Hongwei Hu

Directly conditioning video diffusion on audio wastes massive capacity on static background and identity pixels—routing control transitively through a causal motion latent distilled under a single frozen video teacher achieves real-time streaming at 15.4 FPS with zero fidelity loss.