Nvidia introduced the Synthetic Video Detector NIM microservice as part of its AI for Media platform. Inspect video footage for signs of AI-generated content.
The detector checks each frame of the video for signs of synthetic content and gives the footage a classifier score. Organizations can use this score to review and flag problematic footage.
This technical approach wins first place in ICCV’s 2025 SAFE Synthetic Video Detection Challenge and builds on research presented at NeurIPS 2025.
The detector extracts a 504 x 504 pixel crop from the frame and processes it using an ensemble based on Meta’s DINOv2 and DINOv3 vision transformer models. These models generate representations of the spatial and visual characteristics within each frame, which the detector uses to estimate whether the content contains signs of AI generation.
Each analyzed frame receives a composite likelihood score. The scale ranges from 0, representing real content, to 1, representing synthetic content. The frame-level results are then used to generate an overall classifier score for the video.
Nvidia reports that its synthetic video detector achieved up to 92% accuracy when analyzing uncompressed video. Performance dropped to 87% with 15% compression and 82% with 50% compression.
Performance in real-world scenarios may vary depending on the video generator, resolution, codec, compression level, and other processing applied to the footage. Nvidia does not provide data on false positives, false positives, or performance across individual video generation models.
Nvidia also emphasizes speed of service. The NIM microservice says it can process 1080p video in 22ms on supported RTX systems and around 30ms on Nvidia L40 GPUs.
This detector addresses part of the trust determination of biometric systems: verifying whether the footage contains statistical patterns associated with AI-generated videos. The detector should not be treated as a replacement for liveness attack detection or presentation attack detection (PAD).
Wowza deploys Nvidia detectors to live streaming infrastructure
Wowza integrates Nvidia’s composite video detector into the Wowza Video Intelligence Framework. It shows how to integrate NIM microservices into a production video system.
Nvidia provides the synthetic video detection model, NIM model serving layer, and optimized inference software, while Wowza provides the streaming and orchestration layers. Nvidia classifiers can be applied close to the point where the video is received, processed, or distributed.
This means organizations don’t have to remove suspicious clips from their streams and upload them to another detection website. The output can be integrated into existing workflows, allowing organizations to flag suspicious footage before broadcast.
This integration is designed to support deployment on-premises, edge, hybrid, or approved air-gap configurations. Wowza says customers can run the detector on their existing Nvidia infrastructure while maintaining control over video and model output.
“We’ve built the layers to put it where it’s needed. Together, we’re giving organizations protecting people the ability to trust what they’re seeing in real time on their hardware,” said Krish Kumar, CEO of Wowza.
Wowza says its media server technology supports more than 35,000 video deployments in more than 170 countries.
The company cites government and public sector applications as potential use cases, where synthetic videos could be used to impersonate officials. Public security organizations may subsequently analyze videos submitted as evidence of a crime.
The same approach can be applied to video KYC, account recovery, and high-risk financial transactions. Wowza has identified synthetic executive and synthetic video ID attacks in video calls as potential financial services use cases.
Wowza Video Intelligence Framework is now available.
Article topics
Deepfake detection | Generative AI | NVIDIA | Synthetic data | Video analysis
