[tonygrue]

Paper here: https://arxiv.org/pdf/1903.07291.pdf

Page 14 and beyond show a number of examples vs previous methods.

IIUC, their contribution is a masking normalization layer (the 'SPADE' layer [see pg 4]) that allows the network to preserve the structure of the input segmentation/mask, leading to results quite a bit nicer than previous ones.