Gaussian Blending: Rethinking Alpha Blending in 3D Gaussian Splatting

TL;DR

Gaussian Blending addresses aliasing in 3D Gaussian Splatting by treating alpha and transmittance as spatial distributions within pixel regions rather than scalars. By leveraging per-splat 2D Gaussian moments and an adaptively updated transmittance window, it preserves spatial occlusion and reduces erosion/dilation artifacts across unseen sampling rates, all without extra memory or training; it also functions as a drop-in component for existing NVS pipelines. The approach achieves real-time performance comparable to 3DGS while delivering higher fidelity at multiple sampling rates on Blender and Mip-NeRF 360. This work provides a practical, general anti-aliasing solution for neural rendering that extends to various backbones without retraining and improves robustness to scale variation, with clear pathways for addressing remaining camera-induced blur effects in future work.

Abstract

The recent introduction of 3D Gaussian Splatting (3DGS) has significantly advanced novel view synthesis. Several studies have further improved the rendering quality of 3DGS, yet they still exhibit noticeable visual discrepancies when synthesizing views at sampling rates unseen during training. Specifically, they suffer from (i) erosion-induced blurring artifacts when zooming in and (ii) dilation-induced staircase artifacts when zooming out. We speculate that these artifacts arise from the fundamental limitation of the alpha blending adopted in 3DGS methods. Instead of the conventional alpha blending that computes alpha and transmittance as scalar quantities over a pixel, we propose to replace it with our novel Gaussian Blending that treats alpha and transmittance as spatially varying distributions. Thus, transmittances can be updated considering the spatial distribution of alpha values across the pixel area, allowing nearby background splats to contribute to the final rendering. Our Gaussian Blending maintains real-time rendering speed and requires no additional memory cost, while being easily integrated as a drop-in replacement into existing 3DGS-based or other NVS frameworks. Extensive experiments demonstrate that Gaussian Blending effectively captures fine details at various sampling rates unseen during training, consistently outperforming existing novel view synthesis models across both unseen and seen sampling rates.

Figures (17)

• Figure 1: Scalar alpha blending used in previous novel view synthesis methods (e.g., 3DGS and Analytic Splatting) exhibits aliasing artifacts when rendering at different sampling rates. Specifically, erosion with noisy edges occurs during zoom-in (top), while dilation with staircase artifacts occurs during zoom-out (bottom). In contrast, our Gaussian Blending produces consistent synthesis results at unseen sampling rates (e.g., leaves and bicycle frame) without any priors and additional training.
• Figure 2: Comparison of blending behaviors when rendering overlapping splats. (a) In a real camera system, spatial occlusion is properly handled—the yellow splat is attenuated by closer splats, while the blue splat, which does not overlap in screen space, retains high transmittance and remains visible. (b) Scalar alpha blending ignores spatial occlusion by averaging alpha values across the pixel, causing dilation of the overlapping yellow splat and undesired suppression of the non-overlapping blue splat. (c) Our Gaussian Blending dynamically adjusts the transmittance window within each pixel, preserving high transmittance in regions without splats and effectively maintaining the visibility of the blue splat while reducing the influence of the occluded yellow splat.
• Figure 3: Comparison of the remaining transmittance error $\Delta T$ after rendering two splats symmetrically placed at $\mu_1 = [\mu_x, -0.1]^\top$ and $\mu_2 = [\mu_x, 0.1]^\top$, with $o_1 = o_2 = 1$ and $\sigma_1 = \sigma_2 = \sigma$. (Left) Varying $\mu_x$ while fixing $\sigma = 1$. (Right) Varying $\sigma$ while fixing $\mu_x = 0.5$. A negative $\Delta T$ indicates dilation, where a lower transmittance remains after rendering.
• Figure 4: Qualitative results of zoom-out setting (Single-scale training ($\times 1$ resolution) and multi-scale testing). The training resolution is highlighted in bold. Our Gaussian Blending effectively suppresses pixel-level dilation in zoom-out scenarios (e.g., the disk edge in the top scene at $\times 1/8$ resolution).
• Figure 5: Qualitative results of zoom-in setting (Single-scale training ($\times 1/8$ resolution) and multi-scale testing). The training resolution is highlighted in bold. Our Gaussian Blending effectively prevents erosion in zoom-in scenarios (e.g., the leaves in the top scene at $\times 1$ resolution).