bevy_zeroverse: Pre-training Reconstruction Models with Multi-scale Synthesized Data

We present bevy_zeroverse, a procedural synthetic dataset generator for pre-training reconstruction models. bevy_zeroverse synthesizes datasets from primitive shapes with random materials, augmentations (e.g., normal deformation, boolean manifold operations, thin structures, hierarchical transforms), and multi-scale composition (e.g., object, room, and environment scales). Unlike previous works (e.g., Objaverse, Zeroverse) that target the object-scale 3D domain, bevy_zeroverse extends into the multi-scale 4D domain, enabling the training of models that can generalize to larger, more complex, dynamic scenes. We demonstrate that pre-training on bevy_zeroverse improves the performance of a reconstruction model on real-world data, achieving high-quality, sparse-view reconstructions. Additionally, we analyze the impact of synthetic data parameters on model training and performance. Our work highlights the potential of synthetic data for pre-training reconstruction models and emphasizes the importance of multi-scale, multi-domain data generation for training models that can generalize to complex, real-world scenes.