NoizAI's AMBIT Pairs Executable Scene Plans With a Self-Distilling Transfusion for First-Order-Ambisonics Spatial Audio

AMBIT turns language into executable scene plans specifying source content, activity and 3D trajectories, then renders First-Order Ambisonics audio through a FOA-aware continuous VAE with asymmetric grouped KL divergence and spatial-covariance matching. An on-policy self-distilling Transfusion couples autoregressive planning with rectified-flow rendering, and editing uses contrastive features plus reference latents to keep untouched sources intact across speech, music and SFX.