The codebase implements Causal Forcing and Causal Forcing++, techniques that use Causal ODE or Causal Consistency Distillation to drive asymmetric DMD for video generation. It supports both chunk-wise and frame-wise models for text-to-video and image-to-video tasks, with the latter natively unifying both generation modes. The method aims to achieve high visual quality and motion dynamics while maintaining training budget and inference efficiency.