The point is that the GAN is trained to model the "probability distribution" of the descriptions it's trained on precisely so that the concepts are generalized

The point is that the GAN is trained to model the "probability distribution" of the descriptions it's trained on precisely so that the concepts are generalized and can be synthesized/extrapolated. Once trained, you may fix the random seed so that the same description would generate the same image.