Neural-Style-Transfer

Neural Style Transfer using VGG-19 Network

Description:

Implementation of system proposed by Leon et al. in A Neural Algorithm of Artistic Style, 2015.

In very simple terms, the proposal is of a system that can combine the content of an image and the style of another.

Content Image	Style Image	Resultant Image

Installation:

• Clone the repository

Usage:

• ./main.py [--content] [--style] [--total_step=INT] [--log_step=INT] [--sample_step=INT] [--style_weight=FLOAT] [--lr=FLOAT]

Results:

Initial attemps, with very wrong naming of the model features:

Fixed the naming according to the paper:

• using conv features for both content and style with mse loss for both

• using conv features for both content and style with l1 loss for both

• using conv5 for content and all relu features for the style with mse loss for both

• using conv5 for content and all relu features for the style with l1 loss for both

Other outputs:

Try it yourself

Resources:

• Paper: A Neural Algorithm of Artistic Style
• For understanding Gram Matrices: Neural Style Transfer Tutorial -Part 1
• Intuitive Guide to Neural Style Transfer
• NEURAL TRANSFER USING PYTORCH
• https://github.com/yunjey/pytorch-tutorial/tree/master/tutorials/03-advanced/neural_style_transfer
• Rather helpful improvements to get more visually appealing results:
  • How to Get Beautiful Results with Neural Style Transfer
  • https://github.com/EugenHotaj/pytorch-generative