NVIDIA has announced 'Synthetic Video Detector,' a tool capable of identifying AI-generated videos with up to 92% accuracy.
NVIDIA has announced 'Synthetic Video Detector,' which analyzes potentially AI-generated videos frame by frame. This AI microservice is intended for use by news organizations and broadcasters, and is said to achieve up to 92% accuracy with uncompressed video and process 1080p video in as little as 22 milliseconds.
At SIGGRAPH, NVIDIA Advances Graphics and Simulation With Agentic and Physical AI | NVIDIA Blog
Nvidia's new Synthetic Video Detector can identify fake AI videos with up to 92% accuracy — microservice based on cutting-edge research looks to combat misinformation in broadcasts with just 22ms 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-to-92-percent-accuracy-microservice-based-on-cutting-edge-research-looks-to-combat-misinformation-in-broadcasts-with-just-22ms-processing-time
Synthetic Video Detector is part of 'NVIDIA AI for Media,' which NVIDIA offers to the media industry, and is provided as NVIDIA Inference Microservice, or NIM. It works by examining videos frame by frame and outputting a classification score that indicates the likelihood of AI-generated content being included.
News organizations can use this score to prioritize footage that needs review, isolate questionable footage, or send it for further investigation. NVIDIA positions Synthetic Video Detector not as a replacement for existing fact-checking or video verification, but as a tool to add information to time-pressured editing environments.
Synthetic Video Detector divides the video into 504x504 pixel frames and analyzes their features using DINOv2 and DINOv3, image recognition models developed by Meta. Each frame is assigned a score, with 0 representing the closest representation to reality and 1 representing the closest representation to AI generation, and a final judgment result for the entire video is calculated.
NVIDIA has released a demo page where you can actually 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 'Acknowledge & 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 Synthetic Video Detector.
The video I checked this time is the following. As indicated by the watermark in the bottom right, it is an AI video created by Google's Gemini.
The analysis revealed a 92.6% probability that the video is a composite. The line graph 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, it scored 3.5%, indicating that it was almost entirely real video.
Videos posted on social media and news distribution services may undergo processing such as compression, resizing, trimming, and re-encoding. According to NVIDIA's tests, the accuracy was recorded at up to 92% for uncompressed videos, 87% for videos with 15% compression, and 82% even for videos with 50% compression.
Processing speeds are said to be as fast as 22 milliseconds for 1080p video on NVIDIA RTX-equipped systems and approximately 30 milliseconds on NVIDIA L40 GPUs. It can be deployed on-premises, at the edge, in hybrid environments, and in environments isolated from external networks, allowing for the analysis of highly confidential video while keeping it within the organization.
NVIDIA explains that Synthetic Video Detector is not intended to replace traditional video verification, but rather to provide new information for decision-making in time-constrained news environments. The goal is to enable editorial teams to respond quickly, maintain reporting standards, and protect public trust in video footage by combining AI-based judgment with existing verification methods.
Related Posts:
