NVIDIA announced its Synthetic Video Detector, which analyzes potentially AI-generated videos frame by frame. The AI microservice is intended for use by news organizations and broadcasters and is said to be able to achieve up to 92% accuracy on uncompressed video and process 1080p video in just 22 milliseconds.
At SIGGRAPH, NVIDIA advances graphics and simulation with agent and physical AI | NVIDIA Blog
https://blogs.nvidia.com/blog/siggraph-news-2026/#synthetic-video
Nvidia’s new synthetic video detector can identify fake AI videos with up to 92% accuracy. Based on cutting-edge research, our microservices aim to combat misinformation in broadcasts with just 22 milliseconds of processing time. tom’s hardware
https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidias-new-synthetic-video-detector-can-identify-fake-ai-videos-with-up-tA 92 percent accurate microservice based on cutting-edge research that fights in-air misinformation with just 22 milliseconds of processing time
Synthetic Video Detector is part of NVIDIA’s NVIDIA AI for Media offering to the media industry and is available as an NVIDIA Inference Microservice (NIM). It works by inspecting the video frame by frame and outputting a classification score that indicates the likelihood that it contains AI-generated content.
News organizations can use this score to prioritize footage for review, isolate problematic footage, and send it on for further investigation. NVIDIA positions Synthetic Video Detector not as a replacement for existing fact-checking or video verification, but as a tool that adds information to time-sensitive editorial environments.
The synthetic video detector splits the video into 504×504 pixel frames and analyzes its features using: DINOv2 DINOv3 is an image recognition model developed by Meta. A score is assigned to each frame, with the expression closest to reality being 0 and the expression closest to AI generated being 1, and the final judgment result for the entire video is calculated.
NVIDIA has released a demo page where you can try out Synthetic Video Detector.
Synthetic Video Detector Model by NVIDIA | NVIDIA NIM
https://build.nvidia.com/nvidia/synthetic-video-detector
The terms of service will be displayed, so click “Agree and continue”.
Click “Upload” under “Input” on the left to upload the video you want to check from your local machine, then click “Run” to start the analysis by the synthetic video detector.
The video I checked this time is below. This is an AI video created by Google’s Gemini, as indicated by the watermark in the bottom right.
After analysis, we found that there was a 92.6% chance that the video was synthesized. The line chart shows the likelihood that each frame contains AI-generated content.
When we fed NVIDIA’s non-AI video data into the system and analyzed it, we found that the video was almost completely authentic, with a score of 3.5%.
Videos posted to social media and news distribution services may be processed by compression, resizing, trimming, re-encoding, etc. In NVIDIA’s tests, accuracy was up to 92% on uncompressed video, 87% on 15% compressed video, and 82% on 50% compressed video.
Processing speeds are said to be 22ms for 1080p video on NVIDIA RTX-equipped systems and around 30ms on NVIDIA L40 GPUs. Deploy on-premises, at the edge, in hybrid environments, and in environments isolated from external networks, allowing you to analyze sensitive video while keeping it within your organization.
NVIDIA explains that the Synthetic Video Detector is not intended to replace traditional video verification, but rather to provide new information for decision-making in time-sensitive news environments. The goal is to combine AI-based judgment with existing verification methods to enable editorial teams to respond quickly, maintain journalistic standards, and protect public trust in video footage.
