SceneBloom

SceneBloom is a research framework for generating simulation-ready indoor 3D environments for training embodied agents. It combines visual layout sampling with contract-guided 3D compilation. A hierarchical sampler uses diffusion priors to produce furniture arrangements within functional zones. A compiler then uses relations extracted from each proposal to guide geometric fitting and local pose correction.

The authors report evaluations across five room types (living room, bedroom, dining room, office, bathroom) against Holodeck, SceneWeaver and SAGE. They report the highest visual quality and physical stability among the compared methods. It is the only method with no detected mesh collisions or out-of-bounds objects, and layouts closer to real-room distributions from Matterport3D and Scan2CAD. Generation takes about 2.7 minutes per layout, reported as roughly 27× faster than SAGE.

For downstream use, a SemExp object-navigation policy trained from scratch in SceneBloom houses improved success in held-out real homes from 29.7% to 35.3% on average. The authors say this is 1.7× the gain from training on the same number of real scans. A video review says code, LoRA adapters and a 3D asset library were released, but no repository link was in the provided material. It is aimed at embodied-AI researchers rather than end users.