license: cc-by-nc-4.0
pipeline_tag: image-to-image
tags:
- image-stylization
- mixture-of-experts
- lora
- flux
StyleExpert: Mixture of Style Experts for Diverse Image Stylization
StyleExpert is a semantic-aware framework for diverse image stylization based on a Mixture of Experts (MoE) architecture. It addresses the limitations of existing diffusion-based stylization methods by effectively handling complex semantics and material details, rather than just color-driven transformations.
[📖 Paper] [🏠 Project Page] [💻 Code]
Model Description
StyleExpert employs a unified style encoder trained on the StyleExpert-40K dataset (40,000 content-style-stylized triplets). The model uses a similarity-aware gating mechanism to dynamically route styles to specialized experts within the MoE architecture, allowing it to handle styles ranging from shallow textures to deep semantics.
Installation
We recommend using Python 3.10 and PyTorch with CUDA support.
# Create a new conda environment
conda create -n styleexpert python=3.10
conda activate styleexpert
# Install requirements
pip install -r requirements.txt
Quick Inference
Local Gradio Demo
You can run a local web interface to interact with the model:
python app.py
Single Case Inference via CLI
To stylize a single image using a reference style, use the following command from the official repository:
python infer.py --content_path ./data/content.jpg --style_path ./data/style.jpg
Weights Download
The model requires the base FLUX.1-Kontext-dev and the StyleExpert adapters. You can use the provided script in the GitHub repo to download all necessary components:
bash download_models.sh --token YOUR_HF_TOKEN
Visual Results
Citation
If StyleExpert helps your research, please cite the following work:
@article{zhu2026styleexpert,
title={Mixture of Style Experts for Diverse Image Stylization},
author={Zhu, Shihao and Ouyang, Ziheng and Kang, Yijia and Wang, Qilong and Zhou, Mi and Li, Bo and Cheng, Ming-Ming and Hou, Qibin},
journal={CVPR},
year={2026}
}
License
This model is licensed under CC BY-NC 4.0 for non-commercial use. Underlying model weights (like FLUX.1) may be subject to their own respective licenses.