On July 21st, NVIDIA unveiled its Synthetic Video Detector (SVD) service, seamlessly incorporated into the NVIDIA Inference Microservices (NIM) ecosystem. This cutting-edge service is adept at swiftly and accurately identifying videos crafted by artificial intelligence, boasting an impressive accuracy rate of up to 92%. The SVD service meticulously examines videos on a frame-by-frame basis, dissecting them into cropped frames of a predetermined resolution. It then leverages Meta's sophisticated DINOv2 and DINOv3 models to process these frames. Subsequently, it computes a percentage score derived from the average frame score, serving as a reliable indicator of the video's authenticity.

The accuracy of the SVD service is influenced by the degree of video compression. Specifically, it achieves a 92% accuracy rate for uncompressed videos, 87% for videos compressed at a 15% rate, and even maintains an 82% accuracy level for videos compressed at a 50% rate. When it comes to processing speed, the SVD service truly shines. It requires a mere 22 milliseconds to process 1080p videos on an Nvidia RTX GPU, and just 30 milliseconds on workstation models, showcasing its remarkable efficiency.