What it is
Aligns initialization and DMD targets in autoregressive video distillation, then adds a CD constraint to counter DMD's mode collapse and preserve coverage. Instead of treating init and DMD as separate stages, aligns them to the same teacher distribution and adds CD to DMD so mode-seeking refinement keeps mode coverage.
Key results
- 84.83 VBench, beating Causal-Forcing's 84.38 and Self-Forcing's 84.21
- 0.69 coverage vs 0.53 for Self-Forcing and 0.29 for Causal-Forcing
- 1.3B-teacher variant gets 84.54, topping Wan-14B baselines
Shows that distributional alignment and an explicit mode-covering constraint matter more than simply using a larger DMD teacher.