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

Conference: ICCV 2025 arXiv: 2504.12811 Code: GitHub Area: 3D Vision / Novel View Synthesis / Real-Time Rendering Keywords: 3D Gaussian Splatting, Anti-Aliasing, Artifact-Free Rendering, Frustum Culling, View Consistency

TL;DR¶

By incorporating full 3D evaluation (rather than 2D splat approximations) into every stage of the 3DGS rendering pipeline, this work proposes an adaptive 3D smoothing filter, view-space bounding computation, and frustum-based tile culling to jointly address aliasing, projection artifacts, and popping artifacts in 3DGS. The method substantially outperforms existing approaches under out-of-distribution (OOD) viewpoints while maintaining real-time rendering (>100 FPS).

Background & Motivation¶

3DGS achieves efficient rendering by projecting 3D Gaussians onto 2D splats, but this approximation introduces multiple types of artifacts: (1) distortion artifacts at wide angles and image boundaries due to affine projection approximation; (2) aliasing artifacts caused by the absence of anti-aliasing under scale changes; (3) popping artifacts when Gaussians extend behind the view frustum; and (4) view inconsistency caused by global sorting. Existing methods each address only one or two of these issues and still fail under OOD viewpoints.

Core Problem¶

How to unify the elimination of all rendering artifact types through full 3D Gaussian evaluation while preserving rasterization efficiency? The central challenge is that 2D anti-aliasing approaches (e.g., Mip-Splatting's screen-space Mip filter) cannot be directly applied to 3D evaluation.

Method¶

Overall Architecture¶

Built upon StopThePop's hierarchical rasterizer, the method replaces bounding, culling, depth evaluation, contribution estimation, and anti-aliasing with fully 3D-aware implementations. MCMC densification is adopted. The rendering pipeline proceeds as: 3D bounding (view-space angles) → 3D frustum culling → per-pixel 3D Gaussian evaluation → hierarchical sort-and-blend.

Key Designs¶

Adaptive 3D Smoothing Filter: Replaces Mip-Splatting's 2D screen-space Mip filter. A key finding is that directly using volumetric change as amplitude scaling causes highly anisotropic Gaussians to become excessively transparent. The proposed approach instead scales the amplitude solely based on area variation perpendicular to the viewing ray direction (rather than volumetric change). An efficient closed-form solution (Eq. 10–12) is derived via Schur complement and Hartley analysis, avoiding matrix inversion. An adaptive switching mechanism is applied between the maximum sampling frequency at training time \(\hat{v}_{train}\) and the current rendering frequency \(\hat{v}\).
View-Space Perspective-Correct Bounding: Gaussian bounding is computed in view space rather than screen space. Plane fitting is performed on the ellipsoid to solve for angles \(\theta_{1,2}\) and \(\phi_{1,2}\) (Eq. 14–15). Screen-space methods fail when a Gaussian extends behind the image plane (causing popping), whereas view-space angular computation is inherently stable. The resulting angles are then converted to screen coordinates for tile assignment.
Frustum-Based 3D Tile Culling: Extends StopThePop's 2D tile culling to 3D. A four-plane frustum \(\mathcal{F}\) is constructed per tile; the maximum contribution point within the frustum is located in the Gaussian's canonical space and checked against \(\rho(\mathbf{x})^2 < \tau_\rho\). An implementation optimization projects only the 2 nearest planes and 3 edges (rather than all 4 planes and 4 edges), yielding significant performance gains.

Loss & Training¶

Standard 3DGS training objective (L1 + SSIM) with MCMC densification
Filter parameter \(k=0.3\) (consistent with Mip-Splatting)
Truncation threshold \(\tau_\rho\) used for bounding and culling computations

Key Experimental Results¶

Standard Datasets (In-Distribution Viewpoints)¶

Method	Mip-NeRF360 PSNR	T&T PSNR	Deep Blending PSNR
3DGS	27.443	23.734	29.510
StopThePop	27.304	23.226	29.929
Mip-Splatting	27.540	23.821	29.660
MCMC	28.027	24.642	29.727
Ours	27.835	23.582	30.485

Large-FOV Evaluation (3× FOV, OOD)¶

Method	Mip-NeRF360 PSNR	T&T PSNR
3DGS	17.112	-
MCMC	14.369	-
Taming 3DGS	11.545	-
Ours	23.583	-

OOD quality improvement is substantial: at 3× FOV, PSNR improves from 14.369 (MCMC) to 23.583 (ours), a gain of +9.2 dB.

