SeedThree

Open-source procedural tree & plant generator for the web, built on Three.js (WebGPU).

▶ Use Live NOW!  (WebGPU-capable browser required — Chrome/Edge)

A fully procedural tree and plant generator: pick a species, tune its parameters, and get a unique, textured, wind-animated 3D plant you can drop into a scene or export to glTF.

Status: v0.1.0-alpha. Twenty species, full LOD + export pipeline, and a living scene are in — but it's early and rough in places. Expect sharp edges.

• Twenty species across two biomes
  • Temperate: White Oak · Red Maple · Tulip Poplar · Sweetgum · American Beech · Ponderosa Pine · Loblolly Pine · Douglas Fir · Cultivated Apple · Sweet Cherry · Paper Birch · Quaking Aspen · American Sycamore · Flowering Dogwood · Weeping Willow
  • Desert: Joshua Tree · Saguaro · Creosote Bush · Blackbrush · Big Sagebrush
• Two generators. A Weber–Penn parametric model for broadleaves & conifers, and a from-scratch dichotomous L-system for the desert plants — succulents (merged-tube mesh, rib crests, areole spines) and multi-stem shrubs (root-crown stem splay, spray-card foliage).
• Real fruit. The orchard species hang actual fruit meshes (AI-generated, retopoed, and rebaked down to game weight) from their fruiting twigs — gravity-oriented instances with anti-clip placement, present at every LOD: real geometry near, baked into the branch cards and billboard far.
• Real morphology. Each species' branch angles, taper, gnarl, and crown shape are dialed to reference photos, not generic defaults.
• Foliage as cards. Base-anchored single leaves, needle sprays, and top-anchored hanging vines with backlit translucency (Barré-Brisebois SSS), dome-normal canopy shading, and per-instance wind.
• LOD chain + impostors. LOD0 full geometry → reduced-geometry LOD1 → baked branch-card LOD2 → a 2-plane billboard impostor, baked off-thread in a Web Worker so the viewer never stalls. Per-LOD density & branch-prune dials.
• A living scene. Instanced forest ring with per-instance LOD, wind-animated grass & desert scrub, procedural rocks, fallen dead limbs resting on the terrain, PBR terrain with slope/height material blending, volumetric-ish clouds, and a movable sun.
• Ambient audio. Per-biome wind bed (seamless loop) + randomly interspersed bird calls, with a mute toggle.
• glTF export. One click writes a .glb with merged per-LOD meshes and standard KHR_materials_* extensions (incl. leaf transmission).

A WebGPU-capable browser — recent Chrome or Edge (Chrome 113+). There's an automatic WebGL2 fallback, but WebGPU is the intended path.

npm install
npm run dev      # http://localhost:5390

Then use the control panel to pick a species, reshape it, reseed, tune LODs, and export a .glb. Drag to orbit; the corner speaker button toggles ambient sound.

npm run build    # production bundle in dist/
npm run preview  # serve the built bundle

Textures and audio ship in the repo, so a clone runs out of the box — you don't need to regenerate anything. But they're generated, and the tooling is included:

• Textures (bark albedo, leaf/needle/spine alpha cards) come from OpenAI Codex CLI's $imagegen (gpt-image-2). Scripts in scripts/texture/ chroma-key the alpha, dilate, and derive normal/roughness/translucency maps.
• Wind beds are generated with Stable Audio 3 via a local ComfyUI, then analyzed and flattened into seamless loops by scripts/audio/. Bird calls are trimmed xeno-canto recordings.

It's a cross-tool collaboration: the engine, PBR derivation, and scene are written by Claude Code; Codex paints the textures; Stable Audio scores the wind.

New plants are added by dropping in a preset and a small set of generated textures — no engine changes. This is designed to be done by a coding agent (Claude Code, Codex, etc.) driving an image generator like gpt-image-2. There are three steps.

