Reproduce “Mixture of Style Experts for Diverse Image Stylization” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Shihao Zhu, Ziheng Ouyang, Yijia Kang, Qilong Wang, Mi Zhou, Bo Li, Ming-Ming Cheng, Qibin Hou

The StyleExpert framework enables semantic-aware image stylization by integrating a pre-trained style encoder with a Mixture of Experts architecture, addressing limitations of existing methods that often revert to color transfer. It introduces a new semantically diverse dataset and achieves state-of-the-art performance across multiple metrics for style transfer and content fidelity.

Key insights

  • A pre-trained, semantic-aware style encoder provides explicit style priors that effectively guide the MoE router, significantly stabilizing training and accelerating convergence for diverse style handling.
  • Dynamically routing different styles to specialized experts within an MoE architecture allows the model to handle diverse styles at varying granularities, capturing intricate semantic details beyond simple color transfer.
  • The quality and semantic diversity of training data are crucial for developing advanced stylization models that can accurately transfer complex artistic attributes.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.