StyleNET

A Tensorflow implementation of Gatys et al. "Image Style Transfer with Convolutional Neural Networks" from CVPR 2016 (https://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/Gatys_Image_Style_Transfer_CVPR_2016_paper.pdf)

Background

StyleNET takes two images, a style image and a content image, and renders the content image in the style of the style image. It does so by extracting the representations of each image from the intermediate outputs of the VGG19 network, pretrained on the ImageNet dataset.

Often the style image is a painting; however Gatys et al. do note that their algorithm is capable of photo-realistic style transfer as well. Incredibly textured paintings, like Impressionist pieces, seem to work best, given that their style is very distinct and does not rely as heavily on intricate details. The content image can be of anything, but do note that intricate details often get distorted by the style.

For more details on the intricacies of the algorithm, see the original paper Gatys et al.

Requirements

This project requires the use of Python 3.x and TensorFlow 1.14. I use a TensorFlow version compiled from the binaries, thus I cannot guarantee that the version from PyPi will work.

Other required packages are OpenCV and Numpy, (see requirements.txt for exact versions).

Running StyleNET

Running StyleNET is as simple as running the following command: python styleNet.py --FLAG1=f1 --FLAG2=f2

If you are using TensorFlow GPU, it can be helpful to set the environment variable OPENCV_OPENCL_DEVICE=disabled to keep opencv from hogging the GPU.

Command Line Flags

style_image_dir

A path to the directory where style images are located. Default: "../data/styleImages"

content_image_dir

A path to the directory where content images are located. Default: "../data/contentImages"

style_image_name

The name of the style image file. Default: "style_image.jpg"

content_image_name

The name of the content image file. Default: "content_image.jpg"

num_steps

The number of transfer steps the program takes before writing the final image. Default: 1000

checkpoint_steps

The number of steps between image checkpoints. Default: 100

style_weight

The weighting factor of the contribution of the style loss to the total loss. Default: 1e-2

content_weight

The weighting factor of the contribution of the content loss to the total loss. Default: 1e4

style_layers

A comma separated string (with no spaces) of the layers used for style representations. Default: "block1_conv1,block2_conv1,block3_conv1,block4_conv1,block5_conv1"

content_layers

A comma separated string (with no spaces) of the layers used for content representations. Default: "block4_conv2"

content_resize_factor

The factor by which the content image is scaled. Default: 1.0

output_dir

A path to the directory where the output images are written. Default: ""../outputs"

output_name

The base name of the output images. Default: "output"

output_extension

The file extension for the output images (with no "."). Default: png

print_final_shape

Boolean flag whether to print final output shape before transfer steps are made. Default: False

Acknowledgments

All credit for the algorithm goes to Gaty's et al.