Nvidia Unveils AI Video Detector at SIGGRAPH 2026
At SIGGRAPH 2026 in Los Angeles, Nvidia introduced Synthetic Video Detector (SVD), an AI-powered tool that detects AI-generated videos with high accuracy. The service analyzes video frames, extracting frequency-domain statistical patterns indicative of diffusion-based generation, and is designed to be robust to compression artifacts. SVD crops frames to 504×504 pixels, passes them through two Vision Transformers that score each frame from 0 (real) to 1 (fake), then averages scores into a percentage. On uncompressed 1080p video, SVD achieves 92% accuracy, processing a frame in 22 milliseconds on Nvidia RTX systems. Accuracy drops to 87% at 15% compression and 82% at 50% compression. Nvidia pitches SVD as a microservice for newsrooms, providing an additional signal for time-sensitive editorial decisions. It is part of Nvidia's AI for Media Private Access Program, not publicly available, though a demo exists online.
Key facts
- Nvidia unveiled Synthetic Video Detector (SVD) at SIGGRAPH 2026 in Los Angeles.
- SVD analyzes video frames using frequency-domain statistical patterns to detect diffusion-based AI generation.
- Frames are cropped to 504×504 pixels and scored by two Vision Transformers from 0 (real) to 1 (fake).
- On uncompressed 1080p video, SVD achieves 92% accuracy and processes a frame in 22 milliseconds on RTX systems.
- Accuracy falls to 87% at 15% compression and 82% at 50% compression.
- SVD is offered as a microservice for newsrooms, not replacing but supplementing existing verification practices.
- It is part of Nvidia's AI for Media Private Access Program, with a demo available online.
- Nvidia aims to help editorial teams protect standards and ensure public trust.
Entities
Institutions
- Nvidia
- SIGGRAPH
- Tom's Hardware
- PetaPixel
Locations
- Los Angeles
- United States