OPSD-V Checkpoints

This repository hosts checkpoints for OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators.

• Project page: https://kumapowerliu.github.io/opsd-v/
• Paper: https://kumapowerliu.github.io/opsd-v/assets/opsdv.pdf
• Code: https://github.com/MeiGen-AI/opsd-v

Download

Install the Hugging Face Hub CLI and download the complete checkpoint repository into the root of the OPSD-V code release:

pip install -U "huggingface_hub[cli]"
hf download MeiGen-AI/OPSD-V --local-dir .

Alternatively, use Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="MeiGen-AI/OPSD-V",
    local_dir=".",
)

Files

Path	Description
checkpoints/longlive_base.pt	LongLive base few-step autoregressive generator.
checkpoints/longlive_lora.pt	Original LongLive LoRA adapter used as the starting point.
checkpoints/opsdv_longlive_lora.pt	OPSD-V LoRA checkpoint for the LongLive backbone.
checkpoints/self_forcing_dmd_ema_as_generator.pt	Self-Forcing DMD/EMA generator used as the base model.
checkpoints/opsdv_self_forcing_lora.pt	OPSD-V LoRA checkpoint for the Self-Forcing backbone.

Usage

Place these files under the checkpoints/ directory of the OPSD-V code release and use the provided training or inference configs.

checkpoints/
  longlive_base.pt
  longlive_lora.pt
  opsdv_longlive_lora.pt
  self_forcing_dmd_ema_as_generator.pt
  opsdv_self_forcing_lora.pt

See the code repository README for detailed setup, inference, and training commands.

Citation

@misc{liu2026opsdv,
  title  = {OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators},
  author = {Liu, Hongyu and Wang, Chun and Gao, Feng and He, Xuanhua and Ma, Yue and Wan, Ziyu and Zhang, Yong and Wei, Xiaoming and Chen, Qifeng},
  year   = {2026},
  note   = {Preprint}
}

License

The OPSD-V code release is Apache-2.0. Please also follow the licenses and usage terms of the corresponding upstream models and datasets.