Reproduce “Domain Generalizable Portrait Style Transfer” 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
Xinbo Wang, Wenju Xu, Qing Zhang, Wei-Shi Zheng
Researchers from Sun Yat-sen University and Amazon introduce a diffusion-based framework for domain-generalizable portrait style transfer, integrating semantic correspondence and a novel latent initialization technique. The method achieves high-quality, semantically aligned stylization across diverse portrait domains (photo, cartoon, sketch) while effectively preserving identity and content details.
Key insights
- Integrating dense semantic correspondence extraction with a dual-conditional diffusion model allows for accurate region-specific style transfer across diverse portrait domains.
- Using Haar Discrete Wavelet Transform (DWT) for high-frequency structural guidance in ControlNet effectively preserves content details and facial identity.
- A novel AdaIN-Wavelet Transform for latent initialization robustly blends color statistics from the style reference with content details, balancing stylization strength and content preservation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.