Turning reference photos into code-sculpted Three.js props

What it does Takes an attached image and guides Codex through reconstructing the visible object as a browser-ready, procedural Three.js model built entirely from TypeScript and geometry code. It writes a structured ObjectSculptSpec that decomposes the object into hierarchy, materials, pivots, and animation anchors, then generates a factory skeleton. The output is not a downloaded mesh or photogrammetry scan, but a code-driven asset with meaningful parent-child relationships and transform sockets.

The interesting bit The plugin intentionally slows Codex down. It requires a global surface-topology plan before any visual modules are accepted, runs blockout-form-lookdev passes in order, and locks final generation until AI vision has compared browser renders against the original reference. Unsupported geometry is rejected rather than silently replaced with a box.

Key highlights

• Gated sculpting pipeline with complexity assessment and quality contracts before code generation
• Geometry registry covering tubes, lathes, extrusions, curve sweeps, and sculpted-surface fields; unsupported shapes fail explicitly
• Animation-ready hierarchies with pivots, sockets, destruction anchors, and collider hooks planned upfront
• Adaptive workflow that skips structure and optimization passes for simple objects
• Self-correcting review loop using AI vision comparison, silhouette IoU, and critical-feature scoring
• Procedural PBR evidence generation for albedo, roughness, height, normal, and AO maps
• Python helper scripts using only standard-library modules and shell image tooling; no Playwright or Chromium bundle required

Caveats

• Requires Codex with local plugin support and a separate Three.js browser project to run the generated factories
• Visual acceptance depends on an external AI vision reviewer and rendered screenshots, so the workflow is not fully autonomous
• Physics simulation and raymarched volumes remain out of scope

Verdict Worth exploring if you want Codex to produce reusable, animation-ready 3D props from concept art or reference photos without touching external asset stores. Skip it if you need exact scanned replicas or a fully self-contained 3D pipeline.