Real-Time Motion Synthesis for Previsualization

Previsualization is supposed to be fast. In practice, even rough previs takes days to produce using traditional keyframe animation tools.

We built a prototype we call LiveBlock that connects motion capture data directly to a USD scene graph, with AI-driven procedural motion filling gaps in real time. A director can walk a scene wearing a simple IMU suit and see their path reflected in the previs camera within a few hundred milliseconds.

The gap-filling problem

Sparse IMU capture is fast and cheap but gives you body position estimates with significant drift and no hand or face data. We trained a motion prior on a cleaned version of AMASS — roughly 40 hours of full-body motion capture — and used it to infer plausible limb trajectories from sparse joint estimates in real time. The model runs on a single RTX 4090 at 60fps.

What this changes on set

In the three productions where we have run LiveBlock, the previs phase shortened by about 40%. Real-time feedback changes the creative process, not just the timeline.