash (one black frame inserted mid-video), and White flash (one white frame inserted mid-video). The temporal likelihood is largely unaffected by reversal and shuffling, since adjacent-frame difference statistics are preserved; however, abrupt flash frames cause a strong drop in the score, with white flashes producing a larger degradation than black flashes.
PerturbationImplementationSeverity levels (1 →\rightarrow 5)
Gaussian blurTF.gaussian_blur(r,σ,k)=(r,\sigma,k)=
  {(1,0.5,3),(2,1.0,5),\{(1,0.5,3),(2,1.0,5),
  (3,1.5,7),(5,2.5,11),(3,1.5,7),(5,2.5,11),
  (10,5.0,21)}(10,5.0,21)\}
JPEG compressionPIL.save(...,JPEG quality
 format="JPEG", quality)q∈{80,50,30,10,1}q\in\{80,50,30,10,1\}
Table 10: Image perturbation details. These perturbations are used in the robustness ablation (Sec. D.5). Levels 1–5 span the full range of corruption strength for each type; level 0 corresponds to no perturbation.
Figure 17: Perturbations example. mandrill.
Figure 18: Perturbations example. Sailboat on lake.
Figure 19: Qualitative examples. Each row shows sampled frames from a video clip, with indicators marking whether its spatial and temporal behavior appears natural or unnatural.
Figure 20: Qualitative examples. Each row shows sampled frames from a video clip, with indicators marking whether its spatial and temporal behavior appears natural or unnatural.
DomainMethodMean [sec]Std [sec]
Zero shot imagesAEROBLADE [ 68 ]2.52660.0243
ZED [ 28 ]1.13940.0067
RIGID [ 42 ]0.43630.0024
Supervised videoT2VE [ 1 ]1.99500.0102
AIGVdet [ 5 ]5.42160.0787
Zero-shot videoD3 cos [ 93 ]0.21570.0043
Table 12: Inference time comparison for all methods.
domainmethodmodel loading cpu (MB)model loading gpu (MB)inference cpu peak (MB)inference gpu peak (MB)
zero shot imagesAEROBLADE [ 68 ]7875.93141.029711.182624.43
ZED [ 28 ]101.1116.061707.72310.08
RIGID [ 42 ]142.41327.301373.67567.57
supervised videoT2VE [ 1 ]1852.161271.432795.96140.96
AIGVdet [ 5 ]488.77182.591849.03673.18
zero-shot videoD3 cos [ 93 ]315.641157.721470.08160.30
Table 13: Memory usage comparison for all methods

Explore similar work

CardsList
Sep 22, 2026

TRACE: Trajectory Representation and Consistency Estimation for AI-Generated Video Detection

Recent advances in generative video models have enabled the synthesis of visually realistic content, posing significant challenges to synthetic video detection. Existing detectors often rely on appearance artifacts, semantic inconsistencies, and temporal patterns that may be generator-specific, limitating generalization to unseen synthesis models. We investigate whether responses to a pretrained generative model provide more transferable forensic cues. Our key observation is that real and AI-generated videos exhibit distinct velocity responses under a pretrained Flow Matching video model. This distinction persists when different pretrained video-generation backbones are used as probes, suggesting that velocity responses offer transferable forensic signals beyond visual artificts. Motivated by this observation, we propose TRACE (Trajectory Representation and Consistency Estimation), a generation-process-aware framework for AI-generated video detection. TRACE leverages a pretrained video DiT as a velocity-field probe to extract representations at multiple flow time points, and models cross-frame consistency through velocity differences between adjacent frames. We further introduce a Real-Centered Trajectory Optimization objective that encourages generator-invariant representation learning. Extensive experiments on AIGVDBench demonstrate that TRACE generalizes effectively across diverse generators, substantially outperforming prior state-of-the-art methods on unseen open- and closed-source video generation models.

Huangsen Cao, Hongkang chu, Siyao Yu +3

Zhejiang University · University of the Chinese Academy of Sciences · Nanjing University of Information Science and Technology +1

Deepfake Detection
AskPDF
Feb 19, 2026

EA-Swin: An Embedding-Agnostic Swin Transformer for AI-Generated Video Detection

Recent advances in foundation video generators such as Sora2, Veo3, and other commercial systems have produced highly realistic synthetic videos, exposing the limitations of existing detection methods that rely on shallow embedding trajectories, image-based adaptation, or computationally heavy MLLMs. We propose EA-Swin, an Embedding-Agnostic Swin Transformer that models spatiotemporal dependencies directly on pretrained video embeddings via a factorized windowed attention design, making it compatible with generic ViT-style patch-based encoders. Moreover, we construct the EA-Video dataset, a benchmark dataset comprising 130K videos that integrates newly collected samples with curated existing datasets, covering diverse commercial and open-source generators and including unseen-generator splits for rigorous cross-distribution evaluation. Extensive experiments show that EA-Swin achieves 0.97-0.99 accuracy across major generators, outperforming prior SoTA methods (typically 0.8-0.9) by a margin of 5-20%, while maintaining strong generalization to unseen distributions, establishing a scalable and robust solution for modern AI-generated video detection.

Hung Mai, Loi Dinh, Duc Hai Nguyen +6

N2TP Technology Solution JSC, Hanoi, Vietnam · College of Technology, National Economics University, Hanoi, Vietnam · University of Science, Vietnam National University, Ho Chi Minh City, Vietnam +3

Deepfake DetectionVision Transformer
AskPDF
May 7, 2026

Advancing Reliable Synthetic Video Detection: Insights from the SAFE Challenge

The proliferation of generative video technologies has intensified the need for reliable methods to detect and characterize synthetic media. To address this challenge, we organized the SAFE: Synthetic Video Detection Challenge, co-located with the Authenticity and Provenance in the Age of Generative AI (APAI) Workshopat ICCV 2025. The competition invited participants to develop and evaluate algorithms capable of distinguishing real from synthetic videos under fully blind evaluation conditions with over 600 submissions from 12 teams over a 90 day span. Hosted on the Hugging Face platform, the challenge comprised two primary tasks: (1) detection of synthetic video content generated by diverse state-of-the-art models, and (2) detection of synthetic content following common post-processing operations such as resizing, re-compression, motion blur and others. The challenge data consisted of 13 modern high quality synthetic video models with generated content matched to real videos from 21 diverse and challenge sources, all adding up to 20 hours of 6,000 video samples. This paper describes the challenge design, dataset construction, evaluation methodology, and outcomes, offering insights into the generalization and robustness of contemporary synthetic video detection methods. Our findings highlight measurable progress in cross-generator generalization but also persistent vulnerabilities to post-processing artifacts. https://safe-video-2025.dsri.org

Kirill Trapeznikov, Gabriel Mancino-Ball, Jonathan Li +9

STR · Drexel University · Aptima, Inc. +1

Deepfake Detection
AskPDF