In the paper Texture Networks: Feed-forward Synthesis of Textures and Stylized Images we describe a faster way to generate textures and stylize images. It requires learning a feedforward generator with a loss function proposed by Gatys et. al.. When the model is trained, a texture sample or stylized image of any size can be generated instantly.

Instance Normalization: The Missing Ingredient for Fast Stylization presents a better architectural design for the generator network. By switching batch_norm to instance norm we facilitate the learning process resulting in much better quality.

This also implements the stylization part from Perceptual Losses for Real-Time Style Transfer and Super-Resolution.

• Torch7 + loadcaffe
• cudnn + torch.cudnn (optionally)
• display (optionally)

Download VGG-19.

cd data/pretrained && bash download_models.sh && cd ../..

You can use an image dataset of any kind. For my experiments I tried Imagenet and MS COCO datasets. The structure of the folders should be the following:

dataset/train
dataset/train/dummy
dataset/val/
dataset/val/dummy

The dummy folders should contain images. The dataloader is based on one used infb.resnet.torch.

Here is a quick example for MSCOCO:

wget http://msvocds.blob.core.windows.net/coco2014/train2014.zip
wget http://msvocds.blob.core.windows.net/coco2014/val2014.zip
unzip train2014.zip
unzip val2014.zip
mkdir -p dataset/train
mkdir -p dataset/val
ln -s `pwd`/val2014 dataset/val/dummy
ln -s `pwd`/train2014 dataset/train/dummy

th train.lua -data <path to any image dataset>  -style_image path/to/img.jpg

To achieve the results from the paper you need to play with -image_size, -style_size, -style_layers, -content_layers, -style_weight, -tv_weight.

Do not hesitate to set -batch_size to one, but remember the larger -batch_size the larger -learning_rate you can use.

th test.lua -input_image path/to/image.jpg -model data/checkpoints/model.t7

Play with -image_size here. Raise -cpu flag to use CPU for processing.

You can find a pretrained model here. It is not the model from the paper.

soon

• The code was tested with 12GB NVIDIA Titan X GPU and Ubuntu 14.04.
• You may decrease batch_size, image_size if the model do not fit your GPU memory.
• The pretrained models do not need much memory to sample.

The code is based on Justin Johnson's great code for artistic style.

The work was supported by Yandex and Skoltech.