Code for the paper: "Generalized Rectifier Wavelet Covariance Model For texture Synthesis" (Brochard, Zhang, Mallat, ICLR 2022) https://openreview.net/pdf?id=ziRLU3Y2PN_.
Requirements:
- Pytorch (version >=1.8.0)
- Kymatio (
pip install kymatio) to create the Morlet wavelet filters. - Numpy, scipy, scikit-image, Pillow, tqdm, matplotlib
Create the wavelet filters by running python build-filters.py
To generate a synthesis:
- Grayscale images:
python synthesis/gray.py - Color images:
python synthesis/color.py
Specify the input image with the argument --image (e.g. python synthesis/color.py --image honeycomb). The argument --save stores the synthesis in the 'results' folder, in .npy format. Other arguments, to specify the model, can be found in 'synthesis/gray.py' and 'synthesis/color.py'.
To generate a synthesis using PS model:
- Grayscale images: ps/demo_gray.m
- Color images: ps/demo_color.m
p.s. You need download the original Matlab code (http://www.cns.nyu.edu/~lcv/texture/)
To generate a synthesis using RF model:
- Grayscale images: rf/do_synthesis0.sh
- Color images: rf/do_synthesis01.sh
p.s. You need install the following packages using conda:
- theano, version 1.0.4
- matplotlib, scikit-image
- Lasagne-0.2.dev1, install via pip:
pip3 install --upgrade https://github.com/Lasagne/Lasagne/archive/master.zip # Lasagne-0.2.dev1
- To use GPU, configure ~/.theanorc
[global]
floatX = float32
device = cuda
Releasing under an Apache-2.0 license.