Some good-humored people wrote a paper about [Generative Unadversarial Networks]((https://arxiv.org/abs/1703.02528) striving to end violent behaviour by discriminators.
It reads as if it had been published on April 1st - good fun.
Still, I wonder if they have a point in replacing the Discriminator by a “Motivator” (albeit without the powerpoint slides, please). If D could give more feedback to G by pointing out ways to improve the image or where in the image the problem lies - would it help G to learn faster? So, output of D would be p (probability for image being real) and I (image map of probabilities). G could use I as additional input. Afaik they didn’t write anything like that, but that thought crossed my mind.
I think that’s an interesting idea - see if you can get something working and compare! 
I’ll have to concretize what I exactly looks like. Hm, that goes inline with what my group is planning for this week anyways. Let’s see what I can get.
The tricks used by segmentation networks (Unet, densenet, etc) with their shortcut connections could help there.
Wow, I admire the work of Olaf Ronneberger. Thanks for pointing this out. Currently, I am first getting some simpler GAN to work, but then I’ll look into this idea.