Anti-Aliased 2D Gaussian Splatting

Mae Younes, Adnane Boukhayma · 2025

Read Anti-Aliased 2D Gaussian Splatting by Mae Younes, Adnane Boukhayma (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-11252.txt.

Introduction

Zoom a reconstructed 3D scene in or out and the image can shimmer with aliasing artifacts. Mae Younes and Adnane Boukhayma attack that flaw in 2D Gaussian splatting, a fast method for rendering scenes and recovering surfaces. The technique beats its 3D cousin on consistency and geometry. But it breaks when the rendering sampling rate differs from training, exactly what zooming does.

They trace the artifacts to two causes: no frequency limits in the representation, and weak screen-space clamping. Their fix adds a smoothing kernel in world space, capping each splat's frequency content. A new filter works in the splat's own local space using an affine shortcut. The result keeps the geometric benefits while rendering cleanly across scales, from wide views to close-ups.

Catalog record

CREATORMae Younes, Adnane Boukhayma
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-11252

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Anti-Aliased 2D Gaussian Splatting Mae Younes, Adnane Boukhayma Abstract 2D Gaussian Splatting (2DGS) has recently emerged as a promising method for novel view synthesis and surface reconstruction, offering better view-consistency and geometric accuracy than volumetric 3DGS. However, 2DGS suffers from severe aliasing artifacts when rendering at different sampling rates than those used during training, limiting its practical applications in scenarios requiring camera zoom or varying fields of view. We identify that these artifacts stem from two key limitations: the lack of frequency constraints in the representation and an ineffective screen-space clamping approach. To address these issues, we present AA-2DGS, an antialiased formulation of 2D Gaussian Splatting that maintains its geometric benefits while significantly enhancing rendering quality across different scales. Our method introduces a world space flat smoothing kernel that constrains the frequency content of 2D Gaussian primitives based on the maximal sampling frequency from training views, effectively eliminating high-frequency artifacts when zooming in. Additionally, we derive a novel object space Mip filter by leveraging an affine approximation of the ray-splat intersection mapping, which allows us to efficiently apply proper anti-aliasing directly in the local space of each splat. Code will be available at AA-2DGS . 2DGS AA-2DGS (Ours) Reference 8x Zoom Out Training Resolution 4x Zoom In Figure 1 : 2DGS and AA-2DGS under change of image sampling rate . We trained the models on single-scale images and rendered images with different resolutions to simulate Zoom In/Out. While they achieve similar performance at training scale, strong artifacts appear in 2DGS when changing the sampling rate. Our method (AA-2DGS) shows significant improvement in comparison. 1 Introduction 3D reconstruction from multi-view images has been a fundamental problem in computer vision, graphics, and machine learning for decades. This field has seen renewed interest due to its applications in autonomous driving, medical imaging, gaming, visual effects, and extended reality experiences, which all require high-quality 3D modeling and visualization. Neural Radiance Fields (NeRF) nerf revolutionized this area by introducing a neural representation that models scenes through differentiable volume rendering. Building on this foundation, 3D Gaussian Splatting (3DGS) 3dgs recently revitalized point-based graphics by replacing neural networks with explicit 3D Gaussian primitives. These primitives are rasterized and rendered via volume resampling and their parameters are optimized through gradient-based inverse rendering. With its efficient density control, primitive sorting, and tile-based rasterization, 3DGS achieves state-of-the-art novel view synthesis while enabling real-time rendering and requiring shorter training times. The Gaussian Splatting approach has evolved into two primary variants: 3DGS 3dgs and 2D Gaussian Splatting (2DGS) 2dgs . While 3DGS represents scenes using volumetric 3D Gaussian primitives, 2DGS employs flattened 2D Gaussian disks embedded in 3D space. This distinction is significant: 3DGS projects 3D Gaussians onto the screen to obtain 2D screen-space Gaussians, which are then rendered. In contrast, 2DGS evaluates the Gaussians directly at ray-splat intersections in the local coordinates of each planar primitive. This approach gives 2DGS superior geometric accuracy, particularly for depth and normal reconstruction, making it valuable for applications requiring precise geometry such as mesh recovery surfels , physics-based rendering irgs , and reflectance modeling refgaussian . Despite its strengths, 2DGS faces a significant challenge: its formulation complicates the integration of proper antialiasing techniques. The original 2DGS implementation attempts to address this by employing a screen-space lower bounding approach (clamping) clamp , but our investigation reveals that this approach often exacerbates aliasing artifacts rather than mitigating them. This is particularly evident when rendering at different sampling rates, such as zooming in or out from a scene (See Tab. 1 , Tab. 2 and 3 ). Recent work on Mip-Splatting mipsplat has identified two key sources of aliasing in 3DGS: the lack of 3D frequency constraints and inadequate screen-space filtering. Mip-Splatting addresses these issues by introducing a 3D smoothing filter to constrain the frequency content of primitives based on training view sampling rates, and by replacing the traditional screen-space dilation with a Mip filter that better approximates the physical imaging process. However, these solutions cannot be directly applied to 2DGS due to its fundamentally different primitive representation and rendering approach. In this paper, we present Anti-Aliased 2D Gaussian Splatting (AA-2DGS), which adapts the insights from Mip-Splatting to the flattened 2D Gaussian representation. Our approach makes two key contributions: \bullet We introduce a world space flat smoothing kernel that constrains the frequency content of 2D Gaussian primitives based on the sampling rates of the training views. This addresses high-frequency artifacts when zooming in on a scene by ensuring that the primitives respect the Nyquist-Shannon sampling theorem shannon1949communication . \bullet We derive an object space Mip filter that leverages an affine approximation of the ray-splat intersection mapping used in 2DGS. This allows us to incorporate mip filtering directly in the local space of each splat, where the Gaussian evaluation occurs. The resulting formulation is both mathematically elegant and computationally efficient. Ablative analysis of these two components in the supplementary material. We evaluate AA-2DGS on standard novel view synthesis datasets, including Mip-NeRF 360 mipnerf360 and Blender nerf , as well as the DTU dtu mesh reconstruction benchmark. Our results demonstrate that AA-2DGS consistently outperforms the original 2DGS method, particularly under challenging conditions such as varying sampling rates and mixed resolution training. Importantly, our approach maintains the geometric accuracy that makes 2DGS valuable while significantly reducing aliasing artifacts. Figure 2 : Overview. We constrain the maximum frequency of our 2D Gaussians (Red) to a limit estimated from the training images with world space flat smoothing filter. Next, leveraging an affine approximation of the mapping from screen space to local splat space: m_{\mathbf{J}} where \mathbf{J}=\tfrac{\partial\boldsymbol{u}}{\partial\boldsymbol{x}} , we can express the reconstruction kernel footprint in screen space (Blue). This enables the integration of a screen space anti-aliasing Gaussian filter (Green). Via the affine mapping, we can revert back to a final simpler and computationally lighter expression of our kernel (Orange) defined in local splat space. 2 Related Work Until recently, implicit representations coupled with differentiable volume rendering have been at the forefront of 3D shape and appearance modeling. NeRFs nerf model scenes with an implicit density and view dependent radiance, parametrized with MLPs. Anti-aliasing can be implemented in these reprsentations through cone tracing and pre-filtering the input positional or feature encodings mipnerf ; mipnerf360 ; zipnerf ; trimiprf ; zhuang2023anti . Multiscale volume rendering requires intensive MLP querying thus limiting the rendering frame rate. This issue can be aleviated with grid based representations ingp ; dvgo ; kplanes ; plenoxels ; tensorrf . These can struggle with large unbounded reconstruction, despite Level-of-detail Octrees nsvf . By expressing density as a function of a signed distance field, NeRFs lead to powerful geometry recovery methods neus ; volsdf ; neuralangelo ; voxurf . 3D Gaussian splatting 3dgs subverted this trend lately. Combining volume rendering volrend and EWA splatting ewa within an efficient inverse rendering optimization adam . It has spawned substantial research due to its remarkable novel view performance and high rendering frame rate. Extensions include generalizable models pixsplat ; mvsgaussian , bundle adjustment bad ; colfree , higher dimensional primitives ndim , spatiotemporal models 4d , in addition to several methods to improve density control 3dgsmcmc , model compactness compact and training speed newton . Recent work augmented 3DGS’s anti-aliasing abilities. anal analytically approximates

Read the complete text