Official repository for CaO₂: Rectifying Inconsistencies in Diffusion-Based Dataset Distillation.

(2025.7.7) Upload distilled datasets.

(2025.6.26) Starting to update the repository.

Download the ImageNet dataset, and place it at your intended IMAGENET_PATH.

git clone https://github.com/hatchetProject/CaO2.git
cd CaO2
conda env create -f environment.yaml
conda activate cao2

Download the pretrained DiT model (DiT/XL 256×256), and place it in pretrained_models.

Download the pretrained classification models for RDED evaluation from RDED and place them in data/pretrain_models. We also provide the script for running evaluation without teacher supervision.

We provided an example script in run.sh for the users to run the code. Note that the users will need to modify the variables in the script for it to run properly.

To gain better performance, hyperparameters can be tuned, but the variations are usually marginal.

For direct evaluation and benchmarking, the distilled datasets are uploaded here (>7GB).

• Complete the code upload
• Upload the distilled datasets
• Update code for MAR distillation

Thanks to these amazing repositories: Minimax Diffusion, RDED and many other inspiring works.

If you find this work useful, please consider citing:

@misc{wang2025cao2rectifyinginconsistenciesdiffusionbased,
      title={CaO$_2$: Rectifying Inconsistencies in Diffusion-Based Dataset Distillation}, 
      author={Haoxuan Wang and Zhenghao Zhao and Junyi Wu and Yuzhang Shang and Gaowen Liu and Yan Yan},
      year={2025},
      eprint={2506.22637},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.22637}, 
}