Core idea
Extends rCM to the autoregressive video setting: TF-based continuous-time CMs + SF-based DMD, via a custom masked FlashAttention-2 JVP kernel for efficient causal training. ~10× faster convergence; 2-step generation on a 1.3B model; applied to streaming video + action-conditioned world models (Cosmos). Hits my world-models + inference-efficiency interests at once.