Keras model deteriorates when saved and loaded back in

Hi! So I am working with a GAN architecture with a generator and discriminator (critic). I have written my own training loop and therefore have not used a class version of GAN.

When I save my generator and the load it back in it gives vastly different results, see the picture. Have anyone experienced anything similar? I have done it both saving the full model and also only saving the weights (see picture) with the same outcome. There are three histograms where two is overlapping.