GANs
[short presentation by Jen]
A great detailed walk-through of Training StyleGAN2ADA model in a Colab Notebook by Derrick Schultz
https://www.youtube.com/watch?v=uRLV26zlyZw
Google Colaboratory
Visual explanation of Vector Space by Lia Coleman and Artificial images
Intro to ML Art with RunwayML: Week 2
<aside>
🔥 Tips for gathering data to train StyleGAN
- More pictures gives you better results. Try getting between 500 and 5.000 images (but ideally even more! Max is 25.000).
- StyleGAN generates 1x1 style images as output (for instance 1024x1024 px), so think about cropping your data in a 1x1 format before feeding them to a Colab Notebook.
- Aim for collecting hi-res images (ideally around 1024 px x 1024 px, but realistically perhaps 500x500 pixels).
- If a GAN is going to produce "realistic" looking images, you should try to train it with coherent data, similar looking objects, similar angles, light, crops, backgrounds etc. Otherwise you get nightmare cats (which is also cool, but perhaps not what you want?).
</aside>
<aside>
🌐 Links for collecting data
</aside>