🎯 Overview

We develop CHORDS, a training-free and model-agnostic diffusion sampling acceleration strategy via multi-core parallelism. CHORDS views diffusion sampling as a multi-core pipeline, where slower yet accurate solvers progressively rectify faster solvers through a theoretically justified inter-core communication mechanism.

CHORDS significantly accelerates sampling across diverse large-scale image and video diffusion models, yielding up to 2.1x speedup with four cores, improving by 50% over baselines, and 2.9x speedup with eight cores, all without quality degradation.

Please give us a star ⭐ if you find our work interesting!

🛠 Dependencies

Our code relies on the following core packages:

torch
transformers
diffusers
hydra-core
imageio
imageio-ffmpeg

For the specific versions of these packages that have been verified as well as some optional dependencies, please refer to requirements.txt. We recommend creating a new virual environment via the following procedure:

conda create -n chords python=3.10
conda activate chords
pip install -r requirements.txt

🚀 Running Inference

Prior to running inference pipeline, please make sure that the models have been downloaded from 🤗 huggingface. We provide the download script for some example models for both image and video generation in download.py.

We use hydra to organize different hyperparameters for the image/video diffusion model as well as the sampling algorithm. The default configurations can be found under configs folder. The entries to launch the sampling for image and video generation are run_image_mp.py and run_video_mp.py, respectively.

⭐ Text-to-Image (T2I)

The command below is an example to perform image generation on Flux using CHORDS on 8 GPUs.

python run_image_mp.py \
    model=flux \
    ngpu=8 \
    output_base_path=output_samples_flux \
    prompt_file=prompts/image_demo.txt \
    algo=chords \
    algo.num_cores=8 \
    algo.init_t=0-2-4-8-16-24-32-40 \
    algo.stopping_kwargs.criterion=core_index \
    algo.stopping_kwargs.index=2

For model we currently support:

• flux: Flux
• sd3-5: Stable Diffusion 3.5-Large

ngpu corresponds to the number of GPUs to use in parallel.

output_base_path is the directory to save the generated samples.

prompt_file stores the list of prompts, each per line, that will be sequentially employed to generate each image.

algo.init_t refers to the initialization sequence of CHORDS, as we have elaborated in the paper.

algo.num_cores refers to the number of cores, where we currently designate one core as one GPU since it already almost reaches the compute bound, therefore it should equal to ngpu.

algorithm.stopping_kwargs defines the early-stopping criteria, where algo.stopping_kwargs.criterion can be

• core_index, which forces to return the output produced by the algo.stopping_kwargs.index-th fastest core. We recommend using 1 or 2.
• residual, which adaptively returns the output when the residual falls below certain algo.stopping_kwargs.threshold

For full functionality of the script, please refer to the arguments and their default values (such as the number of inference steps, the resolution of the image, etc.) under the configs folder, which is parsed by hydra.

We also provide the script for the single-core sequential sampling baseline to facilitate comparison as follows:

python run_image_mp.py \
    model=flux \
    ngpu=1 \
    output_base_path=output_samples_flux \
    prompt_file=prompts/image_demo.txt \
    algo=sequential

⭐ Text-to-Video (T2V)

Similarly, the following script can be used for video generation with CHORDS:

python run_video_mp.py \
    model=hunyuan \
    ngpu=8 \
    output_base_path=output_samples_hunyuan \
    prompt_file=prompts/video_demo.txt \
    algo=chords \
    algo.num_cores=8 \
    algo.init_t=0-2-4-8-16-24-32-40 \
    algo.stopping_kwargs.criterion=core_index \
    algo.stopping_kwargs.index=1

where for model we currently support:

• hunyuan: HunyuanVideo
• wan2-1: Wan2.1
• cogvideo: CogVideo1.5X-5B

We recommend using 1 or 2 for algo.stopping_kwargs.index. If the video quality does not meet expectation, you can try increasing algo.stopping_kwargs.index to a slightly larger integer value.

📌 Citation

Please consider citing our work if you find it useful:

@inproceedings{han2025chords,
  title={CHORDS: Diffusion Sampling Accelerator with Multi-core Hierarchical ODE Solvers},
  author={Han, Jiaqi and Ye, Haotian and Li, Puheng and Xu, Minkai and Zou, James and Ermon, Stefano},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={19386--19395},
  year={2025}
}

🧩 Contact and Community Contribution

If you have any question, welcome to contact me at:

Jiaqi Han: jiaqihan@stanford.edu

🔥 We warmly welcome community contributions for e.g. supporting more models! Please open/submit a PR if you are interested!