SynCity 3000: Bootstrapping Scene-Scale 3D Diffusion

Forum topic · 小凯 · 2026-07-08

Summary

SynCity 3000 is a 3D scene generation framework proposed by researchers including Paul Engstler, Iro Laina, Christian Rupprecht, and Andrea Vedaldi (arXiv 2607.05392). It produces globally consistent 3D scenes with fine-grained layout control by adapting an image-to-3D generator into a convolutional operator, extending single-image 3D asset generation to entire scene scale. To address the scarcity of 3D scene training data, the model is fine-tuned on scene-level data produced by a new synthetic data engine. The convolutional generator is applied to an axonometric image of the whole scene rendered from a user prompt, yielding 3D scenes of arbitrary scale and complexity. Across diverse prompts and layouts, SynCity 3000 generates large, coherent, and detailed scenes, overcoming limitations of prior 3D scene generation methods.

Paper Overview

Field: Computer Vision Authors: Paul Engstler, Iro Laina, Christian Rupprecht, Andrea Vedaldi Published: 2026-07-06 arXiv: 2607.05392

Abstract

This paper introduces SynCity 3000, a framework for generating globally consistent 3D scenes with fine-grained layout control. Building on the ability of current image-to-3D generators to produce complex 3D assets from a single image, the authors adapt the generator into a convolutional operator, extending its capability to whole scenes.

To overcome the scarcity of 3D scene training data, the model is fine-tuned on scene-level data generated by a new synthetic data engine. The convolutional generator is applied to an axonometric image of an entire scene generated from a user prompt, producing 3D scenes of arbitrary scale and complexity.

Results

Across diverse prompts and layouts, SynCity 3000 generates large, coherent, and detailed scenes, addressing the shortcomings of previous 3D scene generation methods.

---

*Auto-collected on 2026-07-06*