ProbableTrain
commented
4 years ago If you don't want the input image to affect road direction, this is probably best done in post (i.e. photoshop).
Otherwise, you'd have to do something much more complex involving inferring the tensor field from the image, then clipping the generated city to the tensor field. The original paper (Chen et al, Interactive procedural street modeling) had a 'brush' tool to edit the tensor field manually - that could be a good starting point.
Please import transparent PNG and generate map, something like this: https://www.pinterest.com/pin/266345765444848670/