Show HN: TerrainSR – fast, realistic heightmap upscaling model
4 hours ago
- TerrainSR enhances low-resolution terrain data (e.g., 100m to 10m) for realistic detail generation.
- It is trained without human-made features (e.g., cities, mines) to avoid artifacts in outputs.
- Useful for historical strategy games by adding plausible terrain on compact 100m data.
- Faster and more varied than traditional erosion simulators for large terrains.
- Generates a 50x50km patch in 0.74s on RTX 4070, suitable for real-time gaming.
- To use, download repo, install Python 3.10+, PyTorch 2.6+, and run with example or custom arrays.
- Input requires 100m height array and 10m water mask with 3.2km context; output is 10m heights.
- Model weights are Apache 2.0 licensed; training data has separate terms.