snap3d

One photo in → an editable 3D model out. Powered by the spatial reasoning of Claude Fable 5 — no photogrammetry rig, no multi-angle capture, no manual modeling.

Drop a single 2D photo (a car, a chair, a kettle…) and Fable 5 infers depth, hidden geometry, and textures, then rebuilds the object as a parametric scene of named primitives rendered live in Three.js.

The out-of-the-box feature: Remix with words ✦

Everyone else stops at "photo → mesh". snap3d's scene is semantic — every part has a name ("front-left wheel", "windshield") — so you can keep editing in plain English:

"make the wheels huge" · "turn it into a convertible" · "paint it neon green"

Fable 5 re-edits only the parts you asked about and hands back the updated scene. Plus per-part sliders (scale, color) and one-click .obj export for Blender, Unity, or any render engine.

Run it

node server.js          # → http://localhost:3777

Zero npm dependencies. The brain is resolved in this order:

1. ANTHROPIC_API_KEY set → Anthropic API (claude-fable-5, override with SNAP3D_MODEL)
2. claude CLI installed → your local Claude Code session does the reasoning
3. neither → a bundled demo scene, so the UI always works

How it works

photo (base64) ─→ Claude Fable 5 ─→ {parts:[{name, shape, dim, pos, rot, color}]} ─→ Three.js
       remix text + current scene ─→ Fable 5 ─→ updated scene JSON ──────────────────┘

The whole trick is a strict JSON contract: box/cylinder/sphere/cone/torus primitives, Y-up, ground at y=0. The model does the spatial reasoning; the browser just builds meshes.

Files

• server.js — stdlib Node server + Fable 5 bridge (~100 lines)
• public/index.html — viewer, parts panel, remix box, OBJ export (one file)
• public/demo-scene.json — offline fallback scene

MIT · built by Rohit Raj with Claude Fable 5