SceneBuilder

Existing indoor 3D scene datasets are severely limited in availability and diversity—there are fewer than 10k high-quality scenes in total, of which 99+ percent depict residential environments. Synthetic generation methods solve the quantity aspect, yet still suffer from diversity—most works cannot be tractably extended to output non-residential scenes, and are limited to extremely simplistic floor plans, unsuitable for embodied navigation training.

We present SceneBuilder to show that 3D scene generation using agentic VLM workflows is possible. We contribute SBD-100k, a 3D scene dataset with unprecedented domain diversity and navigation complexity. Using ablation experiments, we show that navigation models trained on SBD-100k outperform respective baselines.

Demo and code available at scenebuilder.github.io.