Nvidia AI for Media is launching as a tool to help newsrooms detect fake videos, or those that were generated by AI or captured in the real world.
More precisely, the AI tool detects whether a video that has been sent to a network or provided by another party is synthetic or based on real world video.
Every day, video brings the world’s biggest stories into view — from breaking news across
continents, to events reshaping communities, to moments that unite people across the
globe. In a news cycle that moves around the clock, trustworthy video is the medium through which people see what’s happening, understand why it matters and stay connected and up to date, Nvidia said in a presentation at Siggraph, the graphics and computer event.
For that reason, public trust in video is more essential than ever. At Siggraph, Nvidia announced the Synthetic Video Detector Nvidia NIM microservice, part of the Nvidia AI
for Media platform, to bring an AI-assisted detection signal into editorial and media workflows.
Rev Lebaredian, vice president of physical AI simulation technology at Nvidia, said in a press briefing that AI video generation is amazing in that the videos are indistinguishable from real video. However, it poses a problem.
“If we cannot tell the difference between a synthetic video and a real one, it can erode public trust when videos are presented as news, as something that came from the real world,” he said. “Fortunately, the same AI technologies we’re developing to generate these videos can also be used to detect when a video is real or synthetic.”
The NIM microservice analyzes video frame by frame to produce a classifier score of whether it contains synthetic content. Editorial teams can use that score to prioritize clips for review, flag or quarantine questionable footage, or escalate it for deeper analysis.
Rather than replacing established verification practices, the microservice provides another
signal for time-sensitive decisions — helping teams move quickly while protecting editorial
standards and ensuring public trust.
Synthetic Video Detector remains effective after the compression, resizing, cropping and
re-encoding steps common in newsroom and social-video workflows. In Nvidia testing, the
model’s accuracy reached up to 92% on uncompressed video, 87% at 15% compression and
82% at 50% compression.
The NIM microservice can process 1080p video in as little as 22 milliseconds on Nvidia
RTX systems and approximately 30 milliseconds on Nvidia L40 GPUs.
The latest model revision has also shown improved internal benchmark results, including AUC
of 0.9614 and accuracy of 0.9453 on the internal Nvidia test set. AUC (Area Under the Curve)
measures how well a classifier ranks positive samples above negative ones independent of
thresholds.Thresholds can be configured to support different review postures, including more
conservative settings that prioritize reducing the chance that synthetic video is missed.
Deploy detection where video lives
Organizations can deploy the NIM microservice closer to where sensitive video is captured,
stored or distributed, including in on-premises, edge, hybrid and approved air-gapped environments. This flexibility helps teams maintain control over video data, access and
operations.
Partner adoption is already helping move Synthetic Video Detector from model capability to deployable media infrastructure. Wowza is embedding the microservice through the Wowza Video Intelligence Framework, bringing real-time synthetic video detection into livestreaming workflows used across more than 35,000 deployments in over 170 countries.
That scale matters because many of the organizations most exposed to synthetic media
risk, including broadcasters, government agencies, financial institutions and critical
infrastructure operators, also face strict requirements around data residency, security and
operational control.
By pairing Synthetic Video Detector with a video infrastructure layer customers already
use, Wowza can help make AI-assisted verification available closer to ingest and streaming
operations, allowing teams to flag questionable video in real time while keeping sensitive
footage inside their own environments.
Try the Nvidia Synthetic Video Detector NIM microservice.
I asked how the service detects whether a video is fake. Nvidia replied that it does not determine whether the event shown is factually true; it estimates whether the video itself was synthetically generated.
The Nvidia Synthetic Video Detector analyzes individual frames with DINOv2 and DINOv3 vision-transformer backbones, the company said. Each produces frame-level classifier scores that are averaged into video-level scores, combined, and mapped to a final synthetic-likelihood score.
It was trained and evaluated on real and synthetic videos—including diffusion-generated content—and is designed to remain useful after compression, resizing, cropping, and re-encoding. Newsrooms should use the score to triage and escalate clips for further verification, not as a definitive verdict or replacement for editorial judgment.