User-Controllable Arbitrary Style Transfer via Entropy Regularization

Ensuring the overall end-user experience is a challenging task in arbitrary style transfer (AST) due to subjective nature of quality. A good practice provide users many instead one AST result. However, existing approaches require run multiple models or inference diversified (DAST) solution times, and thus they are either slow speed limited diversity. In this paper, we propose novel ensuring both efficiency diversity for generating user-controllable results by systematically modulating behavior at run-time. We begin with reformulating three prominent methods into unified assign-and-mix problem discover that entropies their assignment matrices exhibit large variance. then solve an optimal transport framework using Sinkhorn-Knopp algorithm user input ε control said entropy modulate stylization. Empirical demonstrate superiority proposed solution, stylization quality comparable better than significantly more diverse previous DAST works. Code available https://github.com/cplusx/eps-Assign-and-Mix.