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

1. Introduction

• The paper addresses detection of fully generated videos, distinct from deepfakes that manipulate real content, amid risks including misinformation, fraud, and impersonation.The introduction says realistic, controllable video models produce long, high-fidelity sequences and underscores the urgency of reliable detection.Source: §1 · p. 1
• Supervised image detectors need labeled datasets and generalize poorly to unseen generators, while zero-shot image detectors assess videos frame by frame and miss temporal artifacts.The introduction defines zero-shot as requiring no additional training and no generated content.Source: §1 · pp. 1–2
• Video detectors predominantly use supervision, and the first zero-shot video detector, D3, uses temporal cues alone without spatial information or principled theoretical foundations.The introduction identifies a need for a mathematically grounded detector that jointly analyzes spatial content and temporal dynamics.Source: §1 · p. 2
• STALL combines DINOv3-based image likelihoods with temporal likelihoods for frame transitions, measuring alignment with real-video statistics without training or generated samples.Its preprocessing uses a calibration set of real videos.Source: §1 · p. 2
• STALL achieves state-of-the-art results across three challenging benchmarks, including ComGenVid, curated with videos from recent high-performing generators.The paper also describes the method as lightweight and efficient for real-time or large-scale screening.Source: §1 · p. 2
• Across experiments, STALL remains robust to common image perturbations, FPS variation, and a range of hyperparameter settings.The paper characterizes the method as lightweight and efficient for real-time or large-scale screening pipelines.Source: §1 · p. 2

2. Background and Related work

• Image detection spans supervised CNNs [11] [33] [94], data-efficient variants [26, 27, 61, 70, 91], and zero-shot comparisons [14] [28] [42] [68], yet remains frame-only.These image-only approaches miss anomalies that appear only in motion or inter-frame transitions.Source: §2 · p. 2
• For fully generated videos rather than edited deepfakes, supervised detectors struggle on unseen models [76], while D3 and rely on temporal differences alone.Other work pairs GenVideo with DeMamba [23] and VideoFeedback with VideoScore [40]; MLLM detectors [35] [87] offer rationales but require curated training data and tuning.Source: §2 · p. 2
• STALL jointly scores spatial and temporal likelihoods, unlike spatial-only ZED and temporal-only D3, and is robust when either modality alone is insufficient.The qualitative comparison shows ZED missing temporal inconsistencies and D3 failing when spatial realism is misleading.Source: §2 · p. 2, Figure 2 · p. 3
• Gaussian CLIP likelihoods are motivated by embedding-geometry studies [54] [55], whitening-based closed-form estimates [10], and theoretical Gaussian-projection results and InfoNCE analysis [9].STALL normalizes temporal embeddings to promote Gaussian statistics; the approach extends to other vision encoders [47], while whitening has also proved effective for LLM activations [65].Source: §2 · p. 3

3.1. Whitening transform and Gaussian likelihood

• Whitening centers embeddings and yields zero-mean, identity-covariance representations; the transform is nonunique, with PCA-whitening presented as a common choice.Source: §3.1 · p. 3
• Under a Gaussian model for whitened coordinates, the squared whitened norm provides a closed-form likelihood proxy.Source: §3.1 · p. 3

3.2. Asymptotic Gaussian projections

• Uniform unit-sphere vectors have approximately Gaussian coordinates under the Maxwell–Poincaré lemma [31, 32]; more generally, fixed low-dimensional projections are Gaussian-approximated with near-uniform directions and concentrated norms.Source: §3.2 · pp. 3–4

4. Method: STALL

• STALL jointly scores frame embeddings and inter-frame transitions to produce a video-level detection score.Each frame is mapped to an embedding by a vision encoder; the method models spatial and temporal evidence separately before combining them.Source: §4 · p. 4

4.1. Spatial likelihood

• Under a Gaussian assumption for whitened embeddings, STALL assigns each frame a spatial log-likelihood using statistics calibrated from real videos.The real-video calibration is computed once without training; at inference, frames are encoded, whitened, and scored.Source: §4.1 · p. 4

4.2. Temporal likelihood

• STALL normalizes frame-to-frame differences because raw transitions are not Gaussian, then models their whitened directions with a temporal likelihood.Identical consecutive frames are discarded; if all frames are identical, STALL falls back to the spatial score. Generated videos may have low likelihood under the temporal model.Source: §4.2 · pp. 4–5

