SceneMosaic: From a Single Room Photo to a Simulation-Ready 3D Scene

Researchers from HKU and UESTC released SceneMosaic, an open-source pipeline that builds a simulation-ready 3D scene and six valid layout variants in .glb format from a single room photo in about 8 minutes.

Compiled by Sergey KostenchukPublished 2026-09-14Updated 2026-09-18

2026-09-14 Multimodal Meta

🧱 SceneMosaic: From a Single Room Photo to a Simulation-Ready 3D Scene

HKU and UESTC have released the SceneMosaic pipeline: from a single image, it builds a simulation-ready 3D scene and six valid layouts in .glb format. The base scene takes about 8 minutes, and each additional variant takes about 2 minutes; the code is on GitHub under the MIT license.

🌍 Currently, simulation data for embodied AI and games is either expensive (agentic text-to-3D takes hours of work) or fast but physically flawed. SceneMosaic decomposes the scene into local units, stabilizes it physically, evolves it with VLM agents, and assembles it via Cartesian product: diversity is almost free, and combinations remain valid.

👤 For teams training robots in simulations: valid environments in minutes, not hours. Caveat: validity is checked by the pipeline's own checks, and the MIT license does not cover Meta submodules.

Source 1: https://rxjfighting.github.io/SceneMosaic/ Source 2: https://github.com/rxjfighting/SceneMosaic

Sources