MeiGen-AI/OPSD-V

What it solves

OPSD-V addresses the problem of error accumulation and weakened motion dynamics in few-step autoregressive (AR) video generators. In these models, small artifacts or semantic drift in one generated chunk are written back into the KV cache and affect all subsequent chunks, leading to degraded visual quality and unstable motion over long horizons.

How it works

It uses an on-policy self-distillation paradigm for post-training. During training, a "student" model performs the actual deployment rollout, generating chunks and updating its own KV cache. Simultaneously, a "teacher" model is evaluated at the same noisy states as the student but is provided with a cleaner, AR-consistent temporal cache derived from real long-video context. By matching the student's velocity predictions to the teacher's, the model learns to correct its trajectory without changing the original few-step sampler or the inference-time cache mechanism.

Who it’s for

Researchers and developers working on autoregressive video generation, specifically those looking to improve the long-term stability and motion quality of few-step diffusion models.

Highlights

• On-policy student rollout: The student follows the exact trajectory it will use during inference.
• Cleaner teacher context: Uses real-video history to provide high-quality supervision targets.
• Preserved inference path: Maintains the original 4-step AR sampler, ensuring no increase in latency.
• Memory-efficient training: Employs chunk-wise backward passes, FSDP, and LoRA to make long-video rollouts computationally feasible.