README.md

DKDM: Data-Free Knowledge Distillation for Diffusion Models with Any Architecture

Qianlong Xiang¹, Miao Zhang^1✉, Yuzhang Shang², Jianlong Wu¹, Yan Yan³, Liqiang Nie^1✉

¹Harbin Institute of Technology (Shenzhen)    ²Illinois Institute of Technology ³University of Illinois Chicago
^✉Corresponding author  

The repository contains the code for our paper "DKDM: Data-Free Knowledge Distillation for Diffusion Models with Any Architecture" (CVPR 2025). We have released all the codes for our method and baselines. If you have any questions, please feel free to raise an issue or contact us via email (xiangqianlongcs@gmail.com).

Summary of our paper:

• DKDM: A scenario, or a question, which asks: Can we train new diffusion models by using existing pretrained diffusion models as the data source, thereby eliminating the need to access or store any dataset?
• Dynamic Iterative Distillation: Our proposed method to answer the above question.

News

• [2025-03-28] 🚀 We release the code about image generation in latent space. Please refer to latent-diffusion/README.md for details.
• [2025-03-17] 🚀 We release the code about image generation in pixel space. Please refer to guided-diffusion/README.md for details.
• [2025-02-27] 🚀 Our paper is accepted by CVPR 2025 paper!

Dynamic Iterative Distillation

We implement the dynamic iterative distillation in pixel and latent space, respectively. Please refer to guided-diffusion/README.md and latent-diffusion/README.md for details.

Acknowledgement

Our codebase is built upon guided-diffusion and latent-diffusion, which train diffusion models in pixel space and latent space, respectively. Thanks for their great works! We also thank Xingyi Yang, one of the authors of Diffusion Probabilistic Model Made Slim, CVPR 2023, for his help in the implementation of distilling latent diffusion models.

If you find our paper and repository helpful, please consider citing our paper:

@inproceedings{xiang2025dkdm,
  title={Dkdm: Data-free knowledge distillation for diffusion models with any architecture},
  author={Xiang, Qianlong and Zhang, Miao and Shang, Yuzhang and Wu, Jianlong and Yan, Yan and Nie, Liqiang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={2955--2965},
  year={2025}
}