NVIDIA Synthetic Video Detector Arrives, Judging AI Video at 92% / 22ms

What Is the Synthetic Video Detector (What It Does)

The Synthetic Video Detector is a tool announced by NVIDIA at SIGGRAPH in July 2026 that estimates whether a video was created by AI. As AI-generated video (synthetic video) grows rapidly, it appeared as a means of telling it apart. First, let us establish its role and how the judgment works.

Definition (a microservice that judges AI-generated video)

The Synthetic Video Detector is a service that, given a video, returns a number for "is this likely AI-generated, or real?" It is positioned as a GPU-accelerated microservice that analyzes video content and estimates the likelihood that the video is AI-generated (synthetic) versus real.

A microservice is a small program that provides a specific function on its own. You can embed it in an existing system and use it as a video-checking function. Based on the published page's URL structure, this is said to be part of NVIDIA's video-oriented AI suite "Maxine," though the announcement text does not explicitly say "Maxine."

How it judges (forensic method, gRPC / offline)

The idea of the judgment differs from a human seeing through it by eye. It adopts a forensics-oriented method that identifies real from synthetic using the inherent traces (artifacts) that generative AI leaves in the process of making a video.

There are two delivery forms. A bidirectional streaming gRPC API that returns a judgment when you send video data, and offline forensic analysis that examines accumulated video later in bulk. It is designed to handle both cases where you want to judge on the spot, as in live streaming, and where you want to verify after the fact.

Accuracy and Speed (The Numbers)

The reason this tool draws attention is its combination of accuracy and speed. Here we look, in numbers, at accuracy including compression resistance and at a speed that withstands real-time use.

Accuracy (resistance to compression)

First, accuracy. In NVIDIA's testing, it reached up to 92% on uncompressed video, 87% at 15% compression, and 82% at 50% compression. It is honest that the trend of lower accuracy with more compression is clearly shown.

Because on real social media and streaming, video circulates compressed, this "resistance to compression" matters. While it reportedly keeps 82% even at 50% compression, the "up to 92%" is a value under the best conditions—a point to keep in mind. It is a tool to use as a probabilistic estimate, not perfect detection.

Speed (1080p in 22 milliseconds)

Speed is also a big feature. It can process 1080p video in as little as 22 milliseconds on NVIDIA RTX systems and about 30 milliseconds on L40 GPUs—fast enough to withstand real-time use.

A speed of 22 milliseconds is a level that can judge live-stream frames in the flow. But this is a value on RTX systems; if the GPU changes, so does the speed (about 30 milliseconds on L40). Noting that the feel changes depending on which hardware you run it on helps avoid excessive expectations.

What Can It Be Used For (Misinformation Countermeasures)

Finally, we organize where this tool is useful and its limits when using it.

Expected uses (media authenticity checks)

The main use is in settings where you need to confirm whether a video is real. The central use is expected to be media authenticity checks—confirming, in news, broadcasting, and streaming, that incoming video is not an AI-generated fake. It plays the role of mechanically sifting synthetic video that has become hard to tell apart by hand.

AI-generated video gets harder to distinguish year by year, and concerns about its misuse to spread misinformation are rising. Such detection technology is positioned as a countermeasure. If you are interested in telling AI-generated content apart, see also articles on how generative AI works, such as What Is Claude (Anthropic's generative AI).

Cautions when using it (it is a probabilistic estimate)

While useful, its limits must be understood. Detection is only a probabilistic estimate, not a 100% judgment. Accuracy and speed change with the compression rate and hardware, and "up to 92%" is a best-case value. Rather than deciding truth by this tool alone, it is appropriate to use it as one input for judgment.

Detection and generation are in a cat-and-mouse relationship, and if the generation side advances, detection accuracy can fluctuate. Even if it shows high accuracy now, an attitude of not overtrusting it and combining it with other means of confirmation is required.

Conclusion

NVIDIA's Synthetic Video Detector is a practical forensic tool that tells AI-generated video apart with high accuracy and speed. Numbers like up to 92% uncompressed and 1080p in 22 milliseconds give real weight to real-time authenticity checks.

That said, before using it you should understand that accuracy drops with compression, that speed changes with the environment, and that it remains a probabilistic estimate. As the competition between generation and detection continues, such tools will grow in importance as one hand in countering misinformation.

Overseas official announcements and technical documents often come out in English and take effort to read through accurately. When you want to turn an English web page into a form that is easy to read, the following tool helps.

Free ToolURL to Markdown ConverterConvert any public web page URL to Markdown. Preserves headings, tables, lists, and links — perfect for LLM and RAG preprocessing, research notes, and archiving web articles.Try it now →