Spatially-Conditional 3D Furniture Generation Model for Indoor Scene Synthesis

Recent advances in generative models have significantly enhanced the capabilities of 3D indoor scene synthesis, a field that is rapidly gaining interest due to its implications for applications such as embodied AI. These applications often require diverse, large-scale indoor datasets that include realistically modeled 3D furniture. Traditional text-to-3D models, while capable of producing realistic assets, often lack precise control over spatial dimensions, which is critical for ensuring that furniture fits appropriately within the scenes. This paper introduces SC-Shap·E, a spatially conditional 3D generative model that not only enhances the realism of 3D furniture but also ensures it adheres to specified spatial dimensions of height, width, and depth. Built on the pretrained Shap·E model, SC-Shap·E incorporates an additional network that utilizes spatial information alongside textual prompts, offering improved control over furniture sizing within generated scenes. By comparing our model with the original Shap·E, we demonstrate its superior ability to reflect accurate spatial conditions. Additionally, we present a novel three-stage system for 3D indoor scene synthesis that includes floor-plan creation, furniture layout, and 3D furniture mesh production, showing its effective application in creating diverse and realistic 3D indoor scenes.