23M-param diffusion model generates game terrain, trained in 4.5h on one RTX 5060, runs in browser via WebGPU
Old_Cow_6636 · reddit · 2026-10-10
A Reddit user open-sourced Talus, a 23M-parameter diffusion model generating 64x64 game terrain heightmaps conditioned on terrain type and any subset of five measured properties.
- Cheap to train: data from the author's own procedural generator (fBm/ridged noise, erosion models); trained from scratch on one RTX 5060 (8GB) in 4.5 hours
- Architecture: pixel-space U-Net, v-prediction, cosine schedule, 50-step DDIM, CFG 2.0; learned per-property unknown embeddings allow arbitrary conditioning subsets
- Relative heights: normalized shapes placed by the sampler at requested elevation/relief, fixing grainy plains (distance ratio 3.98 → 1.23)
- Rigorous eval: distances normalized by real-vs-real noise floor; current model scores 1.51x (metric W1), 9.1x (spectrum), 1.65x (slopes)
- Browser demo: ONNX fp16 export + ONNX Runtime Web WebGPU, 3s per map; JS sampler matches PyTorch within 0.6m
Apache-2.0 code and weights on GitHub; the author is seeking tips on closing the remaining spectral-band gap.
More from Research
- RFD4-Proteina's 'underrated' docs detail partial diffusion for protein redesign — AllThingsApx · 2026-10-23
- MPI-Sec PhD Thesis on Vulnerability Detection Benchmarks Includes Pre-LLM Classic "Top Score on the Wrong Exam" — mboehme_ · 2026-10-23
- LiveNerf's 29-day data finds no nerfing of Claude Opus 5.5 — TheOnlyVibemaster · 2026-10-23
- New paper makes 9 falsifiable predictions toward a major AI-enabled terror attack by 2027-2028 — StephenLCasper · 2026-10-23
- HumanoidPose recovers humanoid robots' full-body joint states from RGB video alone — cholz · 2026-10-23
- Sophont's Holon-S1 predicts genes and biomarkers from routine pathology slides — iScienceLuvr · 2026-10-23