A short writeup and example images are up on my blog.

In an attempt to learn Tensorflow I've implemented an Image Transformation Network as described in Perceptual Losses for Real-Time Style Transfer and Super-Resolution by Johnson et al.

This technique uses loss functions based on a perceptual similarity and style similarity as described by Gatys et al to train a transformation network to synthesize the style of one image with the content of arbitrary images. After it's trained for a particular style it can be used to generate stylized images in one forward pass through the transformer network as opposed to 500-2000 forward + backward passes through a pretrained image classification net which is the direct approach.

• Changed upscale method to the one mentioned in http://distill.pub/2016/deconv-checkerboard/
• Better padding befoer each convolution to avoid border effects
• For some reason, padding the original image before the transformation image removes a lot of noise.
• "Instance normalization" instead of batch normalization as mentioned in Instance Normalization: The Missing Ingredient for Fast Stylization
• No longer python3.

While the results are now much better, I'm still not sure why the original implementation didn't perform as well as Johnsons original work (Now published here)

First get the dependecies (COCO training set images and VGG model weights):

./get_files.sh

To generate an image directly from style and content, typically to explore styles and parameters:

python neural_style.py --CONTENT_IMAGE content.png --STYLE_IMAGES style.png

Also see other settings and hyperparameters in neural_style.py

To train a model for fast stylizing first download dependences (training images and VGG model weights):

./get_files.sh

Then start training:

python fast-neural-style.py --STYLE_IMAGES style.png --NAME=my_model

Where --TRAIN_IMAGES_PATH points to a directory of JPEGs to train the model. --NAME is used for tensorboard statistics and file name of model weights. The paper uses the COCO image dataset (13GB).

To generate images fast with an already trained model:

python inference.py --IMAGE_PATH=my_content.jpg --NAME=my_model

• Tensorflow
• VGG-19 model
• COCO dataset

• Add a pretrained model
• Add example pictures / videos

• [Chainer implementation] (https://github.com/yusuketomoto/chainer-fast-neuralstyle)
• [Tensorflow Neural style implementation] (https://github.com/anishathalye/neural-style) (Both inspiration and copied the VGG code)