GaussianCity: 60x speed 3D city model generation technology

The first super-large 3D city model generation technology called "GaussianCity" of the Nanyang Technological University research team has attracted a lot of attention. This new framework developed by the research team not only achieves an astonishing 60 times improvement in generation speed, but also breaks through the scale limitations of traditional methods and supports borderless 3D city generation.

This technological achievement has been accepted by CVPR2025 (Top Conference on Computer Vision and Pattern Recognition), becoming a major breakthrough in the fields of virtual reality, autonomous driving and digital twins.

It is understood that GaussianCity has reached the most advanced level in generating 3D urban models from drone perspectives and street perspectives. Its rendering speed is as high as 10.72 frames per second (FPS), which is 60 times faster than the existing CityDreamer solution. Although CityDreamer performs well in the field of 3D city generation, its computing efficiency and scale expansion capabilities have always been limited. And GaussianCity successfully overcomes these difficulties by introducing innovative algorithm design.

The core of GaussianCity lies in two key technological breakthroughs. First, it adopts a compact 3D scene representation method "BEV-Point" (bird's-eye view point), which greatly reduces the memory demand and makes large-scale scene generation no longer limited to hardware resources. Traditional 3D Gaussian Splating (3DGS) technology requires billions of points when dealing with unlimited-scale cities, which often occupies hundreds of GB of video memory. GaussianCity maintains constant video memory usage through BEV-Point, achieving true boundaryless generation. Secondly, the research team developed a space-aware Gaussian attribute decoder, using point serializer to integrate the structure and contextual features of BEV points to ensure that the generated urban model is both efficient and realistic.

It is worth mentioning that the research and development team of GaussianCity announced that the papers, codes and related materials of the project have been fully open source. The emergence of GaussianCity has brought new possibilities to multiple fields. In virtual reality (VR) and augmented reality (AR), it can quickly generate high-quality large-scale urban environments, providing users with an immersive experience; in the field of autonomous driving, GaussianCity can be used to rebuild geometrically accurate 3D scenes, providing realistic digital twin cities for training and testing; in urban planning and game development, its efficiency and scalability will also greatly improve creative efficiency.

Project entrance: https://github.com/hzxie/GaussianCity