AAA-Gaussians: Anti-Aliased and Artifact-Free 3D Gaussian Rendering

Highlighted at ICCV 2025

Michael Steiner* 1   Thomas Köhler* 1   Lukas Radl 1   Felix Windisch 1   Dieter Schmalstieg 1,3   Markus Steinberger 1,2
* Both authors contributed equally to this work!

1 Graz University of Technology
TU Graz Logo
2Huawei Technologies
Huawei Logo
3 University of Stuttgart
University of Stuttgart Logo
Paper Code Pre-Trained Models
Pipeline

We enable artifact free rendering of Gaussian-Splatting models by introducing a novel 3D anti-aliasing filter and a stable view-space bounding method.

Abstract

Although 3D Gaussian Splatting (3DGS) has revolutionized 3D reconstruction, it still faces challenges such as aliasing, projection artifacts, and view inconsistencies, primarily due to the simplification of treating splats as 2D entities. We argue that incorporating full 3D evaluation of Gaussians throughout the 3DGS pipeline can effectively address these issues while preserving rasterization efficiency. Specifically, we introduce an adaptive 3D smoothing filter to mitigate aliasing and present a stable view-space bounding method that eliminates popping artifacts when Gaussians extend beyond the view frustum. Furthermore, we promote tile-based culling to 3D with screen-space planes, accelerating rendering and reducing sorting costs for hierarchical rasterization. Our method achieves state-of-the-art quality on in-distribution evaluation sets and significantly outperforms other approaches for out-of-distribution views. Our qualitative evaluations further demonstrate the effective removal of aliasing, distortions, and popping artifacts, ensuring real-time, artifact-free rendering.

Video

Adaptive, Lower-Bounded 3D Anti-Aliasing Filter

We introduce a novel adaptive 3D anti-aliasing filter specifically designed for evaluating Gaussians in 3D via ray sampling—avoiding the distortion artifacts inherent in affine projection approximations. Previous 3D evaluation methods (e.g., [GOF, Yu et. al. 2024] and [Hahlbohm et. al. 2025]) only apply a fixed 3D filter based on the closest training view to prevent over-sampling (following [Mip-Splatting, Yu et. al. 2024]). However, they exhibit under-sampling aliasing artifacts when observing scene content from further away. In contrast, our filter dynamically adapts to the camera's sampling frequency by adjusting the Gaussian's 3D scale only perpendicular to the viewing ray, and is correctly lower-bounded by the fixed 3D filter based on training views. This preserves fine details and prevents oversmoothing or transparency artifacts, enabling consistent, artifact-free rendering even in challenging out-of-distribution scenarios.

Adaptive 3D Filter

(A stuttering or out-of-sync video is usually resolve by reloading the page)

Fixed 3D Filter (lower-bound based on training views)

Robust View-Space Bounding

To eliminate pop-in artifacts that appear due to incorrect frustum culling and bounding, we propose a novel view-space bounding method for 3D Gaussians. Unlike prior approaches [Hahlbohm et. al. 2025] that fit bounds in screen space and discard Gaussians extending beyond the near or far planes, our method performs bounding directly in view space using angular limits. This allows us to robustly handle Gaussians that partially extend outside the view frustum, preventing excessive and incorrect culling. The result is stable, artifact-free rendering when navigating close to scene content or using wide fields of view, where previous methods suffer from abrupt pop-in effects.

*We used a Mini-Splatting model resulting in larger Gaussians and more noticable artifacts.

Distortion Free Rendering

Distortion artifacts in 3D Gaussian Splatting often stem from the affine projection approximation of 3D Gaussians as 2D splats, which becomes especially problematic under extreme camera settings such as wide fields of view or viewpoints far from the training distribution. These artifacts manifest as stretched or warped details, particularly near image borders, and can be exacerbated by affine projection errors. To address this, we adapt the rendering approach by [Hahlbohm et. al. 2025], which evaluates Gaussians fully in 3D, ensuring geometry is preserved across the entire image plane, even for out-of-distribution views, resulting in stable, projection-accurate rendering without stretching or warping.

Popping

Popping artifacts in 3D Gaussian Splatting often arise from global depth sorting, where changes in camera perspective can abruptly alter the blend order of overlapping Gaussians. This results in distracting flicker or sudden shifts in appearance. AAA-Gaussians resolves this by employing hierarchical per-pixel sorting [Radl et. al. 2024], which more accurately determines the correct blending order on a per-ray basis. To retain high performance, hierarchical sorting is combined with tile-based culling, which efficiently reduces the number of Gaussians that need to be sorted per pixel. Therefore, we adapt our novel 3D frustum bounding to tile-based culling, enabling accurate sorting of Gaussians in 3D while maintaining high rendering performance.

BibTeX


@inproceedings{steiner2025aaags,
    author    = {Steiner, Michael and K{\"o}hler, Thomas and Radl, Lukas and Windisch, Felix and Schmalstieg, Dieter and Steinberger, Markus},
    title     = {{AAA-Gaussians: Anti-Aliased and Artifact-Free 3D Gaussian Rendering}},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    year      = {2025}
}

Related Work

[Mip-Splatting, Yu et. al. 2024] Z. Yu, A. Chen, B. Huang, T. Sattler and A. Geiger, Mip-Splatting: Alias-Free 3D Gaussian Splatting, in Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 19447-19456.

[GOF, Yu et. al. 2024] Z. Yu, T. Sattler and A. Geiger, Gaussian Opacity Fields: Efficient Adaptive Surface Reconstruction in Unbounded Scenes, in ACM Transactions on Graphics, 2024.

[Hahlbohm et. al. 2025] F. Hahlbohm, F. Friederichs, T. Weyrich, L. Franke, M. Kappel, S. Castillo, M. Stamminger, M. Eisemann, and M. Magnor, Efficient Perspective-Correct 3D Gaussian Splatting Using Hybrid Transparency, Computer Graphics Forum, vol.44, no.2, 2025.

[Radl et. al. 2024] L. Radl, M. Steiner, M. Parger, A. Weinrauch, B. Kerbl, and M. Steinberger, StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering, ACM Transactions on Graphics, vol. 43, no. 4, art. no. 64, 2024.