Data cleaning negatively affects the quality of generated images,
confessions of a GAN artist
Show this thread
-
they never converge but could generate interesting stuff (typo) and they can be run much longer than the cleaned datasets , by never converging they can indefinitely produce some images that have potential
models trained on well formed but small datasets tend to get into mode collapse quite quickly
models trained on diverse (small) datasets prob experience mode dropping - they never converge could generate interesting stuff 
