Abstract
3D Gaussian Splatting (3DGS) has recently enabled real-time photorealistic rendering in compact scenes, but scaling to large urban environments introduces severe aliasing artifacts and optimization instability, especially under high-resolution (e.g., 4K) rendering. These artifacts, manifesting as flickering textures and jagged edges, arise from the mismatch between Gaussian primitives and the multi-scale nature of urban geometry. While existing "divide-and-conquer" pipelines address scalability, they fail to resolve this fidelity gap. In this paper, we propose PrismGS, a physically-grounded regularization framework that improves the intrinsic rendering behavior of 3D Gaussians. PrismGS integrates two synergistic regularizers: pyramidal multi-scale supervision and explicit size regularization. Our method is plug-and-play and compatible with existing pipelines, achieving state-of-the-art performance with significant PSNR gains around 1.5 dB against CityGaussian.
Addressing the Fidelity Gap in Large-Scale 3DGS
While 3D Gaussian Splatting (3DGS) has revolutionized real-time photorealistic rendering for small, object-centric scenes, scaling it up to massive urban environments introduces two major challenges: aliasing and geometric instability.
- Aliasing: When viewing a large scene from a distance or at high resolutions (like 4K), textures often "shimmer" or flicker, and straight edges look jagged. This happens because there is a mismatch between the discrete Gaussian primitives and the multi-scale nature of city geometry.
- Geometric Degeneracy: During optimization, the system often creates "needle-like" or "pancake-like" Gaussians. These tiny, thin primitives overfit to specific training views but fail when viewed from a new angle, creating holes or visual "popping" artifacts.
Existing "divide-and-conquer" pipelines focus on the scale of the data but often ignore these fidelity gaps.
The PrismGS Framework
PrismGS introduces a "plug-and-play" regularization framework that can be integrated into existing 3DGS pipelines. It relies on two synergistic components:
Pyramidal Multi-Scale Supervision (MSS)
Instead of only supervising the model at the highest resolution, PrismGS creates a pyramid of ground-truth images using Gaussian blurring and downsampling. During training, the model is forced to render the scene at multiple resolutions simultaneously. This "bakes" anti-aliasing directly into the 3D primitives, ensuring they remain coherent whether viewed from up close or far away.
Physically-Grounded Size Regularization ($L_{size}$)
To prevent the formation of degenerate "needle" Gaussians, PrismGS introduces a lower bound on the dimensions of each primitive. This bound is derived from the Nyquist theorem. By calculating the pixel sampling interval in 3D space based on camera focal length and depth, the framework establishes a physical limit for the smallest resolvable structure. Any Gaussian that tries to shrink below this threshold is penalized.
Experimental Evaluation
The authors evaluated PrismGS on three prominent large-scale benchmarks: MatrixCity, Mill-19, and UrbanScene3D. For standard evaluation, images were downsampled by 4x, but the authors also tested robustness by rendering novel views at full 4K resolution (3840 × 2160).
Key Results
- Quantitative Superiority: PrismGS consistently outperformed SOTA methods like CityGaussian and Momentum-GS. On the Building scene, it achieved a PSNR of 23.516 dB and significantly better LPIPS scores, which correlate more closely with human perception of flickering.
- 4K Robustness: In demanding 4K rendering tests, PrismGS maintained its lead, outperforming competitors across all metrics on the Rubble dataset.
- Visual Clarity: Qualitative results show that PrismGS preserves fine-grained textures on distant buildings where other methods produce a blurry or "shimmery" mess.
Insights and Future Directions
A key insight of this work is that aliasing in 3DGS is an optimization issue. If the model never "sees" how a primitive should look at a lower resolution, it will never learn to be scale-consistent. Furthermore, linking Gaussian size to the Nyquist sampling theorem provides a rigorous mathematical justification for geometric constraints.
The primary limitation is the static environment assumption; PrismGS does not explicitly handle dynamic objects like moving cars. Future work could involve integrating semantic scene understanding to differentiate between static structures and dynamic elements.
Summary updated