Can a world model predict a deepfake video? A Consistency Margin Prior for Deepfake Detection

Deepfake detectors often perform well on the dataset they are trained on, but their performance drops when tested on videos from different datasets. In this paper, we study whether a predictive video model can provide a signal that generalizes better across datasets. Instead of relying only on generator-specific artifacts, as many existing detectors do, we use a model trained to predict natural video and measure the difference between what it predicts and what it observes in a video. Using V-JEPA 2.1, we keep the encoder frozen and adapt the predictor using rank-16 LoRA. We define the difference between the predictor output and the frozen encoder target as the prediction residual. We hypothesize that manipulated clips produce, on average, larger prediction residuals than authentic clips, and use this as our consistency-margin prior. We process these residuals using a lightweight Latent Consistency Head (LCH), which separates residual magnitude from direction for training and classification. To test whether the direction of the consistency margin affects cross-dataset generalization, we compare standard and reversed margin objectives while keeping the remaining training setup fixed. We match the two settings on FF++ validation performance before evaluating them on unseen data. The reversed-margin setting exhibits a larger cross dataset generalization gap on average, providing evidence that the direction of the consistency margin affects cross-dataset transfer. Our complete model adds only $8.25$M trainable parameters, which is 2.47% of the total $334$M parameters. Trained only on FF++, it achieves an AUC of $90.5$ on Celeb-DFv2 and performs well across several other cross-dataset benchmarks. Our ablations show that separating residual magnitude from direction, reducing dataset-specific variation, and using a low-rank projection improve cross-dataset generalization in our experiments.