Skyfall-GS points to a simpler route for building digital cities: start with standard satellite images, then use AI to infer the parts the camera cannot directly see. The result is a walkable 3D city model that is not limited to rooftops.
The system matters because many satellite-based 3D city models have a basic weakness. Aerial views can capture roof shapes well, but they often miss facades, side details, and street-level texture. That gap can leave buildings looking blurry, distorted, or blocky when viewed from lower angles.
Why satellite-only 3D cities are difficult
Older approaches to detailed 3D city modeling can depend on costly 3D scanners or fleets of camera cars. Those methods can gather street-level information, but they add expense and operational complexity.
Satellite imagery is far more broadly available, but it comes with a viewpoint problem. From above, a model can see rooftops and broad layouts. It cannot directly observe the sides of buildings or many of the details that make a street feel navigable.
Skyfall-GS addresses that limitation by treating missing visual information as something an AI model can reconstruct. It first builds a rough 3D outline from satellite images. Then it fills in facades and street-level textures in a way that resembles how image generators complete unfinished pictures.
How Skyfall-GS works
The system combines two AI techniques. The base 3D structure is created with 3D Gaussian splatting, which represents a scene as clouds of light points. Diffusion models, the same kind of models used in popular image generators, then add more realistic visual detail.
The name Skyfall-GS reflects its training strategy. The virtual view begins high above the city and moves downward in steps, refining the scene as if a camera were falling from the sky.
The process runs in five passes. During each pass, the virtual camera changes its viewing angle, moving from 85 degrees down to 45 degrees. The AI creates 54 different views in each round and uses text prompts to guide the cleanup and enhancement process.
Those prompts focus the system on improving visual quality. In the source description, the prompt guidance shifts a flawed input described as a “satellite image of an urban area with distorted areas and blurring artifacts” toward a cleaner target: a “clear satellite image with sharp buildings, smooth edges, and natural lighting.”
What the tests showed
Researchers tested Skyfall-GS with real satellite images from Jacksonville, Florida, and New York City. Against previous methods, the system produced more realistic buildings and cleaner textures.
A user study with 89 participants found strong preference for the new approach. Skyfall-GS was rated best in 97 percent of comparisons for both geometry and overall quality.
Performance is also a major part of the result. Skyfall-GS runs at 11 frames per second on a standard graphics card and up to 40 frames per second on a MacBook Air. CityDreamer, a previous system, manages 0.18 frames per second on more expensive hardware.
That speed matters because walkable 3D environments are not just static maps. They need to be rendered from changing viewpoints as a user, camera, or simulation moves through the scene. A system that can produce better city geometry while running faster could make satellite-derived city models more practical.
Where this could be useful
The most obvious applications are fields that already need large digital environments. Game developers could use this kind of system to create city settings more efficiently. Film productions could generate digital backgrounds without manually building every street-facing surface.
Robotics teams may also find value in simulated real-world spaces. A more complete city model can help create environments where machines can be tested before they face physical streets and buildings.
The larger opportunity comes from the volume of satellite imagery already being collected. WorldView-3 collects around 680,000 square kilometers per day at up to 31 centimeters per pixel. That scale suggests a path toward automated 3D modeling across large areas, if the modeling process can become efficient and reliable enough.
Limits still remain
Skyfall-GS is not presented as a finished answer to every city modeling problem. The researchers acknowledge that it still requires a lot of computing power. They also note that it does not always handle highly detailed street scenes well.
Those limits are important. Inferring street-level detail from aerial views is inherently difficult because the source images do not directly contain every surface the final model needs. The system can improve missing or distorted areas, but highly detailed scenes remain challenging.
The next goals are performance and scalability. The code is available as open source on GitHub, and demos are available on the project website. For now, Skyfall-GS shows how satellite images, 3D Gaussian splatting, and diffusion models can work together to make walkable city reconstruction less dependent on ground-level capture.