Skip to main content
Back to timeline
arXivSource publication:

Trained on only 4 generators, SCRiD lifts average AI-video detection ACC to 72.72% across 88 unseen generator or task subsets

Synopsis

Generalizable AI-generated video detection is hard because real footage already produces strong residuals through fast motion, camera movement and lighting changes, so the size of a discrepancy alone cannot indicate synthesis. SCRiD compares inter-frame residuals at motion-corresponding locations and at nearby original image coordinates, learns a content-conditioned denoising reference from authentic videos to adapt features, and exchanges the two evidence streams through reversible symplectic updates. Trained only on synthetic videos from 4 generators, it reaches an average ACC of 72.72% and AP of 74.23% across 6 benchmarks and 88 generator or task subsets, 3.24 and 0.81 percentage points above the strongest baseline.

Source-provided article image: SCRiD: Spatial and Contextual References for Generalizable AI-Generated Video Detection
arXiv

Interpretation

SCRiD replaces the question of how large a discrepancy is with the question of which reference makes it appear, so the detector can still decide on generators it never saw in training. Earlier frame-level and video-level methods rely on synthesis artifacts, inter-frame differences or motion patterns, but natural motion and content variation in authentic videos trigger the same cues, making synthesis-related discrepancies harder to separate on unseen generators. Training uses only synthetic videos from 4 generators in GenVidBench Pair1; across 6 benchmarks covering 88 generator or task subsets, SCRiD reaches an average ACC of 72.72% and AP of 74.23%, 3.24 and 0.81 percentage points above the strongest baseline, with a two-sided paired t-test over five runs reporting significance.

The two evidence streams answer where residuals should be compared and whether a change agrees with its surrounding content. Matched-Coordinate Residual Response (MCR) samples the next frame's residual along estimated optical flow and also near the original image coordinates, giving both references identical bilinear sampling weights before comparing correlation and distance responses. Real-Conditioned Score Adaptation (RCSA) trains a noise predictor conditioned on surrounding content using authentic videos, probes its responses with symmetric perturbations across four quality views and three noise scales, and uses them to adapt DINOv2 features; Reversible Evidence Exchange (REE) couples the two states through symplectic updates that preserve recoverability of the joint representation.

Removing any design choice costs accuracy, indicating the gain comes from the reference conditions rather than added capacity alone. Replacing matched-coordinate responses with conventional flow-aligned residuals lowers ACC and AP by 3.08 and 1.87 percentage points; removing content modulation lowers them by 6.91 and 4.12 points. Freezing the randomly initialized noise predictor lowers ACC and AP by 4.06 and 1.62 points; allowing classification gradients to update it lowers AP by 4.37 points; replacing symplectic updates with a parameter-matched residual MLP lowers ACC and AP by 1.72 and 2.32 points.

Perspective

Evaluation covers 6 benchmarks and 88 generator or task subsets: GenVidBench Pair2, AIGVDBench, ComGenVid, GenBuster, ViF-Bench and RA-Bench, with training restricted to synthetic videos from 4 generators in GenVidBench Pair1 and video-level ACC and AP as metrics. Robustness tests apply 7 post-processing perturbations on ViF-Bench, ComGenVid and GenBuster, where SCRiD averages 74.24% ACC, a mean drop of 0.76 percentage points from clean videos.

On AIGVDBench, DeMamba has higher ACC than SCRiD, while SCRiD leads by 3.07 percentage points in AP on that benchmark, so the two metrics do not always agree. Hyperparameter analysis shows ACC declining as the coordinate offset radius grows from 1 to 4, and a nonmonotonic trend in the number of REE exchange layers; the source reports one layer setting reaching the highest ACC and AP of 73.09% and 75.64%, while the default uses fewer layers and remains competitive.