Why is it so hard to train GANs?

I will be starting a new project related to my master studies but my supervisor warned me about GAN training. He told me it would be tough if I don’t have any prior experience in training GANs. I read several blog posts, but I am really wondering why is it so hard to train them? What do you suggest to tackle that problem?

From my experience implementing some GANs from scratch (a couple of years ago now), I would recommend to have a lot of patience. :wink:
I had to run a lot of experiments to try to figure out what causes e.g. mode collapse etc.

I think there are a lot of good repositories now, which will work as a starter, also there was a lot of progress in this field.