StyleExpert

Mixture of Style Experts for Diverse Image Stylization

🖼️ Visual Results

🔧 Dependencies and Installation

We recommend using Python 3.10 and PyTorch with CUDA support. To set up the environment:

# Create a new conda environment
conda create -n styleexpert python=3.10
conda activate styleexpert

# Install requirements
pip install -r requirements.txt

⚡ Quick Inference

Tips

StyleExpert uses a Mixture of Experts (MoE) architecture. For the best results on complex semantic styles (like specific brushstrokes or materials), ensure your style reference image clearly showcases those textures. The model uses a pre-trained Style Representation Encoder to guide the router.

Local Gradio Demo

python app.py

Model Download

You can download the base model FLUX.1-Kontext-dev and our StyleExpert adapters directly from Hugging Face:

• Base Model: FLUX.1-Kontext-dev
• StyleExpert LoRA Experts: Hugging Face Link

Alternatively, use the provided script:

bash download_models.sh --token YOUR_HF_TOKEN

This will download these fixed repos into the local default paths used by inference:

• HH-LG/StyleExpert -> ./weights/
• black-forest-labs/FLUX.1-Kontext-dev -> ./models/FLUX.1-Kontext-dev/
• google/siglip-so400m-patch14-384 -> ./models/siglip-so400m-patch14-384/

---

📊 Dataset: StyleExpert-40K

We provide the StyleExpert-40K dataset, containing 40,000 high-quality content-style-stylized triplets. This dataset is specifically curated to balance color-centric and semantic-centric styles.

Download via Script:

python download_dataset.py --token YOUR_HF_TOKEN

# only fetch metadata first
python download_dataset.py --metadata-only --token YOUR_HF_TOKEN

# or
bash download_dataset.sh --token YOUR_HF_TOKEN

Single Case Inference

python infer.py --content_path ./data/content.jpg --style_path ./data/style.jpg

You can directly run inference with the example pairs in assets/examples/:

# Use example_00 pair
./run.sh ./assets/examples/content_00.png ./assets/examples/style_00.png ./outputs/example_00_out.png 

# Use example_01 pair
./run.sh ./assets/examples/content_01.png ./assets/examples/style_01.png ./outputs/example_01_out.png

---

🧪 Method Overview

StyleExpert utilizes a two-stage training approach:

1. Style Representation Encoder: Trained with InfoNCE loss to learn discriminative style features.
2. MoE Fine-tuning: Uses a similarity-aware gating mechanism to route styles to specialized LoRA experts.

---

📜 Citation

If StyleExpert helps your research, please star the repo and cite our work:

@misc{zhu2026mixturestyleexpertsdiverse,
      title={Mixture of Style Experts for Diverse Image Stylization}, 
      author={Shihao Zhu and Ziheng Ouyang and Yijia Kang and Qilong Wang and Mi Zhou and Bo Li and Ming-Ming Cheng and Qibin Hou},
      year={2026},
      eprint={2603.16649},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.16649}, 
}

📧 Contact

For questions, please open an issue or contact Shihao Zhu.

License

Licensed under CC BY-NC 4.0 for non-commercial use.