Nvidia's deepfake detector reaches up to 94% accuracy

As time passes, fake news is becoming harder to spot, with AI-generated videos making it nearly impossible to distinguish real incidents from fabricated ones. Nvidia is now stepping into the mainstream with a tool designed specifically for media companies to fact-check the information they broadcast.

Nvidia unveiled its new Synthetic Video Detector at SIGGRAPH 2026, an NIM microservice that scans every frame of a video or purported footage and, in turn, provides a score or prediction indicating whether the video is AI-generated. The purpose of the NIM microservice is to support Nvidia’s expansion into the media industry by enabling editorial teams to quickly deliver accurate news and flag false information before it reaches national television.

Nvidia’s VP of physical AI simulation technology, Rev Lebaredian, explained the rationale for creating the Synthetic Video Detection system in a press release, stating that the same technology used to create generative AI videos can also be used to detect them.

Currently, Nvidia is already distributing the technology to Wowza, an enterprise-level live video and media server solution used by multiple news networks. The CEO of Wowza, Krish Kumar, has already acknowledged the integration of Nvidia’s Synthetic Video Detector into Wowza’s infrastructure. He stated that the company is moving toward integrating AI into its real-time broadcasts and live streams, noting that “the market is moving toward live video systems that can generate value while the stream is still in motion.”

A promising non-linear drop in accuracy with compression in play

When it comes to real-time performance, Nvidia claims that the Synthetic Video Detector can detect AI-generated content with 92% accuracy on uncompressed video. Accuracy drops to 87% when video is compressed by 15%, while it remains at a relatively respectable, non-linear drop to 82% at 50% compression.

A fast solution with testing indicating potentially better accuracy with tweaks

The most interesting and lucrative aspect of the NIM microservice for media networks is its ability to detect generative AI footage at a speed of 22 ms per 1080p video file on RTX systems and 30 ms on L40 GPUs.

However, Nvidia’s own internal testing with the newest model achieved an AUC of 0.9614 and an accuracy of 94%. News networks and other organizations can tweak and configure their thresholds according to their requirements.

Currently, Wowza is already incorporating the NIM service into its Video Intelligence Framework, which reaches 35,000 deployments across nearly 170 countries, meaning real-time video flagging will extend to live-streaming services beyond traditional newsrooms.