Reproduce “Anime Style Space Exploration Using Metric Learning and Generative Adversarial Networks” 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
Sitao Xiang, Hao Li
Researchers from USC and Pinscreen developed a method to learn explicit, interpretable representations of artistic style using metric learning and generative adversarial networks. Their system successfully disentangles artistic style from image content, demonstrating semantic control over specific style dimensions like hair detail and character maturity in generated images.
Key insights
- Artworks by the same artist exhibit similar styles, which can be used as a supervision signal for learning style representations through metric learning.
- Disentangling style and content can be achieved by explicitly providing the learned style code to the content encoder's decoder, forcing the content encoder to learn only complementary information.
- Using nested dropout during training helps create an interpretable style space where dimensions are ordered by importance and can be semantically manipulated.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.