San Francisco-based AI lab Tavus has introduced Griffin, a video-to-video model designed for real-time interaction. The system integrates perception, generation, movement, and conversation into a single architecture, replacing the company's previous approach of chaining separate systems like Phoenix and Raven.
In a blind study involving 54 participants, 48% of users believed they were speaking with a human during a one-minute video call. This marks an improvement over the company's previous system, which had a 2.4% success rate. Additionally, Griffin scored 3.83 on NVIDIA's VideoFDB benchmark, compared to a human reference score of 3.92.
Tavus claims the model outperforms the next-best system by 37% in real-time reaction metrics. Griffin is built on the existing PAL platform, which has been utilized by over 150,000 developers and businesses. The company currently offers a limited research preview called Griffin-Lite.
CEO Hassaan Raza and Head of Research Ioannis Patras stated that a full release is pending the implementation of safety measures. The technology is intended for use in business applications such as customer support and onboarding, though the company acknowledges potential risks regarding impersonation.
