The question about the loss

[martinliuxy]

Great work for keeping the ability of the original model. However, I am confused about the role of the L_align-sem loss when L_align-layout is already present. Looking forward to your reply!

[ToTheBeginning]

As mentioned in the paper, L_align-layout tends to have a greater degradation on ID fidelity, while in comparison, L_align-sem is comparatively more gentle.