Talus
Talus generates terrain heightmaps for games. You pick a terrain type and any of five measurable properties (mean elevation, relief, slope, water coverage, roughness), and Talus draws a 64x64 map covering 4 x 4 km with heights up to 1,200 m. It is a 23M-parameter conditional diffusion model trained from scratch on 45,000 maps from our own procedural generator. It runs on an 8 GB consumer GPU, and on a CPU if you can wait a few seconds per map.
- Try it in your browser: talus.tersa.tech. The model runs on WebGPU (or the CPU) and exports PNG, Unity RAW and OBJ.
- Code, training and benchmarks: github.com/osfv/talus
Talus-3 is the default model. Talus-3.1 is a variant for flatter maps: better plains, islands and mesas, worse ridges and mountains.
Files
| File | Size | Contents |
|---|---|---|
talus-3.pt |
93 MB | Talus-3 (default): EMA weights, configs and dataset statistics. No optimizer state. |
talus-3.1.pt |
93 MB | Talus-3.1 (flat-terrain variant), same format. |
onnx/talus-3.onnx |
93 MB | Talus-3 denoiser exported to ONNX (fp32). |
onnx/talus-3.fp16.onnx |
47 MB | The same network with fp16-stored weights. The browser demo uses this one. |
onnx/talus-3.json |
25 KB | Sampler metadata for the ONNX files: world scale, conditioning statistics, noise schedule and sampling preset. |
onnx/talus-3.prior.bin |
252 KB | Placement bank the sampler uses when you leave mean elevation or relief open. |
SHA256SUMS.txt |
Checksums. They match the GitHub releases. |
The .pt files load with torch.load(weights_only=True). Sampling needs the nullscape package from
the GitHub repo, which holds the model code and the sampler.
Quick start
You need Python 3.11+ and PyTorch 2.7+. RTX 50-series cards need a CUDA 12.8 build of PyTorch.
git clone https://github.com/osfv/talus
cd talus
pip install torch --index-url https://download.pytorch.org/whl/cu128
pip install -e . huggingface_hub
hf download osfv/talus talus-3.pt --local-dir checkpoints
Generate 16 island maps with 60% water and export them for Unity, Unreal or Godot:
nullscape sample --checkpoint talus-3 --archetype islands --prop water_fraction=0.6 \
--n 16 --formats png16,r16,obj --unity --out exports/islands
Add --device cpu if you have no CUDA GPU. For the flat-terrain variant, download talus-3.1.pt the
same way and use --checkpoint talus-3.1.
Which model
| You are making | Use |
|---|---|
| A mix of terrain, or ridges, hills and mountains | Talus-3 |
| Plains, islands and coasts, or mesas and plateaus | Talus-3.1 |
Results
Scores on the held-out TEST split, 1,000 maps per half. Each score runs from 0 to 100, where 100 means indistinguishable from real procedural terrain at this sample size.
| Benchmark | What it measures | Talus-1 | Talus-2 | Talus-3 | Talus-3.1 |
|---|---|---|---|---|---|
| Realism | 25 terrain metrics vs real maps | 34.4 | 56.5 | 66.2 | 64.5 |
| Spectrum | detail at every scale (power spectrum) | 3.6 | 7.6 | 11.0 | 11.3 |
| Heights | elevation distribution | 88.5 | 98.3 | 99.1 | 98.3 |
| Slopes | slope distribution | 41.2 | 56.4 | 60.5 | 59.0 |
| Control | requested vs measured properties | 81.8 | 90.4 | 92.1 | 92.4 |
| Playability | walkable-route pass rate vs real maps | 98.4 | 97.6 | 99.9 | 99.3 |
| Diversity | variety vs real maps | 96.9 | 94.4 | 95.0 | 94.6 |
| Originality | distance from the training maps | 100 | 98.1 | 99.1 | 98.4 |
| Clean detail | no excess micro-bumps | 58.4 | 97.7 | 99.4 | 99.5 |
| Average | 67.0 | 77.5 | 80.3 | 79.7 |
Per-type realism on TEST (distance to real maps over the noise floor, lower is better):
| plains | islands | mesas | hills | mountains | ridges | |
|---|---|---|---|---|---|---|
| Talus-3 | 1.17 | 1.21 | 1.43 | 1.16 | 1.71 | 1.80 |
| Talus-3.1 | 1.13 | 1.16 | 1.33 | 1.20 | 1.75 | 1.93 |
Both sample about 5.6 to 5.8 maps/s at batch 128 on an RTX 5060 (8 GB), or one map in 1.3 s. Full scorecards and the benchmark protocol are in the GitHub repo.
How it works
- Data. A procedural generator builds each map from fBm, billow or ridged multifractal noise, then runs stream-power erosion, hillslope diffusion and thermal erosion. The dataset has 50,000 maps, split 45,000 / 2,500 / 2,500 into train, validation and test. Talus-3.1 continued training on 96,000 fresh maps from the same generator.
- Model. A pixel-space U-Net denoiser (v-prediction, cosine noise schedule) conditioned on the terrain type and the five properties. Training drops each property at random, so the model learns every subset, and classifier-free guidance strengthens the conditioning at sampling time.
- Sampling. 50 DDIM steps with quadratic spacing and guidance 2.0, stored in the checkpoint as its default preset.
Limitations
- Talus only knows the procedural world it was trained on. It has never seen real elevation data.
- 64x64 cells at 64 m is coarse for close-up play. Treat the output as a base layer for detail noise or upscaling.
- Mountains and ridges remain the weakest types, and the power spectrum is still 9x the noise floor, mostly at the finest scales.
- The roughness control (
spectral_beta) is the loosest of the five properties.
License
Code and weights are under the Apache-2.0 license.