Griffin: The First Human Interaction Model
4 hours ago
- Griffin is the world's first Human Interaction Model (HIM), designed for real-time face-to-face interaction that understands expressions, pauses, and nonverbal cues.
- Griffin passed the real-time video Turing test: 48% of participants believed they were talking to a real human, a huge leap from previous systems (max 2%).
- The model combines real-time perception, conversational decision-making, and expressive video-to-video generation into one unified system.
- Griffin uses a Continuous Conversational Modeling engine and an Audio-Visual Generation engine that operate concurrently for seamless interaction.
- It achieved top scores on NVIDIA's VideoFDB benchmark for both generation and perception tracks, nearing human-level performance.
- Key capabilities include full-duplex conversation, behavior and emotion modeling, full-scene generation, and temporal understanding.
- Griffin-Lite is available as a research preview to select testers; broader release will follow after addressing safety and disclosure concerns.
- Future applications include tutoring, rehearsing conversations, and customer support, enabling natural human-computer interaction.