4.3. Unified score

• STALL combines the maximal spatial likelihood with the minimal temporal likelihood, converts each to a calibration-relative percentile, and aggregates the components into a video score.The minimum/maximum pairing had the lowest cross-domain correlation; percentile normalization makes the terms scale-free, and the unified score performs best across the reported ablation choices.Source: §4.3 · p. 5

4.4. Calibration set

• STALL calibrates whitening statistics and percentile ranges using real videos only, with a set disjoint from evaluation benchmarks and other data used in the paper.No generated samples are used; real-only calibration anchors the spatial and temporal likelihoods.Source: §4.4 · pp. 5–6

5.1. Experimental settings

• Evaluations cover VideoFeedback [40], GenVideo [23], and ComGenVid, with equal numbers of real and generated videos sampled for fair comparisons.ComGenVid includes generated videos from Veo3 and Sora; AUC and AP are reported, using a separate 33k-video VATEX calibration set [82].Source: §5.1 · pp. 6–7

5.2. Results

• STALL achieves the highest average performance on each benchmark, leads per-generator results in most cases, and avoids the below-0.5 AUC failures seen in other methods.It also outperforms supervised video detectors that were partially trained on evaluated generators.Source: §5.2 · p. 7, Figure 6 · p. 7
• At 0.49s per video, STALL is among the fastest methods, alongside D3 at 0.5s and RIGID at 0.6s.ZED and T2VE take 0.92s and 0.97s per video, respectively.Source: §5.2 · p. 7

5.3. Ablation study

• Calibration from GenVideo yields only slightly higher performance than alternatives, and calibration sets of at least 5k samples avoid the significant drop seen at smaller sizes.The size comparison reports mean AUC and standard deviation over five sampling iterations.Source: §5.3 · pp. 7–8
• Image encoders perform strongly, including MobileNet and ResNet-18, whereas video encoders perform poorly when a single video embedding discards frame-wise and transition statistics.Spatial-only and temporal-only variants each perform well, but their combination performs best; results are robust to score ranking and min, max, or mean aggregation. Higher-order temporal differences yield nearly identical results to first-order transitions.Source: §5.3 · p. 8, Table 2 · p. 8
• STALL remains robust across JPEG compression, Gaussian blur, resized crop, additive noise, and tested temporal settings, including variations in frame rate, video duration, and temporal step size.Image corruptions are applied at inference while the calibration set remains unchanged.Source: Figure 7 · p. 7, §5.3 · p. 8, Figure 8 · p. 8

• Across multiple benchmarks, STALL consistently outperforms prior supervised and zero-shot image and video detectors, including on recent frontier models such as Sora and Veo3.Source: §6 · p. 8
• STALL fuses per-frame spatial and inter-frame temporal likelihoods in one zero-shot probabilistic framework, using only real videos to define reference distributions.The method uses no generated samples and requires no training.Source: §6 · p. 8
• STALL is efficient and robust to spatial and temporal perturbations, calibration-data size and source, and aggregation choices.Source: §6 · p. 8
• Although there remains room for improvement, the results highlight modeling real videos’ statistical structure as a promising path for robust detection.Source: §6 · p. 8

A.1. Detailed algorithms

• The supplement provides full algorithms for whitening, scoring, calibration, and inference to support reproducibility.It also expands statistical tests, dataset descriptions, experimental details, and efficiency analysis; code, data, and precomputed whitening parameters are publicly available.Source: §A.1 · p. 13, Abstract · p. 13

A.1.1. Notation

• STALL represents each frame with an image embedding, models temporal change using consecutive-embedding differences, and maintains separate spatial and temporal whitening parameters and scores.The notation defines normalized temporal differences and separate means and whitening matrices for spatial and temporal embeddings.Source: §A.1.1 · p. 13

A.1.2. Algorithms

• Calibration whitens sampled frame embeddings and normalized temporal differences; inference averages their percentile ranks from separate calibration-score distributions.Inference computes spatial and temporal scores and combines their calibration percentiles into the final video score.Source: §A.1.2 · p. 15

A.2. Implementation details

• Implementation uses DINOv3 as STALL’s encoder and relies on official baseline implementations when available, with specified substitutes for some baselines.For example, the authors implement RIGID with DINO and evaluate ZED using the SReC-based lossless-compression setup and its Δ01 criterion.Source: §A.2 · pp. 15–16

