Related: I know the authors used a DCGAN implementation (a pre-trained model, it looks like), but is it known what their approach for up-scaling the generated images is? In art generation I've seen GAN output of 128x128, e.g. that is then upscaled with a super-resolution network. Is something similar being done for the "final painting", or is the GAN somehow efficient enough to do large-format output in a decent training time?
I don't know exactly what was used to upscale here, but Progressive Growing of GANs[0] was a breakthrough last year that proved at least 1024x1024 was viably producable.
The short explanation is: Train and freeze the most basic layer of the model progressively to "let the network" understand higher resolution concepts one piece at the time, and avoid mode-collapse.
The network architecture illustrating this a bit better is shown on page 3 of the paper [1]
I think they’ve upscaled using Lanczos resampling and have then done wavelet deconvolution - the thing is furry with artefacts, and looks just like when I push an astronomical image too far.
