Owen Yuwono open-sources five Three.js procedural graphics systems — ocean, grass, flora, fluid, planet

TL;DR

Token Gremlin’s thread bundles five newly open-sourced Three.js/WebGPU graphics experiments by Owen Yuwono (@owenyuwono) — Poseidon (real-time FFT ocean), Gaia (procedural grass from a deterministic genome + environment), Dryad (procedural trees/flora from physics + seed), Tiamat (~100k-particle SPH GPU fluid), and Demiurge (whole-planet generator coupling tectonics → uplift → erosion → climate → biomes, with seamless orbit-to-surface zoom). All run in the browser on WebGPU, all are physics-/procedural-code driven with no learned neural components, and all are packaged as reusable building blocks rather than one-off demos.

Key claims

Poseidon simulates large swells, ripples, foam, reflections, and choppy displacement using physically-inspired wave spectra entirely in-browser on WebGPU [tweet body].
Gaia generates every grass blade, seed head, and field from a deterministic genome + environmental parameters — no authored grass models [tweet body].
Dryad generates trees and other plant forms from physics + environmental conditions + a seed, again with no authored 3D models or textures [tweet body].
Tiamat runs an SPH fluid simulation with ~100,000 particles and renders the resulting water directly in the browser [tweet body].
Demiurge builds a whole planet by chaining tectonic plate generation → uplift-driven erosion → latitude/climate → biomes/wind/weather, and supports seamless orbit-to-surface navigation [tweet body].
The framing claim from Token Gremlin is that these are reusable primitives for games, simulations, and agent-built 3D environments — not standalone visual demos [tweet body].

Method

The five systems are classical procedural / physics-based graphics, not neural. Poseidon uses the FFT ocean formulation (waves as sums of complex-amplitude spectral components inverse-FFT’d to displacement) on WebGPU compute shaders. Gaia and Dryad are parameterized procedural generators: a seed + environment vector deterministically drives geometry synthesis for grass blades and plant morphologies respectively. Tiamat implements smoothed-particle hydrodynamics on the GPU (position + velocity per particle, kernel-weighted density/pressure updates) with real-time particle-to-surface rendering. Demiurge is a pipeline of coupled procedural systems layered so that upstream state (tectonics) drives downstream state (erosion, climate, biomes). No figure is embedded from the tweet.

Results

No quantitative benchmarks; the artifact is the running code plus the accompanying WebGPU demos. Practical claims from the tweet: browser-real-time on WebGPU for all five, ~100k SPH particles for Tiamat, and seamless orbit-to-surface transition for Demiurge. Each repo lives at github.com/owenyuwono/{poseidon,gaia,dryad,tiamat,demiurge}.

Why it’s interesting

Almost everything in the wiki’s world/scene-generation cluster is learned — diffusion, autoregression, JEPA. These five systems are a reminder that classical procedural + physics primitives cover a large chunk of the same output space (oceans, vegetation, terrain, atmospheres) with different tradeoffs: bit-deterministic, O(1) random access, no training data, no VRAM tail, but no learned realism and no natural-language conditioning. That maps closely onto the “infinite / stateless / learned-realism trilemma” framing in InfiniteDiffusion: Bridging Learned Fidelity and Procedural Utility for Open-World Terrain Generation — Owen Yuwono’s systems occupy the “infinite + stateless” corner that InfiniteDiffusion tries to bridge toward learned realism. They also complement bounded neural scene generators like WorldGen: From Text to Traversable and Interactive 3D Worlds and FlashWorld: High-quality 3D Scene Generation within Seconds: the neural systems produce one high-fidelity scene, procedural systems produce unbounded classical scenes, and a plausible near-term recipe is hybrid — neural conditioning on top of procedural infinite substrates.