I would suggest rewriting the code for your own data/purpose, using the reusable classes/functions from my code, rather than simply trying to run the code as it is. The code has some requirements that I tried to set as defaults in my system (like paths, etc.)
Yes, using a GAN, you should be able to generate images by giving only a word as input. The GAN would learn features corresponding to each label and after training, one could use the generator part of the model to generate images of a given label.
It sounds like reverse of Image Captioning. There is good research going on in that field.