Tavus Unveils Griffin AI Video Model for Real-Time Human Interaction
AI video generation startup Tavus has debuted a real-time conversational video model called Griffin, which generates 720p video in 350-millisecond blocks to interact dynamically with human users, Živě.cz reported. The system processes both speech and visual input simultaneously, allowing the AI to gesture, track conversation, and respond to objects shown to it over a video feed. According to internal benchmarks cited by the company, Griffin currently ranks first on Nvidia’s leaderboard directly behind a reference live human. However, independent verification of these metrics is not yet available.
Security Hurdles Delay Public Release of Human Interaction Model
The underlying technology, which Tavus terms a Human Interaction Model, remains restricted to internal testing and is not currently available to the public. The company stated that public deployment requires extensive work on security safeguards and response fine-tuning because the model can already deceive human observers in limited settings. While AI-generated text and synthetic voice call centers have already blurred the lines of digital communication, Griffin represents a step toward real-time visual impersonation during live video calls. Security researchers point out that such systems heighten the risk of advanced phishing attacks and fraudulent video calls once bad actors adopt the technology.
Practical Applications Beyond Synthetic Video Calls
Beyond security implications, developers envision practical uses across education and technical troubleshooting where contextual visual feedback matters. An AI tutor powered by real-time video generation could observe a student’s facial expressions and adjust its teaching pace when a concept appears unclear. Alternatively, a user could hold a broken mechanical part up to a camera and receive visual and verbal guidance from an AI agent in real time. These use cases aim to bridge the gap between static chatbots and natural human communication, though widespread commercial availability depends on resolving ongoing latency, safety, and reliability challenges.
Frequently Asked Questions About the Griffin Model
When will the Griffin model become available to the public?
Tavus has not announced a public release date for the Human Interaction Model, noting that additional security testing and response calibration are required before the technology leaves internal evaluation.
How does Griffin process user input during a call?
The system simultaneously evaluates audio and visual streams, enabling it to react dynamically to spoken words and objects presented in front of the camera while generating responsive 720p video output.
More on Human Interaction Model
Related reading