The paper introduces a unique methodology for terrain map generation, structured in two primary phases. Initially, established procedural map generation techniques such as Perlin noise and averaged smoothed sums of Gaussian noise are employed. Subsequently, a Neural Style Transfer is implemented, applying style attributes from real-world height maps onto the algorithmically generated image. This process marries the efficiency of procedural generation with the nuanced realism inherent to neural methods. The methodology aims to enable the creation of terrain maps with arbitrary content while retaining the morphological characteristics of a specific real-world region. The approach combines procedural generated images with style elements obtained from real-world height maps by transferring morphological properties using neural style transfer.

In the implementation, traditional procedural generation algorithms and the Neural Style Transfer technique are utilized. The structures of choice for representing the generated terrain are height maps, where each pixel's grayscale value corresponds to a specific elevation in the terrain. The paper details the specific map generation techniques, including the use of Perlin noise for generating patterns with a more organic appearance compared to explicit noise. The Neural Style Transfer technique is explained, utilizing the VGG-19 architecture and defining content and style based on the network's ability to classify images and compute correlations between different filter responses at multiple layers of the network. The final optimization loss is obtained by balancing the content and style components using tunable weights.

The methodology for terrain map generation involves obtaining a height map from the real world suitable for morphological transfer, creating procedural noise maps able to receive the extracted information, and tuning and executing the style transfer algorithm using the procedural noise as content and the height map from the real world as style. Additionally, a technique for including arbitrary features into the map is discussed, where a custom drawn map is averaged with a procedural generated map and used as a source for the style transfer. The computational load is distributed mostly on the style transfer step, while the procedural noise algorithm is noted for its efficiency.