Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation

Researchers introduced Causal Forcing++, a few-step autoregressive diffusion distillation method for real-time interactive video generation.

It uses causal consistency distillation to initialize frame-wise 1-2 step models more efficiently than prior approaches like Causal Forcing.

The method improves benchmark quality, halves first-frame latency, and cuts Stage 2 training cost by about 4x.

It also extends to action-conditioned world model generation, pushing video generation toward faster and more controllable use.