CityGaussian: Real-time high-quality large-scale scene rendering with Gaussians
Real-time rendering of large urban scenes with 3D Gaussian splatting is drawing attention for its near-photorealistic results, but also skepticism over its current dependence on high-end GPUs like Nvidia’s A100. Commenters compare it to traditional meshes and photogrammetry, debating whether splats are a practical long-term alternative or best treated as a “3D photo” or final-render format, especially given challenges around animation, dynamic lighting, and data size. Many expect hardware progress and optimization to make the technique more accessible and see strong potential in virtual production, VR/spatial computing, and hybrid pipelines that mix splats with conventional geometry.
Performance, “Real-Time,” and Hardware
- Reported 36 FPS on an A100 leads to debate over what “real-time” means; several note the traditional graphics-paper pattern of using modest resolutions and very high-end GPUs.
- Some argue A100-level performance will trickle down to consumer GPUs in a few years; others note deliberate market segmentation between “pro” and “consumer” lines.
- There’s discussion about whether memory bandwidth (A100) or raw compute (e.g., 4090) is the bottleneck; unclear from the thread.
- Multiple commenters cite other Gaussian splatting implementations already achieving 60–400+ FPS on consumer hardware or in browsers, though performance is highly device-dependent.
Scaling, LOD, and Optimization
- CityGaussian is praised as one of the first 3D Gaussian Splatting methods to use LOD and spatial blocking for large-scale scenes.
- Many believe there is substantial room for optimization (LOD streaming, pruning, formats, compression), and that research code tends to be far from production-quality.
Use Cases: Maps, Games, VR, and “3D Photos”
- Strong interest in applications to Google Maps/Earth–style city exploration and virtual production, but skepticism about mobile viability in the short term.
- For games, key challenge is dynamics: current methods are mostly static with baked lighting. Efficient animation, object interaction, and consistent lighting in hybrid pipelines remain open problems.
- Some see 3DGS as ideal for “3D photographs” and static environments (VR, virtual sets), with hybrid pipelines mixing splats and traditional meshes.
Comparison to Photogrammetry and Meshes
- Proponents say 3DGS outperforms classical photogrammetry on complex, semi-volumetric, or fine-detail content (foliage, hair, rough surfaces).
- Skeptics argue 3DGS is data-heavy, slow, blurry up close, and poorly suited to non-Lambertian materials due to underlying Structure-from-Motion limitations.
- Several suggest 3DGS may be better as an output/final-render format, with meshes still useful for editing and animation; others point to active work on converting between meshes and splats.
Data, Demos, and Related Work
- The showcased city dataset (MatrixCity) is identified as derived from Unreal Engine’s Matrix city, i.e., synthetic but photorealistic, with perfect camera poses.
- Commenters link related large-scale GS work (e.g., octree-based methods) and simple web demos mixing Google 3D tiles with splats.
Miscellaneous
- Interest in open-source licensing and engines.
- Brief off-topic subthread discusses how motion blur is implemented via motion vectors and post-processing in games.