Bri3D Introduces Novel Framework for Single-Image 3D Scene Generation
Bri3D has unveiled a new image-to-compositional 3D scene generation framework designed to bridge the gap between object-level detail synthesis and scene-level modeling. This framework, detailed in a recent publication, utilizes instance-aware voxel scene anchors to create complex 3D scenes from a single image. The core of Bri3D involves two main trainable modules: a Scene Scaffold Generator and an Anchored Voxel Refiner. The Scene Scaffold Generator predicts a color-coded voxel scaffold for all foreground objects, establishing their coarse geometry and spatial arrangements within a shared scene representation. Subsequently, the Anchored Voxel Refiner converts these scaffold-derived object clusters into high-resolution voxel supports, which are compatible with pre-trained object generators. This approach allows for the creation of detailed 3D objects within a globally coordinated scene context, ensuring that objects are placed and structured realistically relative to each other. The system employs Color-Att...
