> This page is for NVIDIA Synthetic Video Detector NIM, version 2.0.0 (Latest) (default).
> For other versions, use one of these documentation indexes:
> - 2.0.0 (Latest) (default): https://docs.nvidia.com/nim/maxine/synthetic-video-detector/2.0.0/llms.txt
> - 1.0.0: https://docs.nvidia.com/nim/maxine/synthetic-video-detector/1.0.0/llms.txt

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/nim/maxine/llms.txt.

# 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](/nim/maxine/synthetic-video-detector/basic-inference#input-and-output). 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).

- **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 $[0, 1]$ 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](https://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](https://build.nvidia.com/nvidia/synthetic-video-detector/api) feature, which uses the NVIDIA Cloud Function backend.