It looks different because a lot of things in the image are not relevant for planning and so not encoded in the output of the vision model. The GAN doesn't reconstruct those details faithfully. I updated the image caption so it is more clear.