SPLIT: Training-Free AI-Generated and Partially Edited Video Detection via Spatial Patch-Level Incoherence and Temporal Roughness

Hyun, Jongyeop; Kim, Hyounghun

ECCV 2026

SPLIT: Training-Free AI-Generated and Partially Edited Video Detection via Spatial Patch-Level Incoherence and Temporal Roughness

Jongyeop Hyun¹, Hyounghun Kim^1,2

¹Graduate School of Artificial Intelligence, POSTECH
²Department of Computer Science and Engineering, POSTECH

Paper arXiv Code BibTeX

Three radar charts, one per benchmark, plotting Fake Recall at FPR = 0.1 percent for FID, AIGVDet, ReStraV, D3, and SPLIT across FakeParts categories, GenVideo generators, and ViF-Bench generators. SPLIT's outline encloses the others on nearly every axis. — **Fake Recall (%) at FPR = 0.1%** for FID, AIGVDet, ReStraV, D3, and SPLIT across FakeParts, GenVideo, and ViF-Bench. SPLIT (red) achieves the highest and most uniform recall across nearly every manipulation type and generator — without any training.

Abstract

Deploying AI-generated video detectors in real-world services demands an ultra-low false positive rate (FPR) on real videos to avoid falsely rejecting authentic content, a regime where standard metrics such as AUROC fail to reflect actual operating behavior. We introduce Spatial Patch-Level Incoherence and Temporal Roughness (SPLIT), a training-free detector that operates on patch tokens from a frozen vision encoder to detect both fully generated and partially edited videos. SPLIT computes two complementary signals: Two-step Temporal Roughness (TTR), capturing non-smooth patch trajectories via one-step and two-step feature variation contrast, and Local Spatial Motion Incoherence (LSMI), measuring spatially inconsistent temporal changes through gradients of a feature-space motion field. The two are fused multiplicatively with gamma correction to sharpen real–fake separation at strict thresholds. We further propose a service-aligned evaluation protocol based on Fake Recall at fixed FPR with real-only threshold calibration and cross-real threshold transfer. Across three benchmarks — FakeParts, GenVideo, and ViF-Bench — SPLIT achieves the highest Fake Recall at FPR = 0.1%, substantially outperforming supervised and training-free baselines while remaining robust to post-processing with negligible overhead.

At a glance

Fake Recall at FPR = 0.1% — the deployment-relevant operating point, where the threshold is set to reject roughly 1 authentic video in 1,000 on the calibration set.

74.5%

FakeParts

partial edits
best baseline 35.7%

85.2%

GenVideo

10 generators
best baseline 45.5%

84.5%

ViF-Bench

2024–25 models
best baseline 33.9%

Training-free. Two closed-form statistics on frozen encoder tokens — zero learned parameters, no generator labels.
Patch-level. Scoring every patch instead of a pooled clip embedding keeps localized edits from washing out.
Real-only calibration. Thresholds are set from authentic video alone, and stay near target when the real domain shifts.
Negligible overhead. Matches D3's runtime on the same backbone; the encoder forward pass dominates.

Method

A frozen vision encoder turns a clip into patch tokens. Two statistics read them — one along time, one across space — and multiply into a single score.

Temporal

Two-step Temporal Roughness

\mathrm{TTR}(n) = \log_2\!\left(L_1(n) + \epsilon\right) - \log_2\!\left(L_2(n) + \epsilon\right)

Each patch traces a trajectory in feature space. $L_1$ accumulates frame-to-frame displacements; $L_2$ is a length-normalized two-step chord. Constant-velocity motion makes them equal and smooth motion keeps them close — flicker and local misalignment push $L_1$ above $L_2$ .

