3D scene composition

Compose impressive interactive Three.js scenes from real model assets.

Use real GLB/GLTF assets instead of primitive geometry:

arka scene-3d "Museum visitors" \
  --model https://example.com/human.glb \
  --model https://example.com/robot.glb \
  --out museum-scene \
  --preset museum

One-shot impressive scenes (no URLs required)

Arka can auto-resolve verified free assets from the curated three.js examples catalog:

arka scene-3d "Animated robot gallery" --auto --preset gallery --out arka-scenes/robot-gallery

Or from natural language:

arka create an impressive interactive 3d robot gallery scene

Resolution order: user-provided models → curated catalog → optional threejs MCP search → local generation (procedural / HF Shap-E) unless --no-generate is set.

Visual presets

Preset	Look
gallery	Reflective dark floor, rim lights, bloom
studio	Neutral sweep, soft shadows (default)
outdoor	Warm sun, light fog, large ground
racing	Low chase camera, track-style ground
interior	Warm ambient, tighter framing
space	Starfield, bloom
museum	Polished exhibit floor

If --preset is omitted, Arka infers one from the scene plan context.

Scene features

The generated scene includes:

• OrbitControls with damping and double-click camera reset
• Shadow maps, ACES tone mapping, and optional bloom
• Ground plane tuned to the preset
• Loading overlay while models fetch
• Animation clip picker when a model has multiple clips
• Responsive resize handling

Arka does not invent asset URLs: auto mode uses verified catalog entries or configured search providers.

Plan contextual scenes

arka scene-3d "Person typing" --intent "working at a desk" \
  --model ./models/person.glb --model ./models/desk.glb --plan

The plan identifies the setting and required model roles (for example, a desk, keyboard, monitor, and chair) and includes approximate real-world dimensions in meters for proportional placement. It also includes placement rules such as putting a keyboard on the desk, a chair in front of it, and aligning a seated character with the keyboard. To describe a rendered model preview with the configured vLLM/vision backend, add --describe-model preview.png. A missing vision backend produces a structured warning; it does not invent a description.

Arka also chooses semantic orientation defaults before generation. For example, vehicle racing/game prompts such as “Cybertruck vs Ferrari racing game” use a rear three-quarter chase-camera view by default, because racing games usually show the back of the player car. Explicit user wording still wins: “front view,” “side profile,” “top layout,” or “show the back” override the default.

Flags

Flag	Purpose
--auto	Resolve assets from plan roles
--no-generate	Catalog + search only; skip AI mesh generation
--preset	Visual environment preset
--intent	Action/context for planning and layout
--plan	Print the contextual plan JSON
--out	Output directory (writes index.html)