Modelpublic

zhuhz22/Causal-Forcing

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

21likes

Model Card

Causal Forcing

Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation

<p align="center"> <p align="center"> <div> <a href="https://zhuhz22.github.io/" target="blank">Hongzhou Zhu*</a><sup></sup>, <a href="https://gracezhao1997.github.io/" target="blank">Min Zhao*</a><sup></sup> , <a href="https://guandehe.github.io/" target="blank">Guande He</a><sup></sup>, <a href="https://scholar.google.com/citations?user=dxN1X0AAAAJ&hl=en" target="blank">Hang Su</a><sup></sup>, <a href="https://zhenxuan00.github.io/" target="blank">Chongxuan Li</a><sup></sup> , <a href="https://ml.cs.tsinghua.edu.cn/~jun/index.shtml" target="_blank">Jun Zhu</a><sup></sup> </div> <div> <sup></sup>Tsinghua University & Shengshu & UT Austin </div>

</div> </p> <h3 align="center"><a href="https://arxiv.org/abs/2602.02214">Paper</a> | <a href="https://thu-ml.github.io/CausalForcing.github.io">Website</a> | <a href="https://github.com/thu-ml/Causal-Forcing">Code</a> | <a href="https://huggingface.co/zhuhz22/Causal-Forcing/tree/main">Models</a></h3> </p>

Causal Forcing significantly outperforms Self Forcing in both visual quality and motion dynamics, while keeping the same training budget and inference efficiency—enabling real-time, streaming video generation on a single RTX 4090.

Abstract

To achieve real-time interactive video generation, current methods distill pretrained bidirectional video diffusion models into few-step autoregressive (AR) models, facing an architectural gap when full attention is replaced by causal attention. We propose Causal Forcing that uses an AR teacher for ODE initialization, thereby bridging the architectural gap. Empirical results show that our method outperforms all baselines across all metrics, surpassing the SOTA Self Forcing by 19.3% in Dynamic Degree, 8.7% in VisionReward, and 16.7% in Instruction Following.

Quick Start

The inference environment is identical to Self Forcing, so you can migrate directly using our configs and model.

Installation

bash

conda create -n causal_forcing python=3.10 -y
conda activate causal_forcing
pip install -r requirements.txt
pip install git+https://github.com/openai/CLIP.git
pip install flash-attn --no-build-isolation
python setup.py develop

Download checkpoints

bash

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
hf download zhuhz22/Causal-Forcing chunkwise/causal_forcing.pt --local-dir checkpoints
hf download zhuhz22/Causal-Forcing framewise/causal_forcing.pt --local-dir checkpoints

CLI Inference

We open-source both the frame-wise and chunk-wise models; the former is a setting that Self Forcing has chosen not to release.

Frame-wise model (higher dynamic degree and more expressive):

bash

python inference.py \
  --config_path configs/causal_forcing_dmd_framewise.yaml \
  --output_folder output/framewise \
  --checkpoint_path  checkpoints/framewise/causal_forcing.pt \
  --data_path prompts/demos.txt \
  --use_ema
    # Note: this frame-wise config not in Self Forcing; if using its framework, migrate this config too.

Chunk-wise model (more stable):

bash

python inference.py \
  --config_path configs/causal_forcing_dmd_chunkwise.yaml \
  --output_folder output/chunkwise \
  --checkpoint_path checkpoints/chunkwise/causal_forcing.pt \
  --data_path prompts/demos.txt

Training

<details> <summary> Stage 1: Autoregressive Diffusion Training (Can skip by using our pretrained checkpoints. Click to expand.)</summary>

First download the dataset (we provide a 6K toy dataset here):

bash

hf download zhuhz22/Causal-Forcing-data  --local-dir dataset
python utils/merge_and_get_clean.py

Then train the AR-diffusion model:

—Framewise:

bash

    torchrun --nnodes=8 --nproc_per_node=8 --rdzv_id=5235 \
    --rdzv_backend=c10d \
    --rdzv_endpoint $MASTER_ADDR \
    train.py \
    --config_path configs/ar_diffusion_tf_framewise.yaml \
    --logdir logs/ar_diffusion_framewise

—Chunkwise:

bash

    torchrun --nnodes=8 --nproc_per_node=8 --rdzv_id=5235 \
    --rdzv_backend=c10d \
    --rdzv_endpoint $MASTER_ADDR \
    train.py \
    --config_path configs/ar_diffusion_tf_chunkwise.yaml \
    --logdir logs/ar_diffusion_chunkwise

</details>

<details> <summary> Stage 2: Causal ODE Initialization (Can skip by using our pretrained checkpoints. Click to expand.)</summary>

If you have skipped Stage 1, you need to download the pretrained models:

bash

hf download zhuhz22/Causal-Forcing framewise/ar_diffusion.pt --local-dir checkpoints
hf download zhuhz22/Causal-Forcing chunkwise/ar_diffusion.pt --local-dir checkpoints

In this stage, train ODE initialization models:

—Frame-wise:

bash

  torchrun --nnodes=8 --nproc_per_node=8 --rdzv_id=5235 \
    --rdzv_backend=c10d \
    --rdzv_endpoint $MASTER_ADDR \
    train.py \
    --config_path configs/causal_ode_framewise.yaml \
    --logdir logs/causal_ode_framewise

</details>

Stage 3: DMD

This stage is compatible with Self Forcing training, so you can migrate seamlessly by using our configs and checkpoints.

—Frame-wise model:

bash

  torchrun --nnodes=8 --nproc_per_node=8 --rdzv_id=5235 \
    --rdzv_backend=c10d \
    --rdzv_endpoint $MASTER_ADDR \
    train.py \
    --config_path configs/causal_forcing_dmd_framewise.yaml \
    --logdir logs/causal_forcing_dmd_framewise

Acknowledgements

This codebase is built on top of the open-source implementation of CausVid, Self Forcing and the Wan2.1 repo.

References

bibtex

@article{zhu2026causal,
  title={Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation},
  author={Zhu, Hongzhou and Zhao, Min and He, Guande and Su, Hang and Li, Chongxuan and Zhu, Jun},
  journal={arXiv preprint arXiv:2602.02214},
  year={2026}
}