The style image in neural-style does not have to be a work of art. In this example, I will take a part of a topographic map, like this and apply it to two photographs, a building and a street view.
The first save, after 100 iterations, is essentially a deformed image built from the style. The photograph showing a building is not yet discernible. But quite soon the form of a building emerges, the lines and dots get arranged and bent according to an outline of the house. And now that we start to see the house, it is possible to see the form of the house already in the first image.
The final image looks like this. The schoolhouse is drawn entirely using the elements of the topographic map, while at the same time the picture has the feel of a free hand drawing.
Let’s have a look at a street view. In this example I have scaled the style to 0.5 to have a more detailed texture to match the more detailed street image.
The first save again is a deformed image full of features from the topographic map. The emerging street view is there already, but not yet perceptible if you don’t know what to look for.
Quite soon then, the street view emerges…
This, then, is about as clear as it gets. Again, the view is drawn using the elements of the map, and there is also a strong hard drawn feel. Look, for instance, at the windows straight ahead on the right hand side.
We can, of course, also test what the map style does to a portrait:
Lowering style weight from 100 to 50 and then further down to 10, the look of the topographic map starts to disappear while the general feeling of the picture partly remains, partly changes.