A tensorflow implement of Perceptual Losses for Real-Time Style Transfer and Super-Resolution , referenced OlavHN/fast-neural-style, hzy46/fast-neural-style-tensorflow and lengstrom/fast-style-transfer.

• Python 2.7.x
• Tensorflow >= 1.4

Also make sure that you've installed numpy, scipy and pyyaml:

pip install numpy
pip install scipy
pip install pyyaml

The code works well on Ubuntu 18.04, Python 2.7.15 and tensorflow 1.13.1.

You should download Pretrained VGG-16 Model of tensorflow-slim and unpack it to folder model/.

cd <this repo>/model
sudo wget -c http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz
sudo tar -xzvf vgg_16_2016_08_28.tar.gz
sudo rm vgg_16_2016_08_28.tar.gz

Then, you should download COCO2014 Dataset , unpack it, and create a symbol link to the folder train2014/ .

sudo wget -c http://msvocds.blob.core.windows.net/coco2014/train2014.zip
sudo unzip train2014.zip
(If you've not installed unzip, use 'sudo apt install unzip' to install it)
cd <this repo>
sudo ln -s <path to train2014/> train2014

The repo has included a pretrained wave model checkpoint in model/trained_model/wave/model.ckpt-20001. To transfer images with a trained model, type that

cd <this repo>
python eval.py -c <path to source image> -m <path to model> -s <path to save result>

Default value of the para -m is defined as the pretrained model, so you can omit it.

To get more pretrained model you can download them here .

The folder conf/ has included some defined config file of different styles. As an example, to train the model cubist , type that

cd <this repo>
python train.py -c conf/cubist.yml

Then checkpoints could be found at model/trained_model/<folder of your style> .