TerrainSR

TerrainSR is a machine learning model designed to upscale low-resolution heightmap data into high-resolution terrain with realistic features. Trained on undeveloped landscapes, it adds plausible erosion patterns, rocks, and natural details to 100m resolution data without the storage overhead or noise of raw LIDAR datasets.

Originally created for historical game development, TerrainSR reconstructs terrain as it would appear in natural, pre-development states by removing artifacts from manmade objects. The model enables efficient, large-scale terrain generation at 1:1 scale without requiring hundreds of gigabytes of patchy source data.