How Visko Streams Real-Time World Models on Google Cloud TPUs

At a glance

AI startup Visko worked with Google Cloud TPU engineers to bring Orbis, its live video world model, onto TPUs in just 30 days. The model lets viewers change scenes in real time and must generate smooth video and audio without interruption.

Key points

1. 1

   Orbis starts a live video stream from a single prompt, and viewers can change the scene with commands while watching.

2. 2

   On a single TPU device, it runs three models in sequence: video generation, video enhancement, and audio generation.

3. 3

   The goal is to process a segment of video faster than its playback time, at 1.07x speed.

4. 4

   Each scene must follow from the previous one, so segments cannot be created in parallel ahead of time.

Why it matters

This is an example of how real-time, responsive AI video technology works as an actual service.