style_transfer-TF

This is a Tensorflow implementation of papers Texture Synthesis Using Convolutional Neural and A Neural Algorithm of Artistic Style by Leon A. Gatys, Alexander S.Ecker, and Matthias Bethge.

The former presents an texture synthesis algorithm by modeling the statistics feature extracted from Convolutional Neural Networks. The latter algorithm combines the texture feature, that is , the style, of one so called style image with the content of another content image, and generated a style transfer image.

Examples of synthesis texture

Texture image	Synthesized image

Example of Style Transfer

Content Image XJTU

Style Image Starry Night

Stylized Image

Dependency

• python3.5
• tensoflow (tested on r1.4)
• VGG19 model weights download from the TF-slim models
• The code is tested on:
  • Ubuntu 16.04 LTS with CPU architecture x86_64 + Nvidia GeForce GTX 1080

Recommended

• Ubuntu 16.04 with TensorFlow GPU edition

Getting started

Denote the directory of this repository as ./style_transfer-TF/

• Run texture synthesis demo to check configuration

# clone the repository from github
git clone https://github.com/LiaoQian1996/style_transfer-TF.git
cd $style_transfer-TF/

# download the vgg19 model weights from 
# http://download.tensorflow.org/models/vgg_19_2016_08_28.tar.gz
# to style_transfer-TF/vgg19/

# run the texture synthesis demo
sh demo.sh
# the result can be viewed at $./results/

• Synthesis your texture image

cd $style_transfer-TF/

# put your own png images in $style_transfer-TF/imgs/

modify the parameters in run_syn.sh

CUDA_VISIBLE_DEVICES=0 python main.py \
    --task_mode texture_synthesis \
    --texture_shape 512 512 \ # set the synthesized texture shape, which will be same as sample's if set as [-1, -1]  
    --output_dir ./results/ \
    --target_dir ./imgs/your_own_image_name.png \  # your own image, should be .png format and RGB mode 
    --content_dir ./imgs/tomato.png \ # use this image to initialize the systhesized image,                                                                 # --initials should be set as content \
    --initials content \ # noise (option)
    --top_style_layer VGG54 \ # VGG 11 21 31 41 51 54
    --max_iter 500 \
    --W_tv 0.001 \ # weight of total variation loss
    --vgg_ckpt ./vgg19/vgg_19.ckpt

then run the shell script

sh run_syn.sh

• Style Transfer

cd $style_transfer-TF/

# put your own content and style images (.png) in $style_transfer-TF/imgs/

modify the parameters in run_transfer.sh

CUDA_VISIBLE_DEVICES=1 python main.py \
    --task_mode style_transfer \
    --output_dir ./results/ \
    --target_dir ./imgs/starry-night.png \        # path of style image
    --content_dir ./imgs/tubingen.png \           # path of content image
    --initials content \                          # initial synthesized image, noise or content 
    --top_style_layer VGG54 \       
    --max_iter 50 \
    --W_tv 0.001 \
    --W_content 1e-6 \
    #--vgg_ckpt ./vgg19/vgg_19.ckpt

then run the shell script

sh run_transfer.sh