Gaussian Blending: Rethinking Alpha Blending in 3D Gaussian Splatting

Junseo Koo, Jinseo Jeong, Gunhee Kim

Introduction: The Multi-Scale Challenge

3D Gaussian Splatting (3DGS) has revolutionized Novel View Synthesis by enabling high-quality, real-time rendering. However, a significant limitation persists: the method struggles with multi-scale rendering. When a scene is viewed at resolutions or zoom levels different from those used during training, two distinct artifacts emerge. "Erosion" occurs during zoom-in, where fine features like thin wires or leaves appear thinned out or blurry. Conversely, "dilation" occurs during zoom-out, creating artificial "staircase" edges where foreground objects appear to expand and block background details incorrectly.

Visual comparison of rendering artifacts: 3DGS and Analytic Splatting show significant erosion and dilation compared to the proposed Gaussian Blending.

The Root Cause: Scalar Alpha Blending

The authors identify the fundamental flaw as scalar alpha blending. Traditional 3DGS and its variants compute a single alpha value for a splat at the pixel center. This "point-based" approach ignores the spatial distribution of splats within the 2D area of a pixel. Mathematically, this allows a foreground splat to block background splats across the entire pixel, even if they only physically overlap a small fraction of the pixel's area. This lack of spatial awareness within the pixel leads to the aliasing artifacts observed in multi-scale scenarios.

Methodology: Gaussian Blending

To address this, the paper introduces Gaussian Blending, a drop-in replacement for the standard 3DGS rendering kernel. Instead of representing transmittance as a single scalar value, it is modeled as a 2D uniform distribution (a rectangular "window") within the pixel area.

Conceptual comparison: (a) Real camera integration, (b) Scalar alpha blending leading to aliasing, and (c) the proposed Gaussian Blending which approximates area-based integration.

Key components of this method include:

  • Dynamic Windowing: As the renderer processes splats from front to back, it updates the center, size, and intensity of the transmittance window.
  • Distributional Moments: To maintain real-time efficiency, the system tracks the "mass," "mean position," and "variance" of transmittance using zeroth, first, and second-order moments.
  • Spatial Awareness: This allows the renderer to recognize that a foreground splat on one side of a pixel should not block a background splat on the opposite side, effectively simulating area-based integration.

Results and Evaluation

The method was rigorously tested on the Multi-scale Blender and Mip-NeRF 360 benchmarks. Gaussian Blending consistently outperformed all baselines, including Mip-Splatting and Analytic Splatting, across PSNR, SSIM, and LPIPS metrics. The authors demonstrated the versatility of the approach by successfully applying it as a drop-in replacement for not only the original 3DGS but also 2DGS and Scaffold-GS.

Quantitatively, the approach achieved a 5x reduction in transmittance error compared to previous state-of-the-art methods. Despite the increased complexity of the blending logic, the implementation remains highly efficient, maintaining a rendering speed of approximately 123 FPS on an NVIDIA RTX A6000. Additionally, the model converges up to 3x faster than standard 3DGS because it does not need to learn to compensate for aliasing artifacts during the training process.

Conclusion

Gaussian Blending offers a mathematically grounded solution to the aliasing problems inherent in point-based rendering. By rethinking alpha blending as a spatial distribution rather than a scalar value, it preserves fine geometric details and eliminates visual artifacts across various scales. As a framework-agnostic "drop-in" replacement, it is easily integrated into existing 3DGS pipelines, setting a new standard for high-fidelity, multi-scale novel view synthesis.

Summary updated

BibTeX
@inproceedings{koo2026gaussian,
  author = {Junseo Koo and Jinseo Jeong and Gunhee Kim},
  title = {{Gaussian Blending: Rethinking Alpha Blending in 3D Gaussian Splatting}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2026},
  doi = {10.1609/aaai.v40i7.37495},
}
Figure