Research map / 3D vision / Gaussian splatting
Generative Gaussian Splatting for Unbounded 3D City Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GaussianCity, a generative 3D Gaussian splatting framework that uses a compact BEV-Point representation and a Point Serializer-based decoder to synthesize unbounded 3D cities in a single feed-forward pass.
Problem
Unbounded 3D city generation with 3D Gaussian splatting is infeasible because the number of Gaussians grows to billions, causing out-of-memory issues and huge storage for city-scale scenes.
Approach
Decouples Gaussian attributes into position-related BEV maps (extruded into visible BEV points with a constant VRAM budget) and style-related instance codes in a style lookup table, then decodes Gaussian RGB attributes with a point-serialized Transformer and rasterizes them.
Key results
- On GoogleEarth, GaussianCity reaches FID 86.94 and KID 0.090 versus 97.38 FID for CityDreamer, with runtime of 93 ms versus 5580 ms.
- On KITTI-360, GaussianCity achieves FID 29.5 and KID 0.017, improving over UrbanGIRAFFE (FID 39.6).
- Ray-intersection visibility filtering reduces the point count from 1.4M (region) to 31.7K while keeping VRAM at about 1.4 GB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GaussianCity instead of CityDreamer | GoogleEarth city generation · GoogleEarth | FID: 97.38 -> 86.94 | Helped |
| Use GaussianCity instead of CityDreamer | GoogleEarth city generation · GoogleEarth | Runtime (ms): 5580 -> 93 | Helped |
| Use GaussianCity instead of the strongest listed FID baseline | KITTI-360 city generation · KITTI-360 | FID: 31.9 -> 29.5 | Helped |
| Add ray-intersection visibility filtering (ablation) | BEV-Point representation · GoogleEarth | VRAM (G): Out of Memory -> 1.41 | Helped |
| Add a density map with ray-intersection visibility filtering (ablation) | BEV-Point representation · GoogleEarth | VRAM (G): 1.41 -> 1.39 | Helped |
| Add Point Serializer and Point Transformer (ablation) | BEV-Point Decoder · GoogleEarth | FID: 151.27 -> 86.94 | Helped |
| Add Point Serializer (ablation) | BEV-Point Decoder with Point Transformer · GoogleEarth | FID: 119.40 -> 86.94 | Helped |
| Replace Equation 13 with Hilbert serialization (ablation) | Point Serializer · GoogleEarth | FID: 86.94 -> 86.72 | No clear effect |
| Use Equation 13 and Hilbert serialization together (ablation) | Point Serializer · GoogleEarth | FID: 86.94 -> 86.28 | No clear effect |
Surprising findings
- Removing the density map increases point count only by 20%-50%, while dense point sets run out of memory without visibility filtering.
- Different serialization methods (Eq. 13 vs Hilbert curve) have minimal impact on generation quality.
Limitations
- Height-field lifting assumes Manhattan-style buildings and cannot represent hollow structures.
- Decoder predicts only RGB Gaussian attributes, not the full 3D-GS parameter set (scale, rotation, opacity), limiting expressiveness.
Experimental setup
Datasets: GoogleEarth KITTI-360
Benchmarks: GoogleEarth KITTI-360
Metrics: FID KID depth error camera error runtime user study scores
Baselines: SGAM PersistentNature SceneDreamer InfiniCity CityDreamer StyleGAN2 GSN GIRAFFE UrbanGIRAFFE
Builds on: 3D Gaussian splatting CityDreamer StyleGAN2 Point Transformer
Closest papers by meaning
- Lightweight Predictive 3D Gaussian Splats (ICLR 2025 · similarity 0.90)
- EVolSplat: Efficient Volume-based Gaussian Splatting for Urban View Synthesis (CVPR 2025 · similarity 0.90)
- DirectTriGS: Triplane-based Gaussian Splatting Field Representation for 3D Generation (CVPR 2025 · similarity 0.89)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.89)
- CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-Scale Scenes (ICLR 2025 · similarity 0.89)
- Horizon-GS: Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes (CVPR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.