Researchers have proposed SceneMosaic, a framework for generating diverse, physically valid indoor 3D scenes that are ready for simulation. The work, posted as a paper on Hugging Face, targets a persistent bottleneck in embodied AI and interactive entertainment: producing many different, usable scenes at scale.
Existing approaches split into two camps. Agentic text-to-3D pipelines driven by vision-language models (VLMs) can produce high-fidelity scenes, but they rely on costly iterative object placement and refinement. Parametric image-to-3D models are far more efficient thanks to strong priors learned from 2D images, yet they often yield imprecise layouts with physical problems such as collisions or floating objects. Both paradigms also tend to output only a single scene per input, limiting diversity.
SceneMosaic combines the two. It starts from a candidate produced by the learned image-based prior, then evolves that result through VLM agents, aiming to preserve efficiency while enforcing physical validity. The evolution step exploits the locality of natural scenes: a scene is decomposed into independent local units, each evolved separately, and the global scene is then assembled via a Cartesian product. A perception-aware metric with dynamic Max-Min selection is used to isolate the most diverse, high-quality layouts.
On the SceneEval-100 benchmark, SceneMosaic matches the strongest agentic baseline in semantic layout quality while achieving a 24x speedup, substantially reduces physical violations, and receives the highest human ratings, according to the paper. The authors state that the code is publicly available on GitHub.