The LEGO-Anything study evaluated agents reconstructing a scene from a single photograph as Blender code using 208 images across 104 scenes. The six GPT configurations tested mostly produced executable scenes, but even the top-performing GPT-6 Astra scored only 53.4% on indoor geometric reconstruction and 39.6% outdoors. Models’ ability to choose the more accurate of two outputs was near or below chance. LEGO-Plugin, which adds measurement-based corrections and safeguards against regression, improved every model but did not reach the accuracy needed to replace specialized vision models. Executability and fidelity therefore need separate verification.
Back to Tech
LEGO-Bench exposes the gap between generating 3D scenes and reconstructing them accurately
Judging success only by whether Blender code runs can miss geometric errors and revisions that make results worse.