Abstract
We present WonderWorld, a novel framework for interactive 3D scene generationthat enables users to interactively specify scene contents and layout and seethe created scenes in low latency. The major challenge lies in achieving fastgeneration of 3D scenes. Existing scene generation approaches fall short ofspeed as they often require (1) progressively generating many views and depthmaps, and (2) time-consuming optimization of the scene geometryrepresentations. We introduce the Fast Layered Gaussian Surfels (FLAGS) as ourscene representation and an algorithm to generate it from a single view. Ourapproach does not need multiple views, and it leverages a geometry-basedinitialization that significantly reduces optimization time. Another challengeis generating coherent geometry that allows all scenes to be connected. Weintroduce the guided depth diffusion that allows partial conditioning of depthestimation. WonderWorld generates connected and diverse 3D scenes in less than10 seconds on a single A6000 GPU, enabling real-time user interaction andexploration. We demonstrate the potential of WonderWorld for user-drivencontent creation and exploration in virtual environments. We will release fullcode and software for reproducibility. Project website:https://kovenyu.com/WonderWorld/.