random-texture-synthesis

Code for the ICLR 2017 paper “What does it take to generate natural textures?”

How to run

Synthesising textures

The main function is synthesise.py, which allows to synthesise textures using random shallow models described in the paper.

Running

python3 synthesise.py -t texture.jpg -s 200

will synthesise a new texture of size 200x200 using the multi-scale model described in the paper and save it as texture_synthesised.jpg. You can change the sizes of convolutional filters for different scales using the “--scales” keyword. For example,

python3 synthesise.py -t texture.jpg -s 200 --scales 11

will synthesise a new texture using a single-scale model with 11x11 filters.

Additional arguments to synthesise.py are as follows:

$ python3 synthesise.py -h

usage: synthesise.py [-h] -t TEXTURE -s SIZE [-f TARGET_FILE] [-n N_ITER]
                     [-c N_FEATURES] [--scales [SCALES [SCALES ...]]] [-l]

optional arguments:
  -h, --help            show this help message and exit
  -t TEXTURE, --texture TEXTURE
                        Path to the reference texture
  -s SIZE, --size SIZE  Size of a synthesised texture in pixels
  -f TARGET_FILE, --target-file TARGET_FILE
                        File name of a syntesised texture
  -n N_ITER, --n-iter N_ITER
                        Number of L-BFGS optinisation iterations
  -c N_FEATURES, --n-features N_FEATURES
                        Number of feature maps per each scale
  --scales [SCALES [SCALES ...]]
                        Sizes of convolutional filters
  -l, --linear          Use linear model (conv layer without non-linearity)

• -t is a path to a reference textures,
• -s is a size (in pixels) of a synthesised textures,
• -f is a filename of a synthesised texture (if not provided, by default _synthesised.jpg is added to the filename of the reference texture),
• -n is a number of L-BFGS optimisation steps (default 4000),
• -c is a number of feature maps per scale (i.e. for each filter size),
• --scales are the sizes of convolutional filters (if not provided the multi-scale model is used by default with filters sizes 3, 5, 7, 11, 15, 23, 37, 55,
• -l makes the model linear (no non-linearity after the conv layer is used).

Evaluating VGG-loss

vgg_loss.py allows to evaluate the VGG-loss of a synthesised texture. Function main in this file takes two arguments (path to the reference texture and path to the synthesised texture) and returns the value of the VGG-loss:

import vgg_loss

loss = vgg_loss.main('reference_texture.jpg', 'synthesised_texture.jpg')

Figures

Folders for respective figures contain code to generate these figures (e.g. make_figure2.py) using the provided synthesised textures as well as the shell scripts (e.g. synthesise_figure2.sh) to generate new textures.