High Resolution Texture Transfer

Requirements

Visit https://gist.github.com/ksimonyan/3785162f95cd2d5fee77 and download VGG_ILSVRC_16_layers.caffemodel.
Put downloaded file into this directory.

$ python src/create_chainer_model.py

$ python src/run.py content_image style_image output_dir [options]

Example:

$ python src/run.py content_image.png style_image.png output/texture -g 0

Parameters:

content_image: (Required) Content image file path
style_image: (Required) Style image file path
output_dir: (Required) Output directory path
-g (--gpu) <int>: (Optional) GPU device index. Negative value indecates CPU (default: -1)
-w (--width) <int>: (Optional) Image width (default: 256)
--iter <int>: (Optional) Number of iteration for each iteration (default: 2000)
--initial-image <str>: (Optional) Initial image of optimization: "random" or "content" (default: random)
--save-iter <int>: (Optional) Learning rate (default: 1)
--content-layers <str> <str> ...: (Optional) Layer names to use for content reconstruction (default: relu3_3 relu4_3)
--style-layers <str> <str> ...: (Optional) Layer names to use for style reconstruction. (default: pool1 conv3_2)
--content-weight <float>: (Optional) Weight of content loss (default: 5)
--style-weight <float>: (Optional) Weight of style loss (default: 100)
--tv-weight <float>: (Optional) Weight of total variation loss (default: 1e-3)

Experimantal options:

--keep-color: (Optional) Keep color phase
--match-color-histogram: (Optional) Use "Color histogram matching" algorithm in "Preserving Color in Neural Artistic Style Transfer"
--luminance-only: (Optional) Use "Luminance-only" algorithm in "Preserving Color in Neural Artistic Style Transfer"

$ python src/run.py sample/tubingen.jpg sample/block.jpg output/synthesized -g 0 -w 800