MeiGen-AI/OPSD-V

On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators

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, which then degrades the quality of all subsequent chunks, leading to poor long-horizon visual quality.

How it works

It uses an on-policy self-distillation paradigm for post-training. The "student" model follows the exact deployment rollout—generating chunks and updating its own KV cache—while a "teacher" model is evaluated at the same denoising states. However, the teacher is provided with a cleaner, AR-consistent temporal cache built from real long-video context. By matching the student's velocity predictions to the teacher's, the model receives dense corrective targets that improve stability without changing the original few-step sampler or 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 maintains its own generated KV cache during training to mimic real-world inference.
  • Cleaner teacher context: Uses real-video history to provide a high-quality reference for the teacher model.
  • Preserved inference path: Maintains the original 4-step AR sampler, ensuring no increase in latency during deployment.
  • Memory-efficient training: Implements chunk-wise backward passes, FSDP, and LoRA/EMA support to make long-video training feasible.

Related

  • Project
    chrisliu298/awesome-on-policy-distillation

    A curated collection of papers, frameworks, and tools for on-policy distillation, a post-training technique that trains student LLMs on their own generated samples to bridge the train-inference distribution gap.

  • Project
    vvvvvjdy/D-OPSD

    D-OPSD is an on-policy self-distillation framework that allows step-distilled diffusion models to be continuously tuned for new styles and concepts while preserving their fast inference capabilities.

  • Project
    thunlp/OPD

    A research project and toolkit for analyzing and improving On-Policy Distillation (OPD) of LLMs, providing strategies to recover failing distillation and token-level diagnostics.

  • Project
    thu-ml/Causal-Forcing

    A framework for high-quality real-time interactive video generation using Causal ODE or Causal Consistency Distillation to enable efficient few-step autoregressive diffusion.

  • Project
    NVlabs/rcm

    A score-regularized continuous-time consistency model for distilling large video diffusion models, enabling high-quality, diverse video generation in 2-4 steps.