#img2threejs: Build 3D from a single image without knowing Blender

You do not need to know Blender to build a 3D object from a photograph. img2threejs is an open source skill that hands the job to an AI technical artist, which writes Three.js code, renders it, and reviews it against the reference image over multiple passes.

#What it is

img2threejs is an MIT-licensed skill for Claude Code, Codex, or OpenCode. You give it a reference image, and instead of extracting a mesh, the AI analyzes the object, writes a structured specification, and generates a procedural Three.js model built from primitives, procedural shaders, and generated geometry. The output is a TypeScript factory that returns a THREE.Group — code you can read, edit, version control, and reuse.

The important distinction is that this is reconstruction by code, not photogrammetry or one-click mesh extraction. The pipeline enforces a strict quality gate with a detail inventory and a render-compare loop: after each pass, the agent renders the model in the browser, compares it to the original image, and decides whether to refine the spec, refine the code, request more input, or stop. The output also exposes a runtime hierarchy with pivots, sockets, colliders, and destruction groups, preparing the model for animation and interaction.

#Why it matters

• People who are not 3D specialists can create web-interactive props without first learning the entire Blender interface — modeling, UVs, materials, lighting, and rigging.
• Three.js developers receive TypeScript that is readable, editable, and version controllable, not a binary blob.
• 3D artists can use it for rapid blockouts — testing silhouette, proportions, materials, and camera angles before committing to production-quality work.
• The output has interactive structure rather than being a static mesh, so it is ready for web scenes, game prototypes, and product landing pages.
• The pipeline forces the AI to check its own work visually across multiple passes instead of trusting a one-shot result.

#How it works

You install the skill into your agent's skills directory, hand the agent a reference image, and state the intended use — a web game prop, a portfolio piece, or an animation asset. The agent then runs a staged pipeline: blockout, structure, form, material, lighting, and interaction passes, each with its own acceptance criteria. After every pass, it renders the model in the browser, compares it side by side with the reference, and chooses to continue, refine the spec, refine the code, request more input, or stop if the goal is not feasible. Cleaner images and multiple camera angles produce more trustworthy results.

#Caveats

There are three big limits to know. First, a single image cannot reveal the hidden sides of an object — the agent can only infer them. Second, characters currently lean toward a stylized look rather than photorealistic reconstruction. Third, true skeletal rigging, morph targets, and glTF export are still on the roadmap, so this will not replace a production animation pipeline. Quality also depends on the input image, the complexity of the object, and the vision and coding ability of whatever agent you use.

#Who it's for

img2threejs is for non-3D users who want a starting point for web props, game prototypes, interactive portfolios, or product landing pages; Three.js developers who want editable code; and 3D artists who want to block out and test looks faster before doing production work.

Treat it as an AI technical artist for Three.js — useful for hard-surface props and web 3D prototyping, not a replacement for the artist or the production pipeline.

Source: https://github.com/hoainho/img2threejs