How about the effect of L_twin only?

[toyot-li]

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

1. $L_{twin}\ vs. L_{twin} + L_{base}$;
2. $L_{twin}\ vs. L_{DMD}$

[QAQdev]

Hi, sorry for a late reply. L_twin is very similar to DMD loss (https://arxiv.org/abs/2311.18828). Using L_twin only is more like naive DMD without regression loss, it is kind of unstable on Qwen-Image-20B