hzxie/GaussianCity

The official implementation of "GaussianCity: Generative Gaussian Splatting for Unbounded 3D City Generation". (CVPR 2025)

What it solves

GaussianCity addresses the challenge of generating unbounded 3D city environments. It provides a way to create large-scale, realistic 3D cityscapes that are not limited by a fixed boundary, allowing for the generation of expansive urban areas.

How it works

The project uses Generative Gaussian Splatting to synthesize 3D cities. It employs a dual-generator architecture consisting of a Background Generator and a Building Generator to handle different elements of the urban environment. The system can be trained on datasets like OpenStreetMap (OSM) and Google Earth data, utilizing semantic and instance segmentation to structure the generated city.

Who it’s for

This tool is designed for researchers and developers working in 3D scene generation, urban planning simulation, and computer vision, specifically those interested in large-scale generative 3D modeling.

Highlights

  • Unbounded Generation: Capable of creating 3D cities without fixed spatial limits.
  • Dual-Generator System: Separate models for background and building generation to improve detail and structure.
  • Gaussian Splatting: Leverages the efficiency and visual quality of Gaussian Splatting for 3D representation.
  • Pretrained Models: Provides ready-to-use models trained on Google Earth data.

Related

  • Project
    VIAME/VIAME

    VIAME is an open‑source computer‑vision toolkit that provides detection, tracking, annotation, search and many other image/video processing capabilities through a modular, multi‑language pipeline framework. It ships with desktop and web GUIs, command‑line tools, pre‑built binaries and Docker images, and can be extended via C++, Python or MATLAB plugins.

  • Project
    drivelineresearch/openbiomechanics

    A public research dataset providing cleaned motion-capture files, time-series biomechanics, and physical assessment metrics for baseball pitching and hitting.

  • Project
    jcreinhold/intensity-normalization

    A Python package for normalizing the intensities of MR images across scanners and sites to prevent acquisition artifacts from breaking machine learning pipelines.

  • Project
    jeremyipark/vision-demos

    vision‑demos is a set of four concrete Python demos that apply state‑of‑the‑art pose‑estimation (`vitpose-plus-large`) and segmentation (`sam3.1`) models to everyday activities—dance synchronization, chin‑up counting, bouldering route analysis, and running cadence measurement. Each demo includes a README, setup instructions, and visual examples, and the whole repo is released under Apache‑2.0.