SynCity 3000 · Scene-Scale 3D Diffusion

Abstract

We present SynCity 3000, a framework for generating 3D scenes that are globally coherent while enabling fine-grained layout control. Building on the ability of current image-to-3D generators to produce complex 3D assets from a single image, we extend this capability to the scale of entire scenes by adapting the generator to be applicable as a convolutional operator. We achieve this by fine-tuning the model on scene-like data generated by a new synthetic data engine, which we propose to address the scarcity of 3D scene data for training. The convolutional generator is then applied to a dimetric image of the entire scene, generated from the user prompt, resulting in 3D scenes of arbitrary size and complexity. Across diverse prompts and layouts, SynCity 3000 produces large, coherent, and detailed scenes, addressing the shortcomings of prior approaches to 3D scene generation.

Generated Worlds

Diverse scenes created from a single text prompt

a suburban scene with shops, cafés, cars, and trees

a theme park

a mountainous landscape

antarctic sea ice with a red icebreaker ship

a monastery complex

ancient ruins in a desert landscape

Comparison

Stepping beyond tile-based generation

SynCity generates scenes tile by tile, leaving a visible grid structure in the output. SynCity 3000 applies the generator convolutionally across overlapping windows throughout the diffusion process at every step. This new approach allows for producing worlds with organic, globally coherent structures.

100%

prefer our layout control
vs SynCity

63%

prefer our generated scenes
vs SynCity

59.3%

prefer our scenes over 3DTown,
a concurrent approach

a suburban scene with shops, cafés, cars, delivery vehicles, and plenty of trees

SynCity

SynCity 3000 · Ours

a theme park

SynCity

SynCity 3000 · Ours

a mountainous landscape with water

SynCity

SynCity 3000 · Ours

Method

Joint diffusion in a two-stage pipeline

Stage 1 · 2D

Template Generation

A high-resolution dimetric scene template is generated from text prompts. Inspired by MultiDiffusion, the generator operates jointly on overlapping latent windows throughout the denoising process. This leads to more coherent results rather than stitching independent tiles. Optional layout constraints allow precise placement of specific elements within the scene.

Stage 2 · 3D

Scene Conversion

The 2D template is converted into a full 3D Gaussian Splat scene by a fine-tuned model adapted to operate convolutionally on the image. To ensure geomtric and semantic coherence, it receives context surrounding the current region that is being denoised. Trained on 320k procedurally generated scene-scale examples, the model faithfully reproduces fine detail and maintains visual coherence across scenes of arbitrary size and complexity.

Citation

BibTeX

@inproceedings{engstler2026syncity3k,
  title     = {SynCity 3000: Bootstrapping Scene-Scale 3D Diffusion},
  author    = {Paul Engstler and Iro Laina and
               Christian Rupprecht and Andrea Vedaldi},
  booktitle = {European Conference on Computer Vision},
  year      = {2026}
}