Spatial

Local Spatial Motion Incoherence

\begin{aligned} \mathbf{LSMI} = \tfrac{1}{2}\big( \; & \mathbb{E}_{t,i,j}\big[\|\nabla_x M_t(i,j)\|_2\big] \\ + \; & \mathbb{E}_{t,i,j}\big[\|\nabla_y M_t(i,j)\|_2\big] \big) \end{aligned}

Real motion is spatially coordinated: neighbors change together. LSMI takes spatial gradients of the feature-space motion field $M_t = P_{t+1} - P_t$ . Edited regions move unlike their neighbors, raising it.

s(x) = \mathbf{TTR}^{\gamma} \cdot \mathbf{LSMI}, \qquad \gamma = 8

Multiplicative fusion with a fixed sharpening exponent — chosen once on a held-out split, never learned. A clip is flagged when $s(x) \ge \tau_\alpha$ , where $\tau_\alpha$ is calibrated on real videos only to hit a target FPR $\alpha$ .

Service-aligned evaluation

AUROC averages over regions of the ROC curve no deployed service operates in. If false rejections must be rare, only the far-left tail matters.

1

Fake Recall at fixed FPR

Recall on manipulated video at $\alpha \in \{0.1\%, 1\%, 5\%\}$ , instead of a curve-wide average.

2

Real-only calibration

$\tau_\alpha$ is set on real video alone — no access to the generators a deployed detector has yet to meet.

3

Cross-real transfer

Calibrate on ROVI and on MSR-VTT separately, then report the harmonic mean of both Fake Recalls.

Results

Fake Recall at FPR = 0.1% under cross-real threshold transfer, against the strongest of the eleven compared detectors. Full tables for all methods and operating points are in the paper.

Overall, at FPR = 0.1%

Overall Fake Recall at FPR 0.1% by method on FakeParts, GenVideo and ViF-Bench.
Method	Type	FakeParts	GenVideo	ViF-Bench
FID	sup.	4.71	34.03	5.22
FTCN	sup.	5.44	32.80	15.46
AIGVDet	sup.	9.13	45.38	25.69
ReStraV	sup.	35.67	45.52	33.93
D3	free	2.28	5.47	0.97
SPLIT	free	74.45	85.16	84.52

sup. = supervised, trained on GenVideo-100K; free = training-free. Best baseline per benchmark underlined. SPLIT roughly doubles the strongest baseline on FakeParts and ViF-Bench while learning nothing.

SPLIT as the threshold relaxes

SPLIT Fake Recall at FPR 0.1%, 1% and 5% on each benchmark.
Benchmark	0.1%	1%	5%
FakeParts	74.45	83.58	89.70
GenVideo	85.16	93.09	97.35
ViF-Bench	84.52	95.08	98.46

Recall climbs steadily rather than only just clearing the cut — the mark of a score distribution that genuinely separates real from generated content.

Which edits are hard

FakeParts by manipulation category, FPR = 0.1%.

Fake Recall at FPR 0.1% by method across the eight FakeParts manipulation categories and the overall average.
Method	Extrap.	Faceswap	Inpaint.	Interp.	Outpaint.	Style	T2V	TI2V	Overall
AIGVDet	0.42	0.89	0.98	0.43	0.56	3.69	31.65	34.39	9.13
ReStraV	23.27	12.98	19.09	0.18	91.45	18.60	46.65	73.13	35.67
D3	0.00	0.04	0.00	0.00	2.72	0.34	14.46	0.66	2.28
SPLIT	97.86	60.09	2.48	67.81	96.52	98.80	76.74	95.27	74.45

Extrapolation, Outpainting, Style Change and TI2V are near-saturated even at this threshold. Inpainting is the one category where SPLIT is beaten — ReStraV leads with 19.09% — reflecting how little of each frame an inpainting edit touches. On GenVideo, SPLIT clears 90% on six of the ten generators; on ViF-Bench it ranks first on all nineteen, spanning 61.58–98.98%.

What drives it

Three levers — patch-level granularity, the spatial signal, and the gamma exponent — plus the two properties a deployment cares about: cost and calibration stability.

