Set-Coupled Guidance

Official code release for Set-Coupled Guidance (SCG), an ICML 2026 regular paper.

SCG is a plug-and-play auxiliary controller for diffusion-based dataset distillation. This release focuses on the MGD3+SCG guided-sampling path used for the open-source camera-ready code package.

--num-samples remains the actual IPC. --particle-batch only changes execution memory, not the set-level control object.

Setup

conda create -n scg python=3.9
conda activate scg
pip install -r requirements.txt
pip install -e diffusers

The ImageNet root used by the scripts is defined in scripts/run_scg.sh:

IMAGENET_DIR="${IMAGENET_DIR:-${HOME}/datasets/imagenet1k/}"

The direct Python default is defined in sample_scg.py:

parser.add_argument("--imagenet_dir", type=str, default=os.path.expanduser('~/datasets/imagenet1k/'))

Edit these defaults or set IMAGENET_DIR when running the scripts to use a different dataset path.

Sampling and Evaluation

Run sampling and evaluation for ImageNette and ImageWoof at IPC 10/20/50 sequentially:

bash scripts/run_all.sh

Acknowledgements

This project builds upon the MGD³ codebase: https://github.com/jachansantiago/mode_guidance.git

Citation

@inproceedings{
gan2026setcoupled,
title={Set-Coupled Guidance: Set-Level Coordination in Diffusion-Based Dataset Distillation},
author={Ziang Gan and Qi Zhu and Libao Zhang},
booktitle={Forty-third International Conference on Machine Learning},
year={2026},
url={https://openreview.net/forum?id=onosLtwoiK}
}