[CVPR 2025] StyleSSP: Sampling StartPoint Enhancement for Training-free Diffusion-based Method for Style Transfer

Arxiv

Usage

To run the code, please follow these step:

1. Download
2. Setup
3. Run

Download

This project contains contributions from ControlNet and IP-Adapter-Instruct, licensed under the Apache License 2.0. Modifications and additional content added by StyleSSP in 2024. The pre-trained checkpoints from Tile-ControlNet, MistoLine, IP-Adapter-Instruct

# download adapters
huggingface-cli download --resume-download h94/IP-Adapter --local-dir checkpoints/IP-Adapter

# download ControlNets
huggingface-cli download --resume-download TheMistoAI/MistoLine --local-dir checkpoints/MistoLine
huggingface-cli download --resume-download xinsir/controlnet-tile-sdxl-1.0 --local-dir checkpoints/controlnet-tile-sdxl-1.0

# download models IP-Adapter-Instruct
download the models ckpt ip-adapter-instruct-sdxl.bin from: https://huggingface.co/CiaraRowles/IP-Adapter-Instruct and put it in the folder checkpoints/models

Environment Setup

conda env create -f environment.yaml
conda activate StyleSSP
pip install git+https://github.com/openai/CLIP.git

Run

For running StyleSSP, modify content_image_dir and style_image_dir in src/config.py, then run:

python infer_style.py

Evaluation

For a quantitative evaluation, we incorporate a set of randomly selected inputs from MS-COCO and WikiArt in “./data” directory, as InstantStyle-Plus do.

Before executing evalution code, please run infer_style.py to get the results (40 styles, 20 contents -> 800 stylized images), then put the content, style and stylized images in “./data_evl/content”, “./data_evl/style”, and “./data_evl/tar” directory, respectively.

Then, run:

cd evaluation;
python eval_artfid.py --sty ../data_evl/style --cnt ../data_evl/content --tar ../data_evl/tar

Citation

If you find our work useful, please consider citing and star:

@article{xu2025stylessp,
  title={StyleSSP: Sampling StartPoint Enhancement for Training-free Diffusion-based Method for Style Transfer},
  author={Xu, Ruojun and Xi, Weijie and Wang, Xiaodi and Mao, Yongbo and Cheng, Zach},
  journal={arXiv preprint arXiv:2501.11319},
  year={2025}
}