We live in an era where artificial intelligence increasingly dominates the internet, making it difficult to distinguish between real content and fake news. Every photo or video you see today could be generated by AI, and traditional markers are rapidly disappearing. There are tools that can help identify whether something is made with AI or not, and Nvidia, perhaps the main beneficiary of the AI race, just released its own tool called the “Synthetic Video Detector” (SVD).
SVD is part of the company’s AI for Media private access program, Nvidia Inference Microservice (NIM), so it’s not publicly available to consumers, but demos do exist. In any case, SVD’s job is simple. Detect whether a video is real or generated using AI. Analyze videos at scale and break them down frame by frame to identify anomalies. It is based on cutting-edge research that won an award at the Computer Vision Conference ICCV.
Rather than viewing the video file as a whole, SVD splits it into cropped frames, each with a resolution of 504×504. These frames pass through two powerful Vision Transformers created by Meta: DINOv2 and DINOv3. The vision transformer’s job is to learn how to form patterns without the need for human-labeled data. These are commonly used for image classification, image retrieval, object detection, and depth estimation.
So, when a frame passes through these transformers, its unique spatial features are immediately evaluated and each is assigned a score from 0 to 1. 0 represents a completely real image and 1 represents a completely fake image. The scores are aggregated and averaged at the end to tell users whether the video is authentic or not based on a percentage score out of 100.
In this way, news agencies, broadcasters and news organizations will be able to authenticate received footage more quickly and reliably. SVD is also designed to accommodate the realities of social media compression, as imperfections in videos uploaded online are masked. However, transformers can still detect patterns that the human eye cannot detect, looking past surface-level anomalies and focusing instead on inherent artifacts.
As you can see in the AI GVD bench above, uncompressed video gives the best results with SVD, showing an impressive 92% accuracy. At 15% compression, the model’s accuracy drops to 87%, but even at 50% compression, it still achieves a very good 82% accuracy in detecting AI-generated content. Because it’s a microservice, latency is also very good, processing 1080p video in just 22ms on Nvidia RTX GPUs and 30ms on Nvidia’s workstation models.
That being said, SVD requires an NVENC encoder, so data center cards like the B100 cannot run it natively. Nvidia said it is already working with Wowza to bring real-time composite video detection to live streaming workflows. You can try the demo version right away at build.nvidia.com, but be aware that it takes time to process as it happens in the cloud, has a maximum file size limit of only 100 MB, and often times out.
to follow Tom’s Hardware on Google Newsor Add us as a preferred sourceget the latest news, analysis, and reviews in your feed.
