technologybriefs
8:14in productionCh. 1 · Volume-first rendering/ 8:14 · ceiling 15 min
Models · Software

Gaussian splatting

Gaussian splatting ditches ray-marching for splatting—and gains speed by abandoning geometry.

Gaussian splatting is a volume rendering technique introduced by Lee Westover in the early 1990s and revitalised in 2023 by Inria for real-time radiance field rendering. It uses differentiable, anisotropic, explicit 3D Gaussians integrated from sparse camera-calibrated points. It is rasterisation-based, GPU-optimised, and avoids surface or line primitives. It enables fast photorealistic reconstruction from few images—but offers no mesh output, no topological guarantees, and limited editability.

ModelsSoftwareAlgorithmsArtificial intelligenceComputingAlgorithms across the network →
Chapters & takeaways4
  1. 0:44
    Volume-first rendering

    It renders volume data directly—no surfaces, no lines, no conversion.

  2. 1:57
    A 30-year reboot

    Born in the early 1990s, reborn in 2023—not evolution, but repurposing.

  3. 3:45
    Explicit over implicit

    Scenes become clouds of optimised 3D Gaussians—not neural weights, not voxels.

  4. 5:12
    Rasterisation, not marching

    GPU-native splatting replaces ray-marching—fast, visible, differentiable.

Worth your time?

Yes. Study the whole thing.

4/ 5
What works
  • real-time rendering from sparse images
  • differentiable scene optimisation
  • anisotropic radiance field approximation
  • GPU-accelerated visibility-aware splatting
What does not
  • produce meshes
  • guarantee topological consistency
  • support long-horizon Gaussian stability
  • run outside GPU rasterisation pipelines
Study it if
  • real-time radiance field researchers
  • graphics engineers building sparse-view reconstruction tools
  • teams prioritising speed over geometric export
Skip it if
  • CAD or simulation pipelines
  • low-end embedded devices
  • long-term archival scene representation
The written brief1 min read

What it is and the problem it solves

Gaussian splatting is a volume rendering technique that solves the problem of rendering photorealistic 3D scenes from sparse 2D images without mesh reconstruction. It was originally introduced by Lee Westover in the early 1990s. It was revitalised in 2023 for real-time radiance fields.

How it works

It renders volume data directly, without surface or line primitives. It uses unstructured, explicit, differentiable 3D Gaussians. These Gaussians integrate sparse camera-calibrated points. They are anisotropic and optimised interleaved with density control. A visibility-aware GPU rendering algorithm projects them to 2D splats.

What works

It delivers real-time radiance field rendering from sparse inputs. It retains continuous volumetric properties while being rasterisation-based. It supports anisotropic modelling and differentiable optimisation. It runs efficiently on GPUs using visibility-aware splatting.

What does not

It does not convert volume data into surfaces or lines—so it cannot output meshes or CAD-ready geometry. It does not guarantee consistent topology across views. It does not support long-horizon optimisation: Gaussians drift under repeated update. It is not defined for non-rasterisation pipelines.

What it changes

It shifts radiance field representation from implicit (NeRF-style networks) to explicit (Gaussians), enabling direct GPU rasterisation. It replaces volumetric integration with splatting, bypassing ray-marching. It makes photorealistic reconstruction from sparse images feasible at interactive rates—but only when calibrated points and GPU acceleration are available.

Is it worth your time

Yes—if you work on real-time radiance fields and need faster rasterisation than NeRFs offer. No—if you need strict geometric fidelity, long-term scene editing, or hardware-agnostic deployment. It trades memory efficiency and differentiability for explicit control and speed.

Same field · Models4 of 6
10:14
Relational modelCodd · 1969Codd’s 1969 relational model is a formal specification—not software, not SQL, not a product. It defines data as tuples in relations, enforces consistency via logic, and insists all information be explicit. Real databases violate it daily. That gap is where data engineering begins.
9:18
Diffusion modelDiffusion models are a 2015 method for sampling from complex distributions using thermodynamic-inspired noise reversal. They work by learning to undo a diffusion process—turning noise into data, step by step. They succeed in computer vision tasks as of 2024. They do not generalise beyond vision. They require iterative sampling, trading speed for quality. Their mechanism is physical in inspiration, statistical in execution, and narrow in current utility.
9:11
Generative adversarial network2014GANs are an adversarial framework for unsupervised generation. They work by pitting two neural networks against each other: a generator learns to produce data that fools a discriminator, which itself improves at spotting fakes. This indirect training yields samples statistically aligned with the training set—including photorealistic images that pass superficial human inspection. But GANs offer no guarantees on semantic coherence, stability, or interpretability. Their value lies in enabling distribution learning without labels—not in reliability, control, or fidelity.
9:28
LoRA (machine learning)LoRA makes fine-tuning large language models radically cheaper to train without slowing inference—by replacing full-weight updates with low-rank matrix additions. It works. It scales. It merges. But it assumes weight updates are low-rank—and that assumption fails where tasks demand dense, global parameter shifts.
Up next in Technology

Generative adversarial network

2014 · 9:11

GANs don’t model reality—they train a mimic to beat a critic.

Play nextBack to Models
9:11