>All DiscoGAN experiments are on 64x64, this is high resolution. I don't know whether this is an important difference though.
I don't think it is. Working on 64x64 makes global coherence easier and the NNs smaller/faster but shouldn't make a qualitative difference. I believe one guy using DiscoGAN/CycleGAN on my current Danbooru anime dataset ( https://www.gwern.net/Danbooru2017 ) is doing it at 128px without any major changes.