aejion/AccVideo-WanX-I2V-480P-14B

AccVideo: Accelerating Video Diffusion Model with Synthetic Dataset

This repository is the official PyTorch implementation of AccVideo. AccVideo is a novel efficient distillation method to accelerate video diffusion models with synthetic datset. Our method is 8.5x faster than HunyuanVideo.

  

🔥🔥🔥 News

• —Jun 3, 2025: We release the inference code and model weights of AccVideo based on WanXI2V-480P-14B.
• —May 26, 2025: We release the inference code and model weights of AccVideo based on WanXT2V-14B.
• —Mar 31, 2025: ComfyUI-Kijai (FP8 Inference): ComfyUI-Integration by Kijai
• —Mar 26, 2025: We release the inference code and model weights of AccVideo based on HunyuanT2V.

🎥 Demo (Based on HunyuanT2V)

https://github.com/user-attachments/assets/59f3c5db-d585-4773-8d92-366c1eb040f0

🎥 Demo (Based on WanXT2V-14B)

https://github.com/user-attachments/assets/ff9724da-b76c-478d-a9bf-0ee7240494b2

🎥 Demo (Based on WanXI2V-480P-14B)

📑 Open-source Plan

• —[x] Inference
• —[x] Checkpoints
• —[ ] Multi-GPU Inference
• —[ ] Synthetic Video Dataset, SynVid
• —[ ] Training

🔧 Installation

The code is tested on Python 3.10.0, CUDA 11.8 and A100.

conda create -n accvideo python==3.10.0
conda activate accvideo

pip install torch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
pip install flash-attn==2.7.3 --no-build-isolation
pip install "huggingface_hub[cli]"

🤗 Checkpoints

To download the checkpoints (based on HunyuanT2V), use the following command:

# Download the model weight
huggingface-cli download aejion/AccVideo --local-dir ./ckpts

To download the checkpoints (based on WanX-T2V-14B), use the following command:

# Download the model weight
huggingface-cli download aejion/AccVideo-WanX-T2V-14B --local-dir ./wanx_t2v_ckpts

To download the checkpoints (based on WanX-I2V-480P-14B), use the following command:

# Download the model weight
huggingface-cli download aejion/AccVideo-WanX-I2V-480P-14B --local-dir ./wanx_i2v_ckpts

🚀 Inference

We recommend using a GPU with 80GB of memory. We use AccVideo to distill Hunyuan and WanX.

Inference for HunyuanT2V

To run the inference, use the following command:

export MODEL_BASE=./ckpts
python sample_t2v.py \
    --height 544 \
    --width 960 \
    --num_frames 93 \
    --num_inference_steps 5 \
    --guidance_scale 1 \
    --embedded_cfg_scale 6 \
    --flow_shift 7 \
    --flow-reverse \
    --prompt_file ./assets/prompt.txt \
    --seed 1024 \
    --output_path ./results/accvideo-544p \
    --model_path ./ckpts \
    --dit-weight ./ckpts/accvideo-t2v-5-steps/diffusion_pytorch_model.pt

The following table shows the comparisons on inference time using a single A100 GPU:

Inference for WanXT2V

To run the inference, use the following command:

python sample_wanx_t2v.py \
       --task t2v-14B \
       --size 832*480 \
       --ckpt_dir ./wanx_t2v_ckpts \
       --sample_solver 'unipc' \
       --save_dir ./results/accvideo_wanx_14B \
       --sample_steps 10

The following table shows the comparisons on inference time using a single A100 GPU:

Inference for WanXI2V-480P

To run the inference, use the following command:

python sample_wanx_i2v.py \
       --task i2v-14B \
       --size 832*480 \
       --ckpt_dir ./wanx_i2v_ckpts \
       --sample_solver 'unipc' \
       --save_dir ./results/accvideo_wanx_i2v_14B \
       --sample_steps 10

The following table shows the comparisons on inference time using a single A100 GPU:

🏆 VBench Results

We report VBench evaluation results for our distilled models. We utilized the respective augmented prompts provided by the VBench team to generate videos. (HunyuanVideo augmented prompts for AccVideo-HunyuanT2V and WanX augmented prompts for AccVideo-WanXT2V)

🔗 BibTeX

If you find AccVideo useful for your research and applications, please cite using this BibTeX:

@article{zhang2025accvideo,
    title={AccVideo: Accelerating Video Diffusion Model with Synthetic Dataset},
    author={Zhang, Haiyu and Chen, Xinyuan and Wang, Yaohui and Liu, Xihui and Wang, Yunhong and Qiao, Yu},
    journal={arXiv preprint arXiv:2503.19462},
    year={2025}
}

Acknowledgements

The code is built upon FastVideo and HunyuanVideo, we thank all the contributors for open-sourcing.