XGRIDS Builds the World's Largest Air-Ground Integrated 3D Gaussian Model of Minning Town

A 23.5-Square-Kilometer Digital Town: 3D Gaussian Technology at Scale

When spatial intelligence moves from the lab into the real world, the first thing tested is not the elegance of the algorithm but whether it can be delivered reliably in real, complex, large-scale scenarios. A 3D reconstruction project completed by XGRIDS in Minning Town, Ningxia, offers a convincing answer.

This continuous 3D model, covering 23.5 square kilometers (equivalent to 3,291 standard football fields), was trained from 186,000 aerial and ground images and 15 billion radar points, ultimately generating 6.2 billion Gaussian ellipsoids. The model appeared in the China Media Group documentary Minning Avenue, used to present the transformation of this model resettlement town over three decades.

Air-Ground Fusion: 3 Days of Capture and the Engineering Challenge of Billions of Gaussians

From Data Capture to Model Generation

The field data capture for the 23.5-square-kilometer town was completed by 3 operators over 7 days. Aerial drones equipped with capture devices flew over the town in sections, while on the ground the Lixel L2 Pro spatial scanner and Lixel P1 spatial camera moved through streets, roads, and villages. Laser and imagery, overhead and eye-level views—massive multimodal data was gathered within 7 days.

Data processing was completed locally. LCC Studio's "Air-Ground Fusion" feature merged the aerial and ground perspectives into a single space; Multi-SLAM resolved the laser data into a point cloud of 15 billion points, establishing the geometric relationships between eaves, roads, and buildings; LCC Studio then trained 6.2 billion Gaussian ellipsoids from these multimodal data, restoring light, shadow, color, and materials. The final reconstruction achieved centimeter-level geometric accuracy, with reconstruction error better than 2 cm.

The Technical Challenges of a Single Continuous Model

Large-scale scenes are typically reconstructed in blocks and then stitched together, making seams, color differences, and geometric misalignments almost inevitable. Digital Minning Town carries all 23.5 square kilometers in a single continuous model, with no break points from the town entrance to the end of the village. This technical choice implies higher computing and storage thresholds—3 operators, 7 days of field capture, 14 days of processing billion-scale Gaussian data. These numbers would be nearly impossible to achieve under traditional 3D reconstruction approaches.

3D Gaussian Splatting (3DGS) is a 3D scene representation and rendering technique proposed in 2023. Unlike traditional meshes or Neural Radiance Fields (NeRF), it explicitly represents scenes using millions or even billions of 3D ellipsoids with direction, color, transparency, and spatial scale, then achieves real-time rendering through rasterization. Its advantages: training is far faster than NeRF, rendering can reach real-time frame rates, and it restores light, shadow, materials, and details well. The price is that storage and VRAM usage grow linearly with the number of Gaussians, so billion-scale cases place high demands on compression, partitioning, and engine integration. The 6.2 billion Gaussian ellipsoids in this article represent an extremely large case among publicly known projects, which is one reason a dedicated UE plugin had to be developed.

SLAM and Multi-SLAM: SLAM (Simultaneous Localization and Mapping) is a technology that lets a device estimate its own pose while building a map in an unknown environment, and it is the foundation for high-precision point clouds from laser scanning and mobile surveying. Multi-SLAM usually refers to a SLAM approach that fuses multiple sensors, viewpoints, and trajectory segments, enabling unified computation of data captured by drones, backpacks, handheld devices, and more within the same coordinate system, avoiding cumulative drift between multiple captures. The alignment of 15 billion points in a unified space in this article depends precisely on this kind of air-ground integrated multi-source SLAM capability.

LCC Plugin for UE: Bringing 6.2 Billion Gaussians into the Film and Television Production Pipeline

What truly brought the 6.2 billion Gaussians into the film and television production workflow was the LCC Plugin for UE. This is the first Unreal Engine plugin designed for billion-scale 3D Gaussians, and the entire Digital Minning Town entered Unreal Engine in a production-grade manner, with import, editing, relighting, and rendering all completed directly in the engine. According to XGRIDS, there is currently no other UE Gaussian plugin on the market capable of carrying such scale.

From a production perspective, this means the crew could change the street blocks captured during the day to sunset lighting in the engine, and could repeatedly frame the same red rooftops from any height and any angle. The shots in the documentary that sweep over the streets and pass through the archway avenue were completed in this Digital Minning Town.

The precision of the model also preserved the details: the archway at the town entrance, the red-roofed, white-walled blocks, the road leading into the distance, and the wind turbines turning in front of the mountains were all captured in the same continuous model.

