A Sliced Wasserstein Loss for Neural Texture Synthesis

TL;DR

This paper proposes using the Sliced Wasserstein Distance as a new, effective loss function for neural texture synthesis, addressing limitations of the traditional Gram-matrix approach.

Contribution

It introduces the Sliced Wasserstein Loss as a novel, theoretically sound, and practical alternative for texture synthesis in neural networks.

Findings

Sliced Wasserstein Loss outperforms Gram-matrix loss in visual quality.

The method is simple to implement and integrates well with existing neural network frameworks.

Results are superior in both optimization and generative training scenarios.

Abstract

We address the problem of computing a textural loss based on the statistics extracted from the feature activations of a convolutional neural network optimized for object recognition (e.g. VGG-19). The underlying mathematical problem is the measure of the distance between two distributions in feature space. The Gram-matrix loss is the ubiquitous approximation for this problem but it is subject to several shortcomings. Our goal is to promote the Sliced Wasserstein Distance as a replacement for it. It is theoretically proven,practical, simple to implement, and achieves results that are visually superior for texture synthesis by optimization or training generative neural networks.