LLMs are getting better at generating playable 3D games, but coherent game worlds depend on assets that are both visually convincing and structurally correct. Today's models can produce a recognizable object that is still functionally unusable in-game. Parts float or intersect, openings are sealed, materials read incorrectly, text becomes malformed, and spatial relationships described in the prompt are lost. Three.js Asset Generation Benchmark measures these failure modes using deterministic geometry and scene checks alongside multi-view visual evaluation. We expect models to improve by reasoning more precisely about object relationships, inspecting their work from multiple angles, and revising invalid geometry before it reaches a game.
Three.js Asset Generation Benchmark
A verifiable benchmark for structurally correct, visually convincing Three.js assets generated by frontier models.