OPSD-V cuts compounding video generation errors

While few-step autoregressive video generators promise near-real-time long video generation, their practical output is severely hampered by compounding trajectory errors. OPSD-V offers a zero-overhead post-training fix by using real-world video anchors to ground the student's on-policy generations. This approach directly tackles the critical bottleneck of error propagation in autoregressive sequence generation with zero inference overhead. Furthermore, it demonstrates consistent, generalizable improvements in visual quality and motion dynamics across diverse model architectures.