fegemo / multi-domain

Experiments of image to image translation with multi-domain GAN architectures for pixel sides

Challenges of Pixel Art Generation with GANs

This repository contains the source code used by the paper "On the Challenges of Pixel Art Generation with GANs" to be presented at AIIDE 2022.

Installation

The code uses Tensorflow 2.9.1 and requires the following dependencies to be installed:

• jupyter
• tensorflow
• tensorflow_addons
• matplotlib
• scipy
• scikit-image

For convenience, a requirements.txt file is available with the necessary packages and their versions. In a fresh virtual environment, you can:

pip install -r requirements.txt

Running

The notebook experiments.ipynb can be executed from the first to the last cell to execute the training procedure of the four models presented in the paper:

1. baseline (no aug.)
2. baseline
3. palette-indexed
4. with histogram loss

To define which model to train, you can set the value of the model variable in the 3rd cell:

MODELS = ["baseline (no aug.)", "baseline", "indexed", "histogram"]
model = MODELS[0]  # <-- CHOOSE HERE

There is also some configurations over configuration.py, like the batch size, the maximum size of a color palette and so on.

Tensorboard can be used to debug the performance of the models during training. You can point its log directory to temp-side2side to see the logged scalars and images:

tensorboard --logdir temp-side2side