How to compare the efficiency of different GAN models?

I'm comparing different GAN models (CGan, DCGan, WGan, StyleGan) in tensorflow2. In general, I want to use the images that I generate with the generator to train a classifier while being as realistic as possible. At first, I wanted to let them train for 24 hours each, define some early stopping criteria and save the checkpoints with the lowest loss through a callback. But it seems that the lower loss does not always lead to more realistic images. So how do I compare the different models in a scientific way?