[co]ral network is an art experimentation using AI models known as Generative Adversarial Networks or GANs.
To generate the desired images, a GAN needs to be trained with a set of training images so it can create new ones with the same characteristics.
In this project, two GANs have been trained :
- The first one on a dataset of 1463 images of healthy corals for a total of 204 hours of training.
- The second one on a dataset of 1693 images composed of 75% of damaged coral, which is the current state of coral life in Mauritius in 2022.
The results are AI dreams of different state of coral reefs:
- The healthy GAN generates beautiful, colorful and diverse corals.
- The damaged GAN generates dark, bleak and almost menacing coral landscapes.
What do we want for the future?
GAN Training / video editing / sound design and web development by Anthony Bouic - [email protected]
AI model : StyleGAN2 by NVIDIA Research Labs
Computer : Jupyter Notebook on Google CoLab
GPUs : Tesla P100 / Tesla T4