Show HN: TerrainSR – fast, realistic heightmap upscaling model

TerrainSR is a purpose-built terrain upscaling model that converts coarse 100 m-resolution heightmaps into realistic, higher-detail 10 m height data. Trained on paired 100 m→10 m examples and intentionally excluding built-up or heavily human-modified areas, it generates plausible natural terrain detail (and can extrapolate finer detail than 10 m in some cases) while breaking down on inputs coarser than 100 m. It runs extremely fast - on an RTX 4070 it produces a 50×50 km patch (100 m → 10 m) in about 0.74 seconds after loading - making it suitable for realtime and game workflows. It produces more varied results than traditional erosion simulators, though some coarse artifacts (for example from large buildings present in 100 m data) can be converted into hill-like features.

Deployment uses Python 3.10+ and PyTorch 2.6+, with a simple CLI and a Python API: provide a 2D coarse height array (100 m grid) and a 10 m water mask (ten times the resolution), include 3.2 km of padding around the target area and use a crop parameter to remove context. Input dimensions must be even and ≤564 cells; outputs are float32 heights in metres. Weights and inference code are Apache‑2.0 licensed; elevation and land-cover sources include swisstopo, Kartverket, IGN, USGS, Copernicus DEM, ESA WorldCover and OpenStreetMap under their own terms.