Balanced Image Stylization with Style Matching Score

Yuxin Jiang, Liming Jiang, Shuai Yang, Jia-Wei Liu, Ivor W. Tsang and Mike Shou Zheng in ICCV 2025.

Style Matching Score (SMS) is a novel optimization method for balanced image stylization with diffusion models. Unlike existing efforts, SMS reframes image stylization as a style distribution matching problem. The target style distribution is estimated from off-the-shelf style-dependent LoRAs via carefully designed score functions. The optimization formulation naturally extends sylization from pixel space to parameter space, making it readily applicable to lighweight generators for efficient one-step stylization, and offering potential for future 3D stylization applications.

Updates

[03/2026] Support other score distillation methods (VSD, SDS ...) for completeness.
[06/2025] Code is released.
[06/2025] The paper is accepted to ICCV 2025!🎉
[03/2025] Repo is initialized.

🚀 Get Started

Environment Setup

git clone https://github.com/showlab/SMS.git
cd SMS

conda create -n sms python=3.10 -y
conda activate sms
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt --no-deps

Note: Optimization requires ~17GB of GPU memory. A 24GB GPU is recommended.

HuggingFace Cache Configuration (optional)

The program will automatically download pretrained SD models during optimization. If your default disk is low on space, redirect your HuggingFace model cache:

export HF_HOME=/your/large/storage/cache

🎨 Stylization

1. Quick Demos

Each script runs optimization for a specific style, saving both the final result and every-50-iteration previews (as .png and .gif) in the output/ folder. We use off-the-shelf style LoRA from Civitai. Feel free to explore more styles on Civitai or train your own LoRA models to combine with SMS!

Pixel Art Style

wget "https://civitai.com/api/download/models/31228?type=Model&format=SafeTensor&size=full&fp=fp16" -O ./lora_ckpt/3232pixel.safetensors
bash scripts/pixel_art.sh

Oil Painting Style

wget "https://civitai.com/api/download/models/90795?type=Model&format=SafeTensor" -O ./lora_ckpt/fechin.safetensors
bash scripts/oil_painting.sh

Watercolor Style

wget "https://civitai.com/api/download/models/21173?type=Model&format=SafeTensor&size=full&fp=fp16" -O  ./lora_ckpt/Colorwater_v4.safetensors
bash scripts/watercolor.sh

Kids Illustration Style

wget "https://civitai.com/api/download/models/67980?type=Model&format=SafeTensor" -O ./lora_ckpt/COOLKIDS_MERGE_V2.5.safetensors
bash scripts/kids_illustration.sh

2. Custom Stylization

The --lambda_dct flag controls the strength of Progressive Spectrum Regularization, enabling a flexible content-style trade-off based on user preference:

0 results in stronger stylization (more abstract and expressive)
Higher values preserve finer content structure

See stylize.py for the full list of arguments and usage examples.

🎓 Citation

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

@inproceedings{jiang2025balanced,
  title={Balanced Image Stylization with Style Matching Score},
  author={Jiang, Yuxin and Jiang, Liming and Yang, Shuai and Liu, Jia-Wei and Tsang, Ivor and Shou, Mike Zheng},
  booktitle={ICCV},
  year={2025}
}

👏 Acknowledgements

Our idea is implemented based on PDS. Thanks for their great open-source work! We also gratefully acknowledge the creators of the open-source style LoRA models and base SD models used in this work.

SMS

Install / Use

README