DSPO: Direct Score Preference Optimization for Diffusion Model Alignment

This repository is an official PyTorch implementation of "DSPO: Direct Score Preference Optimization for Diffusion Model Alignment" (ICLR 2025)

Environment Preparation

conda create -n DSPO python=3.10 && conda activate DSPO

pip install -r requirements.txt

Model Training

Training on Stable Diffusion v1.5

The hyperparameter $\beta$ can be searched by ${0.001, 0.01, 0.05}$, the best steps can be 500, 1000, 1500

bash sd15.sh

Training on Stable Diffusion XL

The hyperparameter $\beta$ can be searched by ${1000, 2000, 3000, 4000, 5000}$, the best steps can be 600,700

bash sdxl.sh