Pairwise Sample Optimization

This is the official Pytorch implementation of the paper: TUNING TIMESTEP-DISTILLED DIFFUSION MODEL USING PAIRWISE SAMPLE OPTIMIZATION.

Installation

conda env create -f environment.yaml

Fine-tuned Models

Human Preference Tuning

SDXL-DMD2: huggingface models

Usage (Evaluation on PickaPic-Test):

cd human_preference_tuning
accelerate launch eval_sdxl_dmd2.py

Training

Human Preference Tuning

• For SDXL-Turbo

bash Human_Preference_Tuning/online_pso_sdxl_turbo.sh

• For SDXL-DMD

bash Human_Preference_Tuning/online_pso_sdxl_dmd.sh

Concept Customization

check scripts in 'personalization/scripts'

Citation

If you find this work useful in your research, please consider citing:

@article{miao2024tuning,
  title={Tuning Timestep-Distilled Diffusion Model Using Pairwise Sample Optimization},
  author={Miao, Zichen and Yang, Zhengyuan and Lin, Kevin and Wang, Ze and Liu, Zicheng and Wang, Lijuan and Qiu, Qiang},
  journal={arXiv preprint arXiv:2410.03190},
  year={2024}
}

Acknowledgements

This repo is built upon D3PO. We thank the authors for their great work.