Point splats
2000The scene's 23,000 points drawn as opaque squares sized by distance, back to front. You can see the object, and every point.
C = Σ cᵢ αᵢ Π (1 − αⱼ)
1990 to now
No triangles, no distance field, just tens of thousands of soft, coloured, stretched blobs, sorted and blended. Since 2023 it can render captured reality in real time, and the demoscene is only starting to find out what it can do as an effect.
The algorithm
A scene is a cloud of 3D Gaussians, each with a position, a covariance (its size and orientation, like a squashed ellipsoid), a colour and an opacity. Every frame, project each Gaussian to the screen, the Jacobian of the perspective projection turns its 3D covariance into a 2D ellipse, sort them by depth and alpha-blend them back to front. In the 2023 method the Gaussians are optimised against photographs, so the cloud reproduces a real place.
Why it needs today's hardware
It only became practical because of current hardware: millions of splats need a GPU that can sort and blend them every frame. The tricks are a fast radix or counting sort on depth, drawing each splat as one instanced quad sized to 3σ of its ellipse, and premultiplied alpha blending. Nothing is spent on empty space, unlike volume ray marching.
for g in gaussians: // 3D → 2D
c = view · g.position
J = ∂(screen)/∂(camera) at c // perspective Jacobian
Σ₂ = J · R · g.Σ · Rᵀ · Jᵀ + 0.3·I
ellipse = eigen(Σ₂) → quad of ±3σ
sort gaussians by c.z, far to near
for g in sorted: // per pixel of its quad
α = g.opacity · exp(−½ dᵀ Σ₂⁻¹ d)
pixel = pixel·(1 − α) + g.colour·αDocumented milestones, with a source for each. Famous examples, not claims of who did it first.
Lee Westover · SIGGRAPH 1990 · Volume rendering research
‘Footprint evaluation for volume rendering’: instead of casting rays, each data sample is projected and its footprint spread onto the image, the word and the idea of splatting.
SourceSzymon Rusinkiewicz & Marc Levoy · Point-based rendering
Huge scanned models rendered interactively as hierarchies of point splats, points instead of triangles.
SourceZwicker, Pfister, van Baar & Gross · SIGGRAPH 2001 · Point-based rendering
Each point becomes an elliptical Gaussian filtered in screen space, the projection maths that 3D Gaussian splatting still uses.
SourceKerbl, Kopanas, Leimkühler & Drettakis · SIGGRAPH 2023 · Native GPU · CUDA
Anisotropic Gaussians optimised against photos render captured scenes at 100+ fps at 1080p. The original implementation uses CUDA, not the fixed rasterisation pipeline.
Sourcecvlab-epfl (WebGPU) · antimatter15 (WebGL) · Browser · WebGPU / WebGL
Within weeks, open viewers render trained scenes in a browser tab, first in WebGPU (August), then in plain WebGL (September).
SourceAras Pranckevičius · UnityGaussianSplatting · Unity · GPU
Playback of trained splat models inside a general-purpose engine, training and real-time display become separate steps.
SourceEquinor summer interns · Browser · Three.js
Gaussian objects mixed with ordinary 3D objects, including a bowling demo with splat pins, interaction beyond passive viewing.
SourceWith Love · I see you · 4K procedural graphics
A 4K executable still whose author confirms about 450 fitted Gaussian splats, fitting an image, not a 3D fly-through. The scene is only starting to use them.
Source5 live versions, from the oldest platform to today: 3 on the CPU, 2 on the GPU. Retro versions run at their original resolution and palette. Demos pause when off screen.
The scene's 23,000 points drawn as opaque squares sized by distance, back to front. You can see the object, and every point.
Same points, now as proper elliptical Gaussians: each 3D covariance projected with the perspective Jacobian and alpha-blended per pixel, in JavaScript, at 200 pixels wide.
Our own procedural ‘capture’: a knot, a disc, floating spheres, sky and dust, 23,000 anisotropic Gaussians, depth-sorted every frame with a counting sort and drawn as instanced quads.
Splats as a demo effect: 8,300 Gaussians flow from a sphere into a knot and back, each with its own delay, re-sorted every frame as they pass through each other.
An image built from 1,400 stretched 2D Gaussians, added coarse to fine like an optimiser at work: big soft ones for the background, thin ones along the edges.