Introduction to 3D Gaussian Splatting (Dylan Ebert)
June 5, 2026
rasterization - converting mathematical models describing a scene into a grid of pixels
Gaussian splatting - a rasterization technique that uses Gaussians to represent a scene
-learns one model per scene
-renders multiple overlapping Gaussians at once, often in the millions
-each Gaussian has:
1.position
(x, y, z)2.covariance for stretch and scale
3.color
(r, g, b)4.alpha for transparency
Process
1.Structure from Motion (SfM) - estimate a point cloud from a set of images
2.Convert each point into a Gaussian
1.color and position are initialized from SfM
2.the representation is then optimized for better quality
3.Training recipe
1.rasterize the Gaussians into an image
2.compute loss against the ground-truth image
3.update Gaussian parameters
4.apply densification and pruning
1.if the gradient is large and the Gaussian is small, clone it
2.if the gradient is large and the Gaussian is large, split it
3.if the alpha becomes very low, remove it
differentiable Gaussian rasterization - rasterization that stays compatible with gradient-based optimization, so the Gaussian parameters can be learned directly from image reconstruction loss