SCADBench is a crowdsourced benchmark testing how well AI models generate 3D objects from text prompts. The same prompt goes to multiple AI models, each writes OpenSCAD code, and it's automatically rendered into 3D models with PBR materials. You vote in blind side-by-side matchups — same prompt, no labels. Votes feed an Elo rating system. No accounts needed. Just spin, pick, and watch the leaderboard update live.

Humans in the Loop — A free community for people who love using AI to ship faster

A free community for people who love using AI to ship faster

Hey Product Hunt! 👋 Text benchmarks tell you how well an AI can write. But how well does it actually understand the physical world? SCADBench tests something harder to measure — spatial reasoning, structural understanding, and creative interpretation — by asking AI models to build 3D objects from scratch using code. There's no faking it here. The AI has to write valid OpenSCAD code that compiles into a real 3D model. One syntax error and it fails completely. A chess rook needs to actually look like a chess rook. The same prompt goes to every model, the results are rendered automatically, and you vote blind — no labels, no bias. Elo ratings emerge from real human judgment. I'd love for you to try a few votes and see which AI surprises you. Some of the results are genuinely impressive (and some are hilariously wrong).