Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation
Causal Forcing++ proposes a scalable pipeline for real-time video generation by using Causal Consistency Distillation (Causal CD) to initialize few-step autoregressive diffusion models. This method solves limitations in existing initialization techniques, which were either misaligned or too costly. Causal CD efficiently learns the correct AR flow map by using only one online teacher ODE step per training iteration, significantly reducing computational cost and improving generation quality.