What it is
Extends rCM to autoregressive video diffusion by pairing teacher-forcing consistency distillation with self-forcing DMD for few-step streaming video generation and world models. Instead of relying on KD or self-forcing alone, it uses teacher-forcing consistency models as offline, mode-covering initialization, then self-forcing DMD for on-policy refinement in causal video diffusion.
Key results
- 84.63 VBench-T2V with distilled 2-step causal Wan2.1-1.3B
- 84.63 VBench-T2V also at 1-step, leading reported streaming results
- Teacher-forcing sCM/MeanFlow converges 10× faster than dCM
Shows that offline teacher-forcing consistency training is the strongest setup for stabilizing and improving on-policy self-forcing distillation in autoregressive video diffusion.