Neural Style Transfer

This repository is the implementation of A Neural Algorithm of Artistic Style by Leon A.Gatys, et al..

In their paper, they have combined several photographs with famous paintings thereby generating new art. This is done by finding the layers in a Convolutional Neural Network that represent the style and contents of the photograph and the art to be combined.

Once the style layers of the art has been obtained the style layers of the photograph is changed so that it's style represents the style of the art. This is done by using gradient descent where we repeatedly train the style layers of the content image to slowly change it's values to the style layers of the art.

The level upto which one wants to combine is controlled by the number of epochs and the number of steps per epoch. Given below is a combination generated from a Labrador photograph (content image) and a Kandinsky painting (style image). The model has been run for 10 epochs and in each epoch 100 steps have been executed.

You can also check out my article of Neural Style Transfer

The network used here is VGG19 and the model only combines the Kandinski painting for now.
The style layers of the VGG19 architecture have been identified as

• Block1_conv1
• Block2_conv1
• Block3_conv1
• Block4_conv1
• Block5_conv1

There is only one content layer in this architecture and that is Block5_conv2.

Running the model

Libraries required to run the model

1. Numpy
2. Tensorflow
3. Matplotlib
4. PILLOW

Steps to run the model

1. Download the content_style_mixer.py script.
2. Find out either the relative path or absolute path of the image that you want to combine.
3. Open your terminal and activate the environment if you have downloaded the packages in an environments.
4. Run the script using the following command: python content_style_mixer.py [path to the image]. Example: python content_style_mixer.py /home/sathya/Downloads/test1.jpg
5. Mention the number of epochs, number of steps per epoch and the name of the image by which you want your new image to be saved by.
   Note: Higher the epochs*steps_per_epoch value more will the presence of style of the style image in the content image. Most of the values have been set to a good default, you wont't have the need to modify them. In case, you want to look at how the mixed image looks, start with a small value for the number of epochs and the number of steps per epoch.
6. You will find the mixed image in jpeg format in the directory in which the content_style_mixer.py script is present

More famous paintings and high resolution combinations will be made available in the coming weeks