Nvidia brings synthetic video detection to enterprise AI platform

Nvidia has introduced a Synthetic Video Detector NIM microservice as part of its AI for Media platform. It screens video footage for signs of AI generated content.

The detector checks each frame of a video for signs of synthetic content and gives the footage a classifier score. Organizations can then use this score to review or flag questionable footage.

The technical approach is built on research that received first place in the 2025 SAFE Synthetic Video Detection Challenge at ICCV and was published at NeurIPS 2025.

The detector extracts 504-by-504-pixel crops from the frames and processes them 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 then uses to estimate if the content contains signs of AI generation.

Each analyzed frame receives a synthetic-likelihood score. The scale ranges from zero, representing real content, to one, representing synthetic content. The frame-level results are then used to produce an overall classifier score for the video.

Nvidia reports that the synthetic video detector achieved accuracy of up to 92 percent when analyzing uncompressed video. The performance dropped to 87 percent at 15 percent compression and 82 percent at 50 percent compression.

Performance in real-world scenarios could vary with the video generator, resolution, codec, compression level, and other processing applied to the footage. Nvidia does not provide data on false positives, false negatives, or performance across individual video generation models.

Nvidia also emphasizes the speed of the service. It says the NIM microservice can process 1080p video in 22 milliseconds on supported RTX systems and approximately 30 milliseconds on Nvidia L40 GPUs.

The detector addresses one part of the trust decision for biometric systems: verifying whether the footage contains statistical patterns associated with AI generated video. The detector should not be treated as a substitute for liveness or presentation attack detection (PAD).

Wowza brings Nvidia’s detector into live streaming infrastructure

Wowza is integrating Nvidia’s Synthetic Video Detector into the Wowza Video Intelligence Framework. This shows how the NIM microservice can be integrated into an operational video system.

Nvidia supplies the synthetic video detection model, the NIM model-serving layer, and the optimized inference software, while Wowza provides the streaming and orchestration layer. The Nvidia classifier can be applied closer to the point where video is received, processed, or distributed.

This means organizations do not need to remove a suspicious clip from the stream and upload it to a separate detection website. The output can then be incorporated into an existing workflow to allow an organization to flag questionable footage before it is broadcast.

The integration is designed to support deployment on-premise, at the edge, in hybrid, or approved air-gapped configurations. Wowza says customers can run the detector on existing Nvidia infrastructure, while retaining control over the footage and model outputs.

“We built the layer that puts it everywhere it’s needed. Together, we’re giving the organizations that protect people the ability to trust what they’re seeing, in real time, on their own hardware,” says Krish Kumar, chief executive officer at Wowza

Wowza says its media-server technology supports more than 35,000 video deployments across over 170 countries.

The company points to government and public sector applications as a potential use case, where synthetic video could be used to impersonate officials. Public safety organizations could analyze video submitted as evidence of a crime before acting on it.

The same approach could apply to video KYC, account recovery, and high-risk financial transactions. Wowza identifies synthetic executives in video calls and synthetic video identity attacks as possible financial services use cases.

The Wowza Video Intelligence Framework is now available.

Article Topics

deepfake detection  |  generative AI  |  Nvidia  |  synthetic data  |  video analytics