The first ProGAN I ever trained. The first dataset I made using my recombinant style transfer technique. Starting with acrylic paint / pour art.
The dataset was probably too small for ProGAN despite the style transfer amplification–although it did manage to avoid mode collapse all the way to the full 1024x1024 resolution. The quick flashing transitions in the above animation show that the network was memorizing the dataset and wasn’t able to generate plausible interpolations. Perhaps more recent techniques like path length regularization would alleviate this despite the relatively small dataset.