The framework combines hierarchical visual layout sampling with contract-guided 3D compilation. A vision-language model plans functional zones, diffusion models propose arrangements, and geometric fitting plus local refinement reconcile relationships, object placements, and physical constraints before scene export.
Researchers can use SceneBloom to expand environment diversity for navigation and other embodied-learning tasks. The released pipeline provides reproducible scene generation, rendered previews, structured layout data, and Habitat-compatible GLB exports, making the results usable beyond the project's visual demonstrations.
