Hyperbolic Dataset Distillation (NeurIPS 2025)

The first dataset distillation framework that leverages hyperbolic geometry to preserve hierarchical data structures. The repository is based on geoopt , DM and IDM. If you use this code, please consider citing them.

• Dec 2025: Code was released.
• Oct 2025: Project page was released.
• Sept 2025: Our paper has been accepted to NeurIPS 2025!
• May 2025: Preprint was released.

🎯 Key Contributions

• Hyperbolic Geometry for Dataset Distillation Introduces the first dataset distillation framework in hyperbolic space, enabling natural modeling of hierarchical data structures.

• Centroid Alignment via Geodesic Distance Minimizes the Lorentzian hyperbolic distance between Fréchet means of real and synthetic datasets, ensuring semantic and geometric consistency.

• Hierarchical Sample Weighting Derives analytic weighting that prioritizes prototype-like (root-level) samples while attenuating noisy, peripheral ones.

• Efficient Hierarchical Pruning Demonstrates that only 20% of the original data is sufficient to maintain comparable accuracy, highlighting hyperbolic redundancy reduction.

• Broad Compatibility & Superior Results Integrates smoothly with DM and IDM, yielding consistent gains in classification accuracy and training stability across diverse datasets.

Usage

DM with HDD

cd DM-HDD
python main_DM.py  --dataset CIFAR10  --model ConvNet  --ipc 10  --dsa_strategy color_crop_cutout_flip_scale_rotate  --init real  --lr_img 1  --num_exp 5  --num_eval 5 

IDM with HDD

cd IDM-HDD
python3 -u IDM_cifar10.py --dataset CIFAR10 --model ConvNet --ipc 10 --dsa_strategy color_crop_cutout_flip_scale_rotate --init real --lr_img 1 --num_exp 5 --num_eval 5 --net_train_real --eval_interval 100 --outer_loop 1 --mismatch_lambda 0 --net_decay --embed_last 1000 --syn_ce --ce_weight 0.5 --train_net_num 1 --aug

Citing HDD

If you find this project useful for your research, please use the following BibTeX entry.

@inproceedings{li2025hdd,
  title={Hyperbolic Dataset Distillation},
  author={Li, Wenyuan and Li, Guang and Maeda, Keisuke and Ogawa, Takahiro and Haseyama, Miki},
  booktitle={Proceedings of the Advances in Neural Information Processing Systems (NeurIPS)},
  year={2025}
}