In a new research paper, scientists tackle the growing challenge of detecting deepfake videos, which are increasingly realistic and generated by rapidly evolving AI methods. The team introduces FaceVid-Forensics-100K, a comprehensive benchmark comprising 100,000 videos that span 33 synthesis techniques, including recent generators like Seedance 2.0. This dataset is distinguished by its fine-grained textual annotations that describe visual observations and provide forensic explanations, created through a multi-model aggregation and conflict-resolution pipeline.
To leverage this benchmark, the researchers propose ARGUS, a multi-agent forensic reasoning framework. ARGUS employs four specialized domain-expert agents that independently analyze forgery cues from texture, lighting, motion, and physics perspectives. A central judge agent then reconciles these reports to deliver a final prediction along with an explanation, enabling a more holistic and evidence-based assessment.
Extensive evaluations on out-of-domain test sets, which include 20 held-out modern video generators, show that ARGUS consistently outperforms other methods, including closed-source models like GPT-4o and Gemini 3.5-Flash. Despite being composed entirely of small open-source MLLMs, ARGUS achieves top-ranking accuracy (69.87%), recall (81.82%), and F1 score (53.28%), demonstrating its strong generalization to emerging synthetic video techniques.
The project page, code, models, and datasets are publicly available, offering resources for further research and development in deepfake detection.