/multi-scale-style-transfer

Multi-scale style transfer with a pyramid of fully convolutional GANs inspired from "SinGAN: Learning a Generative Model from a Single Natural Image" (ICCV 2019)

Primary LanguagePythonMIT LicenseMIT

multi-scale-style-transfer

Multi-scale style transfer with a pyramid of fully convolutional GANs inspired from SinGAN: Learning a Generative Model from a Single Natural Image (ICCV 2019)

Repository Structure

  • The SinGAN_master directory contains the original, unaltered version of SinGAN.
  • In architectures you will find different networks that can be used for the Pyramid Model.
  • Images (content and style) should be stored in Images.
  • main.py is the ... main file you will want to run using command line arguments similar to the ones used by SinGAN.
  • parse.py contains the necessary command line arguments.
  • training.py handles all of the training for a given Pyramid Model.
  • utilities.py has the necessary functions that don't belong elsewhere.

Usage

The basic structure of a command looks like this:

python main.py --content_input <content input image> --style_input <style input image>

Visualization

For visualization, you will need to install visdom first: pip install visdom. Then run the visdom server in a separate terminal:

python -m visdom.server

and lastly run main.py as usual with the following flag:

python main.py ... --vis True