SceneBloom

SceneBloom proposes an agent framework that automatically generates high-quality, collision-free indoor scenes by combining visual layout sampling with contract-guided 3D compilation, providing a data source for embodied learning.

The hierarchical sampler leverages the prior knowledge of diffusion models to divide rooms into functional zones and samples furniture positions and orientations within these zones to generate furniture arrangements.

The contract-guided 3D compiler extracts inter-object relationships from each layout proposal as constraints (such as "the chair is next to the table" and "the bed is against the wall") and uses them for geometric fitting and local pose correction.

In SceneBloom, a contract refers to the spatial relationship constraints between objects extracted from a layout proposal, such as "the distance between the chair and the table is between 0.3 and 0.8 meters," "one side of the bed is flush against the wall," and "the sofa faces the TV."

In comparative experiments across five room types, SceneBloom achieves the highest level in both visual quality and physical stability, and is the only method in which no mesh collisions or out-of-bounds objects were detected.

The layout distribution generated by SceneBloom is closer to the distribution of real rooms than that of all baseline methods.

An object navigation policy trained in scenes generated by SceneBloom achieved a higher success rate on real scanned data not involved in training than using an equal amount of real training scans.

SceneBloom currently focuses mainly on furniture layout and may not yet cover finer-grained objects (such as small items on a tabletop).

The diversity of scenes generated by SceneBloom is limited by the prior distribution of the diffusion model.

SceneBloom's code has been open-sourced on GitHub, providing a reproducible foundation for future research.