Reconstructing entire cities with 3D Gaussian splatting

What it does

CityGaussian is the official implementation of the ECCV 2024 and ICLR 2025 papers that extend 3D Gaussian splatting from small objects to entire urban environments. It reconstructs massive scenes by partitioning models and data across multiple GPUs with controllable memory costs, then renders them in real time. The codebase also extracts actual geometry via 2DGS-style mesh extraction and evaluates surface reconstruction quality on city-sized canvases.

The interesting bit

Rather than treating a city as one giant Gaussian cloud and hoping your VRAM survives, the system analyzes how to split the model and assign data across an arbitrary number of GPUs. It also jointly optimizes camera poses and Gaussians when your COLMAP results—or poses from 3D foundation models like VGGT-X—are imperfect.

Key highlights

• Multi-GPU reconstruction with no hard limit on GPU count and tunable memory usage
• 2DGS-style mesh extraction with geometric performance evaluation for large scenes
• Trajectory-aligned rendering and mesh video generation with floater removal
• Joint pose and 3DGS optimization for noisy poses from foundation models
• Rebased on Gaussian Lightning v0.10.1, inheriting web viewer, MipSplatting, AbsGS, and StopThePop

Caveats

• Non-commercial CC BY-NC-SA 4.0 license, so commercial use is off the table
• Appearance embedding and V1-style level-of-detail are still on the TODO list, not yet implemented
• The authors warn that COLMAP preprocessing takes considerable time and OOM errors are common without downsampling or cache tuning

Verdict

Worth exploring if you’re building digital twins, urban neural renderers, or anything at scales where a single GPU chokes. Skip it if you need a commercial license or just want to splat a coffee mug.