Overview
- NVIDIA publicly announced the Synthetic Video Detector as an Inference Microservice in its AI for Media Private Access Program on July 20–21 and released an online demo for testers.
- The company says the model inspects 504x504 cropped frames with Meta’s DINOv2 and DINOv3 vision transformers and assigns each frame a 0–1 fake score that is averaged to classify a clip.
- NVIDIA published internal performance figures reporting up to 92% accuracy on uncompressed video, 87% accuracy at 15% compression and 82% at 50% compression, plus processing times as low as 22ms for 1080p on RTX GPUs.
- The service will be distributed privately rather than to general consumers and is being integrated with Wowza’s Intelligence Video framework for deployment across roughly 35,000 installations in about 170 countries.
- Practical limits remain: accuracy drops with heavier compression, the tool requires NVENC so some datacenter GPUs cannot run it natively, and the published benchmarks are from NVIDIA’s internal tests rather than independent evaluations.