mesh-forge

Mesh Forge — directed text/image/multi-view → 3D

The full-fidelity mesh lane: a chain of specialist models (IBM Granite prompt director → FLUX reference render → Microsoft TRELLIS / Tencent Hunyuan3D reconstruction), or direct reconstruction from your own photos.

Tool

mesh_forge (three.ws MCP server, @three-ws/mcp-server).

• Lane: PAID — $0.25 USDC per call over x402. Returns PaymentRequired if no payment payload is attached.
• Modes:
  • Text: Granite rewrites your prompt into an optimized 3D spec, FLUX renders a reference, TRELLIS/Hunyuan3D build the mesh.
  • Image: pass image_url to reconstruct a single reference image directly (prompt-director + text-to-image stages skipped).
  • Multi-view: pass image_urls (1–4 angles, e.g. front/back/left/right) and the backend fuses them — no hallucinated back of the object.

Inputs

Param	Required	Notes
prompt	text mode	3–1000 chars describing one object, e.g. "a worn leather armchair". Optional when an image is supplied (then used as guidance).
image_url	image mode	http(s) URL to a single reference image.
image_urls	multi-view	1–4 http(s) URLs of the same object from different angles. Takes precedence over image_url.
direct	no	Run the Granite prompt-director stage (text mode). Default true.
aspect_ratio	no	Reference image aspect ratio for text mode: 1:1 (default), 4:3, 3:4, 16:9, 9:16.

Outputs

• glbUrl — the generated mesh .glb.
• viewerUrl — https://three.ws/viewer?src=<glbUrl>.
• how many views were fused, which backend handled it, the directed prompt (text mode), and timing.

How to run

1. Choose the mode from the user's input (text / single image / multi-view). Prefer multi-view when several angles of one object are available — it gives the cleanest geometry.
2. Call mesh_forge. On PaymentRequired, surface the $0.25 USDC price and funding path rather than retrying.
3. Return glbUrl + viewerUrl. Feed glbUrl to auto-rig (rig_mesh) for an animation-ready model.

Notes

• For a free, text-only draft, use forge-3d (forge_free) instead — no payment.
• One object per call. For multi-object placed scenes, use the scene server's compose_scene.