Decsription

This is a TensorFlow implementation of the paper A Neural Algorithm of Artistic Style. The Neural Style Transfer algorithm synthesizes new artistic images by combining the content of one image with the style of another image using a pre-trained Convolutional Neural Network (VGG-19) . Below you will find some of the generated images.

Examples

Here, we are using the Department of Informatics and Telecommunications as content image and various well known paintings as style images. Clearly, the algorithm produces visually appealing results:

Next, we will be using the Parthenon as the content image:

Content / Style Tradeoff

One can adjust the trade-off between content and style by modifying the parameters a and b of the algorithm (see Implementation Details below). From left to right the parameter b (emphasis on style) is increased, whereas parameter a (emphasis on content) is fixed. The style image used here is Wassily Kardinsky's Composition VII:

Implementation Details

The generated image is initialized with a percentage of white noise (the percentage for the examples above was 99%) before being fed to the ConvNet. The optimizer of choice in this implemetation is the Adam Optimizer, instead of L-BFGS, which is what the authors of the paper used. As a result, the parameters of the algorithm required some tuning in order to obtain good results. In particular, we experimented with different values of content weight, style weight and learning rate. After some 'trial and error', we ended up with the following:

content weight (alpha) = 5
style weight   (beta)  = 50000/100000/500000 (depending on the style image being used)
learning rate          = 10

With certain images, the style weight might need to be adjusted (choose one of the values shown above).

The layers used for the style reconstructions are, as advised in the paper, conv1_1, conv2_1, conv3_1, conv4_1 and conv5_1 (in general the weight for each "active" layer is 1 / number of active layers, in our case 1/5) . For the content reconstructions layer conv4_2 was used, again, as advised in the paper.

In general, running the algorithm for 1000-1500 iterations produces nice results. However, it is highly advised to run the algorithm on a GPU, since it significantly speeds up the process. For a 400x300 image, 1000 iterations take about 60 minutes on a CPU!

Sidenote: Both the content and style image must be the same dimensions

Setup

Dependencies

• TensorFlow
• NumPy
• SciPy
• Imageio

Data

• The pre-trained VGG-19 network can be found here (see VGG-VD section). After downloading, place it on ./pretrained-model/imagenet-vgg-verydeep-19.mat.