LEGO-Anything: Coding Agents for 3D Scene Reconstruction
The paper frames single-image 3D reconstruction as code: agents write Blender programs, run them, inspect the result, and revise.
LEGO-Anything is evaluated on LEGO-Bench, a simulator-grounded benchmark of 208 images across 104 indoor and outdoor scenes. GPT-6-astra leads the tested agents, but the paper reports large gaps between producing valid scene artifacts and accurately recovering geometry and appearance. The authors identify weak initialization, regressive edits, and unreliable self-evaluation as recurring failure modes. Their LEGO-Plugin harness improves all six evaluated models, while downstream readouts for detection, masks, and depth remain well behind specialized vision systems. Source: HF Daily Papers' note
LEGO-Anything is evaluated on LEGO-Bench, a simulator-grounded benchmark of 208 images across 104 indoor and outdoor scenes. GPT-6-astra leads the tested agents, but the paper reports large gaps between producing valid scene artifacts and accurately recovering geometry and appearance. The authors identify weak initialization, regressive edits, and unreliable self-evaluation as recurring failure modes. Their LEGO-Plugin harness improves all six evaluated models, while downstream readouts for detection, masks, and depth remain well behind specialized vision systems. Source: HF Daily Papers' note
score 5