The current rasterize_to_pixels_2dgs implementation appears to use the Gaussian center z-coordinate (means_c[..., 2]) as the depth value during alpha blending, rather than the actual ray-plane intersection depth as described in the original 2DGS paper.
As demonstrated in the images below with two tilted splats: the original 2DGS implementation correctly captures varying depths across the splat surfaces via ray-plane intersection, whereas gsplat's current implementation renders constant depth values (using only the Gaussian center z-coordinate), which is geometrically incorrect for non-frontal surfaces.


The current rasterize_to_pixels_2dgs implementation appears to use the Gaussian center z-coordinate (means_c[..., 2]) as the depth value during alpha blending, rather than the actual ray-plane intersection depth as described in the original 2DGS paper.
As demonstrated in the images below with two tilted splats: the original 2DGS implementation correctly captures varying depths across the splat surfaces via ray-plane intersection, whereas gsplat's current implementation renders constant depth values (using only the Gaussian center z-coordinate), which is geometrically incorrect for non-frontal surfaces.

