neural_style_synthesizer

INSTALL

The model files of neural networks are not contained in this repository. You can get them from nin_imagenet.caffemodel and VGG_ILSVRC_16_layers.caffemodel.

Dependent libraries are installed with the following script.

pip install numpy
pip install -r requirements.txt

RUN

Whole style transfer

You can transfer whole patch from one to another.

with CPU

python bin/convert_image_multi.py \
  --modelpath=./VGG_ILSVRC_16_layers.caffemodel \
  --iteration=100 \
  --gpu=-1 \
  --xsplit=1 --ysplit=1 --resize=300 \
  input.png \
  style.png \
  --output_image=./converted.png

with GPU

python bin/convert_image_multi.py \
  --modelpath=./VGG_ILSVRC_16_layers.caffemodel \
  --iteration=100 \
  --gpu=0 \
  --xsplit=1 --ysplit=1 --resize=300 \
  input.png \
  style.png \
  --output_image=./converted.png

Partial style transfer

Choose optimal patches from style image and transfer them to another image. Split style image to 2x2

python bin/convert_image_multi.py \
  --modelpath=./VGG_ILSVRC_16_layers.caffemodel \
  --iteration=100 \
  --gpu=0 \
  --xsplit=2 --ysplit=2 --resize=300 \
  --model=vgg_nopad\
  input.png \
  style.png \
  --output_image=./converted_optimal_2x2.png

Style transferred video

Tranfer style on video frame using last frame's result.

python bin/convert_video.py \
  --iteration=100 --model=vgg \
  video.mp4 \
  style.png \
  output_directory

Then you can find the style transferred video at output_directory/out.avi after 100 x frame times calculation.

Optimal Blended Texture Transfer

Please see https://nico-opendata.jp/en/casestudy/neural_style_synthesizer/index.html for technical details.

python bin/convert_image_multistyle.py \
  --model=vgg_nopad \
  --iteration=100 \
  --gpu=3 --xsplit=1 --ysplit=1 --resize=200 \
  /path/to/input/file \
  /path/to/directory/contains/multiple/refarence/files \
  --debug --out_dir=/path/of/output