Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion