One-Forcing

Towards Stable One-Step Autoregressive Video Generation

Jiaqi Feng^1,* · Justin Cui^2,* · Yuanhao Ban² · Cho-Jui Hsieh^2
1Tsinghua University ²UCLA ^*Equal contribution

Paper | Website | Models/Data

---

One-Forcing enables stable 1-step autoregressive video generation by augmenting DMD-based causal distillation with a shared noised-latent adversarial critic, achieving state-of-the-art 1-step VBench performance and efficient framewise generation.

---

Installation

conda create -n one_forcing python=3.10 -y
conda activate one_forcing
pip install -r requirements.txt
pip install flash-attn --no-build-isolation
python setup.py develop

Inference

Download the trained One-Forcing checkpoint:

hf download JiaqiFeng/OneForcing checkpoints/one_forcing.pt --local-dir .

bash scripts/infer.sh \
  --checkpoint_path checkpoints/one_forcing.pt \
  --prompt_path prompts/demos.txt \
  --output_folder outputs \
  --use_ema

Training

Dataset Preparation

hf download JiaqiFeng/OneForcing --include "clean_data/*" --local-dir .

Download ODE initialized checkpoint

hf download JiaqiFeng/OneForcing checkpoints/framewise/causal_ode.pt --local-dir .

You can refer to Causal Forcing Stage1/2 to train your ODE initialized checkpoint

Download Wan2.1 Base Models

hf download Wan-AI/Wan2.1-T2V-1.3B \
  --local-dir wan_models/Wan2.1-T2V-1.3B
hf download Wan-AI/Wan2.1-T2V-14B \
  --local-dir wan_models/Wan2.1-T2V-14B

One Forcing Training

torchrun --nproc_per_node=8 train.py \
  --config_path config.yaml \
  --generator_ckpt checkpoints/framewise/causal_ode.pt \
  --teacher_model_path wan_models/Wan2.1-T2V-14B \
  --data_path mixkit_latents_lmdb \
  --logdir runs \
  --disable-wandb \
  --no_visualize

Evaluation

Export videos first, then run VBench with your local VBench installation:

python scripts/run_vbench.py \
  --videos_path outputs \
  --full_info_path VBench_full_info.json \
  --output_dir eval/vbench \
  --name one_forcing_framewise

Citation

@article{feng2026oneforcing,
  title={One-Forcing: Towards Stable One-Step Autoregressive Video Generation},
  author={Feng, Jiaqi and Cui, Justin and Ban, Yuanhao and Hsieh, Cho-Jui},
  journal={arXiv preprint arXiv:2605.23458},
  year={2026},
  eprint={2605.23458},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2605.23458}
}

Acknowledgements

This codebase builds on Causal Forcing, Self Forcing, CausVid, and the Wan model family.