Nvidia’s AI Detector Can Tell Whether a Video Is AI-Generated in Milliseconds
Nvidia has launched a new tool that identifies AI-generated and altered videos in news and media. Called Synthetic Video Detector, the tool analyzes video frame by frame and assigns a probability score indicating how likely a clip is to be artificial.
The detector looks for specific marks in images that betray the use of generative AI. It also seeks out basic details and patterns that are often overlooked by video generators.
Nvidia says the microservice achieves up to 92% accuracy on uncompressed video, with accuracy dropping to 87% at around 15% compression and 82% at about 50% compression. These numbers are based on internal tests from various systems like COSMOS, Sora, Veo, OmniAvatar, OVI, and Nvidia’s own LipSync and Live Portrait tools, so it’d be fair to expect real-life accuracy to be much lower than this.
At a threshold of 0.3, the detector correctly identifies about 96% of the AI-generated videos. The threshold determines how strict the detector is about identifying AI-generated content. Lower numbers mean it’s more likely to mark something as AI-generated, while higher numbers make it more cautious. A balanced threshold of 0.5 gives 85% accuracy on synthetic clips and 82% on real footage, as reported by Wccftech.
The detector can process 1080p video in about 22 milliseconds on RTX GPUs and roughly 30 milliseconds on L40 data‑center hardware, making it possible to run AI checks in near-real time.
