While the tech industry pours billions into generative models that spit out ever more convincing video, the flip side — telling real from fake — has been quietly underfunded. Nvidia just tried to close that gap. At SIGGRAPH 2026, the company unveiled its Synthetic Video Detector, a tool built to flag AI-generated clips before they slip past unnoticed.
The concept is refreshingly straightforward: the detector chews through footage frame by frame and hands each clip a probability score indicating how likely it is to be artificial. No vague verdicts, no black-box hand-waving — just a number you can act on.
Performance is where things get interesting. On pristine, uncompressed video, the detector reaches up to 92% accuracy. That figure is important because it comes with an honest caveat that Nvidia doesn’t hide: compression degrades detection. At roughly 15% compression, accuracy slips to 87%, and by around 50% compression it drops to 82%. Since most video people actually encounter online has been squeezed and re-encoded several times over, that decline matters — but the tool remains usefully reliable even when the source material is far from perfect.
Speed is the other headline. Running on Nvidia RTX systems, the detector can analyze uncompressed video in as little as 22 milliseconds. That’s fast enough to imagine real-time or near-real-time screening — the kind of throughput newsrooms, moderation teams and rights holders would need if they ever hope to keep pace with the flood of synthetic media.
Here’s the part that tempers the enthusiasm: you can’t just download it. The Synthetic Video Detector isn’t on general sale and has no consumer price tag. Instead, it’s being distributed through early access as part of Nvidia’s AI for Media Private Access Program. In other words, this is a tool aimed at industry partners first — studios, platforms and media organizations — rather than the average person trying to sanity-check a suspicious clip in their feed.
That’s both understandable and slightly frustrating. The people most exposed to deepfakes are ordinary viewers, and a gated professional program does little for them directly. But there’s logic to the approach: detection tools are only as good as their secrecy. Publish the exact detection method too widely, and generative models simply train to evade it. Keeping the detector behind a controlled program buys time before the inevitable cat-and-mouse escalation.
What Nvidia has shown is proof that the same GPU muscle powering generative AI can be pointed back at it. The accuracy figures are strong without being oversold, the honesty about compression is welcome, and 22 milliseconds is a genuinely impressive number. The open question is whether tools like this ever reach the wider public — or stay locked inside the industry that helped create the problem.
