SceneMosaic: Efficient and Diverse Simulation-Ready Scene Generation via Hybrid Agentic Layout Evolution
SceneMosaic claims agent-style scene quality with a 24x speedup.
The paper proposes a hybrid pipeline for generating indoor 3D scenes that are ready for simulation. It starts from an image-based prior, then uses vision-language agents to evolve the layout toward better physical validity. The method splits scenes into local units, evolves them separately, and recombines them to increase diversity. On SceneEval-100, the authors say it matches the strongest agentic baseline on semantic layout quality, reduces physical violations, and gets the highest human ratings. HF Daily Papers' note
The paper proposes a hybrid pipeline for generating indoor 3D scenes that are ready for simulation. It starts from an image-based prior, then uses vision-language agents to evolve the layout toward better physical validity. The method splits scenes into local units, evolves them separately, and recombines them to increase diversity. On SceneEval-100, the authors say it matches the strongest agentic baseline on semantic layout quality, reduces physical violations, and gets the highest human ratings. HF Daily Papers' note
score 4