Generative video has mostly been a wait-and-see affair: type a prompt, go make a coffee, come back to a clip. A new model called Griffin is trying to kill the waiting entirely. It streams 720p video in 320-millisecond chunks starting from a single reference image, with audio-to-video latency averaging 0.43 seconds on Nvidia H100s. That is fast enough to feel live.
What “real-time” changes
Until now, AI video has been a production tool: slow, batch, offline. Real-time generation turns it into something else entirely, an interactive medium. Think live avatars that respond as you speak, game worlds generated on the fly, video calls with a synthetic presenter that reacts in the moment. The 0.43-second audio-to-video figure matters most here, because lip-sync and responsiveness are what separate an uncanny puppet from something that reads as present.
The obvious double edge
Everything that makes this exciting also makes it dangerous. Real-time, low-latency video from a single image is a deepfake engine by another name, and the gap between “impressive demo” and “convincing live impersonation on a video call” is uncomfortably small. It is no accident that Griffin is currently restricted to select early testers pending safety evaluations, with planned disclosure features. The honest version of this story is that the capability and the risk arrived in the same package, as they usually do.
Why it matters for creators and businesses
If you make content, real-time generation eventually means producing video at the speed of conversation, which collapses both cost and turnaround. If you run a business, it means customer-facing synthetic presenters that are cheap and instant. And if you are anyone with a face, it is one more reason the “is this person real” question stops being paranoid and becomes basic hygiene.
What this means
Real-time generative video was the obvious next milestone, and it looks close. Treat the restricted release as a brief window to get your guard up: verification habits, agreed code-words for sensitive calls, scepticism about live video from unknown numbers. The tools to make this are arriving faster than the tools to catch it.
Sources: Griffin technical disclosure and contemporaneous reporting, early October 2026.