DDIS

Official repository for the ICML 2025 paper:

When Model Knowledge meets Diffusion Model: Diffusion-assisted Data-free Image Synthesis with Alignment of Domain and Class

• Paper: https://arxiv.org/abs/2506.15381
• Code: https://github.com/kdst-team/DDIS

Overview

DDIS is a diffusion-assisted data-free image synthesis method that uses a text-to-image diffusion model as a strong image prior for Data-Free Image Synthesis (DFIS). The method extracts knowledge from a pretrained classifier and guides the diffusion process so that synthesized samples align better with the original training distribution.

The framework introduces:

• Domain Alignment Guidance (DAG) to align generated images with the target domain during diffusion sampling.
• Class Alignment Token (CAT) to capture class-specific semantics from the given pretrained model.

Experiments in the paper are reported on PACS and ImageNet.

Repository Layout

• index.html, static/, and the figure assets in the repository root: project website for the paper.
• DDIS_code/main_ddis.py: main training, evaluation, and control-strength entry point.
• DDIS_code/config.py: experiment configuration.
• DDIS_code/utils.py: classifier and diffusion-model setup utilities.
• DDIS_code/p2p/: expected location of Prompt-to-Prompt related helper files.

Setup Notes

This repository currently does not ship a dedicated environment file. The code uses Python with the following libraries, as referenced in the source:

• torch
• torchvision
• diffusers
• transformers
• accelerate
• kornia
• numpy
• matplotlib
• pyrallis

Depending on the experiment setup, additional pretrained checkpoints may also be required.

Running the Code

The main entry point is:

cd DDIS_code
python3 main_ddis.py --class_index 1 --train True --evaluate False --control_strength False

There is also a simple loop script:

cd DDIS_code
bash run.sh <cuda_device_id>

For experiment-specific settings, edit DDIS_code/config.py and inspect DDIS_code/main_ddis.py.

Important Note on p2p

To keep fresh clones and GitHub Actions checkouts stable, the previously broken submodule entry at DDIS_code/p2p was removed.

If you want to run DDIS_code/main_ddis.py, you should place the expected Prompt-to-Prompt source files in DDIS_code/p2p/, including:

• ptp_utils_distG.py
• prompt_to_prompt.py

Citation

If you find this work useful, please cite:

@inproceedings{kim2025ddis,
  title={When Model Knowledge meets Diffusion Model: Diffusion-assisted Data-free Image Synthesis with Alignment of Domain and Class},
  author={Kim, Yujin and Kim, Hyunsoo and Kim, Hyunwoo J. and Kim, Suhyun},
  booktitle={International Conference on Machine Learning},
  year={2025}
}

License

This repository includes an MIT license in LICENSE.