Nvidia has unveiled the Synthetic Video Detector, which is designed to identify AI-generated videos with up to 92 percent accuracy.
In a world full of AI-generated videos, how can you still tell what’s real and what isn’t? This is precisely the question Nvidia aims to address with its new “Synthetic Video Detector” tool.
Nvidia unveiled the tool as part of its in-house platform, Nvidia AI for Media, at the SIGGRAPH conference. It is a NIM microservice—where NIM stands for Nvidia Inference Microservices, which are AI model containers that can be easily integrated into existing software.
NVIDIA has introduced Synthetic Video Detector, an AI tool that detects AI-generated videos with up to 92% accuracy in milliseconds. https://t.co/guNLFksWj4
— Digital Trends (@DigitalTrends) July 20, 2026
Nvidia Synthetic Video Detector: How Detection Works
TheSynthetic Video Detectoranalyzes video footage frame by frame and calculates a classification score—a probability estimate of the proportion of synthetic content in the footage.
The detector is not suitable as a replacement for established verification methodsis not suitable as a replacement for established verification methods—according to Nvidia—but rather serves as an additional verification step when a quick assessment is needed to determine whether a clip should be examined more closely.
- According to the developer, the model is designed to remain practical even if a clip has been previously compressed, cropped, resized, or re-encoded—in other words, exactly the kinds of steps that are routine on social media platforms.
- Nvidia estimates the accuracy rate at up to 92 percent for uncompressed material, 87 percent with 15 percent compression, and 82 percent with 50 percent compression.
On the internal test dataset (via wccftech), the current model version is reported to achieve an AUC of 0.9614 and an accuracy of 0.9453, and to lead the AIGVD Bench comparison test with an average AUC of 98.7 percent.
Speed as a Key Selling Point
In terms of speed, Nvidia cites figures in the low double-digit millisecond range: On RTX hardware, a 1080p clip is said to be processed in about 22 milliseconds; on the L40 series server GPUs, the same analysis takes around 30 milliseconds, according to the developers.
To drive wider adoption, Nvidia is partnering with streaming provider Wowza: The Detector is set to be integrated into live-streaming workflows via Wowza’s Video Intelligence Framework, thereby reaching more than 35,000 deployments in over 170 countries. Nvidia also offers the Detector for on-premises use.