Unreal Engine (UE) is a real-time 3D creation engine developed by Epic Games, widely used in virtual production for film and television, games, and digital twins. In traditional virtual production workflows, real scenes usually enter the engine as panoramic photos, laser point clouds, or manual models; the former lacks真实 geometry, while the latter has a long production cycle. If a 3D Gaussian model can be natively carried by UE, it can be lit, color graded, framed, and rendered directly in the engine, eliminating a large amount of intermediate conversion. However, the engine's real-time rendering imposes strict constraints on data formats, LOD (level of detail), and VRAM management. Therefore, a plugin for billion-scale Gaussians essentially solves the engineering problem of "streaming and real-time rendering of large-scale point primitives," which is also the context for the claim that no comparable plugin on the market can currently handle this scale.

Real Spatial Data: A New Entry Point in the World Model Competition

World models and spatial intelligence are becoming a new focal point of technological competition between China and the United States, and real-world 3D data is regarded as a key entry point for artificial intelligence to understand the physical world. XGRIDS's strategy in this direction is two-pronged: on one hand, building a solid technological foundation through independent innovation; on the other, participating in the global ecosystem with an open approach.

The open LCC format has already been made compatible with dozens of domestic and international platforms, including NVIDIA Isaac Sim, Sketchup, Revit, 3DS Max, and SuperMap. It is rare in the field of spatial intelligence for a data format defined by a Chinese company to enter the global mainstream technology ecosystem.

From raw data to digital assets, and then to AI training corpora, a complete data value chain for real space was explored in this project. The 3D digital assets generated by 3D Gaussian scanning have the characteristics of high realism, fast production efficiency, and low deployment cost, and they have application potential in the digital preservation of cultural relics and ancient architecture, immersive cloud tours, and digital cultural creativity.

World Model refers to an artificial intelligence system capable of internally representing the physical world and, based on that, predicting future states or generating interactive environments. It is often seen as a key path toward embodied intelligence and general intelligence. Training world models requires a large amount of real 3D data with geometric, material, lighting, and physical attributes, while 2D images and text on the internet are not sufficient to provide this information. Therefore, high-precision, measurable, and interactive real-space scan data is becoming a scarce AI training corpus. The "real spatial data value chain" emphasized in this article connects capture, 3D reconstruction, format openness, and engine integration into a pathway from data to assets, so that spatial scanning is no longer merely a surveying result but a foundational resource that can be repeatedly invoked by AI.

From the Gobi to Digital Space: Preserving the Original Appearance for the Era

The story of Minning Town itself is a microcosm of East-West collaboration. In 1996, Fujian and Ningxia established a paired assistance and cooperation relationship. The following year, Minning Village was built on the Gobi Desert at the eastern foot of the Helan Mountains, where the first group of migrants from Xihaigu settled, with an initial resident population of more than 8,000. Thirty years later, it has become Minning Town, comprising 1 community and 6 administrative villages, with a population of more than 60,000.

In the documentary, two kinds of imagery were used for the past and the present: the Gobi and adobe houses that left no 3D records were reconstructed by AI based on historical materials, while today's Minning Town was reconstructed from real scenes using 3D Gaussian technology. With the two types of images placed side by side, the distance of thirty years becomes tangible and visible in the cuts between shots.

As a national-level specialized and sophisticated "key little giant" enterprise, XGRIDS has built a complete technology system spanning spatial capture, 3D Gaussian generation, compression, editing, and application deployment, with more than 100 invention patents, serving more than 8,000 customers in over 80 countries.

For Minning Town, the significance of this record goes beyond a technical demonstration: Minning Town in 2026 has been preserved as an enterable, measurable, and comparable three-dimensional space. Many years later, when the descendants of the first generation of migrants want to see what their hometown looks like at this moment, they can truly "walk into it." Real space is becoming a new content medium, and this application in a national-level documentary provides a reference sample for the large-scale deployment of spatial intelligence technology.

One-line Summary

XGRIDS built a continuous 23.5-square-kilometer 3D Gaussian model in Minning Town, with 6.2 billion Gaussians entering film production and the AI data chain.

Paragraph Summary

XGRIDS completed a large-scale 3D reconstruction in Minning Town, Ningxia: 3 operators completed field capture of 23.5 square kilometers in 7 days, and 186,000 images plus 15 billion radar points were used to train 6.2 billion Gaussian ellipsoids into a seamless continuous model, with reconstruction error better than 2 centimeters. Through the LCC Plugin for UE, the 6.2 billion Gaussians entered Unreal Engine directly for shot production in the China Media Group documentary Minning Avenue. Meanwhile, the open LCC format has been made compatible with dozens of platforms including NVIDIA Isaac Sim, Sketchup, and Revit, demonstrating the complete value chain of real spatial data from capture to digital assets to AI training corpora, and providing a sample for the large-scale deployment of spatial intelligence.