To address the growing challenge of AI deepfake videos, NVIDIA has recently launched a new synthetic video detection service called SVD. This technology is now integrated into NVIDIA's inference micro-platform, and it can accurately identify whether content is generated by artificial intelligence by analyzing each frame of the video.
Frame-by-Frame Scoring to Analyze Image Details
In terms of technical implementation, the service splits the video into specific-sized cropped frames and sends them to an advanced visual model for in-depth analysis. The system assigns a realism score to the spatial features of each image frame, and finally generates an intuitive percentage-based result by calculating the average score of all frames.
Even when dealing with compressed videos, this detection technology still demonstrates strong performance. Test data shows that the accuracy rate reaches as high as 92% when detecting uncompressed original videos, and remains stable at 82% even under a high compression rate of 50%.
Millisecond-Level Response for Efficient Detection
In addition to its high identification accuracy, this technology also offers significant advantages in processing speed. Relying on powerful GPU computing power, the system achieves extremely low latency when processing standard HD videos, truly realizing an efficient real-time detection experience.
According to official estimates, processing a 1080p HD video on a consumer-grade GPU takes only 22 milliseconds, and on a professional workstation device, it takes just 30 milliseconds. This millisecond-level response speed provides a solid technical foundation for the large-scale application of AI video detection in the future.
