Overview
The NVIDIA Synthetic Video Detector NIM is a GPU-accelerated microservice that analyzes video content and estimates the likelihood that a video has been AI-generated (“synthetic”) versus real. The service consumes an input video and produces per-frame detection scores and an aggregated video-level probability, supporting screening at both the frame and video levels.
The service employs a forensic-oriented detection strategy that focuses on identifying intrinsic artifacts introduced by generative video pipelines. Rather than relying on semantic content, the detector analyzes low-level statistical and frequency-domain traces that are characteristic of synthetic generation.
These signals are robust to common video transformations and generalize across different generative architectures; however, heavy compression, aggressive re-encoding, or codec differences may still reduce detection accuracy, so evaluate the selected threshold on representative content.
Designed for high-throughput and low-latency deployment, the service supports use cases such as media authenticity screening, digital forensics, content moderation, and integrity monitoring in real-world environments.
Inputs and Outputs
-
Input
- A video in a supported container and codec combination. Supported containers include MP4/QuickTime (
.mp4,.mov,.m4v), Matroska (.mkv), and WebM (.webm). The service decodes the video into RGB frames internally using the NVIDIA hardware video decoder (NVDEC).
- A video in a supported container and codec combination. Supported containers include MP4/QuickTime (
-
Per-frame or per-segment output (optional)
- Synthetic likelihood scores associated with individual frames or temporal segments, which are useful for analysis and visualization.
-
Final aggregated output
- A single probability in that summarizes how likely the entire video is synthetic.
- Optional summary artifacts (for example, CSV or JSON) for offline analysis.
High-Level Service Architecture
At a high level, the service consists of the following:
-
gRPC front-end: A bi-directional streaming gRPC API that accepts video bytes and streams detection results back to the client.
-
Video decoding layer: A GPU-accelerated NVDEC pipeline that converts supported input video into normalized RGB frames suitable for model inference.
-
Inference engine: A batched inference layer that runs frames through the DINOv2 + DINOv3 ensemble using TensorRT-optimized engines.
-
Results and reporting: Video classification and per-frame analysis.
Try It Out
Try the NVIDIA Synthetic Video Detector NIM at build.nvidia.com/nvidia/synthetic-video-detector.
To experience the NVIDIA Synthetic Video Detector NIM API without having to host your own servers, use the Try API feature, which uses the NVIDIA Cloud Function backend.