NVIDIA has unveiled a new weapon in the fight against undeclared AI-generated content. The Synthetic Video Detector is a NIM microservice built to help editorial teams and broadcasters spot fake footage, announced at the SIGGRAPH 2026 conference in Los Angeles.
AI for Media
The detector runs as a NIM microservice, short for NVIDIA Inference Microservices, inside the company’s AI for Media platform. NIM packages are pre-built software bundles that deploy optimized AI models on NVIDIA GPUs without the usual setup work. AI for Media is the development platform that delivers them, covering video, audio and augmented reality in broadcast and entertainment pipelines.
Frame by Frame
Upload a clip and the microservice analyzes it frame by frame, then returns a classifier score indicating whether the video contains synthetic content. A higher score means stronger evidence the model found something artificial.
Speed and accuracy
NVIDIA says the detector stays effective after compression, resizing and re-encoding, the steps that show up in every newsroom and social-video workflow. In the company’s testing, accuracy hit 92% on uncompressed video, 87% at 15% compression and 82% at 50% compression. Processing runs as fast as 22 milliseconds for 1080p on RTX systems and roughly 30 milliseconds on NVIDIA L40 GPUs.
Those numbers deserve a second look. At 50% compression, the model is wrong about one clip in five. YouTube, TikTok and Instagram all compress uploads aggressively, which strips out the subtle artifacts detection models depend on. The detector degrades exactly where the suspect footage tends to come from.
Time-sensitive decisions
NVIDIA isn’t positioning this as a replacement for established verification practices, but as an extra signal. Editorial teams can use the score to prioritize clips for review, flag or quarantine questionable footage, or escalate it for deeper analysis. In a newsroom racing a clock, that triage is the value. Nobody has to look at everything.
Livestreaming
Organizations can deploy the microservice close to wherever sensitive video is captured, stored or distributed, including on-premises, at the edge, in hybrid setups and in approved air-gapped environments. Suspect footage gets assessed before it travels.
Wowza is the first announced partner, embedding the microservice into its Video Intelligence Framework to bring real-time detection to livestreaming workflows across more than 35,000 deployments in over 170 countries.
What we think
AI video isn’t going back in the box. Plenty of it is harmless, and nobody’s credibility rides on whether a cat video was real. Editorial footage is a different matter.
Channels and platforms all want to break a story first, and now they carry the added risk that the clip driving it was generated. A tool that flags a suspect file in 22 milliseconds fits that pressure well.
Just don’t mistake a score for a verdict. An 82% hit rate on compressed video means a working detector that still misses, and the failure mode cuts both ways. A false negative pushes a fake to air. A false positive kills real footage. This belongs in the workflow as a triage signal with a human at the end of it, which is roughly how NVIDIA is pitching it.
Availability
You can see a demonstration of the Synthetic Video Detector on the NVIDIA website.