Component ablation

Fake Recall on FakeParts Overall for each combination of patch-level tokens, TTR and LSMI.
Patch	TTR	LSMI	0.1%	1%	5%
–	✓	–	54.21	76.72	88.77
✓	✓	–	68.29	82.09	89.37
✓	–	✓	8.04	14.63	25.07
✓	✓	✓	74.45	83.58	89.70

Swapping global [CLS] embeddings for patch tokens is the largest single jump, +14.08 at FPR = 0.1%. LSMI alone is weak, but adds a further +6.16 on top of patch-level TTR. FakeParts Overall, cross-real threshold transfer.

Gamma exponent

Gains rise steadily from $\gamma = 1$ to $8$ , then flatten; $\gamma = 16$ can regress slightly. $\gamma = 8$ is used for every benchmark and every FPR target.

Cost

Left: bar chart of FakeParts Overall Fake Recall at FPR 0.1 percent. Right: end-to-end inference time on 1000 FakeParts T2V samples on a log scale, where SPLIT matches D3.

Recall at FPR = 0.1% against end-to-end time on 1,000 FakeParts T2V samples (log scale). Encoding dominates, so TTR and LSMI add almost nothing — SPLIT matches D3's runtime while taking recall from 2.28% to 74.45%.

Calibration stability

Actual false positive rate on a held-out real domain when the threshold was calibrated on a different real domain, per method and direction.
Method	MSR-VTT → ROVI	ROVI → MSR-VTT
AIGVDet	13.22	0.00
ReStraV	0.00	15.30
D3	0.00	1.18
SPLIT	0.03	0.56

Actual rejection rate (%) on a held-out real domain when the threshold was calibrated for a 0.1% target on a different one — closer to 0.1 is better. SPLIT stays near target both ways, while ReStraV rejects nothing in one direction and 15.30% of real videos in the other.

Post-processing

Fake Recall at FPR 0.1% on FakeParts Overall under blur, JPEG compression and axis flips.
Method	Clean	Blur σ2	JPEG 80	Flip y
ReStraV	35.67	54.73	24.42	38.50
D3	2.28	1.89	2.78	2.09
SPLIT	74.45	55.42	64.40	66.92

FPR = 0.1%, the harshest setting of each perturbation. Blur costs the most yet still leads; JPEG is gentler, so the signal does not rest on fragile high-frequency artifacts. At FPR = 5% every condition stays within 87.12–89.70%.

Vision encoder

Fake Recall at FPR 0.1% for SPLIT versus D3 across representative frozen vision encoders.
Encoder	FakeParts		GenVideo		ViF-B.
Encoder	D3	SPLIT	D3	SPLIT	D3	SPLIT
XCLIP-B/16	2.28	74.45	5.47	85.16	0.97	84.52
CLIP-B/16	1.66	71.47	2.37	85.87	0.74	85.29
DINOv2-B	4.00	47.77	16.32	68.03	3.97	56.64
ResNet18	1.93	61.23	1.68	76.10	0.94	73.53
MobileNet-v3	1.77	49.81	1.60	67.15	1.02	61.07

FPR = 0.1%. SPLIT beats D3 with every backbone tested — five of ten shown — transformer and CNN alike. Finer patch grids help, but even MobileNet-v3 lifts FakeParts recall from 1.77% to 49.81%: the gain comes from patch-level scoring, not a particular representation.

BibTeX

@misc{hyun2026splittrainingfreeaigeneratedpartially,
      title={SPLIT: Training-Free AI-Generated and Partially Edited Video Detection via Spatial Patch-Level Incoherence and Temporal Roughness},
      author={Jongyeop Hyun and Hyounghun Kim},
      year={2026},
      eprint={2607.02886},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.02886},
}

Supported by IITP grant funded by the Korea government (MSIT) (No. RS-2019-II191906, Artificial Intelligence Graduate School Program, POSTECH).