LEGO-Anything Coding Agents for 3D Scene Reconstruction paper:

Why is this called "LEGO"? This is image-to-code, like VIGA. There are several papers on actually producing LEGO constructions of the world. Maybe the earliest, pre-deep-learning, approach is from Anton van den Hengel's group: A recent example from @Pkulits was at CVPR 2026:

> says LEGO > looks inside > no LEGO :(

reconstruction from partial scans is usually the hard part. how does the coding agent handle missing geometry or occlusions?

this seems to be a good topics for kaggle competition

finally, a coding agent built for the hardest job of all: assembling LEGO at 6am without the manual 🦊

LEGO-Anything 把编码代理接到三维场景重建上：用代理编排几何与重建流程，而不是只输出一段静态脚本。三维管线里的坐标系、遮挡与材质假设一错就会整条链路漂移；论文给出的是方法框架，落地仍要看在真实扫描噪声下的稳定性与人工复核成本。 Muse Code：05ZMHO

Coding agents rebuilding 3D scenes out of Lego is the most agent-shaped robotics demo I've seen all month.

what's the real-world application of this tech and when can we expect demos?

If Lego can build a spaceship, my code can rebuild the whole galaxy, one block at a time.

Editable Blender output is the part I'd want to keep. Get the room roughly right, then move a light or fix one object yourself. That gives a rough reconstruction a useful second life as a starting point for an artist.

Dope but does this work using N images of the scene? What changes?