img2obj: Encoding Images into 3D Code

To address inefficiencies in web 3D asset management, the Codex plugin img2obj has been released, reconstructing reference images as Three.js procedural code rather than binary meshes. Instead of using photogrammetry to directly convert pixels into meshes, this tool analyzes images to create a part hierarchy and geometry specification (ObjectSculptSpec), then generates TypeScript factory functions based on that specification.

Key Features and Workflow

Since the output of img2obj is code, pivot axes and sockets can be defined at the specification stage, facilitating the addition of rigging or interactions later. It follows a controlled reconstruction workflow consisting of four stages: blockout, shape, lookdev, and interaction. At each stage, the browser rendering result is compared with the original image, and the user must approve it before proceeding to the next stage.

Scope and Limitations

It is suitable for reproducing structured objects such as stylized props, mechanical devices, and plants, but is not suitable for photogrammetry, precise dimension extraction, transparent glass, or complex character depictions. The generated code is explicitly stated as a starting point, not a replacement for finished production assets.

Usage and Installation

Clone the GitHub repository to load it as a local Codex plugin; Python 3.10 or higher and a Three.js project environment are required. The specification generation, validation, and code generation processes can be controlled directly from the command line via scripts/sculpt.py.