This paper addresses the challenge of generating 3D models of geographically vast areas (thousands of square kilometers). To address this, we present Aerial-Earth3D, a large-scale 3D aerial dataset consisting of 50,000 600m x 600m aerial photographs of the entire continental United States. The dataset contains multi-view images, depth maps, normals, semantic segmentation, and camera position information, and is quality-controlled to ensure terrain diversity. Based on this, we propose the EarthCrafter framework for large-scale 3D Earth generation using sparse-decoupled latent diffusion. EarthCrafter reduces computational costs by separating structure and texture generation using dual sparse 3D-VAE, which transforms high-resolution geometric voxels and 2D Gaussian splats (2DGS) into a compressed latent space. In addition, we model latent geometric and texture features independently and flexibly using condition-aware flow matching models trained on inputs that are either semantic, image, or a combination of both. Experimental results show that EarthCrafter excels at large-scale generation, supporting a wide range of applications from semantically guided city layout generation to unconditional terrain synthesis.
