AMBIT Pairs Scene Plans With Self-Distilling FOA Rendering
TL;DR
- AMBIT unifies First-Order Ambisonics generation and editing under one Transfusion backbone that couples autoregressive planning with rectified-flow rendering.
- Executable ScenePlans specify source content, activity, and 3D trajectories; for editing, contrastive features and reference latents preserve untouched sources.
- Joint on-policy self-distillation refines planning on student-generated prefixes and rendering at student-reached sampling states, with target-conditioned teachers.
AMBIT, a unified framework for native First-Order Ambisonics generation and editing, couples autoregressive scene planning with rectified-flow rendering under one Transfusion backbone, according to a preprint posted to arXiv on October 25 by Kam Man Wu, Yu Zhang, Pengjun Fang, Mingfei Chen, Weijia Chen, Zeyue Tian, and Qifeng Chen.
Executable "ScenePlans" specify "source content, activity, and 3D trajectories," and for editing the plan "describes the full target scene, including retained sources" while contrastive features ground the requested change and "full reference latents supply residual acoustic detail for preservation." The spatial side rests on an FOA-aware continuous VAE that combines "shared-decoder omnidirectional recovery, asymmetric grouped KL, and spatial covariance matching to preserve content and inter-channel structure."
The third piece, joint on-policy self-distillation, has "target-conditioned teachers supervising both modes of the shared student" across planning and rendering. The abstract reports gains in generation content and spatial quality, acoustic fidelity during instruction-guided editing, and unedited-source preservation across speech, music, and sound. No named baselines, datasets, or metric numbers appear in the abstract on the arXiv page.
Originally reported by arxiv.org
Read the original article →Original headline: NoizAI's AMBIT Pairs Executable Scene Plans With a Self-Distilling Transfusion for First-Order-Ambisonics Spatial Audio