Framework identifies which AI model generated videos

SAGA (Source Attribution of Generative AI videos) represents the first comprehensive framework designed to identify which specific generative AI model produced a given synthetic video. Rather than stopping at detection—determining whether a video is real or fake—SAGA provides forensic attribution at five distinct levels of granularity. It identifies authenticity (real or synthetic), the generation task used (such as text-to-video conversion or image-to-video synthesis), the model version, the development team behind it, and the precise generator itself. This multi-level approach offers far richer insights than existing binary real/fake detectors.

The framework's core innovation is a novel video transformer architecture that effectively captures spatio-temporal artifacts—the subtle fingerprints left by different video generators across both spatial and temporal dimensions. To enable practical deployment, the researchers introduced a data-efficient pretrain-and-attribute strategy that achieves remarkable efficiency: SAGA attains state-of-the-art attribution performance using only 0.5% of source-labeled data per class, matching the accuracy of systems trained on fully labeled datasets. This efficiency is critical for scaling across many generator models and versions that emerge over time.

For interpretability, the framework proposes Temporal Attention Signatures (T-Sigs), a novel visualization method that reveals the learned temporal differences distinguishing one video generator from another. This provides the first transparent explanation for why different generators are forensically distinguishable, addressing a key limitation of earlier black-box approaches. Extensive experiments conducted on public datasets, including cross-domain scenarios where training and test data come from different sources, demonstrate that SAGA establishes a new benchmark for synthetic video provenance, delivering the interpretable, fine-grained attribution insights necessary for forensic investigation and regulatory oversight in an era of hyper-realistic AI-generated media.

The rise of hyper-realistic synthetic videos created by generative AI has outpaced the ability of existing detection tools to handle the threat. Binary real/fake detectors, which simply classify whether a video is authentic or synthetic, lack the granularity needed for forensic investigation and regulatory enforcement. SAGA addresses this gap by going beyond detection to identify the source—not just that a video is fake, but which specific model, version, and development team generated it. This shift from binary classification to multi-level attribution reflects a maturation in how the field thinks about synthetic media accountability.

The framework's technical innovations center on two key advances. First, a novel video transformer architecture leverages features from a robust vision foundation model to capture spatio-temporal artifacts—the telltale traces left by different generators across both space and time. Second, the data-efficient pretrain-and-attribute strategy allows SAGA to match the performance of fully supervised systems while using only 0.5% of the labeled data per class, a significant practical advantage for scaling deployment across many generator models. The introduction of Temporal Attention Signatures provides interpretability by visualizing the learned temporal differences that distinguish one generator from another, offering the first transparent explanation for why the framework's attributions are possible.