NVIDIA introduced the Synthetic Video Detector NIM microservice at SIGGRAPH, adding an AI-assisted signal that flags whether a video clip contains synthetic content. It is part of the NVIDIA AI for Media platform and is aimed at editorial and media teams making fast calls on questionable footage.
The tool produces a classifier score, not a verdict. It analyzes video frame by frame and outputs a probability that the clip is synthetic, which teams can use to prioritize, flag, quarantine, or escalate.
Accuracy reached up to 92% on uncompressed video in NVIDIA testing, holding at 82% even after heavy compression.
Wowza is already embedding it for real-time detection across more than 35,000 livestreaming deployments.
A frame-by-frame score built to survive newsroom compression
The microservice analyzes video frame by frame to generate a classifier score for synthetic content. Editorial teams can use that score to prioritize clips for review, flag or quarantine questionable footage, or escalate it for deeper analysis. NVIDIA positions the tool as an additional signal for time-sensitive decisions rather than a replacement for established verification practices.
Detection quality usually degrades once a clip is compressed, resized, cropped, or re-encoded, which is exactly what happens to video moving through newsroom and social pipelines. NVIDIA says the model stays effective through those steps. In its testing, accuracy reached up to 92% on uncompressed video, 87% at 15% compression, and 82% at 50% compression.
The latest model revision reported an AUC of 0.9614 and accuracy of 0.9453 on NVIDIA's internal test set. AUC, or Area Under the Curve, measures how well a classifier ranks positive samples above negative ones independent of thresholds. Those thresholds can be tuned, including more conservative settings that reduce the chance a synthetic clip slips through unflagged.
Built for real-time review at newsroom speed
Speed matters when a clip needs a decision before it airs or posts. The microservice can process 1080p video in as little as 22 milliseconds on NVIDIA RTX systems and roughly 30 milliseconds on NVIDIA L40 GPUs.
The detector ships in NVIDIA's NIM microservice format, the packaged deployment model we covered when NVIDIA first brought it to production media pipelines. That packaging is what lets editorial teams drop the model into an existing workflow instead of standing up a separate detection stack.
On-prem and air-gapped deployment for sensitive footage
Organizations can run the microservice closer to where sensitive video is captured, stored, or distributed, including on-premises, edge, hybrid, and approved air-gapped environments. NVIDIA frames that flexibility as a way for teams to keep control over video data, access, and operations.
That deployment model targets a specific set of buyers. NVIDIA names broadcasters, government agencies, financial institutions, and critical infrastructure operators as the groups most exposed to synthetic media risk, and also the ones facing strict requirements around data residency, security, and operational control.
Partner adoption is where the detector moves from a model to deployable infrastructure. Wowza is embedding the microservice through its Video Intelligence Framework, which we saw the company introduce at NAB, bringing real-time synthetic video detection into livestreaming workflows that span more than 35,000 deployments across over 170 countries. Pairing the detector with a video layer customers already run puts AI-assisted verification closer to ingest and streaming operations while keeping footage inside their own environments.
Detection tooling catches up to the generation curve
The release lands as synthetic video quality keeps climbing and detection shifts from a research topic toward a working newsroom requirement. We covered how quickly a single AI clip can trigger a public trust crisis, the kind of scenario a frame-level score wired into ingest is meant to catch earlier. NVIDIA is positioning the Synthetic Video Detector as one signal inside a broader verification process, which sets a realistic bar for what automated detection can do as generation tools continue to improve.
