Geometry-Aware Dataset Condensation for Diffusion Model Training

Preparation

Environment

For convenience, you can create the Conda virtual environment through:

conda env create -f environment.yaml
conda activate GADC

Downloading Dataset

Download ImageNet-1K from https://www.image-net.org/challenges/LSVRC/index.php and locate it at:

../imagenet/

Downloading Reference Batches

Download reference batches for evaluation from https://openaipublic.blob.core.windows.net/diffusion/jul-2021/ref_batches/imagenet/256/VIRTUAL_imagenet256_labeled.npz and https://openaipublic.blob.core.windows.net/diffusion/jul-2021/ref_batches/imagenet/512/VIRTUAL_imagenet512.npz, then locate them at:

evaluation/reference/

Data Selection

Run the following command:

cd selection
sh run.sh

Diffusion Model Training

Run the following command:

cd training
sh run.sh

Image Generation and Evaluation

Run the following command:

cd evaluation
sh run.sh
sh test.sh

Then you can get the FID, Inception Score, Precision and Recall.

BibTeX

@article{cui2026geometry,
  title={Geometry-Aware Dataset Condensation for Diffusion Model Training},
  author={Cui, Xiao and Qin, Yulei and Zhu, Mo and Zhou, Wengang and Li, Hongsheng and Li, Houqiang},
  journal={arXiv preprint arXiv:2606.05883},
  year={2026}
}

Acknowledgement

Thanks to Rui Huang et al. for their nice work, Accelerating Diffusion Model Training under Minimal Budgets: A Condensation-Based Perspective！