Live output runs about 2.4 seconds behind your camera. That number isn’t a guess: it’s the sum of four measured stages on an H100.
The four stages
- Filling a 12-frame chunk: about 0.45 s
- Generating it on the GPU: about 0.98 s
- Network and playout: about 0.84 s
- Capture, upload and pose: about 0.15 s
The live model draws 12 frames at a time, so it has to wait for those frames before it can start. Generating them is the largest single step, and the network carries the rest.
Why not one frame at a time? Chunks keep the target steady from frame to frame. Smaller chunks would cut the wait but make identity and lighting drift more.
What comes next
Shorter chunks and lighter models will bring the delay down over time. We’ll publish the new numbers when they’re measured, not before.