A.3.1. Video Preprocessing

• Videos shorter than 2 seconds or below 8 FPS are filtered; retained videos are sampled at 8 FPS and, unless stated otherwise, truncated to 2 seconds, yielding 16 frames.For non-integer original-to-target frame-rate ratios, sampling uses approximately uniform temporal spacing.Source: §A.3.1 · p. 16

A.3.2. Pairwise Comparison Protocol

• The pairwise comparison protocol balances real and generated videos within each generative-model comparison to address class imbalance.The supplement’s reproducibility material also outlines its sections on statistical tests, datasets, experimental details, and efficiency analysis.Source: Abstract · p. 13, §A.3.2 · p. 16

A.4. D3 Baseline Evaluation

• D3’s reported performance gap reflects two evaluation-protocol differences, including frame duplication that creates temporal redundancy in real videos.GenVideo’s MSR-VTT real videos are stored at 3 FPS and D3 duplicates frames to reach 8 FPS, while generated videos are already at 8 FPS or higher; the authors instead use high-frame-rate real videos and uniformly downsample them to 8 FPS.Source: §A.4 · p. 16, §A.4.2 · p. 17

A.4.1. Unbalanced test set

• When GenVideo has fewer than 1,000 synthetic videos, comparing them with 1,000 real videos can inflate AP; pairwise evaluation matches real and generated video counts.For each generator, the protocol samples an equal number of authentic videos; Videofeedback’s real videos are drawn equally from two datasets.Source: §A.3.2 · p. 16, §A.4.1 · p. 17

A.4.3. D3 Ablation Results

• Each D3 protocol difference inflates mean AP on GenVideo, and their combination explains the gap from D3’s reported results; X-CLIP16 also gives a small increase.The settings vary embedder, test-set balancing, and sampling.Source: Table 3 · p. 17, §A.4.3 · p. 17

B.1. Normality and uniformity tests

• The embeddings’ Gaussianity is assessed coordinate by coordinate with independent one-dimensional Anderson–Darling and D’Agostino–Pearson tests.The tests are applied to each embedding coordinate independently.Source: §B.1 · p. 18, §B.1.1 · p. 18, §B.1.2 · p. 18

B.1.1. Anderson-Darling normality test

• Anderson–Darling refines the Kolmogorov–Smirnov goodness-of-fit test by weighting discrepancies in distribution tails more heavily.It compares an ordered sample with a reference normal CDF whose parameters are estimated from the data.Source: §B.1.1 · p. 18
• The study treats an Anderson–Darling statistic A2 < 0.752 as evidence to accept normality and compares test values with tabulated critical values.The test is run with SciPy’s stats.anderson function.Source: §B.1.1 · p. 18

B.1.2. D’Agostino-Pearson Test

• D’Agostino–Pearson combines sample skewness and kurtosis, transforms them into approximately standard-normal statistics, and uses a statistic that follows approximately chi-square with two degrees of freedom under normality.The test is based on a univariate sample’s central moments.Source: §B.1.2 · p. 18
• For D’Agostino–Pearson, p > 0.05 is treated as compatible with Gaussianity; positive or negative skewness indicates right or left skew, while kurtosis indicates heavy or light tails.The test is run with SciPy’s stats.normaltest function.Source: §B.1.2 · p. 18

B.1.3. Results

• Raw frame embeddings and whitened frames show high normal-feature pass rates, raw temporal differences nearly none, while normalized and whitened differences pass nearly all normality thresholds.Tests use 40 groups of 250 VATEX embeddings and evaluate coordinate-wise Anderson-Darling and D’Agostino-Pearson normality.Source: §B.1.3 · p. 18, Table 4 · p. 19

C. Datasets

• The calibration normality analyses use VATEX, while the real-versus-generated frame-embedding histograms use GenVideo [23].VATEX is also the source of adjacent-frame pairs for the normalized temporal-difference histograms.Source: §B.1.3 · p. 18, §B.2.2 · p. 19, §B.2.1 · p. 19

B.2.1. Raw vs. normalized temporal differences

• ℓ2 normalization makes temporal-difference histograms closely match Gaussian fits, unlike raw differences, which deviate substantially.Figure 9 compares the first four DINOv3 dimensions over adjacent frame pairs in VATEX.Source: §B.2.1 · p. 19, Figure 9 · p. 21

B.2.2. Real and AI-generated videos embeddings

• Real and generated video frame-embedding dimensions are approximately Gaussian, although their means and variances differ slightly.The histograms use randomly sampled frames from GenVideo [23].Source: §B.2.2 · p. 19, Figure 10 · p. 22

