Diffuse materials
Matte paint, stone, wood and many plastics have broad reflections. Current splats can reproduce them well when the training views cover the intended use, which makes these materials a useful baseline.
Master's thesis / research archive
What does a neural scene keep, and what does it lose?
My master's research asked whether a detailed product asset could become an interactive neural scene without losing the material behaviour that made the original convincing. Synthetic captures made that question measurable because the cameras, lighting and reference images were known.
Explore the live experiment ↓01Can rendered ground truth replace expensive photo capture?
02Where do Gaussian splats stop behaving like real materials?
03Can error guide the next capture and the next representation?
Thesis abstract
Neural rendering for high-fidelity product visualisation
High-fidelity product visualisation usually needs substantial manual preparation before a detailed 3D asset can run interactively.
I investigated whether synthetic neural rendering could remove part of that preparation. I trained Gaussian splats from rendered images with known cameras and controlled material ground truth, rather than photographs and estimated camera poses. This removed several common reconstruction variables and made failures easier to attribute to the representation itself.
The results produced strong image metrics and convincing diffuse materials. Metals remained sensitive to view direction, while transparent and translucent materials exposed more fundamental limits. Synthetic training was clearly useful, but the experiment also showed that one representation is not enough for every surface, viewing direction or lighting condition. That unresolved problem became the starting point for Neural Prism.
Original interactive experiment
Inspect the trained result from every side.
This PLY and SOG 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 see where the reconstruction holds together and where a new view exposes its limits.
Delivery research
Delivery affects what people can explore.
The viewer can load PLY files or Spatially Ordered Gaussians (SOG). SOG packages the same interactive scene into a much smaller download, so the research can be examined in a browser instead of remaining on a workstation. The table shows the practical reduction for four of the test assets.
| 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
Image metrics need material context.
A single score can summarise average image similarity, but it cannot explain why a surface fails. Looking at diffuse, reflective and transmissive materials separately revealed which errors were local and which pointed to a limit in the representation.
Metals
Conductive materials depend on accurate, view-dependent reflections. Small directional or colour errors remain visible, so a good average score can still conceal an unconvincing product surface.
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.
Case study / Cannelle et Fromage
A synthetic scene isolates the representation failure.
This scene tested SuperSplat's video editing and export options. It also makes several failure modes of the basic INRIA Gaussian-splat approach easy to see.
The source is synthetic. The camera poses are known and the training process does not depend on an estimated sparse point cloud. The visible faults are therefore not explained by bad pose estimation or weak reconstruction input. They provide direct evidence that the appearance model itself needs further work.
RESEARCH → SYSTEM
Where this leads
The next question needed a workbench.
The master's research exposed the limits of the spherical-harmonic appearance model, but it did not provide a controlled way to test the alternatives. A useful next step needed matched cameras, scene-linear reference images, error maps and several renderers under one colour contract. I am building that system as Neural Prism, so each new representation can be compared against the same source and the same evidence.
Enter Neural Prism ↗