The project adapts a pretrained bidirectional diffusion transformer into an autoregressive transformer that generates frames on-the-fly for interactive applications. It extends distribution matching distillation (DMD) to video, compressing a 50-step diffusion model into a 4-step generator. The approach introduces student initialization based on teacher’s ODE trajectories and asymmetric distillation to mitigate error accumulation in autoregressive generation, enabling high-quality long-duration video synthesis.