TerrainSR
huggingface.co- Category
- AI Tools
- Rank
- No. 2435Tools index
Previous survey · No. 2331 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Data Processing
- Date
About
TerrainSR is a model that generates higher-detail terrain height data from lower-detail height data, trained on 100m to 10m pairs from undeveloped land so outputs avoid artifacts like mines and buildings. On an RTX 4070 it upscales a 50x50km patch in 0.74 seconds. It is available for PyTorch and ONNX, and was built for a real-scale historical strategy game.
What it does
The site says TerrainSR is a super-resolution model for elevation maps: it takes coarse height grids plus a fine water mask and adds plausible terrain detail. The author says it was trained to avoid human-made features like roads and cities, and describes it as faster and more varied than a traditional erosion simulator.
Stated on the product site
- License
- Apache 2.0 for weights and inference code
- Interfaces
- Command-line script and Python class
- Runtime
- PyTorch, GPU or CPU
- Input range
- Intended for data at 100m resolution or finer
- Training data sources
- swisstopo, Kartverket, IGN and USGS elevation data
Not stated on the site
- How output quality compares against measured elevation data is not reported on the page.
- Whether the model is tested on terrain outside the regions covered by the training sources is not stated.
- The page does not say how the separate dataset terms affect use of generated outputs.
Written from the product site at huggingface.co.
Tags
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.