Code accompagning the article 3D Geological Image Synthesis From 2D Examples Using Generative Adversarial Networks, Guillaume Coiffier, Philippe Renard and Sylvain Lefebvre
Generative Adversarial Networks (GAN) are becoming an alternative to Multiple-point Statistics (MPS) techniques to generate stochastic fields from training images. But a difficulty for all the training image based techniques (including GAN and MPS) is to generate 3D fields when only 2D training data sets are available. In this paper, we introduce a novel approach called Dimension Augmenter GAN (DiAGAN) enabling GANs to generate 3D fields from 2D examples. The method is simple to implement and is based on the introduction of a random cut sampling step between the generator and the discriminator of a standard GAN.
For pretrained models, training images and sample of outputs, see https://github.com/randlab/DiAGAN_Examples
python (>3.5) with the following librairies :
- Pillow
- Numpy
- mpstool : https://github.com/UniNE-CHYN/mps_toolbox
- py-vox-io (optionnal) : https://github.com/gromgull/py-vox-io
- Tensorflow (for the TF version)
- Keras (for the TF version)
- Pytorch (for the pytorch version)
See the HOW_TO_USE.md files in the Pytorch/ and Tensorflow/ folders
| Training examples | Output images | Output images |
|---|---|---|
  |
  |
  |
  |
  |
  |
 |
 |
 |
 |
 |
 |