LEGO-Anything Turns 3D Reconstruction Into Blender Code Loops
TL;DR
- LEGO-Anything reframes 3D scene reconstruction as an iterative loop where an agent writes, executes, inspects and revises Blender code.
- On the new LEGO-Bench of 208 images across 104 indoor and outdoor scenes, GPT-6-astra tops the field at 53.4% indoor and 39.6% outdoor.
- A training-free LEGO-Plugin lifted six models by up to 62.7% relative on overall score, without any retraining.
LEGO-Anything treats 3D scene reconstruction as an iterative Blender-scripting task: a coding agent writes program code, renders it, looks at the result, and rewrites. On the arxiv paper's own LEGO-Bench, the strongest model tested, GPT-6-astra, reaches 53.4% indoor and 39.6% outdoor.
The pipeline is described plainly. The authors write that "a coding agent iteratively writes and executes Blender code, inspects scenes and renderings, and revises the program." The benchmark that sits under those numbers is small but deliberately structured, with 208 images drawn from 104 indoor and outdoor scenes, and it "separately scores artifact validity, visible-surface geometry, and rendered appearance" rather than collapsing everything into one composite.
The paper does not oversell. Its own summary calls program-constructed scenes "a promising but not yet sufficiently precise representation of natural images," and concedes that "substantial gaps remain between delivering valid scene artifacts and faithfully recovering scene geometry and appearance." A training-free add-on the authors call LEGO-Plugin, tested on six models, delivered "relative gains of up to 62.7% in overall score," which suggests the current ceiling is not fixed.
Originally reported by paper
Read the original article →Original headline: LEGO-Anything Encodes 3D Scenes as Runnable Blender Code, Introduces LEGO-Bench