Anti-Aliased 2D Gaussian Splatting

TL;DR

We address aliasing in 2D Gaussian Splatting (2DGS) by introducing AA-2DGS, which enforces frequency constraints via a world-space flat smoothing kernel and applies an object-space Mip filter derived from an affine approximation of the ray-splat mapping. The world-space kernel limits primitive frequencies based on training-view sampling, mitigating magnification artifacts, while the per-splat Mip filter transfers screen-space anti-aliasing into the local splat space for efficient, per-pixel filtering. Extensive experiments on Blender, Mip-NeRF 360, and DTU show AA-2DGS consistently improves multi-scale rendering quality over vanilla 2DGS and rivals or surpasses 3DGS-based anti-aliasing approaches, with maintained geometric accuracy. The approach preserves 2DGS’s strengths in depth and normal fidelity, enabling more reliable mesh reconstruction and view-consistent rendering across zooms and varying FOVs, at a modest rendering-time overhead. Limitations remain around the inherent planar nature of 2D Gaussians and fixed filter parameters, which can cause residual artifacts in extreme magnification or grazing-view scenarios.

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 anti-aliased 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.

Figures (7)

• 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.
• Figure 2: Overview. We constrain the maximum frequency of our 2D Gaussians (Red) to a limit estimated from the training images with a 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 to a final simpler and computationally lighter expression of our kernel (Orange) defined in local splat space.
• Figure 3: Single-scale Training and Multi-scale Testing on the Blender Dataset nerf. All methods are trained at full resolution and evaluated at different (smaller) resolutions to mimic zoom-out. Our method (AA-2DGS) consistently demonstrates improved quality across all sampling rates compared to the baseline 2DGS method.
• Figure 4: Single-scale Training and Multi-scale Testing on Mip-NeRF 360 dataset mipnerf360 All models are trained on 1/8 resolution and tested at different upscaling factors. Our AA-2DGS method maintains high fidelity when rendering at resolutions higher than the training resolution, reducing magnification artifacts compared to the baseline 2DGS method.
• Figure 5: 2DGS and our method's RGB and normal rendering under different image sampling rates than the training views. We show results of simulating Zoom In (2x) and Zoom Out (4x). In addition to anti-aliased color rendering, our method also improves other attributes rendering.