SCADBench – Benchmarking AI-generated 3D models from text prompts

Summary: SCADBench evaluates AI models by testing their ability to generate 3D objects from text prompts through OpenSCAD code. It renders models automatically and uses blind human voting to rank models with an Elo rating system, providing a measure of spatial reasoning and structural understanding.

What it does

SCADBench sends identical text prompts to multiple AI models that generate OpenSCAD code, which is compiled and rendered into 3D models with PBR materials. Users vote on side-by-side blind matchups, and votes update a live Elo leaderboard.

Who it's for

It is designed for researchers, developers, and enthusiasts interested in assessing AI capabilities in 3D spatial reasoning and code-based model generation.

Why it matters

It provides a practical benchmark for evaluating AI’s understanding of physical and spatial concepts by requiring valid, compilable 3D code and human judgment rather than text-based metrics.