ATW-GS

Abstract

3D Gaussian Splatting (3DGS) enables real-time novel view synthesis but suffers from aliasing artifacts when rendered at sampling rates that differ from those used during training. Existing anti-aliasing methods mitigate this issue by prefiltering Gaussian primitives or integrating their response over the pixel area; however, they rely on scalar alpha composition, which treats per-pixel transmittance as a single value and therefore discards the spatial structure within each pixel. More recent spatially aware approaches maintain a per-pixel transmittance window, but represent it as an axis-aligned box, preventing the model from capturing the true orientation of projected splats. To address this limitation, we propose Anisotropic Transmittance Windows for 3D Gaussian Splatting (ATW-GS), which models per-pixel transmittance as a full covariance Gaussian that preserves splat orientation throughout the front-to-back composition sequence. Furthermore, ATW-GS computes each splat's contribution weight using the exact error-function integral of the pixel footprint projected onto the window's marginal axes. This ensures consistency between the anisotropic window state and the per-splat weight computation. Extensive experiments on the NeRF Synthetic (Blender) and Mip-NeRF 360 datasets demonstrate that ATW-GS consistently outperforms state-of-the-art anti-aliasing methods across both single-scale and multi-scale evaluation regimes.