Tensorflow implementation of the paper "A Neural Algorithm of Artistic Style".
The algorithm renders an image that keeps the content of one reference image while copying the style of another. For example:
There's plenty of great explanations on the web regarding how this works so I won't go into detail here...
- Tensorflow (version > 1.0)
- tensorflow-vgg
Most parameters can be configured when running using command line, though some parameters can currently be configured only through default_params.py.
To run using command line:
python main.py
All parameters are optional and have default values:
--iterations <number of iterations>, --out_width <width of output image>, --content_path <content file path>, --style_path <style file path>, --result_path <Path of output image (with extension). If only a directory path is given, an automatic file name will be generated>, --noise_ratio <init image is noise_ratio*noise + (1-noise_ratio)*content_image>, --gamma <content/style ratio>, --beta <style weight>, --theta <tv loss weight>, --optimizer <'adam' or 'lbfgs'>.
For the trained VGG19 I have used the implementation of mechrisaa.