Single-image 3D urban generation

CityLiftCompositional 3D Urban Generation
from a Single Image

Haoran Gong1,2Jie Feng2,†
1Xi’an Jiaotong University2Zhongguancun Academy

† Corresponding author

One image. An urban scene of individual, editable 3D buildings.

From a single image to a 3D urban sceneVisual results coming soon
CITYLIFT / SHOWCASE
CityLift generates an explicit mesh for each building, bringing scene layout and local geometric detail into one composition.
01Single-image inputGuided by building instance masks
02Compositional outputIndividual buildings in a shared scene
03Explicit geometryBuilding meshes for further editing

Abstract

Generating urban scenes as individual 3D assets enables flexible scene editing and provides geometry for simulation and urban world modeling. CityLift addresses this task from a single image and building instance masks, generating a separate mesh for each building within a shared scene.

Our framework first generates an urban scaffold that jointly represents scene layout and coarse building structure. Generative spatial inheritance then transfers coarse geometric support and scene features into dedicated local building volumes. Conditioned building regeneration produces detailed local structure, which is converted to meshes and assembled in the scene through the inherited spatial mapping. We also introduce CityLift-5K, a collection of 4,997 urban scenes and 36,338 unique building assets, pairing images and instance masks with building geometry and scene placement.

Explore the results

Interactive Results

A scene at a glance. Each building in detail.

Scene 01Preview coming soon
Input image
Input imageComing soon

One image + building instance masks

Generated scene
Explore the generated sceneInteractive results coming soon

Scene previews will be added here.

Side-by-side results

Comparisons to Other Methods

Six methods in one synchronized orbit. Choose an input image below to compare.

Loading rendered previews…

Input images

Select a scene to compare

The framework

Scene layout. Local detail.

Three stages connect a shared urban scaffold to individual building meshes.

01

Urban Scaffold
Generation

Jointly generate scene layout and coarse building structure with shared scene–instance modeling and scene guidance.

02

Generative Spatial
Inheritance

Map coarse support and scene features into each building’s local volume through a deterministic spatial interface.

03

Conditioned Building
Regeneration

Generate local building structure, convert it to a mesh, and assemble it in the scene using the inherited mapping.

Method overviewFramework figure coming soon
A scene-to-building generation framework with a shared spatial interface for local generation and scene assembly.
Paired urban scenes

CityLift-5K

CityLift-5K brings together 4,997 synthetic urban scenes and 36,338 unique building assets, pairing RGB images and instance masks with textured building meshes and scene placement annotations.

Dataset download coming soon
Urban scenes
4,997
Unique buildings
36,338
Scene imagesInstance masksBuilding meshesScene transforms
Explore CityLift-5KDataset overview video
A one-minute overview of scene diversity, paired RGB images and instance masks, and building annotations.
Follow the project

Research resources

Read the paper on arXiv. Code and dataset are coming soon.