I’m still tinkering with the code and the central issue is that I’d like to get better coverage of the area near the horizon, which seems to mean more than 5,000,000 iterations of training, and thus an increase of cell size. The following images show some of these explorations. I’ve also used an even more pointed Gaussian to soften the boundaries.
The top image uses the same parameters are previous, but using a exponential increase of neighbourhood size and 10px cells. The bottom image is the same, but with a linear increase of neighbourhood size. I think the middle image is the most successful because it is less dominated by the sky, and the spires are less dominant. The middle image is the result of a combination of large cell size (20px) and exponential increase of neighbourhood size. For the next iteration I’ve stuck with the 20px cells, exponential increase, and I’ve moved the horizon lower since the neighbourhood increase is initially very slow. With 20px I should also be able to get beyond the 5,000,000 iteration limit.