Breaking

Shanghai AI Lab and others release Code-as-Room, building 3D rooms from top-down images

June 29, 2026 at 09:14 EDT

AI Agents
Media Generation
Research & Papers

A research team led by the Shanghai Artificial Intelligence Laboratory has released Code-as-Room, an agentic framework that reconstructs a 3D indoor scene by generating executable Blender Python code from a single top-down image of a room. The paper was posted to arXiv on May 18, 2026, with code and video demos distributed via a project page and GitHub. Code-as-Room

May 18, 2026 · Shanghai AI Lab

Code-as-Room: One Top-Down Photo In, a Full 3D Room Out

An MLLM agentic framework reads a single floor-plan-like image and writes an executable Blender Python script — complete with geometry, materials, and lighting — to rebuild the room in 3D.

top-down image as the only input

pipeline stages from image to scene

difficulty tiers: simple / medium / special

The 5-Stage Pipeline

1 · Scene structuring

→

2 · Layout code

→

3 · Object profiling

→

4 · Object code

→

5 · Interior decoration

A cross-stage memory module carries information between stages to fight context forgetting and runaway loops.

Why it stands out

Assembles an entire room from a flat image into coherent code
Beats the VIGA baseline and other MLLMs on layout accuracy and placement
Replaces repetitive manual modeling work in a short time

Still open

No large-scale real-world usage reports yet
Pricing, release date, and licensing undisclosed
Benchmark coverage and field validation still to expand

Models put head-to-head

Evaluated on a new benchmark across GPT-5.5, Gemini 3.1, and Gemini 3 Flash — targeting interior design, VR, gaming, and Embodied AI.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.