Repository navigation
Talus-3.1 (flat-terrain variant)
Talus-3.1 is a variant of Talus-3 for flatter maps. It is closer to real maps on plains, islands and mesas, and further from them on hills, mountains and ridges. Talus-3 (v3.0.0) stays the default. Use Talus-3.1 when your maps are mostly flat ground, coastlines or plateaus.
gh release download v3.1.0 --repo osfv/talus --pattern talus-3.1.pt --dir checkpoints
nullscape sample --checkpoint talus-3.1 --archetype mesas --n 16 --formats png16,r16,obj --unityHow it differs from Talus-3
Per-type realism on the held-out TEST split (distance to real maps over the real-vs-real 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 |
Scorecard (0-100, higher is better): Realism 64.5 (Talus-3: 66.2), Spectrum 11.3 (11.0), Slopes 59.0 (60.5), Control 92.4 (92.1), Playability 99.3 (99.9), Originality 98.4 (99.1). Average 79.7 (80.3).
How it was made
Talus-3 continued for 12,000 steps on 96,000 procedural maps it had never seen (same generator, new seed). We picked it on the validation split as a variant before running TEST, and TEST could not change the default. Training, selection and memorization checks: docs/TALUS31.md. Full scorecard: docs/TALUS31_SCORECARD.md.
Files
talus-3.1.pt(93 MB): EMA weights, configs and dataset statistics. No optimizer state. It loads withtorch.load(weights_only=True).SHA256SUMS.txt: checksum.
Code and weights: Apache-2.0. The browser version at talus.tersa.tech runs Talus-3.