2026-10-09 · America/Los_Angeles · 社区动态 · #8

Talus: a 23M-parameter diffusion model for game terrain, evaluated against a real-vs-real noise floor, running in the browser on WebGPU [P]

Talus generates 64x64 heightmaps (4 km, up to 1,200 m) conditioned on a terrain type and any subset of five measured properties: mean elevation, relief, mean slope, water fraction and spectral slope. I trained it from scratch on one RTX 5060 (8 GB), about 4.5 hours in total across versions. The data is 45,000 maps from my own procedural generator (fBm/ridged noise, stream-power erosion, hillslope diffusion, thermal erosion). Model Pixel-space U-Net, v-prediction, cosine schedule, 50-step DDIM with quadratic spacing, classifier-free guidance 2.0 Each property has a learned "unknown" embedding…

热度 49.0 / 100；排名与评分保留该期记录。