img2threejs

img2threejs (github.com/hoainho/img2threejs) tackles a genuinely hard problem from an unexpected angle: it rebuilds an object from a single reference photo as a procedural Three.js model written entirely in code — primitives, generated geometry, and shaders — rather than extracting a mesh or running photogrammetry. The output is browser-ready and animation-ready with zero external assets, and the design is deliberately token-thrifty: plain Python scripts (standard library only, Python 3.10+) handle all the mechanical validation, so the agent spends its judgement where it counts — visually comparing each render against the source.

What impressed us is the discipline of the pipeline. A run is gated through staged “sculpting” phases — suitability probe, a mandatory detail inventory, spec authoring, strict-quality validation, then locked build passes (blockout → structural → form → material → surface → lighting → interaction → optimization) — and each pass must render and pass a visual score before the next unlocks. After every pass the agent chooses to continue, refine the spec, refine the code, request input, or stop, so the model self-corrects toward the reference. You get a TypeScript factory returning a THREE.Group, a structured ObjectSculptSpec, and a sculptRuntime hierarchy of pivots, sockets, and colliders that is ready to animate.

It is honest about its limits: strongest on hard-surface mechanical objects, with a newer anatomy-aware character track that produces stylized reconstructions, and a built-in “honesty clause” that flags when output is approximate or low-poly. Install is a single git clone into ~/.claude/skills/, and it is agent-agnostic (Claude Code, Codex, OpenCode, or your own screenshots). The live showcase — earbuds, a shotgun, a Gerber knife, a loot chest — is worth a look. See our companion Claude Guide to Rebuilding Images as 3D Models for a full walkthrough.