GoKawiilNew AI model TerrainSR upscales 100m heightmaps to 10m terrain detail
An independent developer released TerrainSR, a model trained on paired 100m-to-10m elevation data that generates higher-resolution terrain heightmaps from coarser input. The model excludes built-up and human-modified land from its training data so outputs avoid artifacts like roads or buildings, and on an RTX 4070 it can process a 50x50km patch in 0.74 seconds.
GoKawiil's interpretation of the reporting above, not reported fact.
The developer says the tool was built to solve a specific problem in historical strategy game design: getting realistic terrain detail across a continent-scale map without the storage cost of true 10m data or the manual labor of scrubbing modern infrastructure from it. Its speed compared to traditional erosion simulators suggests potential use in other realtime applications such as gaming or simulation, though broader adoption would depend on how well it generalizes beyond the creator's test cases.
- TerrainSR upscales 100m-resolution heightmap data to 10m detail, extrapolating further where input resolution is higher.
- It was trained without cities or human-modified terrain, so outputs avoid artifacts like roads or mines, though it can occasionally reinterpret steep building shapes as hills.
- It generates a 50x50km terrain patch in under a second on consumer GPU hardware, making it fast enough for realtime use such as games.
Source: huggingface.co, 2026-10-07
Published there as: “Show HN: TerrainSR – fast, realistic heightmap upscaling model”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.