Demystifying Neural Style Transfer

TL;DR

This paper offers a new theoretical interpretation of neural style transfer as a domain adaptation problem, showing that matching Gram matrices aligns feature distributions via MMD, which clarifies the underlying principle.

Contribution

It introduces a novel perspective by linking neural style transfer to distribution matching using MMD, supported by experiments with alternative distribution alignment methods.

Findings

Matching Gram matrices is equivalent to minimizing MMD with a polynomial kernel.

Feature distribution matching is the core of neural style transfer.

Alternative distribution alignment methods also produce appealing results.

Abstract

Neural Style Transfer has recently demonstrated very exciting results which catches eyes in both academia and industry. Despite the amazing results, the principle of neural style transfer, especially why the Gram matrices could represent style remains unclear. In this paper, we propose a novel interpretation of neural style transfer by treating it as a domain adaptation problem. Specifically, we theoretically show that matching the Gram matrices of feature maps is equivalent to minimize the Maximum Mean Discrepancy (MMD) with the second order polynomial kernel. Thus, we argue that the essence of neural style transfer is to match the feature distributions between the style images and the generated images. To further support our standpoint, we experiment with several other distribution alignment methods, and achieve appealing results. We believe this novel interpretation connects these two important research fields, and could enlighten future researches.