Neural Network Terrain Generator trained on handmade 3D models.


[Update] ML Terraform was developed in 2018 using a DCGAN. The project has been deprecated due to general lack of interest in this type of tool.

Real world terrain is not random. Even using Perlin noise and similar technics won't output a realistic looking terrain as erosion and sedimentation are difficult to do programmatically. Desert dunes, snow, rocky mountains, snow mountains all have a different type of characteristics that would take a lot of time to define and imitate programmatically. Most terrain generation tools require artistic work after a terrain is generated randomly to make it look more convincingly real. A neural network can help this task by learning the characteristics that make a realistic looking terrain, given a good set of terrains to learn from.

ML Terraform approaches this problem by using a neural network to learn from handmade 3D models. It is easier to use than other terrain generators, not requiring so much artistic skill and messing with terrain stamps and texture spawners.

The Unity Plugin allows to automatically merges adjacent terrains and generate landscapes of terrains with a single click, such as 5x5 or any other dimension. All individual terrains inside the landscape remain independently configurable.



Does not support .Net 3.x used in older versions of the Unity Editor.


Bug reports or business inquiries please contact me at