Why Real Time

The future of video models is heading towards instant, real-time generation to reduce creation time and enable new applications. This approach focuses on optimizing for time-to-first-frame by post-training base foundation models to generate video causally and autoregressively. The technical method involves a two-stage distillation process, using teacher forcing to make the model causal and student forcing to speed it up. On-policy distillation is particularly crucial for reducing error accumulation that can compound over a video sequence. These real-time capabilities are foundational for interactive experiences in gaming, robotics, and simulation.