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.