One image in, one procedural 3D model running on the web out. That is the compact promise of img2threejs, an open source skill for coding agents.
One image in, one procedural 3D model running on the web out. That is the compact promise of img2threejs, an open source skill for coding agents.
img2threejs is an open source skill for Claude Code, Codex, and OpenCode. You give it a reference image, and it rebuilds the object as a code-only, procedural Three.js model — editable TypeScript, not a flattened image or a baked mesh. The output can be read, modified, and version controlled with Git.
The agent analyzes the reference, builds geometry, materials, and lighting, then renders, compares against the source image, and refines across multiple passes. It is an iterative loop, not a one-shot generator.
The workflow is a gated loop. The agent looks at the reference image, writes a structured specification, generates geometry with primitives and procedural shaders, renders the result in the browser, and compares it to the original. Based on that comparison, it decides whether to refine the spec, refine the code, request more input, or stop. Each pass is checked before the next is allowed to proceed.
This is not a one-click production mesh generator. Hidden sides of an object can only be inferred from a single image. Rigging and glTF export remain on the roadmap. For hard-surface web 3D and rapid prototyping it is worth a look, but it will not replace a production 3D pipeline or a skilled artist.
img2threejs is for developers building interactive websites, game props, quick prototypes, portfolios, or fast 3D blockouts who want code they can control rather than a black-box asset.
Think of it as an AI technical artist that helps you start, not a machine that replaces the artist.