Neural-Style

This is a Keras/Tensorflow implementation of the neural-style algorithm from the paper A Neural Algorithm of Artistic Style by Gatys et. al. The code is based on Francois Chollet's implementation as laid out in his book Deep Learning with Python, and incorporates some modifications from recent work such as Improving the Neural Algorithm of Artistic Style. I wanted a python-based executable for quick experimentation along the lines of this Torch implementation; this repo is the result.

What It Does

The algorithm combines the content of an image with the styles and textures of another image using features derived from a convolutional neural network. For example:

Implementation Differences

• Gatys et. al. use a vgg19 network with normalized weights. Here, we use non-normalized weights. The normalized weights are available in caffe format and a script to download them is included, but not used yet.
• The Chollet implementation uses max-pooling layers. Here, max-pooling layers are converted by default to average-pooling layers.

Basic Usage

python neural_style.py --style_image=<style_image.jpg> --content_image=<content_image.jpg>

Better results can often be obtained by specifying the same style and content layers (and more of them), as described in Improving the Neural Algorithm of Artistic Style:

python neural_style.py --style_layers block1_conv1 block1_conv2 block2_conv1 block2_conv2 block3_conv1 block3_conv2 block4_conv1 block4_conv2 block5_conv1 block5_conv2 --content_layers block1_conv1 block1_conv2 block2_conv1 block2_conv2 block3_conv1 block3_conv2 block4_conv1 block4_conv2 block5_conv1 block5_conv2 --style_image=<style_image.jpg> --content_image=<content_image.jpg>

Options

Images and Directory Options

• --image_size: Maximum side length of output image; default is 512.
• --style_image: image to use for style and texture.
• --content_image: image to use for content.
• --output_dir: default is tempdir/neural_style_output.

Model Options

• --num_iters: number of iterations, default 20.
• --avg_pool: convert max_pool layers to avg_pool, as in original paper? Default is True.
• --content_weight: weight on content image, default is 0.001
• --style_weight: weight on style image, default is 3.0
• --total_variation_weight, default is 0.001
• --style_layers: layers to use to extract style features. Defaults to blocki_conv1 for i=1,...,5.
• --content_layers: layers to use to extract content features. Defaults to block4_conv2.

Improving the Quality of the Generated Images

Fork the repo and try incorporating the results in this paper!.

Tested on Ubuntu 17.10, Keras 2.1.3, Tensorflow 1.4.0, Cuda 9.0, CuDNN 7.