Reproduce “Non-Parametric 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
Jeong-Sik Lee, Hyun-Chul Choi
This research introduces an Exact Distribution Matching (EDM) layer for arbitrary image style transfer, utilizing Exact Histogram Specification (EHS) and channel-correlation awareness. The method generates high-quality stylized images with improved style uniformity and reduced artifacts by precisely matching feature map distributions within a feed-forward network, achieving practical computational efficiency with the EDM-Hourglass architecture.
Key insights
- Exact Histogram Specification (EHS) with spatial filters substantially improves the uniformity and consistency of style pattern application by enabling a unique ordering of feature map elements.
- Integrating channel-correlation handling via covariance matching with EHS is important for preserving stylistic coherence and achieving comprehensive multi-variate distribution alignment in feed-forward networks.
- Sliced Wasserstein Distance (SWD) is an effective style loss for training distribution-matching networks, yielding visually superior results and lower overall style loss compared to Central Moment Discrepancy (CMD).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.