About identity loss weights in CycleGAN #99

Professor, in the lecture we learned that the total loss of CycleGAN is composed of GAN loss, cycle consistency loss, and identity loss, and that the identity loss mainly contributes to preserving color and overall appearance.

From a theoretical point of view, what phenomena occur when the weight of the identity loss is set too large versus too small? For example, how does the trade-off manifest between preserving color/content? (e.g., images staying too close to the input) and achieving sufficient style change (e.g., color shifts becoming too strong or style changes becoming excessive) Could you elaborate on how this trade-off appears in practice and in theory?