The output of the discriminator is the reconstruction cost. When we drop γ, we pay less attention to have the real image to easily spot against the generated image. Instead, it improves the reconstruction.

This turns out to improve image quality but with less diversity. Does that equivalent to what you said, I am not so sure? Because there are some researches that try to emphasize the generator training more with some simple mechanism but it does not see improvement. So I do not want to draw that conclusion for now and stay with the explanation in the paper instead.