I’m launching a new hobby deep learning project ThisBeachDoesNotExist.com – a synthetic beach image generator. StyleGAN2-ADA network trained in non-square HD resolution 1280×768 on a dataset with ~20.000 images of the beach.
The biggest problem seems to be the trunks and leaves of the palm trees. I see two main reasons why the details are not drawn out:
– too small dataset 20k (NVidia uses datasets like 40-70k+)
– StyleGAN likes images with one distinct object in the center. Landscapes are the exact opposite.
Training progress
The video shows the learning process of the network. The learning progress is measured in kimg (kilo images). The network was trained for 25,000 kimg (until it had seen 25,000,000 images).
Download
You can download the network for your experiments or retrain the network with your own dataset (transfer learning): https://bit.ly/thisbeachdoesnotexist-v1-pkl