Prototype / Technical Research

Gaussian Splatting Prototype

Exploring Gaussian Splatting as a medium for spatial perception, lighting, and immersive visual storytelling.

Overview

This prototype explores Gaussian Splatting (GS) as a method for translating historical perspective into interactive digital space. Using Hans Holbein’s The Ambassadors (1533) as a case study, the project investigates how point-based rendering can reinterpret a static painting as a navigable spatial experience.

The original artwork is known for its anamorphic skull, which only becomes visible from a specific viewing angle. This built-in requirement for viewer movement aligns naturally with GS, which allows scenes to be explored dynamically through real-time viewpoint changes.

Digital copy of Hans Holbein’s The Ambassadors from the National Gallery
Hans Holbein’s The Ambassadors — reference image captured at the National Gallery.

Capture and Reconstruction

The painting was reconstructed using two parallel approaches:

  • 3D scanning (Polycam) — mesh-based reconstruction exported as GLTF
  • Gaussian Splatting (Postshot Desktop) — point-based reconstruction exported as PLY

Both datasets were captured in person at the National Gallery, London, using the same mobile phone camera and identical image sequences to ensure a comparable workflow.

The comparison revealed key differences:

  • Mesh-based scanning produced reflective artefacts from the protective glass.
  • Gaussian Splatting preserved spatial detail more effectively and reduced visible glare.
3D scan result exported from Polycam
3D scan result exported from Polycam, showing reflective artefacts and glare from the protective glass.
Gaussian Splatting result exported from Postshot
Gaussian Splatting result exported from Postshot, preserving more spatial depth and reducing visible reflections.

Spatial and Lighting Behaviour

Unlike traditional polygon models, GS represents scenes as volumetric point clusters containing colour, opacity, and depth information. This structure enables:

  • smoother lighting transitions
  • more continuous depth perception
  • subtle atmospheric gradients
  • light diffusion through spatial clusters

Lighting experiments showed that GS produced softer and more immersive illumination compared with mesh-based rendering, which tends to generate sharper shadows and isolated highlights.

3D scan result exported from Polycam
Adjusted noise and clusters for the GS output (Blender)
Gaussian Splatting result exported from Postshot
3D scan (mesh-based) lighting test - harder shadow edges
3D scan result exported from Polycam
3DGS (point-based) lighting test - softer gradients and smoother transitions
Gaussian Splatting result exported from Postshot
3D scan (mesh-based) lighting test - backlight cannot penetrate
3D scan result exported from Polycam
3DGS (point-based) lighting test - backlight can penetrate

Artistic Interpretation

The GS output required manual refinement, including the removal of noise and stray point clusters. This process revealed an important shift in authorship: instead of modelling surfaces, the artist curates density, light, and spatial emphasis within a volumetric field.

Creative decisions therefore occur across several stages:

  • image capture and camera consistency
  • dataset cleaning and point refinement
  • lighting configuration and atmosphere
  • spatial composition and viewer perspective

GS automates reconstruction, but artistic judgement remains central.

Reflection

This prototype demonstrates how Gaussian Splatting can extend historical perspective techniques into immersive digital environments. By combining physically grounded rendering with spatial exploration, GS provides a new framework for visual storytelling that connects technical reconstruction with artistic perception.