AI Video Detector: NVIDIA Says Up to 92% Accurate

NVIDIA AI video detector: a NIM microservice for synthetic clips

NVIDIA's Synthetic Video Detector is a GPU-accelerated NIM microservice that scores a video for the likelihood that it is AI-generated, and NVIDIA announced it at SIGGRAPH in July 2026. A microservice is a small program that offers one function on its own, so the detector can be built into an existing video system rather than run as a separate app. Here that function is checking video.

What NVIDIA documentation says the microservice does

The documentation defines the detector as a service that estimates whether a video is AI-generated ("synthetic") or real. NVIDIA's blog puts it inside the NVIDIA AI for Media platform.

The documentation adds that the service takes an input video and returns detection scores for individual frames plus one aggregated probability for the whole clip. The blog calls the result a classifier score.

A streaming gRPC API that returns results as video arrives

The front end is a bi-directional streaming gRPC API, which means you send video data and the detection results come back in a stream. gRPC is a protocol built for this kind of two-way, continuous exchange. The documentation also lists optional summary files, such as CSV or JSON, for offline analysis.

AI video detection accuracy: 92% down to 82% under compression

NVIDIA reports accuracy of up to 92% on uncompressed video, 87% at 15% compression and 82% at 50% compression, and speed of 22 to about 30 milliseconds for 1080p depending on the GPU. Each number belongs to a condition, so "up to 92%" describes the uncompressed case only. NVIDIA also says the detector remains effective after the compression, resizing, cropping and re-encoding steps common in newsroom and social-video workflows.

Accuracy at 0%, 15% and 50% compression in NVIDIA testing

Accuracy in NVIDIA testing falls as compression rises: 92% uncompressed, 87% at 15% compression and 82% at 50% compression. Going from uncompressed video to 50% compression costs 10 points, so a heavily re-encoded clip gets a less reliable score.

Speed: 22 ms on RTX systems, about 30 ms on NVIDIA L40 GPUs

For 1080p video, NVIDIA reports as little as 22 milliseconds on NVIDIA RTX systems and about 30 milliseconds on NVIDIA L40 GPUs. The figure changes with the GPU, so read it together with the hardware named. NVIDIA gives no other hardware figures in the announcement.

Deepfake detection for broadcasts and Wowza live streaming

NVIDIA presents the detector for newsrooms and media workflows, and names Wowza as a partner bringing it into livestreaming. The blog's own wording is "synthetic video", and "deepfake" is the more common name for the same problem. Broadcasters, government agencies, financial institutions and critical infrastructure operators are the organizations NVIDIA describes as most exposed to synthetic media risk.

Broadcast misinformation: newsrooms use the score to triage clips

In NVIDIA's blog the use case is editorial: the score helps a newsroom decide what to look at first. Editorial teams can use it to prioritize clips for review, flag or quarantine questionable footage, or escalate it for deeper analysis.

NVIDIA adds that the microservice does not replace established verification practices. The score is one more signal for time-sensitive decisions.

Wowza live streaming infrastructure and the detector

According to NVIDIA's blog, Wowza is embedding the microservice through the Wowza Video Intelligence Framework to bring real-time synthetic video detection into livestreaming workflows. That places the detector at the ingest and streaming layer that customers already run.

NVIDIA also says organizations can deploy the microservice in on-premises, edge, hybrid and approved air-gapped environments. Sensitive footage can stay inside the customer's own environment.

How the AI video detector works: NVIDIA's forensic method

NVIDIA's documentation describes a forensic-oriented detection strategy: the detector looks for traces that generative video pipelines leave behind, not for what the picture shows. In practice it works more like file forensics than scene analysis.

Frequency-domain traces that survive compression

The documentation says the detector analyzes low-level statistical and frequency-domain traces that are characteristic of synthetic generation. It adds that these signals are resilient to compression and re-encoding and generalize across different generative architectures.

This is NVIDIA's description of its own method. The 92%, 87% and 82% accuracy figures also come from NVIDIA's own testing.