Multi-Resolution Evaluation (Bicycle Scene)¶

Resolution	MCMC PSNR	Ours PSNR	Ours w/o AA PSNR
1/2×	21.15	26.73	20.99
1×	28.98	32.12	28.54
2×	25.69	25.74	25.65

Rendering Performance (RTX 4090, ms/frame)¶

Method	Mip-NeRF360	T&T
Ours	7.72	5.81
MCMC	6.79	4.43
Ours w/o culling	14.40	8.88

Ablation Study¶

Removing hierarchical sorting: marginally improves in-distribution PSNR (overfitting) but introduces severe popping artifacts.
Removing AA: minor impact in-distribution, but significant quality degradation at low/high resolutions.
Removing 3D evaluation: produces projection distortions at increased FOV.
Removing frustum culling: substantial performance degradation (7.72→14.40 ms) with no quality impact.
Components are complementary: artifact-free rendering requires all components in combination.

Highlights & Insights¶

Perpendicular area vs. volume for amplitude scaling: A key insight for the 3D filter — scale changes along the viewing ray should not affect amplitude; only area variation perpendicular to the ray is relevant. The mathematical derivation via Schur complement is elegant.
View-space angles vs. screen-space coordinates: Fundamentally eliminates instabilities when Gaussians cross the near plane, representing the principled approach to bounding computation.
Unified resolution of all artifact types: The first rasterization method to address aliasing, projection distortion, popping, and culling within a single framework.
Value of OOD evaluation: In-distribution metric differences are modest, but OOD scenarios (large FOV, resolution variation) reveal the true robustness of the proposed approach.

Limitations & Future Work¶

Limited improvement on in-distribution metrics (comparable to MCMC on standard benchmarks).
View-space bounding remains tied to the pinhole camera model; fisheye and other camera types are not supported.
Hierarchical sorting introduces additional performance overhead (nearly 2× slower than no sorting, which however produces popping).
Stronger view consistency reduces overfitting capacity, potentially necessitating richer view-dependent encoding (e.g., higher-order spherical harmonics).

vs. Mip-Splatting: Both target anti-aliasing, but Mip-Splatting operates in 2D screen space and cannot support 3D evaluation; the proposed 3D filter provides substantially greater robustness under OOD viewpoints.
vs. StopThePop: Shares the hierarchical sorting framework, but StopThePop still employs 2D bounding and 2D culling; this work elevates all operations to 3D.
vs. Hybrid Transparency (Hahlbohm et al.): Both perform 3D evaluation and precise bounding, but Hybrid Transparency still exhibits pop-in at boundaries (by discarding out-of-bounds Gaussians), whereas this work fully resolves the issue through view-space bounding.

The view-space operation paradigm is particularly relevant for VR and large-FOV applications. A promising future direction is to transfer the "orthogonal decomposition perpendicular to the ray" concept from the adaptive 3D filter to LOD/anti-aliasing in other 3D representations.

Rating¶

Novelty: ⭐⭐⭐⭐ — Novel amplitude scaling derivation for the 3D AA filter; elegant view-space bounding computation.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Comprehensive in-distribution and OOD evaluation, detailed ablations, multiple datasets and scenes, and runtime profiling.
Writing Quality: ⭐⭐⭐⭐⭐ — Rigorous mathematical derivations, thorough problem analysis, and clear illustrations.
Value: ⭐⭐⭐⭐ — Substantial improvement in 3DGS rendering quality and robustness; open-source code; an important reference implementation for the 3DGS rendering pipeline.