Reproduce “Domain-Aware Universal 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
Kibeom Hong, Seogkyu Jeon, Huan Yang, Jianlong Fu, Hyeran Byun
Researchers from Yonsei University and Microsoft Research created Domain-aware Style Transfer Networks (DSTN), an end-to-end model that unifies artistic and photo-realistic style transfer by adaptively inferring and responding to a style image's domain. The model achieves superior qualitative results and strong quantitative performance, including up to 92.42% user preference for content preservation with photo-realistic styles and 80.62% for artistic style expression.
Key insights
- The "domainness" of a style image can be implicitly learned and represented as a continuous factor, allowing a single network to adapt its behavior across diverse stylistic domains.
- Adaptive modulation of high-frequency information via skip connections, guided by the learned domainness, is crucial for unifying robust content preservation with expressive artistic stylization.
- Training with mixed samples and specific loss functions (DLow Loss for indicator, Contextual Similarity for reconstruction) is vital for learning continuous domain representations and preventing blurriness in outputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.