UrbanGen: Urban Generation with Compositional and Controllable Neural Fields.

Despite the rapid progress in generative radiance fields, most existing methods focus on object-centric applications and are not able to generate complex urban scenes. In this paper, we propose UrbanGen, a solution for the challenging task of generating urban radiance fields with photorealistic rendering, accurate geometry, high controllability, and diverse city styles. Our key idea is to leverage a coarse 3D panoptic prior, represented by a semantic voxel grid for stuff and bounding boxes for countable objects, to condition a compositional generative radiance field. This panoptic prior simplifies the task of learning complex urban geometry, enables disentanglement of stuff and objects, and provides versatile control over both. Moreover, by combining semantic and geometry losses with adversarial training, our method faithfully adheres to the input conditions, allowing for joint rendering of semantic and depth maps alongside RGB images. In addition, we collect a unified dataset with images and their panoptic priors in the same format from 3 diverse real-world datasets: KITTI-360, nuScenes, and Waymo, and train a city style-aware model on this data. Our systematic study shows that UrbanGen outperforms state-of-the-art generative radiance field baselines in terms of image fidelity and geometry accuracy for urban scene generation. Furthermore, UrbenGen brings a new set of controllability features, including large camera movements, stuff editing, and city style control.