Skip to main content

NVIDIA's new tool spots AI fakes with 92% accuracy

NVIDIA just dropped a new weapon in the fight against AI deepfakes: a synthetic video detection service called SVD. It's now part of NVIDIA's inference micro-platform, and it works by breaking down videos frame by frame to figure out if they're real or generated by artificial intelligence.

Image

How it works: frame by frame

The service chops a video into small, cropped frames and feeds them into a powerful visual model. That model gives each frame a "realism score" based on spatial features. Then it averages all those scores into a single percentage — the higher the number, the more likely the video is genuine.

Even with compressed videos, SVD holds up well. On uncompressed originals, it hits 92% accuracy. Crank up the compression to 50%, and it still stays steady at 82%. That's pretty impressive for a tool that's meant to catch fakes in the wild.

Speed matters

Accuracy is one thing, but speed is another. NVIDIA's service relies on GPU muscle to keep latency low. On a consumer-grade GPU, processing a 1080p HD video takes just 22 milliseconds. On a professional workstation, it's about 30 milliseconds. That's fast enough for real-time detection, which could be a game-changer for platforms trying to flag deepfakes as they appear.

Key Points

  • SVD analyzes each video frame and assigns a realism score.
  • 92% accuracy on uncompressed videos; 82% at 50% compression.
  • Processes 1080p video in 22–30 milliseconds on standard GPUs.
  • Integrated into NVIDIA's inference micro-platform for easy deployment.