Marionette World Model Decouples 3D Geometry from Neural Appearance
August 13, 2026
Marionette predicts a 276-dimensional 3D world state including articulated skeletons to maintain long-horizon consistency. It delegates geometric computation to a zero-parameter renderer while using a neural model solely for appearance synthesis.
HOW THIS AFFECTS YOU
●
builderYou can build more controllable interactive environments by leveraging explicit 3D state representations.
●
researcherYou can explore ways to improve long-term world model stability by decoupling physics from rendering.