B.3. Maxwell Poincare Lemma

• The Maxwell–Poincaré results establish that fixed-dimensional coordinates of a uniform unit-sphere vector converge toward a Gaussian as dimension increases.The appendix also notes that the results extend to arbitrary coordinate selections and orthonormal projections.Source: §B.3 · p. 19
• With norm concentration, the sphere-to-Gaussian result extends to vectors with concentrated radii by combining radial concentration and the Maxwell–Poincaré limit via Slutsky’s theorem [77].Figure 11 illustrates the first-coordinate distribution approaching standard normal as sphere dimension increases.Source: §B.3 · p. 20, Figure 11 · p. 23
• Figure 12 compares cosine-similarity distributions for unordered pairs of normalized VATEX temporal differences and uniformly sampled points on the unit sphere.The passage specifies the compared distributions but does not report a numerical comparison or which one better fits the other.Source: §B.3 · p. 20, Figure 12 · p. 23

C.1. ComGenVid benchmark

• 1.7k MSVD videos of 2 seconds each were matched by 1.7k videos each from Sora and Veo3 for ComGenVid.Veo3 videos were randomly selected from ShareVeo3, while Sora videos were manually collected from distinct users on the public explore feed.Source: §C.1 · p. 24

C.2. GenVideo

• GenVideo [23] AI videos came from ModelScope, while original MSR-VTT videos were downloaded to obtain higher-frame-rate versions.Videos were sampled at 8 FPS for 2 seconds (16 frames); HotShot-XL and MoonValley were compared with real 1-second videos at 8 FPS, excluded from ablations, and 1.4k MSR-VTT videos were selected for the pairwise protocol.Source: §C.2 · p. 24

C.3. VideoFeedback

• VideoFeedback [40] evaluation retained only clips with dynamic-degree scores of 3–4 and required at least 2 seconds, except for HotShot-XL.HotShot-XL’s 1-second clips were compared with real 1-second videos at 8 FPS (8 frames).Source: §C.3 · pp. 24–25

C.4. VATEX (Calibration set)

• VATEX was downloaded in three parts for the calibration set.The paper points to Table 8 for the calibration set’s full statistics.Source: §C.4 · p. 25, Table 8 · p. 25

C.5. Other datasets

• Approximately 1.5k real videos from Kinetics400 [50] and PE [13] formed an alternative calibration set used exclusively for the calibration-set ablation in Fig. 6(c).Kinetics400 clips came from the test split, and PE videos were downloaded from the Facebook PE Hugging Face page.Source: §C.5 · pp. 25–26

D. Additional results and experimental details

• Unless otherwise stated, ablations used GenVideo [23] generative videos sampled at 8 FPS for 2 seconds, with DINOv3 as frame-level embedder and VATEX for calibration.HotShot-XL and MoonValley, which produce clips shorter than 2 seconds, were omitted from ablations.Source: D. Additional results and experimental… · p. 26

D.1. Derivative order ablations

• Across derivative orders D ∈{1, 2, 3, 4}, STALL scores are strongly correlated and performance is very similar.The ablation changes only temporal derivative order, fits a separate whitening transform for each order, and averages temporal and spatial likelihood percentiles.Source: §D.1 · p. 26, Figure 13 · p. 27

D.2.1. Components ablation

• The components ablation compares spatial-only, temporal-only, and combined detectors using raw or percentile-ranked scores, and mean- or product-based fusion for the combined model.Average AP and AUC are reported across all three benchmarks for each configuration.Source: §D.2.1 · p. 27, Figure 14 · p. 27

D.2.2. Aggregation ablation

• The aggregation ablation sweeps min, mean, and max frame-level operators for both branches and reports average AUC and AP across three benchmarks as heatmaps.Each heatmap compares a spatial and temporal aggregation pair while other pipeline components remain at their default settings.Source: §D.2.2 · p. 27, Figure 15 · p. 28

D.3. Sampling a single frame difference for temporal whitening

• Using one sampled frame difference instead of all transitions for temporal whitening yields average AUC 0.8105 vs. 0.8110 and average AP 0.8044 vs. 0.8046, with Pearson 0.9994 and Spearman 0.9992 correlations.The spatial branch and its whitening transform are held fixed while temporal whitening is recalibrated on the VATEX calibration set.Source: §D.3 · pp. 27–28

