Scalability of FlowScene to Larger and More Diverse Environments

Determine the scalability and generalization of FlowScene when applied beyond SG-FRONT/3D-FRONT to larger-scale and more diverse datasets, including outdoor scenes, assessing whether the tri-branch Multimodal Graph Rectified Flow maintains scene-level style consistency, controllability, and realism under these broader settings.

Background

FlowScene is evaluated primarily on SG-FRONT/3D-FRONT indoor datasets, demonstrating improvements in realism, controllability, and style consistency over baselines. The method tightly couples three rectified-flow branches (layout, shape, texture) conditioned on a multimodal scene graph.

The authors explicitly note that despite strong results on the chosen dataset, it is unresolved how the approach scales to larger and more diverse environments. They point to extending FlowScene to larger and outdoor scenes as a future direction, indicating the need for systematic evaluation of scalability and generalization.

References

While our current experiments are conducted on a single dataset, the scalability of FlowScene to larger and more diverse environments remains to be further explored, this does not diminish the generality and promise of the proposed interaction framework. Future directions include extending FlowScene to larger and outdoor scene, and integrating more deeply with interactive design and planning tools.

— FlowScene: Style-Consistent Indoor Scene Generation with Multimodal Graph Rectified Flow  (2603.19598 - Yang et al., 20 Mar 2026) in Section 6 (Discussion)