A Flexible Convolutional Solver with Application to Photorealistic Style Transfer

  • 2018-06-13 22:05:35
  • Gilles Puy, Patrick PĂ©rez
  • 23

Abstract

We propose a new flexible deep convolutional neural network (convnet) toperform fast visual style transfer. In contrast to existing convnets thataddress the same task, our architecture derives directly from the structure ofthe gradient descent originally used to solve the style transfer problem [Gatyset al., 2016]. Like existing convnets, ours approximately solves the originalproblem much faster than the gradient descent. However, our network is uniquelyflexible by design: it can be manipulated at runtime to enforce new constraintson the final solution. In particular, we show how to modify it to obtain aphotorealistic result with no retraining. We study the modifications made by[Luan et al., 2017] to the original cost function of [Gatys et al., 2016] toachieve photorealistic style transfer. These modifications affect directly thegradient descent and can be reported on-the-fly in our network. Thesemodifications are possible as the proposed architecture stems from unrollingthe gradient descent.

 

Quick Read (beta)

loading the full paper ...