[CVPR2025] Dataset Distillation with Neural Characteristic Function: A Minmax Perspective

Official PyTorch implementation of the paper "Dataset Distillation with Neural Characteristic Function" (NCFM) in CVPR 2025.

🔥 News

• [2025/03/02] The code of our paper has been released.
• [2025/02/27] Our NCFM paper has been accepted to CVPR 2025 (Rating: 555). Thanks!

🚀 Pipeline

Here's an overview of the process behind our Neural Characteristic Function Matching (NCFM) method:

🔍 TODO

We are currently organizing all the code. Stay tuned!

• Distillation code
• Evaluation code
• Sampling network
• Config files
• Pretrained models
• Distilled datasets
• Continual learning code
• Project page

🛠️ Getting Started

To get started with NCFM, follow the installation instructions below.

1. Clone the repo

git clone https://github.com/gszfwsb/NCFM.git

2. Install dependencies

pip install -r requirements.txt

3. Pretrain the models yourself, or download the pretrained_models from Google Drive.

cd pretrain
torchrun --nproc_per_node={n_gpus} --nnodes=1 pretrain_script.py --gpu={gpu_ids} --config_path=../config/{ipc}/{dataset}.yaml

4. Condense

cd condense 
torchrun --nproc_per_node={n_gpus} --nnodes=1 condense_script.py --gpu={gpu_ids} --ipc={ipc} --config_path=../config/{ipc}/{dataset}.yaml

5. Evaluation

cd evaluation 
torchrun --nproc_per_node={n_gpus} --nnodes=1 evaluation_script.py --gpu={gpu_ids} --ipc={ipc} --config_path=../config/{ipc}/{dataset}.yaml --load_path={distilled_dataset.pt}

📘 Example Usage

1. CIFAR-10

#ipc50
cd condense
torchrun --nproc_per_node=8 --nnodes=1 --master_port=34153 condense_script.py --gpu="0,1,2,3,4,5,6,7" --ipc=50 --config_path=../config/ipc50/cifar10.yaml

2. CIFAR-100

#ipc10
cd condense
torchrun --nproc_per_node=8 --nnodes=1 --master_port=34153 condense_script.py --gpu="0,1,2,3,4,5,6,7" --ipc=10 --config_path=../config/ipc10/cifar100.yaml

📮 Contact

If you have any questions, please contact Shaobo Wang([email protected]).

📌 Citation

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

@misc{wang2025datasetdistillationneuralcharacteristic,
      title={Dataset Distillation with Neural Characteristic Function: A Minmax Perspective}, 
      author={Shaobo Wang and Yicun Yang and Zhiyuan Liu and Chenghao Sun and Xuming Hu and Conghui He and Linfeng Zhang},
      year={2025},
      eprint={2502.20653},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2502.20653}, 
}

Acknowledgement

We sincerely thank the developers of the following projects for their valuable contributions and inspiration: MTT, DATM, DC/DM, IDC, SRe2L, RDED, DANCE. We draw inspiration from these fantastic projects!