Taming Diffusion for Dataset Distillation with High Representativeness

Authors: Lin Zhao, Yushu Wu, Xinru Jiang, Jianyang Gu, Yanzhi Wang, Xiaolin Xu, Pu Zhao, Xue Lin

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our comprehensive experiments demonstrate that D3HR can achieve higher accuracy across different model architectures compared with state-of-the-art baselines in dataset distillation. 5. Main Results 5.1. Experimental Details 5.2. Comparison with State-of-the-art Methods 5.3. Cross-architecture Generalization
Researcher Affiliation Academia 1Northeastern University 2The Ohio State University. Correspondence to: Pu Zhao <EMAIL>, Xue Lin <EMAIL>.
Pseudocode Yes Algorithm 1 D3HR Algorithm
Open Source Code Yes Source code: https://github. com/lin-zhao-reso Lve/D3HR.
Open Datasets Yes For small-scale datasets, we use CIFAR-10 and CIFAR-100 (Krizhevsky et al., 2009) with 32 32 resolution. For large-scale datasets, we use Tiny Image Net (Le & Yang, 2015) with 200 classes (500 images per class, 64 64 size) and Image Net-1K (Deng et al., 2009) with 1, 000 classes (1M images, 224 224 resolution).
Dataset Splits Yes Experiments are conducted on both small-scale and large-scale datasets. For small-scale datasets, we use CIFAR-10 and CIFAR-100 (Krizhevsky et al., 2009) with 32 32 resolution. For large-scale datasets, we use Tiny Image Net (Le & Yang, 2015) with 200 classes (500 images per class, 64 64 size) and Image Net-1K (Deng et al., 2009) with 1, 000 classes (1M images, 224 224 resolution).
Hardware Specification Yes All experiments are conducted on Nvidia RTX A6000 GPUs or Nvidia A100 40GB GPUs.
Software Dependencies No We adopt the pre-trained Diffusion Transformer (Di T) and VAE from Peebles & Xie (2023) in our framework, originally trained on Image Net-1K. (No specific software versions for libraries like PyTorch, TensorFlow, or Python are mentioned.)
Experiment Setup Yes For validation, the parameter settings vary slightly across methods. We adhere to the configurations in (Sun et al., 2024), as detailed in Table A1. Parameter CIFAR-10 CIFAR-100 Tiny-Image Net Image Net-1K Optimizer Adam W Learning Rate 0.01 Weight Decay 0.01 Batch Size 128 Augmentation Random Resized Crop + Horizontal Flip LR Scheduler Cosine Anneal Tempreture 20 Epochs 400 400 300 300