Export a species object and register it in src/species/index.js (import + add to the SPECIES map). The preset chooses which of the two generators runs — pick the one that matches the plant:

A) Broadleaves & conifers → Weber–Penn parametric model. Copy the closest existing species (white-oak.js is the template; pine.js for conifers) and retune the parameters to reference photos.

export const redcedar = {
  name: 'Eastern Redcedar',
  latin: 'Juniperus virginiana',
  bark: 'redcedar_albedo.png',            // → assets/bark/  (+ derived _normal/_roughness)
  leaf: 'redcedar_needle_albedo.png',     // → assets/leaves/ (+ derived _normal/_roughness/_translucency)
  biome: 'temperate',                     // 'temperate' | 'desert' (picks the environment + soundscape)
  tileWorldSize: 1.2,                     // bark tile size in metres
  controls: broadleafControls,            // sliders exposed in the UI
  foliage: { mode: 'leaves', leavesPerBranch: 9, size: 0.7, downAngle: 60, tint: 0xbfcbb0 },
  params: {                               // Weber–Penn — arrays are [trunk, L1, L2, L3]
    scale: 12, levels: 3, ratio: 0.02, baseSize: 0.15, shape: 0 /* 0=conical…5=flame */,
    attractionUp: 0.0, length: [1.0, 0.4, 0.3, 0.25], downAngle: [0, 80, 74, 68],
    curve: [2, 14, 18, 0], branches: [0, 46, 16, 0], radialSegments: [12, 6, 4, 3],
    /* …see white-oak.js / pine.js for the full field list… */
  },
};

Rules of thumb: length[childLevel] is a fraction of the parent's length (small terminal values → short leafy twigs, not bare whips); high downAngle + attractionUp: 0 reads as a conifer, up-swept angles as a broadleaf; shape sets the crown silhouette.

B) Desert succulents (saguaro, Joshua tree, yuccas) → dichotomous L-system (Lindenmayer). These aren't branch-and-leaf trees, so they run a completely different generator: set foliageType: 'rosette' (or cactus: true) and describe the plant as an L-system — a branching grammar of fork rules (probability, depth, split angle), segment length/taper, rib count, arm gating, and rosette-leaf or areole-spine placement — which builds a single merged-tube mesh. Copy saguaro.js / joshua-tree.js; the grammar, parameters, and mesh construction are documented in docs/dichotomous-generator.md.

Generate two source images with your image model, then run them through the pipeline in scripts/texture/. The loader auto-derives map names by suffix, so naming matters: <x>_albedo.png → <x>_normal.png, <x>_roughness.png (+ _translucency.png for leaves).

Bark — prompt for a seamless, tileable bark albedo, then derive PBR:

node scripts/texture/derive-pbr.mjs assets/bark/redcedar_albedo.png   # writes _normal + _roughness

Leaf / needle card — generate the foliage on a flat magenta #FF00FF background (for a clean alpha key), then cut it out and derive maps:

node scripts/texture/chroma-key.mjs        raw.png assets/leaves/redcedar_needle_albedo.png
node scripts/texture/dilate-alpha.mjs      assets/leaves/redcedar_needle_albedo.png --passes 4
node scripts/texture/derive-pbr.mjs        assets/leaves/redcedar_needle_albedo.png
node scripts/texture/derive-translucency.mjs assets/leaves/redcedar_needle_albedo.png  # → *_albedo_translucency.png (rename to _translucency.png)

Art-direction gotchas learned the hard way:

• A needle spray must be a single feather/frond branchlet — one central woody axis with needles emanating ~45° on both sides, fully inside the frame. A radial burst from one point reads as a grass tuft on the tree, not a conifer.
• Bark must tile with no visible seam (offset-check it); a leaf card should fill the frame with a little margin so alpha-dilation and mip-mapping don't clip it.

If you use OpenAI Codex CLI: it can't always save into the workspace, so prompt it to "generate the image only — do not save/read/list/search files", tag the prompt with a unique marker, then harvest the bytes with scripts/texture/harvest-codex-image.mjs --match <marker> <out.png>.

npm run dev    # pick your species in the panel; check LOD0→billboard, wind, and GLB export

That's the whole loop: preset in → textures generated → registered → it's a first-class species with LODs, forest instancing, wind, and export, same as the built-ins.

src/
  core/        generators (weber-penn, dichotomous), meshing, LOD, cards, impostor, wind, terrain, grass…
  species/     one preset file per plant
  audio/       ambient soundscape (wind bed + bird scheduler)
  ui/          control panel
scripts/
  texture/     Codex image → alpha cutout → PBR/translucency maps
  audio/       Stable Audio generation + seamless-loop tooling
assets/        committed textures & audio
docs/          spec notes + media

MIT ©

By SkyeShark (Utah Teapot) and Claudes.

Not originally inspired by EZ-Tree but subsequent post-release features were.