How about the effect of L_twin only? #25
Description
Hi @QAQdev Thanks for your great work!
May I know the performance of keeping L_twin only?
From my understanding, it is a kind of “network reusage”, the generator network also serves as "real score" & "fake score", compared to the original DMD's three separate networks.
Therefore, I am curious about two ablations
-
$L_{twin}\ vs. L_{twin} + L_{base}$ ; $L_{twin}\ vs. L_{DMD}$