notice how the lighting in the style-transferred images isn't physically plausible according to the target images. For example, in the 5th example the house is still lit as in the original, as if by sunlight not by spotlights. Maybe that's why they've chosen to showcase examples with flat or ambiguous lighting (like the nighttime scenes or the autumn scene). The DNN doesn't model the physical reality of the scene, it doesn't get that it's a 3d world, it simply transports a high-dimensional vector (the 2d image) from one space to another. What our imaginations do is map that 2d image back into 3d before transforming it.
And the city images that are styled into night were originally taken near dusk so the lights inside the buildings were already on but not very pronounced. It will not take a full daylight scene and decide where to turn the lights on at night.
