SynCity 3000: Bootstrapping Scene-Scale 3D Diffusion

Summary

SynCity 3000 is a 3D scene generation framework by Paul Engstler, Iro Laina, Christian Rupprecht, and Andrea Vedaldi that produces globally consistent 3D scenes with fine-grained layout control. Building on modern image-to-3D generators that can reconstruct complex 3D assets from a single image, the authors adapt the generator into a convolutional operator to scale generation from objects to entire scenes. To address the scarcity of 3D scene training data, the model is fine-tuned on scene-level data produced by a novel synthetic data engine. The convolutional generator is applied to axonometric (isometric-style) images of a full scene rendered from a user prompt, yielding 3D scenes of arbitrary scale and complexity. Across diverse prompts and layout conditions, SynCity 3000 generates large, coherent, and detailed scenes, addressing limitations of prior 3D scene generation methods. The paper is available on arXiv (2607.05392).