Real world performance lacking?
#3
by Ahin23 - opened
Are there specific preprocessing steps / transforms that need to be used before passing masks and images to the places or celebahq variant? So far the real world performance seems very lacking compared to the benchmarks shown on the paper. I implemented it as stated
I can second that, used the parameters as stated in the paper and the results are very disfigured (ffhq / celebahq) / full of artifacts (places2) on anything but the benchmark photos. Is there something we are missing?