D.4. Temporal likelihood under temporal perturbations

• On 400 real MSR-VTT videos, reversal and shuffling leave temporal likelihood largely unaffected, whereas flash frames sharply reduce it, especially white flashes.The authors attribute stability under reversal and shuffling to preservation of adjacent-frame difference statistics; abrupt flashes introduce temporal inconsistency.Source: §D.4 · p. 28, Figure 16 · p. 28

D.5. Image perturbation experiment details

• On a reduced GenVideo [23] subset, STALL maintains strong separation across four image-corruption types and five severity levels, with level 0 representing unperturbed frames.The experiment applies Gaussian blur, JPEG compression, random resized crops, and Gaussian noise at inference while keeping calibration fixed.Source: §D.5 · pp. 28–29, Table 10 · p. 29

D.6.1. Step size ablation

• The step-size ablation subsamples videos at 8, 4, 2.67, and 2 FPS by applying steps s ∈{1, 2, 3, 4} to both VATEX calibration and GenVideo test videos.The combined spatial-temporal score is evaluated on these sequences; average AUC results are shown in Fig. 8(a) of the main paper.Source: §D.6.1 · p. 29

D.6.2. FPS ablation

• Performance remains essentially unchanged from 2 to 24 FPS when inference videos are subsampled and whitening stays fixed at 8 FPS.The downsampling retains every n-th frame; the passage reports robustness across the tested frame rates but gives no numerical scores.Source: §D.6.2 · p. 29

D.6.3. Length of video ablation

• STALL remains robust on clips truncated to 1, 2, 3, or 4 seconds with whitening fixed at 2 seconds.The passage reports no numerical scores for the duration comparison.Source: §D.6.3 · p. 30

D.6.4. Backbone encoders ablation

• The backbone ablation evaluates five pretrained encoders: DINOv3, MobileNetV3, ResNet-18, VideoMAE, and ViCLIP.For image encoders, features are extracted per frame and applied independently; results are presented in Table 2, whose values are not supplied here.Source: §D.6.4 · p. 30

D.6.5. Calibration set size

• The calibration-size ablation tests VATEX subsets from 1,000 videos through the full 33,976-video dataset, using five random seeds per size.Each subset estimates the whitening parameters before evaluation on GenVideo, with other pipeline components fixed; numerical results are not supplied.Source: §D.6.5 · p. 31

D.6.6. Calibration set sources

• Calibration choices show stable average AUC across all three benchmarks, but performance may degrade when the test domain is absent from calibration.Choosing an appropriate calibration set can adapt the detector to a target domain; stable behavior is reported within the general video regime studied. Table 11 reports average AUC by calibration source, but the supplied passage gives no values.Source: §D.6.6 · pp. 31–32, Table 11 · p. 32

D.7. Additional qualitative results

• The qualitative examples mark whether sampled clips appear natural or unnatural in spatial and temporal behavior.The supplied captions describe the indicators but do not report the examples’ outcomes.Source: Figure 19 · p. 32, Figure 20 · p. 32

E. Efficiency analysis

• The efficiency analysis measures detection methods’ inference time and memory usage under controlled conditions.Source: E. Efficiency analysis · p. 33

E.1. Infrence time analysis

• Inference timing uses 20 videos, five repetitions per method, and includes video loading for end-to-end measurement.Methods are initialized before timing; the complete comparison is in Table 12, but its values are not supplied.Source: §E.1 · p. 33, Table 12 · p. 33

E.2. memory analysis

• Memory profiling measures CPU and GPU use during model loading and video inference, repeating measurements across 10 videos with 3 repetitions per method in isolated processes.CPU memory was tracked with psutil and GPU peaks with PyTorch CUDA tracking; garbage collection and CUDA cache clearing were used between measurements.Source: §E.2 · p. 33
• STALL and D3 maintain moderate, balanced CPU and GPU use across loading and inference, with significantly fewer resource requirements than supervised alternatives.Their efficiency is comparable to an existing temporal-consistency zero-shot method.Source: Table 13 · p. 33, §E.2 · p. 34
• AEROBLADE [68] has the highest memory footprint among zero-shot image methods, while ZED [28] is the most memory-efficient.Source: §E.2 · p. 34
• T2VE uses more GPU memory than AIGVdet during loading, but AIGVdet consumes more during inference.Source: §E.2 · p. 34
• All methods receive video data through CPU memory, contributing to their inference-time CPU peaks.Source: §E.2 · p. 34