MetaEarth3D: Unlocking World-scale 3D Generation with Spatially Scalable Generative Modeling

A team of researchers from multiple institutions has developed MetaEarth3D, the first generative foundation model capable of spatially consistent 3D generation at planetary scale. Unlike existing AI models that are confined to bounded environments, MetaEarth3D treats spatial scale as a core scaling axis, enabling it to model geographic environments across thousands of kilometers. The model was trained on 10 million globally distributed real-world images and can generate multi-level, unbounded, and diverse 3D scenes, spanning large-scale terrains, medium-scale cities, and fine-grained street blocks.

MetaEarth3D demonstrates both strong visual realism and geospatial statistical realism, meaning its outputs are not only visually convincing but also statistically accurate in terms of real-world geography. Beyond generation, the model serves as a generative data engine for ultra-wide spatial intelligence, providing virtual environments for Earth observation and simulation. This breakthrough addresses a critical gap in generative AI, which has previously focused on scaling model parameters and training data while overlooking spatial scale as a key dimension of intelligence. The researchers argue this work could empower next-generation spatial intelligence for Earth observation.