##Neural Art

This is a tensorflow implementation of the paper A Neural Algorithm of Artistic Style by Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge.

Usage

Basic usage:

neural_style.py --style_image <image.jpg> --content_image <image.jpg>

Custom Weights definition:

neural_style.py -c ./images/brad_pitt.jpg -s ./images/picasso_selfport1907.jpg -cw 10 -sw 100 -n pitt

<img src=https://github.com/ioanachelu/neural-art/blob/master/output/'pitt'_step_999.png width=256>

Options

• --content [-c] : Content image path. Default is './images/content.jpg'
• --style [-s] : Style image path. Default is './images/style.jpg'
• --iters [-i] : Number of steps/iterations. Default is 1000
• --output_dir [-o] : Output directory. Default is './output'
• --content_weight [-cw] : Content weight. Default is 5e0
• --style_weight [-sw] : Style weight. Default is 1e2
• --tv_weight [-tvw] : Total variation denoising weight. Default is 1e-3
• --output_image [-n] : Output image name. Default is 'neural_art'

Implementation details

Images are initialized with white noise and optimised with Adam Optimizer.

We perform style reconstructions using the conv1_1, conv2_1, conv3_1, conv4_1, and conv5_1 layers and content reconstructions using the conv4_2 layer. The style layers have equal weights.

The feature maps are extracted using a pretrained VGG network from Caffe. The weights are imported using caffe-tensorflow after updating the models from Model Zoo with upgrade_net_proto_text and upgrade_net_proto_binary