Master's thesis / research archive

Testing neural rendering for high-quality product visualisation

My master's research asked whether a detailed product asset could become an interactive neural scene without losing the materials and lighting that made the original convincing.

Explore the live experiment ↓

01Can rendered ground truth replace expensive photo capture?

02Where do Gaussian splats stop behaving like real materials?

03Can measured error guide the next capture and representation?

Background

WHY THIS PROBLEM MATTERS

Why preparing CAD-derived assets has been expensive

My background is in optimising tessellated CAD for interactive use. That includes automotive and furniture assets, as well as full vehicle assemblies that need to run in WebGL or WebGPU for e-commerce.

This work used to take technical artists weeks. A vehicle assembly could contain hundreds of millions of triangles, much of it in sheet metal, drive systems, wiring harnesses and other parts that a customer would never see.

The first job was to remove that hidden geometry. The visible surfaces still needed accurate materials, lighting and environments, usually created in VRED, Octane or another high-end digital content creation tool.

The assembly then went through repeated rounds of decimation, remeshing and texture baking. Each round had to reduce the cost of the asset without taking it below the quality bar expected in a product experience.

Texture baking creates another trade-off. Low-resolution textures lose the detail that makes the product convincing. High-resolution textures can exhaust memory, particularly on mobile devices.

KTX2 and GPU-friendly texture compression help, but they add another stage to an already difficult pipeline. Textures must remain detailed enough to look right while staying small enough for the browser and GPU memory.

The result is an expensive process with many iterations. That is understandable because companies care about how their products and brands are presented, but the cost and effort remain substantial.

Neural rendering offers another route. It may preserve more of the materials and lighting produced in the high-end renderer while making the result practical to deliver in a browser.

This research tests where that promise holds and where it does not.

Thesis abstract

Neural rendering for high-fidelity product visualisation

The experiment tested whether synthetic neural rendering could reduce the manual preparation needed to put a detailed 3D asset in a browser.

I trained Gaussian splats from rendered images with known cameras and controlled material ground truth. Using synthetic captures removed the uncertainty of estimated camera poses and made failures easier to trace back to the representation.

The results had strong image metrics and reproduced diffuse materials well. Metals were more sensitive to view direction. Transparent and translucent materials exposed more fundamental limits.

Synthetic training was useful, but one representation was not enough for every material, viewing direction or lighting condition. That unresolved problem became the starting point for Neural Prism.

Original interactive experiment

Inspect the trained Gaussian splats

This viewer preserves part of the original master's research. Each object is a Gaussian field trained from synthetic images. Drag to orbit, scroll to zoom, and change the model to inspect it from a new direction.

The bike opens as the original PLY result. The selector also includes compressed SOG versions and the other test assets, which makes the quality and delivery trade-offs easier to compare.

Drag to orbit · Scroll to zoom

Delivery research

Smaller files make browser delivery practical

The viewer loads PLY files and Spatially Ordered Gaussians (SOG). SOG stores the same interactive scene in a much smaller download, which matters when the work must leave a workstation and run in a browser.

The table shows the reduction for four local test assets. Compression does not solve the material problem, but it removes one practical barrier to examining the result on the web.

Local source file sizes
Asset	PLY	SOG	Reduction
Bike	98.7 MB	7.8 MB	13×
Damaged helmet	27.2 MB	3.4 MB	8×
Flight helmet	26.2 MB	3.6 MB	7×
Lounge chair	20.3 MB	3 MB	~7×

What the work shows

The material matters more than the average score

A single score can summarise image similarity, but it cannot explain why a surface fails. Separating diffuse, reflective and transmissive materials showed which errors were local and which came from the representation itself.

WORKS WELL

Diffuse materials

Matte paint, stone, wood and many plastics have broad reflections. Gaussian splats reproduce them well when the training views cover the intended use, so these materials provide a useful baseline.

NEEDS CARE

Metals

Metal depends on accurate, view-dependent reflections. Small errors in direction or colour remain visible, which means a good average score can still hide an unconvincing product surface.

OPEN PROBLEM

Glass and transmission

Refraction, transmission and internal scattering expose the limits of traditional third-order spherical harmonics. These materials need a richer account of direction and light transport.

EMBEDDED SUPERSPLAT SCENE · SYNTHETIC SOURCE

Case study / Cannelle et Fromage

A synthetic scene makes the failure easier to isolate

This scene tested SuperSplat's video editing and export options. It also makes several limits of the basic INRIA Gaussian-splat method easy to see.

The source is synthetic, the camera poses are known, and training does not depend on an estimated sparse point cloud. Bad pose estimation or weak reconstruction input therefore cannot explain the visible faults.

That leaves the appearance model itself as the main source of error and gives the next experiment a clear target.

Explore the interactive SuperSplat scene ↗ View the original LuxCoreRender scene ↗

RESEARCH → SYSTEM

Where this leads

The research led to Neural Prism

The master's work identified limits in the spherical-harmonic appearance model, but it did not provide a controlled way to compare alternatives.

The next step needed matched cameras, scene-linear reference images, error maps and several renderers working under one colour contract. I am building that system as Neural Prism.

It gives each new representation the same source data and the same test conditions, so an improvement can be measured rather than assumed.

Enter Neural Prism ↗