Philip Cadoux ITP Blog

Machine Learning for the Web - Week 08

November 02, 2020

CoralGAN

For this week’s assignment, I decided to use Runway to create my own Image Generation model.

Using the existing StyleGAN, I wanted to generate reef fish that do not exist.

First, I found this website, which had a ton of different coral reef pictures from all over the world.

I tried to choose the ones that had a bit more color and life to them, I also had to remove a ton of images because they were shot with land or in odd shapes (circles). But I still had about 400 images to use.

So, I got to training. I trained using StyleGAN2 against the Scenery pre-trained dataset. It worked out quite well.

After it was completed I was incredible happy with the results. I wish I trained it with a couple thousand pictures, but with 400 images and 5000 steps, I got a mix of hyper-realistic, artistic, and haunting reefs. Here are some examples.

Realistic Image A Realistic Image B Artistic Image A Artistic Image B Haunting Image A Haunting Image B


Written by Philip Cadoux, current ITP student and Creative Technologist. Follow me on Instagram