Nowadays, spotting the difference between AI and real is getting harder to do with the human eye but NVIDIA believes AI can help solve that problem.
The company has introduced a new Synthetic Video Detection model that examines videos and predicts if they are real or AI generated. It is designed for content creators, media authentication, digital forensics and services that need to verify video content before using or publishing it.
The announcement came during this year’s SIGGRAPH conference in Los Angeles, where NVIDIA also shared new research on graphics, simulation and AI. This all comes just after the chipmakers announced their new creation: an AI model.
What Does The Detector Actually Do?
So, instead of giving one answer for an entire clip, the model checks every frame individually.
NVIDIA says each frame receives a score between 0 and 1. A score of 0 means the frame is predicted to be real and a score of 1 means it is predicted to be synthetic. This lets users identify exactly where AI generated content appears instead of receiving a single verdict for the whole video.
The model has been built to identify videos generated using diffusion models and has also been designed to work well after video compression, which often happens when clips are uploaded to websites and social media platforms.
The good news is that the software is ready for both commercial and non commercial use.
What Powers The Model?
The detector accepts MP4 video files and uses a Vision Transformer architecture built from an ensemble of DINOv2 and DINOv3 backbones.
Before analysis begins, each video frame is cropped to 504 by 504 pixels and normalised. The model contains 172 million parameters and runs on NVIDIA GPU accelerated systems using TensorRT.
The software supports Linux and works with NVIDIA Ampere, Blackwell, Turing and Lovelace GPU architectures. It needs NVENC and NVDEC hardware, which means products such as the A100, H100, B100 and B200 cannot run it.
More from Artificial Intelligence
- Will The EU’s Mandatory AI Watermarks Actually Stop Deepfakes?
- What’s The Difference Between Open Weight And Open Source AI?
- Frontier AI Or Budget Models – Which Strategy Makes Sense For Your Business?
- Nvidia Has Its Own AI Model – How Will This Affect Competition In The Industry?
- AI Is Supposedly “Replacing” Writers. So Why Is Anthropic Hiring Editors For $295,000 A Year?
- Should AI Literacy Be Taught Alongside Math And English?
- Should AI Have An Appreciation Day When Other Industries Don’t?
- The Rise Of AI-Powered Product Governance: From Institutional Knowledge To Scalable Decision-Making
NVIDIA also says anyone using the model should do their own testing before putting it into real world products.
If you’d like to test it out, the website is pretty straightforward and easy to navigate. I took a look and immediately found the input/output section. The company also warns not to upload videos that contain sensitive information – this should be standard on every AI platform.
“The integration of foundation and fine tuned models into AI systems requires additional testing using use case specific data to ensure safe and effective deployment. Following the V model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment,” the company says.
How Well Did It Perform?
According to NVIDIA’s model card, the detector recorded an accuracy score of 85.64% during evaluation.
The benchmark used 4,000 videos. Training used fewer than 10,000 hours of footage collected from a hybrid dataset. Real videos came from Panda70M, while AI generated videos came from Pyramid Flow, Open Sora Plan, Allegro and cosmos predict1. NVIDIA says the training data contained 1,400 real videos and 1,400 synthetic videos for each generator.
The company says the output appears as a score for every frame, giving users a detailed view of where synthetic content may appear throughout a video.
Is NVIDIA Becoming An AI Company For Everything?
The video detector was one of many AI announcements during the conference, actually.
During SIGGRAPH, NVIDIA also spoke about AI for weather forecasting, robotics, autonomous vehicles, digital twins, world models, simulation and graphics. The company also introduced Cosmos 3 Edge, a world model designed to run in real time on devices, and Cosmos Dreams, which can generate an entire world from a single frame for autonomous vehicle simulations.
But all of this made me wonder: is NVIDIA becoming an AI company that wants to build tools for almost every corner of the market?
From the company’s point of view, the answer seems to be yes.
NVIDIA founder and chief executive Jensen Huang said, “Creators and designers need tools powerful enough to expand their imagination, malleable enough to give them the freedom to shape ideas and precise enough to realize their vision exactly as intended. Whether for games, cinema, robotics or factory digital twins, the goal is the same: to create virtual worlds that behave with the fidelity and realism of the physical world.”
Edward Liu, director of applied deep learning research at NVIDIA, spoke about the company’s latest neural rendering work.
“Control in artistic direction is a really exciting research direction for us. Simulation defines the world, generation enriches its appearance and artists direct the outcome…AI extending graphics the same way programmable shaders and ray tracing have extended graphics before.”
Neil Ashton, distinguished engineer at NVIDIA, spoke about AI models used in scientific research.
“The challenge that we have is to translate the success of using these AI models in weather and climate into the world that we live in, to be able to design the planes and cars and data centers of the future. One of the things that NVIDIA is really focused on is helping to achieve that vision.”
Ming Yu Liu, vice president of Cosmos Lab at NVIDIA, also spoke about the company’s Cosmos platform.
“What makes a foundation model a foundation model is its capability to consume enormous amounts of diverse data. We trained Cosmos with data for diverse physical tasks, including world understanding, prediction, simulation and action. We use one backbone to solve all the different tasks.
“Every embodiment speaks a different language. Our solution is to build a common vocabulary.”
It would be quite interesting to see how far the company goes and what else they’re adding to its list of ventures. It’d also be interesting to see whether having a hand in everything might eventually impact the quality of products and offerings…
