AI Scene3D — 3D Mesh Generation Foundation

• Area: Intelligence
• Path: services/ai/scene3d
• Kind: 3D mesh generation foundation (text-to-mesh + image-to-mesh; GLB/USD output)
• Status: v0.1.0 — foundation shipped (2026-05-24); handcrafted GLB 2.0 writer, Converter interface, async Mesh store, Go SDK

Role in the stack

scene3d consolidates 3D mesh generation server-side. v1 is proxy-first (Tripo, Meshy, Luma Genie via services/ai/gateway) because self-hosted 3D quality lags 6–12 months behind frontier proprietary models. Hunyuan3D is the most credible local candidate for v2 G4 evaluation.

It is the Koder analog of Tripo, Meshy, Luma Genie, Common Sense Machines and Stability 3D. Bootstrapping the sector early (even with proxy-only v1) avoids debt when 3D becomes tablestakes (Apple Vision and Meta Quest are already in market).

Boundary vs neighbors

• services/ai/imaging is the 2D source for image-to-mesh.
• services/ai/video is the temporal sibling (different generation modality).
• 3D rigging, animation and editing are explicitly out of scope (would be a separate sector).

Features (v1 target)

• Text-to-mesh via 3 proxied providers (Tripo cheap, Meshy mid, Luma premium)
• Image-to-mesh via same providers
• GLB canonical output (webARthree.js/Quest compat)
• USD alternate output (Apple Vision native, pro pipelines) via tinyusdz
• Polycount-tier quotas (free 10k / pro 100k / enterprise 500k)

Primary couplings