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.
- Jun 10, 2025: CivitAI and Huggingface: WanX-14B LoRA extracted by Kijai, compatible for use with other LoRAs.
- 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.
teaser.mp4
wanx-14B.mp4
wanx_i2v.mp4
- Inference
- Checkpoints
- Multi-GPU Inference
- Synthetic Video Dataset, SynVid
- Training
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]"
To download the checkpoints (based on HunyuanT2V), use the following command:
# Download the model weight
huggingface-cli download aejion/AccVideo --local-dir ./ckptsTo 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_ckptsTo 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_ckptsWe recommend using a GPU with 80GB of memory. We use AccVideo to distill Hunyuan and WanX.
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.ptThe following table shows the comparisons on inference time using a single A100 GPU:
| Model | Setting(height/width/frame) | Inference Time(s) |
|---|---|---|
| HunyuanVideo | 720px1280px129f | 3234 |
| Ours | 720px1280px129f | 380(8.5x faster) |
| HunyuanVideo | 544px960px93f | 704 |
| Ours | 544px960px93f | 91(7.7x faster) |
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 10The following table shows the comparisons on inference time using a single A100 GPU:
| Model | Setting(height/width/frame) | Inference Time(s) |
|---|---|---|
| WanX | 480px832px81f | 932 |
| Ours | 480px832px81f | 97(9.6x faster) |
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 10The following table shows the comparisons on inference time using a single A100 GPU:
| Model | Setting(height/width/frame) | Inference Time(s) |
|---|---|---|
| WanX-I2V | 480px832px81f | 768 |
| Ours | 480px832px81f | 112(6.8x faster) |
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)
| Model | Setting(height/width/frame) | Total Score | Quality Score | Semantic Score | Subject Consistency | Background Consistency | Temporal Flickering | Motion Smoothness | Dynamic Degree | Aesthetic Quality | Image Quality | Object Class | Multiple Objects | Human Action | Color | Spatial Relationship | Scene | Appearance Style | Temporal Style | Overall Consistency |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| AccVideo-HunyuanT2V | 544px960px93f | 83.26% | 84.58% | 77.96% | 94.46% | 97.45% | 99.18% | 98.79% | 75.00% | 62.08% | 65.64% | 92.99% | 67.33% | 95.60% | 94.11% | 75.70% | 54.72% | 19.87% | 23.71% | 27.21% |
| AccVideo-WanXT2V | 480px832px81f | 85.95% | 86.62% | 83.25% | 95.02% | 97.75% | 99.54% | 97.95% | 93.33% | 64.21% | 68.42% | 98.38% | 86.58% | 97.40% | 92.04% | 75.68% | 59.82% | 23.88% | 24.62% | 27.34% |
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}
}The code is built upon FastVideo and HunyuanVideo, we thank all the contributors for open-sourcing.