comparison to MG

[toyot-li]

Hi @hanyang-21 Thanks for your great work!

I think an intuitive understanding is similar to a recent work "Momentum Guidance: Plug-and-Play Guidance for Flow Models"?

Both use a momentum-style update for the CFG's diff term, while MG explicitly removes the double inference.

[hanyang-21]

Thanks for your interest in our work! The motivation and formulation of CFG-Ctrl are quite different from MG. Our method is derived from a control-theoretic view of CFG, whereas MG is motivated by reusing past velocities to approximate guidance without extra inference. At the same time, both methods are driven by a closely aligned intuition about correcting oversmoothing along the generative trajectory, and the resulting perspectives are largely consistent and mutually reinforcing, suggesting a common direction for more principled guidance design.